Dual-channel feature extraction and weather-guided two-stage clustering for short-term photovoltaic power prediction

Short-term photovoltaic (PV) power prediction is critical for grid stability and renewable integration. However, existing models struggle to capture both low-frequency trends and high-frequency fluctuations. Thus, this paper proposes a hybrid framework integrating dual-channel meteorological feature extraction, two-stage clustering, and residual correction. Meteorological data are transformed using the Yeo–Johnson method and decomposed via STL to extract eight statistical and dynamic measures from original and residual series. Daily power data are clustered using a Gaussian Mixture Model and merged adaptively by these measures, distinguishing low- and high-volatility regimes while avoiding over-segmentation. Prediction strategies are tailored: XGBoost is applied to low-volatility (LV) clusters, whereas high-volatility (HV) clusters combine XGBoost baseline prediction with BiLSTM-Attention residual correction. Experiments on datasets from Xinjiang (China) and Alice Springs (Australia) show that the integrated model reduces MAE and RMSE by 24–40 % and increases R2 by up to 1.9 %. Compared with advanced models such as PatchTST and TimesNet, errors decrease by up to 30 %, with R2 approaching 0.99. The framework demonstrates strong cross-site generalization, effectively mitigates over-smoothing and phase lag, and provides engineering value for scheduling, dispatch, and renewable reliability assessment.