Dual attention LSTM for building power load forecasting based on feature selection

Abstract

Load forecasting plays an important role in ensuring the safe operation of the power grid. Accurate load forecasting is mainly influenced by historical load data, PV, precipitation and other factors. A load forecasting model with double attention LSTM based on feature selection is proposed, which is designed to comprehensively solve the impact of multiple factors on load forecasting. This model uses Recursive Feature Elimination to remove redundant influencing factors and outputs features with high correlation to the real load. Based on the Long-Short Term Memory network, the feature selection and timing dual attention mechanisms are introduced to merge and dynamically explore the connection between load and input features, which then improves the accuracy of load forecasting. The experimental results show that the accuracy of the proposed model for load forecasting is significantly improved compared with the traditional model.