Multi-Frequency Residual Learning with Dual-Stage Attention LSTM applied in Energy Time Series Forecasting

Energy time series forecasting plays a critical role in a wide range of real-world applications, including smart grid management, power load balancing, and electricity consumption planning. Within this framework, accurate long-term forecasting remains an important issue due to the complex, non-linear, and multi-scale nature of energy usage. To address these challenges, a novel deep learning architecture, named Multi-Frequency Residual Learning with Dual-Stage Attention Long Short Term Memory (MRL-DA-LSTM) is introduced in this paper. This model integrates three core components: (i) multi-frequency decomposition, which isolates signal components at different temporal scales and allow the model to better capture both short- and long-range patterns; (ii) a dual-stage attention mechanism that includes temporal attention, to focus on critical time steps, and feature attention to weigh frequency components, enhancing the interpretability and effectiveness of the learning process; and (iii) a residual learning framework that models the forecasting error and refines predictions in a two-stage process. Unlike the traditional LSTM based models, the proposed approach systematically separates and emphasizes frequency specific features while refining output through residual correction. Extensive experiments on household energy consumption data and the Tunisian Energy and Environment dataset demonstrate that MRL-DA-LSTM significantly outperforms baseline models in both short- and long-horizon forecasting tasks, providing a robust solution for complex time series prediction in the energy domain.