An Optimized Seq2Seq Attention Network Considering Multivariate Temporal Correlation for Short-Term Electricity Price Interval Prediction

In a deregulated electricity market, participants need accurate electricity price forecasting tools in order to maximize their profits and utility. However, accurate electricity price prediction has become a challenging task with increasing renewable energy penetration and the extension of the power system scale. The complexity of electricity market information and insufficient model training limit the prediction accuracy of the existing electricity price forecasting methods. This article proposes a Sequence-to-Sequence Attention algorithm based on the Double Deep Q Network optimization method for short-term electricity price prediction. The article first conducts the maximum information coefficient correlation analysis to select the input sequence from historical electricity price and electrical load. Then, the Sequence-to-Sequence Attention model is proposed for short-term electricity price interval prediction. Finally, the hyperparameters of the model are optimized by the Double Deep Q Network method, which improves the prediction accuracy and the generalization ability of the prediction model. Simulations are carried out on Pennsylvania–New Jersey–Maryland market data and the New South Wales electricity market to validate the proposed method. Numerical results show that the proposed interval prediction method of electricity price has improved the prediction interval coverage probability by up to 10.98% and reduced the prediction interval normalized average width by up to 42.87% compared to four benchmark models. The results suggest that the proposed interval prediction method has good prediction accuracy and generalization ability, providing a powerful decision-making basis for market participants and regulators. © 2024, Taiwan Ubiquitous Information CO LTD. All rights reserved.