Machine Learning Model Boosts Microgrid Power Forecasting Accuracy in Simulation

Researchers report that a hybrid machine learning framework combining swarm-intelligence optimization with a Random Forest ensemble can forecast power generation in IoT-monitored manufacturing microgrids with higher accuracy than standard baselines — at least in simulation. The study, published as open access research in Discover Internet of Things, describes an Adaptive Tunicate Swarm Optimized Random Forest (ATSO-RF) that reached a coefficient of determination (R²) of 96.85 percent, a mean absolute error (MAE) of 3.01, and a root mean square error (RMSE) of 4.81 on a held-out test set. By comparison, a Light Gradient Boosting Machine (LGBM) baseline achieved R² of 84 percent, MAE of 3.93, and RMSE of 5.77, while a conventional Random Forest reached R² of 95.2 percent but with a much larger MAE of 22.8 and RMSE of 31.7.

The work addresses a practical tension in modern power systems. Microgrids — small networks of generation resources and loads that can operate connected to the main grid or in isolation — increasingly depend on variable renewable sources such as solar and wind. That variability complicates generation forecasting, control, and management, particularly in manufacturing settings where power reliability carries direct economic consequences. The authors argue that existing IoT-enabled monitoring systems suffer from limited forecasting accuracy, inefficient hyperparameter optimization, and inadequate adaptation to dynamic renewable energy conditions, and they position their framework as a response to those gaps.

The methodology follows five sequential stages. First, raw IoT sensor data are collected from the microgrid. Second, Z-score normalization standardizes all features, converting measurements from heterogeneous sensors onto a common scale by expressing each value as a number of standard deviations from the mean. Third, Linear Discriminant Analysis (LDA) reduces dimensionality from 13 features to 2 discriminant components. Fourth, the ATSO-RF model is trained on the processed features to forecast power generation. Fifth, performance is evaluated using R², MAE, and RMSE. The implementation ran in Python 3.10.1 on a Windows 11 workstation with an Intel Core i7 processor and 8 GB of RAM.

The dataset deserves particular attention, because it shapes what the results can and cannot show. The researchers compiled a primary-source dataset of 8,760 hourly observations — one full year — covering 13 parameters including temperature, solar radiation, wind speed, humidity, solar power, wind power, battery storage, power demand, power supply, grid supply, error, efficiency, and cost reduction. However, the paper states that the data were generated through a systematic protocol drawing on 43 years of concurrent wind and solar generation modeling over renewable-energy plant locations in the contiguous United States, with plant specifications drawn mainly from U.S. Energy Information Administration Form 860. No conventional physical calibration was performed because the data come from regional atmospheric climate-model simulations and power-generation models. Validation instead compared the modeled output against independent records: the dataset showed less than 5 percent bias versus EIA-923 monthly generation and less than 7 percent bias versus hourly generation from balancing authorities.

The optimization component is where the study claims its main novelty. The tunicate swarm algorithm is a metaheuristic that simulates jet propulsion and swarm aggregation behavior; here it searches for the best Random Forest hyperparameters — number of estimators, maximum depth, and maximum features. Each candidate configuration is scored by a fitness function equal to the average MAE across five cross-validation folds, and a dynamic step size balances broad exploration early in the search against focused exploitation later, with search boundaries recalculated each iteration from the current population extremes. After convergence, the selected configuration of 150 estimators, a maximum depth of 30, and a maximum feature fraction of 0.7 was used to train the final ensemble on the complete training set. The authors note that this optimization adds computational overhead relative to a standalone Random Forest, but that the burden is controlled by the small population size of 10 agents, a maximum of 20 iterations, and only three tuned parameters. Because dedicated wall-clock training-time measurements were not recorded, the study does not claim any specific runtime reduction.

The experimental setup used a stratified 80/20 train/test split preserving class distribution across three labeled energy states — fault, normal, and high-efficiency — yielding 7,008 training and 1,752 test samples. All results were averaged over 10 independent runs with different random seeds. Statistical significance was assessed with a paired Wilcoxon signed-rank test over 10 independent random train/test splits, in which ATSO-RF achieved significantly lower MAE and RMSE than LGBM (p < 0.05) and RF (p < 0.01). An ablation study comparing RF alone, RF with ATSO, RF with LDA, and the full combination indicated that both the optimization step and the LDA preprocessing contributed independently to the performance gains.

Beyond forecasting accuracy, the paper reports a set of operational indicators: energy efficiency of 86 percent, a fault detection rate of 93 percent, system uptime of 94.5 percent, energy savings of 20 percent, energy loss reduction of 10 percent, and data transmission efficiency of 95 percent. The authors explicitly caution that these operational figures describe system-level outcomes rather than predictive accuracy, and that the forecasting metrics should not be interpreted as direct measures of efficiency or uptime. Energy efficiency, for instance, is defined as the ratio of useful power delivered to the load versus total available or generated power, evaluated from the model’s predicted generation rather than measured on physical hardware.

The study is candid about its limits. The current work rests entirely on simulation data and has not been validated on a physical microgrid hardware testbed. Real-world deployment, the authors write, may introduce challenges not represented in the dataset, including sensor noise, communication latency, cybersecurity constraints, and non-stationary load profiles. Variations in sensor accuracy, missing data, equipment faults, weather conditions, load fluctuations, and grid disturbances could all affect forecasting performance in practice. The paper identifies physical or hardware-in-the-loop validation as essential for confirming whether the observed forecasting performance and system-level benefits can be maintained during practical operation, and lists future work priorities including hardware validation, dataset extension to cover seasonal variability, fault events, and multi-site deployments, integration of LSTM-based deep learning hybrids for temporal modeling, and federated learning for privacy-preserving distributed training across microgrid nodes.

Within the broader research landscape, the authors situate their contribution against recent IoT-enabled microgrid studies — including work on IoT-aided SCADA systems for renewable energy monitoring, cost-effective smart household distribution, and reviews of communication protocols such as MQTT, Zigbee, LoRaWAN, and 5G for smart-grid deployments. They also contrast ATSO-RF with other hybrid swarm-and-ensemble approaches, such as particle swarm optimization of LSTM-XGBoost models and bee swarm optimization combined with LSSVM, arguing that their framework’s distinctiveness lies in integrating swarm-based hyperparameter tuning, Z-score normalization, and LDA reduction of 13 heterogeneous IoT indicators into a single adaptive forecasting pipeline. If the promised hardware validation materializes, the approach could offer manufacturing microgrids a comparatively lightweight path to more accurate renewable generation forecasting; for now, the evidence remains confined to modeled data and simulated conditions.

Subject of Research: Machine learning-based power generation forecasting for IoT-monitored renewable energy microgrids

Article Title: Research on microgrid power supply equipment system based on intelligent IoT remote monitoring and management

Article References: Huang, Z., He, L., Li, K., & Chen, H. (2026). Research on microgrid power supply equipment system based on intelligent IoT remote monitoring and management. Discover Internet of Things, 6(1), Article 194. https://doi.org/10.1007/s43926-026-00522-4

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00522-4

Keywords: microgrid, Internet of Things, machine learning, random forest, swarm optimization, renewable energy, power forecasting, LDA, energy management, manufacturing, simulation study, Research