Separate LSTM pathways improve single-household load forecasting

Researchers at Tallinn University of Technology have developed a specialized deep learning architecture that improves the accuracy of day-ahead electricity demand forecasts for individual homes. The study, published in Discover Artificial Intelligence, introduces a multi-input, multi-feature Long Short-Term Memory (MIMF-LSTM) framework designed to handle the complex, nonlinear nature of residential consumption. By processing daily and weekly historical patterns through independent recurrent pathways, the model achieved the lowest mean root mean squared error among the evaluated methods. This approach addresses a significant challenge in power systems engineering, where predicting the demand of a single household is considerably more difficult than forecasting aggregated system-level loads due to the absence of smoothing effects from multiple consumers.

The core innovation of the MIMF-LSTM framework lies in its explicit separation of temporal scales. Instead of concatenating all historical data into a single input sequence, the model uses two independently parameterized LSTM encoders. One pathway processes the immediately preceding 24-hour profile, capturing recent daily behavior, while the other processes the corresponding 24-hour profile from one week earlier, capturing recurrent weekly patterns. These two latent representations are then fused with cyclic calendar features, such as the day of the week and day of the year, to generate a 96-step forecast at 15-minute resolution. This design allows the network to learn distinct representations for recent consumption changes and periodic behavioral recurrence without mixing them within a single latent space.

The study was conducted using approximately one year of high-resolution measurements from a single residential household in Tallinn, Estonia. The dwelling, a two-story house occupied by two adults and one child, provided data from 29 August 2021 to 29 August 2022. Electricity demand was measured at 15-minute intervals, resulting in over 35,000 raw observations. The total household load was calculated by combining alternating-current and direct-current channels, ranging from 0.0856 kW to 19.6801 kW with a mean of 1.2480 kW. The dataset exhibited strong seasonal fluctuations, with peak demand occurring during winter months due to electric heating, and daily peaks typically between 17:00 and 20:00 when household occupancy was highest.

To ensure a controlled comparison, the proposed model was evaluated against a diverse set of benchmark algorithms, including seasonal-naive models, Ridge regression, Random Forest, XGBoost, CatBoost, standard LSTM, GRU, CNN-LSTM, and attention-LSTM. All models were provided with the same daily and weekly historical information, meteorological variables, and chronological data partitions. For the single-stream neural baselines, the daily and weekly profiles were concatenated into a common input sequence, whereas the MIMF-LSTM processed the same information using two separate pathways. This protocol isolates the effect of the temporal encoding architecture rather than differences in access to historical data. The independent test period consisted of 89 non-overlapping day-ahead forecasts, covering the summer months from June to August 2022.

Across three independent training runs, the MIMF-LSTM achieved a mean RMSE of 0.848 ± 0.005 kW and a mean absolute error of 0.464 ± 0.006 kW. It also recorded a mean absolute percentage error of 63.86 ± 2.69%, a metric that is often inflated in residential forecasting due to low-load periods where small absolute errors result in large percentage deviations. The model achieved the lowest mean RMSE among all evaluated methods. Ridge regression produced a marginally lower MAE of 0.463 kW, but statistical analysis showed that the differences between Ridge regression and the MIMF-LSTM were not significant. This suggests that a substantial portion of the predictable structure in residential demand can be captured by linear combinations of lag-aligned observations.

Statistical significance testing using the Wilcoxon signed-rank test with Holm correction confirmed that the MIMF-LSTM significantly reduced both daily RMSE and MAE relative to GRU, standard LSTM, attention-LSTM, CNN-LSTM, Random Forest, XGBoost, and CatBoost. The improvement was particularly notable compared to the standard single-stream LSTM, where the MIMF-LSTM reduced RMSE by approximately 2.17% and MAE by 11.49%. Since both architectures received identical input information, these results indicate that the independent encoding of daily and weekly temporal contexts provides a more effective representation than requiring a single recurrent encoder to learn both relationships simultaneously.

An ablation study was conducted to quantify the contribution of each temporal pathway. The results demonstrated that the combined daily–weekly configuration outperformed both the daily-only and weekly-only variants. The weekly-only model performed better than the daily-only model, indicating that the corresponding profile from seven days earlier contains useful information for this household. However, the combination of recent daily conditions and weekly recurrence yielded the lowest errors. When a third pathway representing a 28-day lag was added, the model complexity increased by approximately 34.1%, but forecasting accuracy did not improve. In fact, RMSE and MAE increased slightly, suggesting that additional long-lag pathways may introduce redundant or weakly relevant information that does not justify the added computational cost.

The forecasting error was also analyzed across the 96-step horizon to understand how prediction difficulty varies throughout the day. Since all test forecasts were issued at midnight, the forecast horizon directly corresponds to the time of day. The lowest horizon-specific RMSE occurred in the early morning, when residential demand was relatively stable. The largest errors occurred around 14:30, coinciding with periods of higher and more irregular household activity. This pattern indicates that forecast difficulty is strongly time-dependent and that residential demand peaks remain substantially more difficult to predict than low-load periods. All evaluated models struggled to accurately reproduce abrupt, short-duration high-demand events, such as a peak of 4.44 kW observed on 20 August 2022, which was significantly underestimated by the neural networks.

The study highlights several important limitations that must be considered when interpreting the results. First, the findings are based on data from a single household, which limits the generalizability of the conclusions to other buildings or occupants. Residential demand patterns vary considerably across different households, climates, and usage behaviors. Second, the independent test period was concentrated in the summer months, meaning the model’s performance during winter, when electric heating drives much higher and more variable demand, was not independently established. Third, the model used historical meteorological observations from the lagged input profiles rather than operational weather forecasts for the prediction day. In practical deployment, weather-forecast uncertainty would likely affect performance, and the model would require periodic retraining to adapt to changes in occupancy or appliance ownership.

Despite these limitations, the study provides valuable insights into the design of deep learning architectures for residential load forecasting. The results suggest that explicitly separating daily and weekly temporal information can improve forecasting accuracy without requiring additional long-term input pathways. The MIMF-LSTM architecture achieves competitive performance with fewer trainable parameters than standard LSTM or attention-based models, making it a potentially efficient solution for residential energy management applications. Future research should evaluate this framework using multi-household and multi-year datasets, rolling-origin validation, and operational weather forecasts to better characterize its robustness and practical utility in diverse residential settings.

Beyond electricity measurements, the experimental setup integrated 8,784 hourly meteorological records spanning the same monitoring window. The research team selected four specific environmental features: air temperature, relative humidity, wind speed, and precipitation. Sea-level and station barometric pressures were deliberately omitted from the feature matrix to eliminate redundant data. Because the weather observations were logged hourly while electricity meters sampled at fifteen-minute intervals, the researchers mapped meteorological inputs to the finer resolution using linear time interpolation. The wind speed record contained twenty-five missing hourly observations, which were resolved through linear interpolation between adjacent valid points, supplemented by forward or backward filling strictly at sequence boundaries so that no missing values remained.

To prepare these streams for modeling, the authors applied a logarithmic transformation to the raw electricity load to compress extreme demand spikes while preserving non-negative values. All continuous input features and the log-transformed target were standardized using the mean and standard deviation calculated exclusively from the training dataset, preventing future data leakage into the validation or testing phases. Chronological splitting divided the year into training targets concluding before April 1, 2022, a validation window running through May 31, 2022, and the summer testing phase. Forecast origins were generated every two hours during training and validation, producing 2,232 training examples and 720 validation examples while preserving temporal sequencing.

The neural models were optimized using the AdamW algorithm with a learning rate of 0.001, a weight decay coefficient of 0.00001, and gradient norm clipping at 1.0 to prevent training instability. Mini-batches contained 64 samples during training and 128 samples during evaluation, with early stopping triggered if validation loss failed to improve across six consecutive epochs within a 35-epoch maximum. Structurally, each of the two recurrent pathways in the framework employed a 32-unit LSTM encoder. Their combined 64 latent features merged with four cyclic sine and cosine calendar values before routing through a 64-neuron rectified linear activation layer with ten percent dropout.

Subject of Research: Multi-input LSTM architecture for day-ahead residential electricity load forecasting

Article Title: Improved residential load forecasting using multi-feature, multi-layer LSTM exploiting trend and periodic patterns in historical data

Article References: Shabbir, N., Shahid, A., & Rosin, A. (2026). Improved residential load forecasting using multi-feature, multi-layer LSTM exploiting trend and periodic patterns in historical data. Discover Artificial Intelligence, 6(1), Article 1851. https://doi.org/10.1007/s44163-026-02729-4

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02729-4

Keywords: residential load forecasting, LSTM architecture, day-ahead prediction, electricity demand, deep learning, time series analysis, Tallinn University of Technology, smart grid, energy management, recurrent neural networks, ablation study, meteorological variables, high-resolution data, power systems, statistical significance, seasonal patterns, forecasting accuracy, Estonia, machine learning, household consumption