Short-term power load forecasting based on AC-BiLSTM model

Abstract

The practice of ultra-short-term power load forecasting serves as a critical strategy for enabling rapid response and real-time dispatch in power systems. By improving the accuracy of load forecasting, both the safety of power systems and the efficiency of electricity usage can be significantly enhanced. Addressing the challenges posed by the non-linear and temporal characteristics of grid load data, this study introduces a novel ultra-short-term power load forecasting model, integrating Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory networks (BiLSTM), and an Attention mechanism, referred to as the AC-BiLSTM model. This innovative approach harnesses the power of CNN and BiLSTM to extract spatio-temporal features of load data, while the Attention mechanism allocates optimal weights to the hidden states of the BiLSTM model, thereby amplifying crucial historical load sequence data and minimizing information loss. The final output of the model is then determined through a fully connected layer. To validate the efficacy of this approach, an empirical study was conducted using real load data from a specific region. The results, obtained from two contrasting experimental scenarios, demonstrate a significant enhancement in forecasting accuracy. This finding underscores the potential of the AC-BiLSTM model as a reliable tool for both strategic planning and maintaining operational stability in power systems.