Adaptive Feature-Based Fusion Recurrent Neural Network for Predicting Day-Ahead Prices

In the electricity spot market, day-ahead price is one of the important reference factors for market decision-making. Accurate day-ahead price forecasting is conducive to power market dispatch and market parties to make wise decision-making strategies. With the emergence and development of deep neural networks, a large number of prediction methods based on deep learning models have been used for day-ahead price prediction and achieved good results. However, the current prediction methods based on recurrent neural networks are not sensitive to the changes of different time scales, and the input feature selection of neural network models requires tedious attempts. In view of the existing problems, a fusion recurrent neural network based on adaptive features is proposed to improve the prediction accuracy of day-ahead prices. The adaptive feature uses the attention coefficient learned by the neural network as the weight of the weighted sum of different input features, so that it can make corresponding adjustments according to the training target. The fusion recurrent neural network is trained by the fusion loss function, which can fuse the output results of neural networks at different time scales to obtain the optimal prediction results. The model is evaluated on the market data of Zhejiang Province, China, and compared with several models. From the experimental results, compared with other baseline models, the prediction error of the fused recurrent neural network is further reduced, and the prediction accuracy is improved.