Hierarchical BiLSTM-LSTM Network for Short-Term Load Forecasting

Accurate residential load forecasting is essential for demand response in the evolving intelligent and interactive power system framework. Short-term load forecasting for individual electric customers has become increasingly critical for future grid planning and operations. However, predicting the electric load of a single consumer presents significant challenges due to the highly stochastic nature of individual usage patterns. This study proposes an enhancement to the widely adopted long shortterm memory (LSTM) networks by incorporating hierarchical Bi-directional LSTM (BiLSTM) and LSTM layers, and comprehensive feature engineering to improve model performance in capturing individual consumer behavior. We optimize feature selection using the chi-squared test, and the proposed approach, compared to its counterparts, demonstrates improvements in prediction accuracy with a reduction of $8.31 \%$ in mean absolute percentage error compared to the baseline gated recurrent unit method. These results underscore the potential of integrating advanced feature engineering techniques with deep learning methods for more precise individual residential load forecasting.