Ultra-short-term Power Load Forecasting Considering Multidimensional Feature Fusion and Bi-directional Information Learning

Accurate power load forecasting is crucial for power system operation decisions, but power load is easily affected by various factors to different degrees. In order to solve the problem of poor prediction accuracy caused by existing methods relying only on historical load data or indiscriminately fusing multiple influencing factors for prediction, an ultra-short-term power load prediction method considering multidimensional feature fusion and bi-directional information learning is proposed. Initially, multidimensional feature fusion is used to enrich feature information, and an Efficient Channel Attention (ECA) module is introduced to adaptively assign corresponding weights to each channel to distinguish the importance of different influencing factors and highlight the key information, which facilitates the Convolutional Neural Network (CNN) to better learn multidimensional features. Subsequently, the task of ultra-short-term power load forecasting is accomplished by further extracting temporal information from the multi-dimensional load features using a Bidirectional Long Short-Term Memory (BiLSTM) neural network. Finally, the effectiveness of the proposed model is verified through case studies based on actual power load data and its related influential feature parameters.