Multi-source meteorology and cloud image joint coding algorithm for photovoltaic power prediction under complex weather disturbances

Introduction Rapid cloud movement, irradiance mutation, and meteorological uncertainty can increase prediction errors and delay ramp tracking in short-term photovoltaic power forecasting. To address these challenges, this paper proposes a multi-source meteorological and cloud-image joint coding algorithm for short-term photovoltaic power prediction under complex weather disturbances. Methods The proposed method uses ground meteorological observations, numerical weather prediction variables, satellite cloud-image sequences, and historical photovoltaic power as inputs. A meteorological temporal encoder and a cloud-image spatial encoder are constructed to extract the temporal evolution of irradiance, temperature, humidity, wind speed, and historical power, as well as cloud texture, cloud boundary, and inter-frame cloud variation. An adaptive cross-modal gated fusion module is introduced to dynamically adjust the contribution of meteorological and cloud-image features under different weather states. To improve the learning of rare but operationally important disturbance samples, a disturbance-weighted loss and physical ramp and solar-altitude constraints are further incorporated. Results Experimental results based on chronologically partitioned measured photovoltaic and meteorological data show that the proposed model reduces RMSE, MAE, MAPE, and ramp tracking error compared with baseline models. Discussion The study focuses on forecasting accuracy and ramp-tracking capability, while grid-control effects are discussed only as potential applications requiring further closed-loop validation.