Adaptive forecasting of photovoltaic power based on dual-type models’ ensemble and online error correction

Solar energy has become an important part of clean energy. Reliable photovoltaic (PV) power forecasting can effectively increase solar energy utilization and reduce costs, which is crucial to the optimal scheduling of PV power generation. However, accurate forecasting of PV power is facing enormous challenges owing to the inherent intermittency and volatility of solar energy. To address the aforementioned issues, a novel adaptive forecasting method for PV power based on dual-type models’ ensemble and online error correction (DMEOEC) is proposed. Firstly, to improve the quality of the basic model, DMEOEC introduces the temporal-aware block to construct a regularity ensemble module. Besides, it also employs the fuzzy entropy (FE) method to refine the generation of small sample data, thereby building a specificity ensemble module. Secondly, an adaptive Bayesian aggregation algorithm achieves the selective fusion of basic models in different modules. Thirdly, DMEOEC establishes two hierarchical weights for the ensemble modules to enhance the robustness and adaptability of the model in a dynamic environment, and the errors obtained from predictions are used to train an online incremental learning model to correct the hierarchical weights in real-time. Lastly, DMEOEC adopts a temporal prioritized experience replay (TPER) and buffer mechanism to cope with model performance degradation due to concept drift and catastrophic forgetting. Compared with traditional methods, DMEOEC not only captures real-time PV power information but also adapts to environmental changes, thus significantly enhancing the accuracy and reliability of PV power forecasting. The effectiveness and superiority of the proposed DMEOEC method are verified using five real PV power plant datasets.