Self-Attentive Bidirectional GRU Time-Series for Electricity Load Forecasting

Precise power load forecasting serves as a fundamental requirement for the effective execution of energy management and strategic planning. Traditional methods, including LSTM and standard GAN models, often fail to adequately address the inherent nonlinearity and rapid temporal changes in electricity load data, struggling to adapt dynamically to new patterns and anomalies. This study introduces a novel time-series generative adversarial network (TimeGAN) that integrates bidirectional gated recurrent Unit (Bi-GRU) networks with self-attention mechanisms designed for electricity load forecasting. The proposed model leverages a bidirectional GRU to fully exploit the temporal characteristics of electricity usage, enabling it to extract relevant contextual information from both historical and future time steps. In addition, the integration of a self-attention mechanism into the network architecture allows the model to selectively attend to key temporal features in the time series, thus enhancing its predictive accuracy. Finally, experiments were conducted based on a self-constructed power load dataset. The results demonstrate that our proposed model significantly outperforms various other network architectures in terms of performance.