Accurate photovoltaic (PV) power forecasting is critical for efficient grid integration and operational planning, yet it remains a challenging task due to the inherent variability of solar irradiance. In this paper, we propose a hybrid forecasting approach that combines a Kalman filter for noise reduction with a Kolmogorov-Arnold network (KAN) for time series prediction. The model’s architecture and hyperparameters are optimized using the tree-structured Parzen estimator, aiming to minimize forecasting error across multiple horizons. We perform extensive benchmarking against statistical, naive, and deep learning models, both with and without denoising filters. The local interpretable model-agnostic explanations (LIME) is applied to highlight important timesteps contributing to the prediction, improving explainability. With a mean absolute percentage error of 5.73%, our optimized hybrid Kalman filter-KAN model significantly outperforms state-of-the-art models, having explainability using LIME. The proposed method proves to be robust, given a statistical assessment, and effective for both short and long-term PV forecasting.
Carvalho et al. (Sat,) studied this question.