Spatiotemporal Prediction of Sandstorm Utilizing Multi-Source Data and ConvLSTM Model: A Case Study in Northwest China

Frequent dust storms pose significant threats to photovoltaic (PV) power generation in arid regions. Effective operation of PV plants requires accurate forecasting of the spatial and temporal distributions of dust aerosols. This study examines dust characteristics in Gansu Province, Northwest China, by combining surface PM10 concentrations with Aerosol Optical Depth (AOD) data from the Himawari-8/9 satellite. The results indicate that dust storms primarily occur in spring, with significantly higher concentrations found in the northwestern regions (e.g., Jiuquan, Wuwei) compared to the southeast. To address the limitations of traditional numerical models in terms of temporal resolution and the spatial dependence gaps in single-source machine learning approaches, we developed an innovative hybrid model that integrates the Ensemble Kalman Filter (EnKF) and Convolutional Long Short-Term Memory (ConvLSTM) networks. By assimilating multi-source data using the EnKF and employing multi-scale convolutional modules with a high-value-focused loss function, the model achieves accurate forecasts for 4 h ahead with an RMSE of 30.1 μg m−3 and MAE of 16.1 μg m−3. Validated against a regional dust event in April 2024, the EnKF-ConvLSTM model effectively captured the evolution of dust storms, with rapid computation times of less than 30 s per simulation. This approach provides a solid foundation for proactive maintenance of PV plants and enhances the stability of renewable energy systems in areas prone to dust storms.