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

Accurate day-ahead forecasting of individual residential electricity demand remains challenging because household consumption is highly variable, nonlinear, and influenced by recurrent behavioral and environmental patterns. This study proposes a multi-input, multi-feature Long Short-Term Memory (MIMF-LSTM) framework for direct 24 hour residential load forecasting at 15 minute resolution. The proposed architecture independently processes two lag-aligned temporal profiles: the immediately preceding 24 hour period and the corresponding 24 hour period one week earlier. Each pathway contains total household load and historical meteorological variables, including air temperature, relative humidity, wind speed, and precipitation. The resulting representations are fused with cyclic calendar features to generate a 96-step day-ahead forecast. The framework was evaluated using approximately one year of measurements from a residential household in Tallinn, Estonia. Under a controlled experimental protocol, all benchmark models were provided with the same daily and weekly historical information, input variables, chronological data partitions, and forecasting horizon. Comparisons included seasonal benchmarks, Ridge regression, Random Forest, XGBoost, CatBoost, LSTM, GRU, CNN-LSTM, and attention-LSTM. Across three independent runs, the proposed MIMF-LSTM achieved an RMSE of 0.848 ± 0.005 kW, an MAE of 0.464 ± 0.006 kW, and a MAPE of 63.86 ± 2.69%, yielding the lowest mean RMSE among the evaluated models. Ridge regression achieved a marginally lower MAE of 0.463 kW, although the difference was not statistically significant. Paired daily comparisons showed that MIMF-LSTM significantly reduced both RMSE and MAE relative to GRU, standard LSTM, attention-LSTM, CNN-LSTM, Random Forest, XGBoost, and CatBoost after Holm correction. Ablation analysis further demonstrated that the combined daily–weekly configuration outperformed the individual pathways, whereas adding a 28-day pathway increased model complexity without improving forecasting accuracy. The results indicate that explicit separation of daily and weekly temporal information can improve day-ahead residential load forecasting without requiring additional long-term input pathways.

Authors

Institutions

Publication Details

Journal
Discover Artificial Intelligence
Published
2026-11-12
DOI
https://doi.org/10.1007/s44163-026-02729-4
Primary Topic
Energy Load and Power Forecasting
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

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

Argo Rosin, Arqum Shahid, Noman Shabbir
Discover Artificial Intelligence
Energy Load and Power Forecasting
article

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

Argo Rosin, Arqum Shahid, Noman Shabbir
article en

Abstract

Accurate day-ahead forecasting of individual residential electricity demand remains challenging because household consumption is highly variable, nonlinear, and influenced by recurrent behavioral and environmental patterns. This study proposes a multi-input, multi-feature Long Short-Term Memory (MIMF-LSTM) framework for direct 24 hour residential load forecasting at 15 minute resolution. The proposed architecture independently processes two lag-aligned temporal profiles: the immediately preceding 24 hour period and the corresponding 24 hour period one week earlier. Each pathway contains total household load and historical meteorological variables, including air temperature, relative humidity, wind speed, and precipitation. The resulting representations are fused with cyclic calendar features to generate a 96-step day-ahead forecast. The framework was evaluated using approximately one year of measurements from a residential household in Tallinn, Estonia. Under a controlled experimental protocol, all benchmark models were provided with the same daily and weekly historical information, input variables, chronological data partitions, and forecasting horizon. Comparisons included seasonal benchmarks, Ridge regression, Random Forest, XGBoost, CatBoost, LSTM, GRU, CNN-LSTM, and attention-LSTM. Across three independent runs, the proposed MIMF-LSTM achieved an RMSE of 0.848 ± 0.005 kW, an MAE of 0.464 ± 0.006 kW, and a MAPE of 63.86 ± 2.69%, yielding the lowest mean RMSE among the evaluated models. Ridge regression achieved a marginally lower MAE of 0.463 kW, although the difference was not statistically significant. Paired daily comparisons showed that MIMF-LSTM significantly reduced both RMSE and MAE relative to GRU, standard LSTM, attention-LSTM, CNN-LSTM, Random Forest, XGBoost, and CatBoost after Holm correction. Ablation analysis further demonstrated that the combined daily–weekly configuration outperformed the individual pathways, whereas adding a 28-day pathway increased model complexity without improving forecasting accuracy. The results indicate that explicit separation of daily and weekly temporal information can improve day-ahead residential load forecasting without requiring additional long-term input pathways.

Discover Artificial IntelligenceVol. 6(1)
Tallinn University of Technology (EE)
Affordable and clean energy, Climate action
Energy Load and Power Forecasting
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon