Short-term photovoltaic power forecasting using DTW-based K-Medoids clustering and a hybrid VMD-CEEMDAN-1DCNN-S-Mamba model

Abstract

Abstract The short-term prediction of photovoltaic (PV) power is critical for grid stability and efficient renewable energy dispatch. However, the strong volatility of PV generation due to weather changes poses major challenges for accurate forecasting. This paper proposes a hybrid model combining DTW-based K-Medoids clustering, two-stage VMD-CEEMDAN decomposition, and a 1DCNN-S-Mamba network for multi-step PV power forecasting.Historical PV power and irradiance sequences are clustered into sunny, cloudy, and rainy patterns using K-Medoids with Dynamic Time Warping (DTW), which is more robust to outliers and temporal shifts than traditional K-Means. Variational Mode Decomposition (VMD) extracts the main trend components, while Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is applied to the residual to further mitigate high-frequency noise. The decomposed components, along with meteorological variables and weather-type codes, are input into the 1DCNN-S-Mamba model, where the 1DCNN captures local multiscale features and the bidirectional S-Mamba models long-range dependencies.Tested on a two-year dataset with a 5-min resolution from three Australian PV stations, the model achieves an $$R^2$$ of 0.9822, an MAE of 0.1188 kW, and an RMSE of 0.2451 kW. It outperforms eleven benchmark models, especially under cloudy and rainy conditions. Ablation studies confirm the effectiveness of the two-stage decomposition and the 1DCNN module. The proposed framework provides a practical solution for high-accuracy PV power forecasting.