Hybrid Neural Network Day-ahead Electricity Price Prediction Model Embedded with Multi-Head Attention Mechanism

• Published in: 2024 IEEE 4th International Conference on Electronic Technology, Communication and Information (ICETCI) pp. 156 - 160
• Main Authors: Li, Xitang, Li, Pengcheng, Yang, Mingsheng
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.05.2024
• Subjects: Accuracy; Attention mechanisms; Electricity; electricity price forecast; Feature extraction; GRU neural network; Heuristic algorithms; multi-head attention mechanism; Neural networks; Predictive models; TCN neural network
• DOI: 10.1109/ICETCI61221.2024.10594577

Accurate prediction of electricity prices is an essential requirement for effective power market dispatch and informed decision-making strategies. Aiming at the problem of poor short-term electricity price prediction accuracy caused by coarse-grained factor characteristics, a hybrid neural network day-ahead electricity price prediction model embedded with multi-head attention mechanism is proposed. First, a hybrid neural network electricity price forecasting model structure is designed. By fusing the TCN layer for short-term local change feature extraction and the GRU layer for global trend feature extraction, we solved the problem of difficult feature extraction. This difficulty arises from the different patterns and trends in electricity price data on various time scales. Secondly, a similar day selection algorithm based on dynamic time warping distance is proposed. The accuracy of similar day selection is improved by fusing feature vector geometric similarity and distance similarity. Finally, a multi-head attention mechanism weight calculation method is designed. The feature sequence output by the prediction model is mapped to the attention head through a linear transformation. The scaled similarity scores are transformed into attention weights based on the softmax function. The weight calculation enhances the comprehension and representation ability of the prediction model, and improves the accuracy of electricity price prediction. The effectiveness of the proposed model has been verified through experimental validation.