Regional wind-photovoltaic combined power generation forecasting based on a novel multi-task learning framework and TPA-LSTM

Existing renewable power generation forecasting methods mainly focus on a single energy source and fail to effectively capture the spatio-temporal correlation between different power generation resources. Furthermore, the current single-site power forecasting no longer fulfills the demands of grid dispatch. This paper introduces an innovative framework for multi-task learning and uses it to achieve regional wind-photovoltaic combined power generation forecasting. First, this paper employs Maximum Information Coefficient (MIC) to identify the crucial meteorological features affecting power generation and analyze the complementarity and correlation between wind and photovoltaic power generation. Then, an innovative multi-task learning framework is proposed that separates task-specific components and shared components, allowing each task to select adaptive information that benefits itself. Besides, this paper proposes a loss optimization strategy to balance the loss magnitude and training velocity of different tasks. In order to effectively share the coupling information among the two kinds of power generation, the proposed framework is adopted to construct the regional wind-photovoltaic combined power generation forecasting model based on Temporal Pattern Attention LSTM (TPA-LSTM) algorithm. Finally, the efficiency and superiority of the proposed method are validated through several verification and comparison case studies.