The research tested whether a compact model could improve short-term load forecasts specifically during heat waves and cold snaps, rather than allowing exceptional events to be diluted within annual averages. The California dataset contained 20,422 hourly records from June 2020 through September 2022, including 1,228 records flagged as extreme by a rule combining Sacramento temperature above 35 degrees Celsius or 10-meter wind speed above 20 meters per second. The New England dataset comprised 5,813 hourly records from July 2023 through February 2024, with 210 extreme records identified by a Boston temperature below minus 5 degrees Celsius or wind speed above 15 meters per second. In California, median load during flagged periods was 37,149 MW, compared with 24,007 MW otherwise. The corresponding New England medians were 14,101 MW and 12,963 MW.

Each forecast used 24 hours of history to predict the following six hours. The input combined grid load with air temperature, relative humidity, precipitation, wind speed, a derived Weather Severity Index and a binary extreme-weather flag. The severity index condensed the meteorological variables into a value from zero to one, with temperature entering in the opposite direction for the heat-sensitive California case and the cold-sensitive New England case. This index served a dual purpose: it provided the network with an explicit measure of climatic stress and increased the penalty applied to errors during severe windows. The model was trained separately for the two operators but without dataset-specific changes to the network architecture or shared hyperparameters.

The CE-LSTM itself contained 2,107,718 trainable parameters across three stacked recurrent layers with 512, 256 and 128 hidden units. An additive temporal-attention layer assigned different relevance to the 24 input hours, allowing the decoder to emphasize periods associated with unusually high or low temperature. The weather-weighted loss used a penalty coefficient of 0.35, selected by validation-set grid search. This design targeted the study’s central concern: an ordinary mean-squared-error objective can allow abundant routine observations to dominate training even when forecasts during the most consequential events remain imperfect.

Across both extreme-weather test partitions, the CE-LSTM produced the lowest errors among six evaluated models. On California, its MAE of 471.8 ± 19.3 MW was 21.1% below the next-best model, an iTransformer with an MAE of 598.3 ± 29.4 MW. Its RMSE was 693.2 ± 24.7 MW and its MAPE was 1.55 ± 0.07%. A two-sample t-test comparing the five runs of the CE-LSTM and iTransformer gave t(8) = 6.53, p = 0.0002 and Cohen’s d = 5.09. On New England, the CE-LSTM’s 329.6 ± 17.9 MW MAE was 19.4% below the iTransformer’s 408.9 ± 27.1 MW; the statistical comparison gave t(8) = 12.46, p < 0.0001 and d = 3.45. The paper reported significance tests only against the iTransformer, not the other four baselines.

Deployment measurements showed the benefit of converting the trained FP32 model to static INT8 precision with TensorFlow Lite. On an NVIDIA Jetson Nano, the quantized CE-LSTM used 8.1 ± 0.2 MB of runtime memory, completed an inference step in 9.2 ± 0.5 milliseconds, drew 3.4 ± 0.2 watts and occupied 23 ± 2% of the CPU during sustained testing. By comparison, the FP32 version used 29.8 MB, required 33.6 milliseconds and consumed 5.7 watts. Quantization therefore produced a 3.65-fold latency speedup and 3.68-fold memory reduction for a 0.98% relative increase in MAE, from 467.2 to 471.8 MW in the reported accuracy-latency comparison. Four-bit quantization was faster still, at 5.6 milliseconds, but increased MAE by 26.2%, a trade-off the authors considered unsuitable for the proposed safety-critical use.

Ablation experiments indicated that explicit weather information supplied most of the accuracy gain. On the California extreme partition, removing all weather features increased MAE by 48.0%, from 471.8 to 698.2 MW. The binary extreme-weather flag had the largest individual effect: removing it raised MAE by 31.7%. Removing temperature increased error by 21.6%, while removing humidity, precipitation or wind speed raised it by 10.2%, 9.3% and 8.6%, respectively. Removing temporal attention increased MAE by 15.2% to 543.6 MW, while replacing the weather-weighted loss with ordinary mean squared error increased it by 6.4% to 501.9 MW. Perturbing any one Weather Severity Index coefficient by 20% changed MAE by at most 1.5%, suggesting the reported performance was not highly sensitive to the selected coefficient values.

The forecast was connected to a four-stage demand-response decision engine rather than used only as a prediction display. The system compared the six-step forecast with an operator-configured capacity threshold and assigned Normal, Advisory, Warning or Critical status. In three illustrative California cases, a forecast peaking at 50,100 MW against a 48,000 MW threshold produced a warning to activate demand curtailment; a 48,500 MW peak produced a conservation advisory; and a forecast declining to 38,200 MW generated a recommendation to restore deferred load. These examples show the engine’s logic, but they are not evidence that the alerts were dispatched or that demand-response programs achieved savings. The engine remained rule-based, and its threshold was fixed rather than supplied dynamically by a grid state estimator.

Several limitations constrain how broadly the results can be interpreted. The California study period covered 28 months, while the New England period covered eight months; the extreme-weather test partitions contained 389 and 58 records, respectively. Only 18 extreme California records appeared in validation, and the New England training split contained just eight flagged records out of 4,069, leaving the model with limited direct exposure to the cold-snap conditions it later predicted. The weather data for each grid came from one reference location—Sacramento for California and Boston for New England—rather than spatially resolved observations. Hyperparameters were also tuned once and shared among the deep-learning baselines, potentially disadvantaging architectures such as the iTransformer. The two US grids demonstrate performance in a summer-heat and winter-cold setting, but they do not establish generalization to other operators, renewable-heavy systems, monsoon-driven demand or live grid operations. The study nevertheless shows that carefully encoded weather stress, recurrent sequence modeling and aggressive but measured quantization can coexist in an edge forecasting system designed to remain responsive when grid demand changes abruptly.

Subject of Research: Weather-augmented edge AI for extreme-weather electricity-load forecasting and demand-response alerts

Article Title: ClimateEdge: a lightweight weather-augmented LSTM framework for real-time demand response in climate-stressed smart grids

Article References: A, H. K., Murugan, V., R, D., Prathiba, S. B., & Asai, S. R. (2026). ClimateEdge: a lightweight weather-augmented LSTM framework for real-time demand response in climate-stressed smart grids. Frontiers in Artificial Intelligence, 9, Article 1918025. https://doi.org/10.3389/frai.2026.1918025

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1918025

Keywords: climate resilience, smart grids, load forecasting, extreme weather, edge AI, long short-term memory, temporal attention, model quantization, demand response, Jetson Nano, ClimateEdge, lightweight

The central innovation is not a single machine-learning model but an “intelligent service hub” that coordinates several task-specific algorithms. Load forecasting used a long short-term memory network with two recurrent layers, while a deep neural network handled binary fault classification and a random forest estimated whether equipment required maintenance. A convolutional neural network extracted visual features, natural-language processing transformed maintenance records into numerical representations, and a weighted late-fusion module combined the results. The authors assigned 50% of the fusion weight to sensor and time-series features, 25% to text and 25% to images or video, reflecting the continuous availability of electrical measurements and the contextual value of reports and visual inspections.

Information moves through the framework in a closed loop. Smart meters, Internet of Things sensors, cameras, drones, operational logs and inspection records are collected, cleaned and synchronized before being sent to specialized processing modules. The service hub then combines those inputs with a knowledge base containing grid topology, operating rules, maintenance records, constraints and safety limits. Its outputs can include load forecasts, fault classifications, equipment-failure probabilities, maintenance schedules, power-rerouting recommendations, renewable-energy dispatch, demand-response adjustments and cybersecurity alerts. Results from the resulting actions are fed back into the system so that predictions and control decisions can be revised as operating conditions change.

The experimental dataset contained 20,000 time-ordered samples assembled from public smart-grid resources. It included electrical measurements, environmental variables, renewable-generation information, maintenance records and fault labels, with text and visual inputs included for equipment monitoring. Because the measurements were time dependent, the researchers used the earliest 80% for training, the next 10% for validation and hyperparameter tuning, and the final 10% as an independent test set. They also conducted walk-forward validation with five sequential folds to assess performance under a more operationally realistic sequence of training on earlier observations and testing on later ones. Ten independent runs were reported with means, standard deviations and 95% confidence intervals.

For fault detection, the deep neural network produced probability scores for faulty and non-faulty states, which were assessed across decision thresholds using a receiver operating characteristic curve. The resulting ROC-AUC was 0.970 ± 0.012, with a 95% confidence interval from 0.961 to 0.979. The predictive-maintenance random forest combined historical maintenance data with equipment age, temperature, vibration, loading conditions and previous faults. On the 2,000-sample test set, it correctly classified 1,348 normal cases and 562 maintenance-required cases, while misclassifying 52 normal cases and missing 38 maintenance-required cases. Its precision for the maintenance-required class was 91.53%, recall was 93.67% and the F1-score was 92.59%.

The same predictive system attained 95.5% overall accuracy and 94.98% balanced accuracy, with a Matthews correlation coefficient of 0.894. The authors also evaluated a Weibull reliability model to represent failures associated with early life, random events or aging, using available component-failure histories where possible. Its initial scale and shape parameters were based on engineering assumptions when records were unavailable, and the researchers said they examined parameter sensitivity. In the operational configuration, components with a predicted failure probability above a validation-selected threshold of 0.60 were marked as requiring maintenance.

p>The load-forecasting module used 24-step sliding windows to predict the next demand or equipment-risk state. Across the reported repeated experiments, it achieved a MAPE of 4.8% ± 0.5%, with a 95% confidence interval of 4.4% to 5.2%. In an example daily profile, demand ranged from about 350 MW at night to 450 MW at midday. In another comparison, consumption ranged from 280 to 320 MW while generation ranged from 300 to 340 MW, illustrating the mismatch the optimization system was intended to manage by adjusting conventional generation, renewable use, storage and flexible demand. The paper describes this as a benchmark evaluation and does not report deployment in an operating utility.

For energy optimization, the framework worked over a 24-step horizon, maintained a 10% reserve margin and allowed demand-response adjustments of up to 15% of non-critical controllable load. The experimental baseline included conventional generation scheduling and reactive adjustment, while the optimized condition jointly coordinated load forecasts, renewable-generation predictions, storage, power allocation and demand response. Under identical test-horizon conditions, total modeled operating cost fell from $12,000 to $8,400. The largest reported saving was in conventional generation, which declined from $9,200 to $6,450 as renewable generation was prioritized and dispatch was adjusted; smaller reductions appeared in grid losses, renewable curtailment, peak demand and storage-control costs.

An ablation analysis indicated that the unified design mattered more than any isolated component. Removing sensor fusion produced the largest degradation among the component-removal tests, lowering fault-detection AUC from 0.970 to 0.931, predictive-maintenance accuracy from 95.5% to 91.6%, load-forecasting MAPE from 4.8% to 6.7%, and cost reduction from 30.0% to 23.5%. Removing energy optimization lowered the modeled savings to 16.2%, while a sensor-only model reached an AUC of 0.926 and 22.7% cost reduction. A broader ablation table reported an even larger cost reduction of 18.6% without multimodal fusion and 21.8% without the service hub, although the paper does not reconcile every numerical difference among its tables and figures.

Human oversight is built into the proposed control hierarchy. Routine, lower-risk actions—such as load balancing, energy allocation and maintenance scheduling—can be automated, whereas severe faults, major voltage deviations or cybersecurity alerts are routed to an operator dashboard for approval. The cybersecurity design contemplates false-data injection, replay attacks, denial of service, unauthorized commands and poisoned training data, using checks such as timestamp verification, role-based access, command authentication, anomaly filtering and cross-modal consistency. These safeguards remain architectural proposals within the evaluated framework. The study itself identifies noisy or incomplete sensor data, renewable intermittency, model generalization to larger grids and extreme conditions, cybersecurity exposure, implementation expense and regulatory standardization as unresolved limitations. Large-scale field testing with live streaming data will therefore be required before the reported benchmark gains can be treated as demonstrated utility-scale benefits.

Subject of Research: Evaluation of an autonomous, operator-controllable AI framework for multimodal smart-grid management

Article Title: An artificial intelligence based framework for smart power grid management using an autonomous and controllable intelligent service hub with multimodal document processing

Article References: Zeng, F., Zheng, H., Li, C., Rong, J., & Nong, Y. (2026). An artificial intelligence based framework for smart power grid management using an autonomous and controllable intelligent service hub with multimodal document processing. Frontiers in Energy Research, 14, 1910556. https://doi.org/10.3389/fenrg.2026.1910556

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1910556

Keywords: smart grids, artificial intelligence, multimodal data, fault detection, predictive maintenance, load forecasting, energy optimization, renewable energy, demand response, grid cybersecurity, artificial, intelligence

The core problem the researchers set out to solve is a tension familiar to anyone working in critical infrastructure cybersecurity. Traditional anomaly detection systems for electricity theft and intrusion detection rely on centralized training, where consumption records from thousands of households are pooled into a single server. That approach gives the detection model a broad view of user behavior, but it also concentrates sensitive personal data in one place, creating a honeypot for attackers and raising serious privacy concerns. FED-XID flips this model on its head by using federated learning, a training paradigm in which the model travels to the data rather than the other way around. Each edge node trains a local detection model on its own consumption records, and only compact model structures are shared upstream for aggregation.

FED-XID introduces three technical innovations that distinguish it from earlier federated anomaly detection frameworks. The first is an embedded, lightweight intrusion detection system based on XGBoost, a gradient-boosted decision tree algorithm prized for its accuracy and computational efficiency. XGBoost builds an ensemble of decision trees sequentially, with each new tree correcting the errors of the previous ones, making it well suited to edge deployment where memory and processing power are limited. The second innovation is FedXgbBagging, a novel tree-level aggregation strategy that combines local models through tree concatenation rather than averaging model parameters. Instead of blending the numerical weights of different local models, the aggregation server stitches together the actual decision trees learned by each node, producing a global ensemble that reflects the collective knowledge of the entire network.

The third innovation addresses a subtle but critical weakness in federated pipelines: data quality. Smart meter data is notoriously messy, with missing readings, sensor faults, and temporal inconsistencies that can degrade model training. FED-XID integrates a hybrid TCN-LSTM module directly into the federated pipeline to perform local data reconstruction before training begins. Temporal convolutional networks excel at capturing short-range patterns in time series through parallelizable convolutional operations, while long short-term memory networks handle long-range dependencies through gated memory cells. By combining both architectures, the module reconstructs missing or corrupted consumption data, ensuring that the federated models learn from high-quality, temporally consistent records rather than noisy raw input.

To evaluate the framework, the researchers trained FED-XID on the SGCC dataset, a widely used benchmark containing consumption records from 42,372 smart grid users. The results are striking. FED-XID achieved an area under the ROC curve of 93.60 percent and an F1-score of 88.80 percent, metrics that measure how well the system distinguishes genuine anomalies from normal behavior. For comparison, a centralized XGBoost baseline trained on the same data reached 98.10 percent AUC and 92.65 percent F1-score. The gap between federated and centralized performance is modest, and the researchers argue that the trade-off is justified by the significant reduction in privacy exposure. Rather than sharing raw consumption data, FED-XID transmits only compact tree structures upstream, keeping sensitive user records local.

Speed is another area where FED-XID shines. Training cycles complete in under 26 seconds, a figure that matters enormously for real-world deployment on smart grid edge nodes. These devices are typically resource-constrained microcontrollers or single-board computers with limited memory and processing power, and any detection system that requires minutes or hours of training per cycle would be impractical to operate. The sub-26-second training time means the federated pipeline can run continuously, updating the global detection model as new consumption data arrives without overwhelming the hardware. This makes FED-XID a scalable, lightweight, and deployable solution for anomaly detection in environments where computational resources are scarce.

The implications for smart grid security are significant. Electricity theft costs utilities billions of dollars annually, and cyber intrusions into grid infrastructure pose risks that extend beyond financial losses to public safety and national security. As grids incorporate more distributed energy resources, electric vehicle charging stations, and Internet of Things devices, the volume of consumption data and the number of potential entry points for attackers will only grow. FED-XID demonstrates that it is possible to detect these threats across a distributed network without concentrating sensitive data in a single server, offering utilities a path toward stronger security that respects user privacy by design.

The research also contributes to a broader conversation about federated learning in critical infrastructure. Earlier federated anomaly detection frameworks have struggled with non-IID data, where consumption patterns differ dramatically across nodes, and with the computational overhead of training deep neural networks on edge hardware. FED-XID sidesteps some of these challenges by using gradient-boosted trees rather than deep networks for the core detection task, and by handling data quality issues locally through the TCN-LSTM reconstruction module. The tree-level aggregation strategy preserves the interpretability of the underlying models, since decision trees can be inspected and audited more easily than the opaque weight matrices of neural networks.

Looking ahead, the researchers suggest that FED-XID could serve as a foundation for more sophisticated federated security systems in smart grids and beyond. The combination of lightweight edge detection, tree-level aggregation, and local data reconstruction could be adapted to other domains where privacy and computational constraints collide, from industrial control systems to healthcare networks. As the digital transformation of critical infrastructure accelerates, frameworks like FED-XID point toward a future where security and privacy are not competing priorities but complementary design principles, built into the very architecture of the systems that keep the lights on.

Subject of Research: Federated learning brings privacy-first anomaly detection to smart grids

Article Title: FED-XID: a novel federated learning based architecture for distributed anomaly detection in smart grids

Article References: Goudjil, L., Nasri, K., Talhi, C., & Dehli, N. (2026). FED-XID: a novel federated learning based architecture for distributed anomaly detection in smart grids. Cluster Computing, 29(15), Article 866. https://doi.org/10.1007/s10586-026-06677-7

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06677-7

Keywords: federated learning, smart grids, anomaly detection, XGBoost, privacy, cybersecurity, edge computing, electricity theft, machine learning, TCN-LSTM, FED-XID, novel

The problem the team tackles is subtle but consequential. Traditional maintenance plans are built on engineering experience and rules of thumb, and they work well when a single piece of equipment is taken out of service. In such cases the grid generally satisfies the N-1 criterion, meaning one outage does not interrupt the load supply. But when many outages overlap, the topology and operating mode of the system change considerably. Experience-based plans treat each outage in isolation and control risk only through time-window allocation and backup paths. They cannot capture the interaction between simultaneous outages, which may redistribute power flow locally, push equipment beyond its ratings, and eventually force load shedding.

Identifying weak links under multiple outage events is therefore essential for preventing voltage violations and equipment overload. Existing identification methods fall into three groups: those based on complex network theory, which reduce the grid to nodes and edges and examine topological metrics such as betweenness centrality; those based on the operating state, which build an index system from operational and equipment parameters; and comprehensive methods that mix subjective judgment with objective factors. Few of these methods, however, are designed for multiple-outage maintenance scenarios, and the risks that accumulate during such periods are rarely analyzed systematically.

The core of the new method is a multi-time-scale joint optimization framework in which forecast-informed weak-link evaluation, cumulative risk quantification, and schedule optimization are coupled. The large language model serves as an enabling tool inside this framework. Its forecasts make the vulnerability weights time-varying, which most existing risk-based scheduling studies do not consider. The first component is an adaptive load forecasting model in which frequency-domain features extracted by the Fast Fourier Transform and convolutional neural networks are fed into a pre-trained large language model.

Electricity load series contain periodic fluctuations at several time scales, such as daily and weekly periods, and extracting these features is important for predicting load variations during maintenance periods. The normalized load window, containing 168 hourly points, is divided into overlapping patches with a length of 24 and a stride of 12, yielding 13 patches that are transformed into the frequency domain. The frequency bins are grouped into three bands by k-means clustering, reflecting the daily fundamental component, the sub-daily harmonics, and the high-frequency stochastic fluctuations of load series. Convolutional neural networks with a kernel of size 3 and 32 filters then extract high-dimensional features from the three frequency bands, and linear layers fuse the periodic and random characteristics into a single feature map.

The forecasting backbone is a locally deployed GPT-2 model with 117 million parameters, chosen for its generalization capability. The fused features are converted back to the time domain by inverse Fourier transforms and packed into token vectors before entering the large language model, whose backbone consists of stacked multi-head self-attention and feed-forward network components. Rather than adjusting all of these parameters, which would increase training cost and compromise generalization, the team adopts Parameter Efficient Fine Tuning, fine-tuning only specific layers, primarily the Add & Norm and Embedding layers. This approach enhances the model’s accuracy in predicting electricity loads during maintenance periods, which can last from several hours to several days.

Based on the forecast load, the method then evaluates the weak links of the system through the electrical betweenness of lines and nodes. The weight of a generation node is its share of the total available supply capacity, and the weight of a load node is its share of the total forecast demand produced by the adaptive forecasting module. The greater the electrical betweenness of a line, the larger the transmission load it bears during the maintenance period, and the greater the impact if it is disconnected by a fault. On the 35-node test system, a 10 kV distribution network with a peak demand of about 60 MW, five nodes exceed the threshold of 0.7, sitting on the main transfer paths of the feeders, so maintenance events involving them are prioritized in the risk model.

To quantify the cumulative effect of simultaneous outages, the team establishes a Risk Overlapping Degree metric that measures how overlapping maintenance events reduce the supply redundancy of the same object. The risk overlapping degree is large when two overlapping events simultaneously increase the failure probability of one supply source and the dependency on that source. A multi-objective optimization model then minimizes load losses caused by maintenance while evenly distributing maintenance tasks across the month, avoiding concentration on a few days that would leave too little time for risk-prevention measures. The weights of the three objectives, determined through the analytic hierarchy process with dispatching engineers scoring pairwise importance, prioritize operational risk over economic loss, and a sensitivity analysis confirms the conclusion is robust to the weight selection.

The results are striking. The forecasting module achieves a mean absolute percentage error of 2.31%, outperforming ARIMA, LSTM, CNN, and standard Transformer baselines, with ablation tests showing that removing either the LLM backbone or the frequency-domain module degrades performance. The risk overlapping degree falls from 38.1 under the original schedule to 2.6 under the optimized plan, load loss decreases from 57.5 MWh to 49.2 MWh, and workload variance drops from 23.9 to 16.3. With 30 concurrent maintenance tasks, the proposed method keeps the risk overlapping degree at about 16, while the traditional method reaches 198. The day-ahead model then generates risk-prevention measures, combining load transfer through interconnection switches, upstream supply re-dispatch, and feeder topology reconfiguration, which together lower load loss by 28.1% and remove all remaining voltage violations, keeping every node within the range of 0.95 to 1.05 per unit.

Tests on IEEE standard systems up to 118 buses confirm that the method applies to larger grids, with computation times growing from 5.2 seconds to 245.3 seconds as system size increases, still acceptable for offline monthly planning on a standard workstation. The average solution time of the base case is 28.5 seconds, sufficient for practical use. The team notes that future work will extend the method to integrated energy systems, considering the coupling between different energy carriers during concurrent maintenance events and establishing risk-prevention and optimal decision-making models for such systems. For now, the proposed method provides a practical decision-making tool for maintenance scheduling under multiple power outage events, offering grid operators a way to plan hundreds of overlapping outages without gambling on the combined risk they create.

Subject of Research: Power grid maintenance scheduling

Article Title: Optimal decision-making method for power grid maintenance scheduling during multiple power outage events based on large language models

Article References: Lei, W., Chen, K., Su, C., Li, C., Hu, Q., Dai, C., Chen, J., & Li, Y. (2026). Optimal decision-making method for power grid maintenance scheduling during multiple power outage events based on large language models. Frontiers in Energy Research, 14, Article 1889135. https://doi.org/10.3389/fenrg.2026.1889135

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1889135

Keywords: large language models, power grid maintenance, outage scheduling, load forecasting, risk overlapping degree, electrical betweenness, smart grids, GPT-2, Fourier transform, particle swarm optimization, IEEE test systems, risk prevention

The research, led by Zhongyuan Yan and colleagues at Henan Jiuyu Longyuan Energy Service Co., Ltd. in Zhengzhou, China, addresses a persistent blind spot in power system planning. In most existing studies, renewable power forecasting and storage planning are treated as separate problems. Forecasting researchers focus on shaving prediction errors, while storage planners simply plug in predetermined renewable power profiles as fixed inputs. The consequence is that forecasting errors propagate silently into storage sizing decisions, potentially leading to systems that are either too small to absorb surplus renewable generation or too large to justify their capital cost.

The team’s solution begins at the data level with a forecasting model they call VMD-EO-SVM, a hybrid that chains together three distinct techniques. Variational mode decomposition, or VMD, first breaks down the raw wind or solar power time series into a set of narrow-band intrinsic mode functions, each capturing fluctuations at a different frequency scale. This decomposition tames the non-stationary, erratic character of renewable output by separating fast spikes from slow trends, effectively turning one hard prediction problem into several easier ones.

Next comes the equilibrium optimizer, a nature-inspired metaheuristic that mimics the physics of mass exchange in a control-volume system seeking equilibrium. The EO searches for the optimal values of the penalty parameter C and the kernel width gamma in a support vector machine, the third component of the pipeline. SVM regression maps input samples into a high-dimensional feature space where a linear regression function is constructed, using a radial basis function kernel prized for its wide convergence region and high resolution. By letting EO tune these hyperparameters for each decomposed component, the framework avoids the guesswork that typically undermines SVM performance. The component-level forecasts are then reconstructed into a single final prediction.

The benchmark results are striking. Using 15-minute-interval measured wind and solar power data from Belgium covering 2020 to 2021, the VMD-EO-SVM model achieved a root mean square error of 42.60 for solar power with a coefficient of determination of 0.9776, and an RMSE of 71.83 for wind power with an R-squared of 0.9217. Compared against deep learning benchmarks, the hybrid model cut solar RMSE by 38.79 percent relative to LSTM and 55.09 percent relative to CNN, and reduced wind RMSE by 37.95 and 50.68 percent against the same rivals. Even more telling is the component-level ablation: for solar power, RMSE fell from 195.21 for plain SVM to 56.08 for EO-SVM and then to 42.60 for the full VMD-EO-SVM stack, demonstrating that decomposition and parameter optimization perform complementary functions rather than redundant ones.

Notably, the study found that signal decomposition alone does not guarantee improvement. For wind power, VMD-SVM actually failed to outperform the original SVM in RMSE terms. It is only when VMD is paired with EO-optimized SVM that the combination delivers its full benefit, suggesting that the two techniques must work in concert: VMD reduces the complexity of the prediction task while EO ensures the SVM parameters are properly matched to each frequency-separated component.

With reliable forecasts in hand, the second stage of the framework turns to the planning question: how big should a shared energy storage facility be? Unlike storage systems configured independently for individual distributed energy systems, shared energy storage allows multiple generation and consumption entities to draw on a common facility. Because different subsystems have partially complementary charging and discharging patterns, aggregation reduces redundant capacity and improves utilization. The capital burden is also distributed among participants rather than borne by any single user, a mechanism the authors argue is essential for making storage investment economically viable at cluster scale.

The multi-objective planning model minimizes total system cost, which includes the daily-equivalent investment cost of storage, operation and maintenance expenses, thermal generation fuel costs, and penalties for renewable curtailment, while simultaneously maximizing the renewable energy utilization ratio. Constraints enforce power balance at every time step, limit curtailment to predicted renewable output, keep the storage state of charge between 0.1 and 0.9, and respect the ramp-rate and minimum on-off-time requirements of seven thermal units ranging from 55 to 130 MW. The optimization determines both the rated power and the energy capacity of the shared storage system.

The case study results reveal a configuration of 947 MW rated power and 5,615 MWh of energy capacity, corresponding to an energy-to-power ratio of approximately 5.93 hours. That duration is significant: it indicates the storage system is not merely smoothing second-to-second fluctuations but shifting energy across several hours, absorbing midday solar surpluses and overnight wind output for release during peak demand. The optimized plan achieves a renewable energy accommodation rate of 90.3 percent with a total system cost of 66,956,684 CNY, of which 64.3 million represents the daily-equivalent storage investment. Operational simulations confirm that the storage charges during periods of renewable surplus and discharges when generation falls short or load peaks, with the state of charge never straying outside safe bounds.

The thermal fleet plays a supporting role in this choreography. Units 1 and 2 operate stably as base-load generators, while units 3 through 7 act as peaking units that shut down during high renewable generation to cut coal consumption, keeping startup and shutdown costs to just 41,720 CNY. Wind and solar output show clear temporal complementarity, with solar dominating daytime hours and wind carrying the night, allowing the combined system to maintain real-time power balance across all 96 fifteen-minute periods of the scheduling day.

The authors are candid about limitations. The forecasting model relies solely on historical power data and does not explicitly incorporate meteorological variables such as wind speed, solar irradiance, or temperature. While historical series implicitly encode past weather patterns, this indirect representation may prove inadequate when future conditions diverge sharply from historical norms, particularly during extreme weather events. Future work, they note, will integrate numerical weather prediction directly into the forecasting framework and examine how extreme conditions ripple through to storage planning decisions.

Beyond the technical achievements, the study carries a broader message for the energy transition: forecasting and storage planning should be treated as a single coupled problem, not sequential silos. Better forecasts yield more reliable storage sizing, which in turn unlocks higher renewable penetration at lower cost. As wind-solar-thermal clusters proliferate under carbon peaking and neutrality targets worldwide, frameworks that bind prediction quality directly to infrastructure investment may well become the template for how distributed grids of the future are designed.

Subject of Research: Optimal planning of shared energy storage for wind-solar-thermal distributed energy system clusters using improved renewable power forecasting

Article Title: Optimal planning of shared energy storage for wind-solar-thermal distributed energy system clusters based on improved renewable energy forecasting

Article References: Yan, Z., Xue, W., Zhang, Y., Du, X., & Zheng, X. (2026). Optimal planning of shared energy storage for wind-solar-thermal distributed energy system clusters based on improved renewable energy forecasting. Frontiers in Energy Research, 14, Article 1946325. https://doi.org/10.3389/fenrg.2026.1946325

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1946325

Keywords: shared energy storage, renewable energy forecasting, wind power, solar power, variational mode decomposition, equilibrium optimizer, support vector machine, distributed energy systems, multi-objective optimization, renewable energy accommodation, energy storage planning, smart grids

The study, published in Frontiers in Energy Research, confronts a subtle problem that has plagued energy storage planners for years. A stationary battery, no matter how well sized and sited, can only shift energy in time at its one connection point. Yet the operational stress caused by high solar penetration does not sit still. Reverse-power-flow violations, voltage rise, and photovoltaic curtailment migrate across different transformer service areas over the course of a day, as load patterns and solar output change from one neighborhood to the next. A battery bolted to one transformer simply cannot follow that moving target.

Mobile shared energy storage, abbreviated MSES, offers an elegant answer: a single storage unit that can drive or be transported between connection interfaces at different distribution transformers. But mobility is not free. Every relocation consumes traction energy drawn from the battery itself, occupies a full scheduling period during which no charging or discharging can occur, and requires connection-preparation time after arrival before the unit can resume grid service. Previous research on mobile storage often glossed over these engineering frictions, and frequently compared mobile units against stationary batteries that were never properly optimized, making the apparent benefits of mobility difficult to trust.

The new framework introduces a two-part solution. The first element is a diagnostic tool called the effective mobility service index, or MSIeff, computed from the location and magnitude of reverse-power-flow exceedances in a network operating without any storage. The index captures two things at once: how much solar stress exists in total, and how far that stress migrates among service areas over time. If violations stay concentrated at a single transformer, a well-placed stationary battery captures nearly all the available benefit, and mobility adds nothing. If the dominant violations hop from one area to another across the day, mobility becomes genuinely valuable. The index therefore acts as a pre-assessment screen, telling planners before any detailed optimization whether a mobile unit is likely to pay off.

The second element is a rigorously fair comparison design. The team optimized stationary storage and mobile storage within an identical finite candidate domain of power and energy ratings, spanning 100 to 200 kilowatts and 200 to 800 kilowatt-hours, subject to a common two-to-four-hour duration rule. The stationary benchmark’s connection location, rating, and daily dispatch schedule were all independently optimized, eliminating the common trap of benchmarking mobility against a deliberately weak fixed alternative. Both schemes shared the same total boundary for connection-infrastructure investment, and the mobile model explicitly represented travel-induced unavailability, traction consumption, interface limits, state-of-charge dynamics, and end-of-day closure constraints on both energy and location.

Solving the coupled problem required a nested optimization architecture. An outer layer enumerated feasible power-energy configurations and, for the stationary case, candidate fixed locations. For each candidate, an inner 24-hour dynamic program minimized daily operating cost, tracking stored energy, current service area, and connection-preparation status in its state space for the mobile unit. The daily costs were annualized and combined with investment and maintenance costs to produce a single equivalent-annual-cost metric for every configuration, from which the best option in each scheme was selected.

The results, drawn from 30 parameterized scenarios varying solar capacity, load levels, and temporal shifts in photovoltaic output, paint a nuanced picture. Mobile storage reduced photovoltaic curtailment in 16 of the 30 scenarios and lowered equivalent annual cost in 15, achieving both benefits simultaneously in 13. In the remaining cases, the mobility premium in travel, startup, and traction costs ate up the physical gains. Critically, the effective mobility service index proved to be a strong predictor of real-world benefit: its Pearson and Spearman correlation coefficients with realized curtailment reduction reached 0.867 and 0.870, respectively, confirming that a simple no-storage diagnostic can reliably flag the scenarios where a traveling battery will actually earn its keep.

The starkest illustration comes from a representative high-adaptability scenario in which reverse-power-flow violations first emerged in one service area and then migrated to two others. Both storage schemes converged on the same 200-kilowatt, 800-kilowatt-hour rating, isolating the effect of location flexibility alone. The mobile unit departed its depot, spent the morning absorbing surplus solar in the hardest-hit area, and returned home by midday. The result was a 14.54 percent reduction in daily photovoltaic curtailment, a 2.02 percentage-point increase in solar accommodation, and a 2.04 percent drop in equivalent annual cost compared with the optimally sited stationary battery. The cost breakdown showed that savings came from slashing curtailment penalties and network losses, not from cheaper equipment.

Sensitivity analyses sharpened the practical guidance. The mobile scheme’s cost advantage eroded as the unit price of mobile connection interfaces rose, breaking even at roughly 280,000 to 300,000 yuan per connection point and collapsing entirely at about 600,000 yuan, beyond which the optimizer chose no storage at all. Longer connection-preparation times also chipped away at the benefit, cutting the annual cost advantage from about 59,800 yuan to roughly 23,000 yuan when preparation stretched to two hours. An idealized mobility model that ignored travel time and traction energy overstated performance, overstating curtailment reduction and understating cost relative to the engineering-constrained design.

The broader lesson is that mobility is a conditional asset, not a universal upgrade. The authors emphasize that their framework rests on an aggregated voltage-sensitivity model and deterministic representative-day profiles, with only three service areas and a single mobile unit, leaving three-phase unbalanced power-flow validation, year-round uncertainty-aware planning, and multi-unit coordination to future work. Yet the message for distribution planners is already actionable: before investing in either fixed or roaming batteries, measure whether your grid’s solar stress actually wanders. If it does, a single battery on wheels can outperform an optimally placed stationary one; if it does not, the open road is just an expensive detour.

Subject of Research: Coordinated planning and operation of mobile shared energy storage in photovoltaic-rich low-voltage distribution networks

Article Title: Service adaptability assessment and coordinated planning and operation of PV-rich low-voltage active distribution network clusters

Article References: Zhou, S., Zheng, M., Huang, X., Li, Y., Yue, Y., Huang, X., Liang, K., Huang, Y., Yang, L., & Zhang, W. (2026). Service adaptability assessment and coordinated planning and operation of PV-rich low-voltage active distribution network clusters. Frontiers in Energy Research, 14, Article 1963215. https://doi.org/10.3389/fenrg.2026.1963215

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1963215

Keywords: mobile energy storage, photovoltaic curtailment, low-voltage distribution network, reverse power flow, battery storage planning, grid flexibility, voltage rise, energy storage sharing, smart grids, rooftop solar, service adaptability index, distribution transformers

The team, led by Biaolong Su of NARI Technology Co., Ltd. together with colleagues from State Grid Henan Electric Power Company, set out to solve a problem that has long frustrated grid engineers. Composite power quality disturbances, formed by the simultaneous coupling of transient events such as voltage sags and continuous distortions such as flicker, exhibit severe feature overlap and a masking effect in which key discriminative signatures are submerged beneath the dominant waveform. Traditional identification pipelines, which first extract features using tools like the Fourier transform, S-transform, wavelet transform, or empirical mode decomposition and then feed hand-picked features into classifiers such as decision trees or support vector machines, work reasonably well for single disturbances. But they struggle when several disturbances coexist, and their performance degrades sharply as background noise increases.

Deep learning has already been applied to this challenge, with convolutional neural networks, LSTM hybrids, and image-based approaches that convert time series into two-dimensional representations. Yet these methods share a common weakness: when confronted with disturbance types or combinations absent from the training data, they must be retrained with new examples incorporated, which severely limits practical deployment. The new approach takes a different path. It processes raw voltage waveforms directly, without any manual feature engineering or explicit signal transformation, and it is explicitly designed to generalize to unseen composite disturbances through a mechanism the authors call confidence correction.

The architecture is built on a three-dimensional synergistic framework. First, a dual-branch fully convolutional network extracts multi-scale spatial and temporal-frequency representations from the raw time-series signal. The two branches use different depths and kernel sizes: a shallower branch captures local, abrupt features that characterize transient disturbances, while a deeper branch expands the receptive field to capture global variations. Their outputs are concatenated and fused into multi-scale feature maps. Second, a gated recurrent unit models the long-term dynamic evolution of the disturbance across adjacent fundamental cycles, using update and reset gates to control what historical information is retained, which also helps alleviate the vanishing gradient problem that plagues conventional recurrent networks. Third, an attention mechanism computes dynamic weights across the hidden states, concentrating probability mass on the narrow temporal windows where transient anomalies occur while assigning near-zero weight to the long stretches of ordinary sinusoidal waveform that would otherwise dilute the signal.

This structural synergy addresses a fundamental imbalance in disturbance data. Transient events are highly localized, their anomalous signatures confined to brief windows, while continuous disturbances such as harmonics and flicker persist across the entire record. Conventional deep architectures processing such heterogeneous signals tend to let the dominant normal cycles wash out the weak representations of localized transients. By letting the attention layer dynamically highlight the informative segments, the framework neutralizes this information sparsity and learns balanced representations across complex composite patterns.

The most distinctive element, however, is the disturbance recognition enhancement mechanism built on confidence correction. The model treats disturbance identification as a multi-label problem: each of seven disturbance types, including voltage sags, swells, interruptions, flicker, harmonics, impulsive transients, and oscillatory transients, is a label that can be present or absent, and a single signal can carry several labels at once. After the network produces initial confidence scores for each category, an enhanced label distinguishes single from compound disturbances, and calibrated confidence enhancement factors adjust the prediction scores. Transient disturbances with concentrated, localized energy receive an explicit factor of 1.25, while steady-state periodic distortions receive an implicit factor of 1.10. These factors were determined through a constrained grid search on a validation set, with a strict safeguard: an inactive disturbance whose initial confidence falls below the 0.5 decision threshold must never be boosted across that boundary, eliminating false-positive pseudo-disturbances.

To test the system, the researchers synthesized an extensive dataset conforming to IEEE Std 1159 and IEC 61000-4-30 standards, generating 48,000 samples across 48 categories: seven single disturbances, seventeen dual composites, seventeen triple composites, six quadruple composites, and normal signals. Each signal comprised 1,280 sampling points covering ten cycles at 50 hertz, with Gaussian white noise added at signal-to-noise ratios uniformly distributed between 10 and 50 decibels. The results were striking. The model achieved 100 percent accuracy on normal and single disturbances, 99.75 percent on dual composites, 99.71 percent on triple composites, and 98.60 percent on the highly coupled quadruple cases, with Hamming losses as low as 0.0004.

Even more impressive was the zero-shot generalization. In one experiment, the model was trained without ever seeing nine triple disturbance categories and all six quadruple categories, yet it identified the completely unseen quadruple composites with 99.45 percent accuracy. When the training set was expanded to include all triple disturbances, accuracy on unseen quadruple cases rose to 99.83 percent. The authors attribute this combinatorial generalization to the fact that the network learns disentangled, noise-invariant representations of each constituent disturbance signature from lower-order examples, so a quadruple combination is not treated as an unlearnable monolithic pattern. The confidence correction mechanism then calibrates the marginal detection probabilities to produce reliable multi-label outputs for novel configurations.

The method also proved its mettle against noise and rival approaches. Benchmarked against classical pipelines such as wavelet transform with decision trees, S-transform with support vector machines, and empirical mode decomposition with support vector machines, the proposed model maintained 95.3 percent accuracy at a harsh 10 dB signal-to-noise ratio, where the traditional methods collapsed to between 68.5 and 72.1 percent. Against deep learning competitors including pure CNNs, CNN-LSTM hybrids, ResNet with Gramian angular fields, and Transformers, it likewise dominated under heavy noise, while requiring only 4.3 million floating-point operations, 4.5 minutes of training time, and 10 seconds of testing time, a fraction of the computational cost of its rivals. Validation on 1,000 field-recorded waveforms from actual distribution network monitoring terminals yielded an overall average accuracy of 99.7 percent in noise-free conditions and 96.1 percent at 10 dB, including successful identification of triple compound events in real data.

Ablation experiments confirmed that each component earns its place: removing attention dropped accuracy to 97.82 percent, removing temporal modeling to 96.45 percent, and removing the convolutional feature extraction to 95.96 percent, with all differences statistically significant at p below 0.01. The researchers envision next steps that include deploying the lightweight framework on embedded edge hardware such as DSP and FPGA-based monitoring devices for real-time inference, incorporating open-set learning to detect emerging unmodeled disturbance patterns, and extending the single-channel model into wide-area spatiotemporal networks capable of concurrent disturbance identification and fault source localization. For grids racing toward high renewable penetration, a compact model that reads raw waveforms, untangles overlapping disturbances, and generalizes to combinations it has never encountered could become an essential guardian of power quality.

Subject of Research: Deep learning identification of composite power quality disturbances in distribution networks

Article Title: Composite power quality disturbance identification method considering confidence correction

Article References: Su, B., Luo, D., Dai, S., Dong, B., Wang, Y., & Ni, F. (2026). Composite power quality disturbance identification method considering confidence correction. Frontiers in Energy Research, 14, Article 1936537. https://doi.org/10.3389/fenrg.2026.1936537

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1936537

Keywords: power quality disturbances, deep learning, FCN-GRU-Attention, confidence correction, smart grids, distribution networks, renewable energy integration, attention mechanism, multi-label classification, noise robustness, voltage sags, harmonics

The central question the researchers posed is deceptively simple: when every weather sensor on a station is reporting, which of those readings genuinely carry information about incoming solar radiation, and which are noise, redundancy, or artifacts of imperfect instrumentation? Conventional practice often throws every available variable into a machine learning model and lets the algorithm sort it out. That approach can work, but it carries a cost. Models trained on irrelevant or collinear predictors tend to be fragile, harder to interpret, and more likely to break down when a sensor malfunctions or conditions shift outside the training range. In renewable energy applications, where forecasts feed directly into grid scheduling decisions, that fragility is not an academic concern—it is an operational liability.

To address the challenge, the team built a multi-stage evaluation pipeline that deliberately combines three complementary families of analysis. The first stage probes linear structure in the data. The researchers applied three rank- and product-moment correlation methods—Pearson, Spearman, and Kendall—to quantify how tightly each meteorological predictor co-varies with solar irradiance, and they deployed principal component analysis (PCA) to map the dominant axes of variance running through the whole dataset. PCA is a classical dimensionality reduction technique that rotates the data into new coordinates ordered by the amount of variance each captures; when a handful of components account for most of the spread, it signals that many original variables are telling overlapping stories. In this study, the first three principal components together explained 62.88 percent of the total variance, a clear indication of substantial multicollinearity among the sensor readings.

Linear analysis alone, however, cannot capture the full picture. Atmospheric variables interact in ways that are stubbornly nonlinear: humidity modulates cloud formation, which alters irradiance, but the relationship bends and saturates depending on temperature, season, and pressure. To test nonlinear sensitivity, the team trained a bilayer neural network—a shallow architecture with two hidden processing layers—capable of registering interactions that straight-line correlation would miss. Crucially, the neural network served not as a final forecasting product but as a probe: by measuring how prediction error responds when each input is perturbed, the researchers could rank variables by their functional importance rather than their statistical association alone. This dual-track strategy, pairing classical statistics with neural sensitivity testing, forms the hybrid core of the framework.

The third and arguably most novel stage brought explainable artificial intelligence into the mix. The researchers applied SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), two techniques borrowed from the machine learning interpretability literature that decompose a model’s prediction into contributions from each input variable. SHAP, grounded in cooperative game theory, allocates to each feature a Shapley value representing its average marginal contribution across all possible coalitions of predictors, yielding a global picture of which variables drive the model overall. LIME, by contrast, fits simple surrogate models around individual predictions to explain why a specific forecast took the value it did. Combining the two gives both the bird’s-eye view and the close-up, allowing the team to verify that the variables flagged by correlation and neural sensitivity are the same ones the trained model actually leans on when making its estimates.

The results delivered a clear hierarchy of predictors. The most influential single variable turned out to be average battery voltage, abbreviated ABV, which achieved a root mean square error of 0.203 when used to predict solar irradiance. The finding may look surprising at first glance—an electrical quantity appearing to outperform classic atmospheric measurements—but it makes physical sense in context: at a ground-based monitoring station, battery voltage tracks the charging cycle driven by the panel’s own exposure to sunlight, making it a sensitive, integrated proxy for the very quantity being forecast. Relative humidity followed as the second-strongest predictor, with an RMSE of 0.239, consistent with its well-documented role in cloud and aerosol processes that scatter and absorb incoming radiation. Wind direction ranked third at 0.244, plausibly reflecting its association with air mass origin and the transport of moisture and particulates.

The explainability analysis reinforced this ranking from an independent angle. Global SHAP assessment confirmed that ABV and relative humidity contributed most significantly to the model’s predictions, while rainfall emerged as a weak global influence—a result the researchers attribute to its sparse occurrence in the record. Rainfall events, though meteorologically dramatic, occupy so few rows in a typical time series that the model has little opportunity to learn a robust mapping from rainfall to irradiance, and the feature’s Shapley values remain correspondingly small. This kind of diagnostic insight is precisely what the framework is designed to surface: rather than silently retaining a predictor that adds little, reliability-aware selection recognizes that a rarely active sensor contributes maintenance cost and potential failure modes without commensurate predictive benefit.

Beyond the rankings themselves, the study’s methodological contribution lies in its explicit handling of the data-quality hazards that plague real-world meteorological records. The authors frame the problem as one of identifying physically meaningful predictors under conditions that are simultaneously non-normal, multicollinear, and sensor-sensitive—three threats that conventional single-method feature selection handles poorly. Non-normal distributions undermine parametric tests; multicollinearity inflates the apparent importance of redundant variables and destabilizes coefficient estimates; and sensor sensitivity means that measured values may reflect instrument behavior as much as atmospheric reality. By triangulating across correlation families, PCA, neural sensitivity, and interpretable machine learning, the framework forces a candidate predictor to earn its place through convergent evidence before it is accepted into a forecasting model.

The practical implications extend across the renewable energy value chain. More reliable irradiance forecasts translate directly into better solar farm siting, more accurate energy yield assessments for investors, and smoother integration of photovoltaic capacity into smart grids, where dispatch decisions must anticipate the variable output of solar assets. The framework’s emphasis on lean, interpretable predictor sets also lowers computational and maintenance burdens, since models built on a handful of well-justified inputs are cheaper to run and easier to audit than sprawling ones. For operators in data-scarce regions, the finding that a simple, routinely logged variable such as battery voltage carries exceptional predictive power suggests that meaningful forecasting capability may be achievable with minimal instrumentation—a potentially significant result for expanding solar infrastructure across developing regions where dense sensor networks are not yet in place.

The study, which received funding support from a Universiti Sains Malaysia Research University Grant, arrives at a moment when the global build-out of solar capacity is accelerating and the marginal value of forecast accuracy is rising accordingly. It also joins a broader movement in environmental machine learning toward transparency: rather than treating predictive models as black boxes, the authors demonstrate how explanation tools can be woven into the model-building process itself, serving as a quality check on the science as much as a window on the model. For a field where forecasting errors carry real economic and grid-stability consequences, the message is clear—the path to trustworthy solar prediction runs not through bigger models with more inputs, but through disciplined, reliability-aware scrutiny of the measurements that feed them.

Subject of Research: Hybrid feature selection for ground-based solar irradiance prediction using meteorological data and explainable machine learning

Article Title: Hybrid feature selection framework for reliable solar irradiance prediction using ground-based meteorological data

Article References: Maiwada, A. D., Ahmad, M. I., Abba, S. I., & Shakir, M. A. (2026). Hybrid feature selection framework for reliable solar irradiance prediction using ground-based meteorological data. Discover Sustainability. https://doi.org/10.1007/s43621-026-04977-x

Image Credits: AI Generated

DOI: 10.1007/s43621-026-04977-x

Keywords: solar irradiance, feature selection, machine learning, SHAP, LIME, PCA, neural networks, renewable energy, meteorology, forecasting, multicollinearity, smart grids

The core challenge is one of sparse sensing. In an ideal world, every node in a hydrogen transportation network or an IoT-enabled smart grid would stream continuous, reliable measurements — pressures, flow rates, injection levels, power draws. In practice, distributed edge devices are often battery-constrained, intermittently connected, or simply too expensive to deploy at full density. The result is a patchwork of incomplete and irregular observations, and the underlying energy flow structure of the system becomes hidden. Traditional physics-based models, which rely on differential equations and precise knowledge of system parameters, demand full observability to function at all. Purely data-driven models, on the other hand, can interpolate from whatever data they see but routinely violate the conservation laws and transport physics that govern real energy systems, producing reconstructions that look plausible statistically but are physically nonsensical.

Liang’s framework, SPEL-VFL — short for physics-embedded latent energy flow learning — is designed specifically for this gap. It combines three tightly coupled components. The first is a sparse-aware encoder, a neural module that learns to extract maximum information from fragmentary, irregularly sampled measurements, effectively treating missingness as a first-class feature of the input rather than an annoyance to be imputed away. The second is a variational latent structure model, which places the hidden energy transmission relationships of the network into a probabilistic latent space. Rather than guessing a single fixed topology of how energy moves between nodes, the model maintains a distribution over plausible latent structures, capturing the uncertainty that sparse sensing inevitably introduces.

The third and arguably most important component is a physics-consistency regularization mechanism. During training, the framework penalizes solutions that drift away from physical law — the same philosophy that underpins the rapidly growing field of physics-informed neural networks, which gained prominence with the 2019 work of Raissi, Perdikaris, and Karniadakis on embedding partial differential equations into deep learning objectives. In SPEL-VFL, this means that the latent energy flows the model infers must remain consistent with the governing dynamics of the physical network, whether that network is an electrical distribution grid or a hydrogen pipeline system with injection points, line-pack storage effects, and pressure-driven transport. The regularization acts as a scientific conscience for the neural network, steering it away from mathematically convenient but physically impossible reconstructions.

The setting for this work is the broader world of IoT-enabled cyber-physical energy systems, and the study appeared in a special issue on Machine Learning and Big Data Analytics for IoT Security and Privacy, reflecting the growing recognition that energy infrastructure monitoring is as much a data and security problem as a thermodynamic one. State estimation — the classical task of inferring the true operating state of a network from noisy measurements — and missing data reconstruction have historically been treated as separate, supporting processes. The paper argues that under sparse sensing these tasks become tightly coupled and mutually dependent: you cannot reliably estimate states without reconstructing missing data, and you cannot reconstruct missing data without a coherent estimate of the system state. Conventional pipelines that handle them sequentially often fail to recover the underlying energy flow structure at all.

What makes the work particularly timely is its explicit application to hydrogen energy transportation and hydrogen intelligent mobility systems. As countries scale up hydrogen production through power-to-gas conversion and build dedicated pipeline networks, operators face a monitoring environment far less mature than the electrical grid. Hydrogen leakage detection remains an active area of sensor research, and the dynamic behavior of natural gas networks into which hydrogen is blended — including line-pack effects where gas stored in the pipeline itself buffers supply and demand — adds layers of complexity that fixed physics models struggle to track in real time. A learning framework that can infer hidden flow structures from sparse IoT measurements, while respecting the physics of gas transport, offers a way to bring hydrogen infrastructure up to the observational standards of modern smart grids.

The technical machinery behind SPEL-VFL draws on several strands of recent machine learning research. Variational autoencoders, popularized in the machine learning literature for learning probabilistic latent representations, provide the backbone for the latent structure model, allowing the framework to reason about uncertainty rather than committing to a single point estimate. Graph neural networks, which have become the standard tool for learning on network-structured data, inform the way the model captures relationships between nodes in the energy system. Attention mechanisms, drawn from the transformer architecture that reshaped natural language processing, offer a means of weighting which sparse measurements matter most for a given reconstruction. The framework also connects to classical techniques: matrix completion methods from convex optimization, particle filters for nonlinear Bayesian tracking, and recurrent neural network approaches for multivariate time series with missing values all appear in the paper’s intellectual lineage.

According to the experimental results reported in the study, SPEL-VFL achieves superior reconstruction accuracy compared with competing approaches, demonstrates stronger robustness to missing data, and produces reconstructions with improved physical plausibility. That last metric deserves emphasis. In energy systems, a reconstruction that is accurate on average but violates conservation of energy at individual nodes can mislead operators into unsafe or economically suboptimal decisions. By enforcing physics-consistency during learning rather than as an afterthought, the framework ensures that its outputs can be trusted not just statistically but physically — a property that matters enormously when the system in question is a pressurized hydrogen pipeline or a distribution grid serving critical loads.

The implications extend beyond hydrogen. Smart grids, district heating networks, integrated energy systems that couple electricity with gas and heat, and the growing fleet of IoT devices monitoring them all face the same fundamental tension: the physics is well understood, but the observations are sparse, noisy, and irregular. A framework that learns latent energy flows under exactly these conditions could improve state estimation for distribution networks, support anomaly and leakage detection, and enable more aggressive deployment of low-cost sensing by compensating algorithmically for what the hardware cannot measure. It also speaks to the security dimension highlighted by the special issue in which the paper appears — resilient reconstruction from partial observations is a natural defense against sensor failures, communication outages, and adversarial data manipulation.

There are, of course, open questions. The study’s datasets are available from the corresponding author upon reasonable request, and real-world deployment at the scale of a national hydrogen network would demand extensive validation across diverse operating conditions, sensor failure modes, and adversarial scenarios. The tension between model flexibility and physical fidelity — how strongly to regularize, and which physical laws to embed — will need careful tuning for each application domain. But the direction of travel is clear. As the energy transition multiplies the number of coupled physical networks that must be monitored in real time, and as hydrogen moves from laboratory curiosity to industrial commodity, the ability to learn the hidden flow of energy from sparse, imperfect sensors while never violating the laws of physics may prove to be one of the most consequential applications of machine learning in the infrastructure of the coming decades. Liang’s SPEL-VFL framework offers a concrete, technically grounded step in that direction, and a preview of how the next generation of energy systems will be seen — even when half their sensors are dark.

Subject of Research: Physics-embedded latent machine learning for hydrogen energy flow reconstruction under sparse IoT sensing

Article Title: Physics-embedded latent hydrogen energy flow learning under sparse sensing conditions

Article References: Liang, X. (2026). Physics-embedded latent hydrogen energy flow learning under sparse sensing conditions. Neural Computing and Applications, 38(19), Article 787. https://doi.org/10.1007/s00521-026-12320-8

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12320-8

Keywords: hydrogen energy, sparse sensing, physics-informed neural networks, energy flow modeling, state estimation, variational autoencoders, IoT, smart grids, graph neural networks, missing data reconstruction, hydrogen pipelines, cyber-physical energy systems

At its core, the book explores the integration of high-performance computing and artificial intelligence as complementary engines of sustainable digital transformation. High-performance computing, or HPC, refers to the aggregation of computing power across thousands of processors working in parallel, enabling scientists and engineers to tackle problems far beyond the reach of ordinary desktop machines. These systems simulate climate dynamics, model molecular interactions for drug discovery, optimize energy networks and process enormous datasets generated by satellites and sensors. Artificial intelligence, meanwhile, supplies the analytical layer that can extract patterns and actionable insights from the flood of data such simulations and sensors produce. The volume argues that when these two domains are deliberately combined with cloud computing platforms, organizations can scale their computational capacity up or down as needed, avoiding the waste associated with permanently provisioned hardware that sits idle much of the time.

The energy dimension of this convergence is among the most pressing themes the editors address. Large-scale computing facilities consume substantial electricity, both to power their processors and to cool the heat they generate. As artificial intelligence models grow larger and training runs stretch across weeks on vast GPU clusters, the carbon footprint of the digital economy has come under increasing scrutiny. The book positions HPC and AI not merely as sources of this footprint but as instruments for reducing it. Computational models can identify inefficiencies in manufacturing processes, optimize logistics routes to cut fuel consumption, and accelerate the design of new materials for batteries, solar cells and carbon capture technologies. In this framing, every kilowatt-hour spent on intelligent computation is an investment that can yield disproportionate savings in energy, emissions and materials elsewhere in the economy.

One of the most concrete application areas the volume examines is agriculture. AI- and AIoT-enabled sustainable agriculture, sometimes called smart farming, combines internet-connected sensors deployed across fields with machine learning systems that interpret the resulting data streams. Soil moisture probes, weather stations, drone-mounted cameras and satellite imagery together generate a continuous picture of crop health that no human observer could assemble manually. Machine learning models trained on this data can recommend precise irrigation schedules, detect pest infestations before they spread, and predict yields with enough accuracy to guide planting and harvesting decisions. The result, as the book describes it, is a form of farming that uses less water, applies fewer chemicals and produces less waste, all while maintaining or improving output. For a world facing both population growth and climate volatility, such precision techniques represent one of the most tangible ways computational technology contributes to sustainability.

The protection of critical cyber-physical infrastructure receives equally detailed treatment. Smart power grids exemplify the category: electricity networks that use digital sensors, automated controls and two-way communication to balance supply and demand in real time, particularly as intermittent renewable sources such as wind and solar are added to the generation mix. These systems are extraordinarily valuable and, by the same token, extraordinarily attractive targets. A successful cyberattack on a national grid could disrupt hospitals, water treatment plants, transportation and communications simultaneously. The book surveys cybersecurity solutions designed for such environments, including frameworks for risk assessment, continuous monitoring and the auditing of systems where a single vulnerability can cascade into physical consequences. Artificial intelligence plays a defensive role here as well, with anomaly detection models trained to recognize the subtle signatures of intrusions in network traffic before human operators could possibly respond.

Cloud security and data protection form a third pillar of the volume’s coverage. As organizations migrate workloads and sensitive data to cloud platforms, the traditional perimeter-based model of security dissolves. Data now crosses organizational, national and jurisdictional boundaries as a matter of routine, raising questions about who is responsible for protecting it and under which legal regime. The book addresses strategies for cloud security, risk assessment and auditing, alongside the governance frameworks needed to keep data handling accountable. These concerns are not abstract. Breaches of cloud-hosted databases have exposed the personal information of hundreds of millions of people, and regulators worldwide have responded with increasingly strict data protection requirements. The volume’s treatment of these topics reflects a growing consensus that security and sustainability are intertwined: infrastructure that cannot be trusted cannot be relied upon to deliver long-term environmental or economic benefits.

Beyond the technical chapters, the editors devote significant attention to the ethical, governance and policy considerations surrounding AI-driven computing. Intelligent systems increasingly make or inform decisions with real consequences for people, from credit assessments to medical diagnoses to the operation of autonomous machinery. Questions of algorithmic bias, transparency, accountability and the concentration of computational power in a small number of corporations are now inseparable from questions of system design. The book engages with these issues as part of a holistic view of sustainable computing, arguing that a digital infrastructure is only genuinely resilient if the societies that depend on it trust its operation and its oversight. Policy frameworks, the volume suggests, must evolve as quickly as the underlying technology if the benefits of AI and HPC are to be distributed broadly rather than captured narrowly.

The readership the editors envision is correspondingly broad. Researchers, graduate students, computer scientists and professionals working in artificial intelligence, high-performance computing, cloud computing, cybersecurity and sustainable computing are all named as the intended audience. This interdisciplinary reach reflects the nature of the challenge itself. No single discipline can deliver sustainable digital infrastructure: it requires computer architects who understand energy-efficient hardware, data scientists who can build trustworthy models, security specialists who can defend distributed systems, and domain experts in agriculture, energy and public policy who know where computational tools can do the most good. By gathering these perspectives in a single reference, the book functions as a bridge between communities that have often worked in parallel rather than in concert.

The volume is edited by Jagdish Chandra Patni and Nilesh Bhaskarrao Bahadure, who bring together research and professional perspectives on emerging computational technologies, artificial intelligence, cybersecurity and sustainable digital systems. Their editorial approach emphasizes the convergence of previously separate fields, treating sustainability not as an afterthought to computing but as a design objective that shapes architecture, software, security and governance alike. In an era when the digital sector’s share of global electricity consumption continues to climb even as the world races to decarbonize, that reframing carries real weight. The book’s central message is that the tools of computation and the goals of sustainability are not adversaries to be balanced but allies to be integrated, provided the engineering, security and ethical foundations are laid deliberately. For researchers and practitioners navigating that integration, the volume arrives as a timely and comprehensive map of the terrain.

Subject of Research: Convergence of high-performance computing, artificial intelligence and cloud technologies for sustainable and secure digital infrastructure

Article Title: High performance computing and sustainability through artificial intelligence

Article References: High performance computing and sustainability through artificial intelligence. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: high-performance computing, artificial intelligence, cloud computing, sustainability, cybersecurity, smart agriculture, AIoT, smart grids, data protection, carbon reduction, digital infrastructure, AI governance