Abstract: A model-data joint prediction method was proposed. The object dynamic model was established through mechanism analysis, and the future solar radiation intensity and user load prediction data were introduced for immediate model prediction. The data prediction model was established through the convolution - short and long time memory hybrid neural network improved by attention mechanism, and the historical data was introduced for rolling data prediction. Then, Kalman filter was used to combine the output of the two prediction models to realize the joint prediction of energy storage. Results show that the combined prediction has the advantages of both methods, which can solve the problem of accumulated energy storage prediction errors over time and timely characterize the changes of energy storage when meteorological factors suddenly change and system operation mode changes. The proposed method has good prediction accuracy under various weather conditions.