Multi-model fusion method for reconstructing the dry-wet time series of daily precipitation and its application

To effectively enhance the accuracy of precipitation forecasts generated by CMIP6 climate models, we propose a daily precipitation reconstruction method based on a multi-model fusion approach. This method utilizes multi-model precipitation data as input and employs a two-layer deep learning fusion model (LCNN) that integrates a LSTM with CNN to reconstruct and correct daily-scale precipitation estimates. The Hanjiang River Basin (HRB) has been selected as the experimental area for evaluating accuracy and analyzing the effectiveness of precipitation data from 96 uniformly distributed surface rainfall gauges within the basin. The evaluation results indicate that the constructed LCNN model for classifying and quantitatively predicting precipitation dry-wet time series can effectively correct precipitation errors. This enhancement is reflected in the improvement of key performance metrics: the CC, NSE, POD, and HSS increase from original values of 0.08, −0.30, 0.31, and 0.06 to 0.57, 0.16, 0.99, and 0.62, respectively. Concurrently, the FAR has been reduced to 0.44, and the RB values for most gauges have decreased from an original ±40 % to within ±18 %. These improvements significantly enhance the model's ability to detect dry-wet time series and reduce the quantitative estimation error of precipitation. Additionally, this method effectively captures the spatial distribution characteristics of precipitation in the HRB, which exhibit a pattern of high in the southwest and low in the northeast. This approach enhances the model's spatial detection capability and improves the accuracy of precipitation predictions. However, the method may show limited improvement in the quantitative estimation error for small-scale precipitation events (daily precipitation below 2 mm), suggesting the need for further research. The findings of this study can address the shortcomings of existing climate models, further enhance the predictive capabilities of precipitation models, and provide better foundations support for hydrological forecasting, flood control, and disaster mitigation in river basins.