Enhancing PV power forecasting accuracy through nonlinear weather correction based on multi-task learning

Accurate short-term photovoltaic (PV) power forecasting is critical for optimizing energy management and maintaining grid stability within the rapidly growing renewable energy sector. However, the inherent high sensitivity of PV systems to varying weather conditions poses significant challenges to achieving reliable predictions. Existing research endeavours to enhance short-term forecasting accuracy through two primary approaches. On the one hand, some studies incorporate weather variables as input features to improve prediction precision, yet this method often falls short of fully capturing the intricate and dynamic interactions between diverse weather factors and PV output. On the other hand, most correction methods utilize error correction (EC) techniques that adjust initial PV forecasts based on predicted errors. Nonetheless, the highly volatile nature of error sequences substantially restricts the effectiveness of EC, as these unpredictable errors compromise the reliability of the corrective adjustments. To this end, we propose a novel two-stage framework that leverages weather information from multiple perspectives to enhance short-term PV power forecasting accuracy. In the first stage, a customized multi-task learning (MTL) framework employs a task interaction matrix to differentiate between task-specific and shared features, thereby facilitating meaningful interactions between PV output and weather variables while providing interpretability. Additionally, a dynamic loss weighting mechanism ensures balanced training across tasks. In the second stage, we implement a nonlinear weather correction (WC) module using neural networks, which refines the initial PV predictions by effectively incorporating the predicted weather variables, thereby enhancing both accuracy and robustness. Experimental validation using real PV data from the Northern Territory, Australia, demonstrates that our framework consistently outperforms baseline models across various seasons and confirms the effectiveness of each component within the framework through ablative experiments.