Short-term Load Forecasting by BiLSTM Model Based on Multidimensional Time-domain Feature

Power load forecasting is vital in modern power systems and is crucial for the stable operation and real-time scheduling of power systems. This paper proposes an innovative power load forecasting method, which integrates signal decomposition methods (EMD, EEMD, VMD) and Shapley values feature selection, as well as a bidirectional long and short-term memory network (BiLSTM) model. First, different signal decomposition methods decompose the historical load data, and the multidimensional time-domain features are combined with the original features, which enhances the ability to capture the complex correlations in the power load data. Then, Shapley values analyzes the correlations, and the most appropriate thresholds are determined to select the best input features, improving the model’s efficiency and performance. Finally, a BiLSTM model is used for load prediction. Extensive experimental validation on public data shows that the method in this paper achieves significant performance improvement on validation and test sets compared to traditional methods and single decomposition methods. In addition, the framework in this paper also shows higher prediction accuracy compared to other state-of-the-art models.