A novel multivariate electrical price bi-forecasting system based on deep learning, a multi-input multi-output structure and an operator combination mechanism

With the liberalization and deregulation of the power industry, the power market presents increasingly invisible dynamics and uncertainty. Establishing an effective price forecasting model has always been an important issue for power market participants to avoid risks and maximize economic benefits. However, due to the inherent characteristics of high frequency and high volatility of electricity price data, it is still a difficult and challenging problem to achieve high-precision forecasting. To overcome these difficulties, this study constructed a novel multivariate electric price bi-forecasting system based on the proposed linear operator combination mechanism and deep learning model. Specifically, improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is used for data preprocessing, while the multiobjective golden eagle optimization algorithm (MOGEO) optimized deep belief network and bidirectional long short-term memory network are used as the main prediction models for realizing two forecasting modes of point forecast and interval forecast by a multi-input multi-output structure simultaneously. To verify the model forecasting ability and effectiveness, four datasets from Australia (New South Wales) from 2020.7 to 2021.1 are selected for experiments. According to the empirical results, the mean absolute percentage error results of the designed forecasting strategy in point forecasting are within 4% for the four datasets, and the average coverage width criterion values in interval forecasting are all approximately 0.3. The results show that the system has good electricity price forecasting ability and can help policy-makers better understand market dynamics and develop more effective market regulation and intervention measures.