An explainable photovoltaic power forecasting method for output based on multi-dimensional similarity game fusion and ISCSO- XGBoost

The randomness and volatility inherent in photovoltaic power generation present significant challenges to grid scheduling, underscoring the critical importance of accurate and interpretable power forecasting. This paper proposes an explainable photovoltaic power forecasting method based on multidimensional similarity game fusion and an improved sand cat swarm optimization (ISCSO)-extreme gradient boosting (XGBoost). Firstly, core meteorological features are screened using the maximum information coefficient (MIC), and weather classification is realized via kernel fuzzy C-means clustering. Secondly, Euclidean distance and grey relational analysis are integrated through a cooperative game-theoretic strategy to construct a comprehensive similarity metric, with the optimal number of similar days for each weather category automatically optimized in conjunction with a benchmark model. Thirdly, an improved complete ensemble empirical mode decomposition with adaptive noise and permutation entropy are employed to decompose the power sequence into trend, low-frequency, and high-frequency components, while a dynamic feature matrix is constructed using time-lag MIC. Finally, cubic chaotic mapping, a spiral search strategy, and a sparrow alert mechanism are introduced to enhance the sand cat swarm optimization algorithm, which is then utilized to optimize XGBoost hyperparameters; feature contributions are subsequently quantified via SHAP value analysis. Experimental results indicate that the mean absolute error under rainy, sunny, and cloudy conditions are 1.396, 0.530, and 1.263, respectively, with corresponding root mean square errors of 2.108, 0.747, and 1.656. These results demonstrate a significant improvement in forecasting accuracy compared to conventional machine learning and deep learning models. The ISCSO outperforms six mainstream optimization algorithms on the CEC2022 standard test functions, validating its superior optimization efficiency. Notably, the proposed model achieves interpretability through a fully mathematical framework, circumventing the black-box nature of deep learning models, and thus satisfying the requirements for reliability and decision-making transparency requirements of real-time power grid scheduling.