Multi-component framework for improved photovoltaic power forecasting performance

Abstract

Solar energy is a promising renewable resource, offering considerable potential for the advancement of photovoltaic (PV) systems. Accurate PV power forecasting is essential for efficient grid integration and optimal system planning. However, environmental factors—such as seasonal variations and social behavior—can introduce major challenges, reducing the reliability of forecasts. This study proposes the CRAK model, a novel approach to PV forecasting, which integrates causal convolution, recurrent structures, attention mechanisms, and the Kolmogorov–Arnold Network (KAN). These components work synergistically to extract key features from historical data—specifically, causal convolution identifies essential features, recurrent structures model temporal dependencies, attention mechanisms prioritize critical features, and the KAN facilitates nonlinear feature extraction. Experimental results demonstrate that the CRAK model outperforms existing forecasting models in terms of the MAPE, RMSE, MAE, and R2, thereby confirming its superior PV-forecasting performance.