BiTCSAtt-TCN: A hybrid model for efficient electricity load forecasting in smart grids

Abstract

With the rapid development of smart grids, electricity load forecasting and anomaly detection have become essential tasks for ensuring the stable operation of the grid and optimizing energy management. The large-scale time-series data generated by smart grids contain complex patterns and various dependencies, which pose challenges for traditional forecasting methods, especially in handling long-term dependencies and high-dimensional data. This paper proposes the BiTCSAtt-TCN model, which cleverly combines Temporal Convolutional Networks (TCN), Bidirectional Long Short-Term Memory networks (BiLSTM), and the self-attention mechanism to leverage the advantages of each component. TCN effectively extracts local time features, BiLSTM enhances the modeling of bidirectional dependencies within the sequence, and the self-attention mechanism dynamically focuses on critical time steps. This combination not only compensates for the shortcomings of individual models but also provides significant advantages in handling long-term dependencies and improving prediction accuracy. Extensive experiments were conducted on the SGRT-LMD, ELF, NREL, and EIA datasets, and the results show that BiTCSAtt-TCN outperforms existing models across multiple evaluation metrics, such as MSE, RMSE, R${}^2$, and MAE.