For the purpose of achieving a scientifically rigorous and methodologically sound clustering partition of photovoltaic clusters, this study incorporates the synergistic effects of meteorological factors and geographical positioning on the clustering of photovoltaic power stations. Employing output characteristics as the basis for subgroup delineation, the paper introduces a K-medoids clustering algorithm based upon an enhanced Manhattan distance metric. The two pivotal dimensions in photovoltaic cluster analysis encompass the robust identification of meteorological features and the precise quantification of inter-station similarity. Accordingly, an integrated approach combining the coefficient of variation method with the rank-sum ratio method is utilized to derive meteorological input features, thereby augmenting the efficacy of feature discrimination. Moreover, the refined Manhattan distance is embedded within the K-medoids framework to comprehensively elucidate the dynamic attributes inherent in the dataset. Three internally validated clustering metrics are leveraged to optimize cluster cardinality, effectively mitigating the dispersion and systematic bias endemic to conventional partitioning approaches while enhancing overall clustering robustness. Subsequent predictive validation is performed based on the established clusters. To resolve the challenge of parameter calibration within the forecasting model, metaheuristic optimization techniques are integrated. Specifically, an improved grey wolf optimization （IGWO） algorithm is deployed to fine-tune the hyperparameters of both the iTransformer and extreme gradient boosting （XGBoost） architectures. The power forecasting results verify the operational efficacy and superior predictive accuracy of the proposed clustering methodology.

photovoltaic power  /  cluster analysis  /  power forecasting  /  improved gray wolf optimization algorithm  /  iTransformer  /  extreme gradient boosting