Hybrid Downscaling Models for Real-Time Predictions of Summer Precipitation in China on a Monthly–Seasonal Scale

Large intermonth variations in summer precipitation tend to cause alternations or transitions of extreme drought and flood in China; however, seasonal averages may cover alternations on a monthly scale and influence prediction skills on a seasonal scale. Therefore, improving monthly climate forecasts is imperative, contributing to prediction enhancement on the seasonal scale. This study focuses on real-time predictions of monthly precipitation at 160 stations in China during the summer season (June, July, and August) using the year-to-year increment method and field information coupled pattern method, and seasonal precipitation is calculated using monthly predictions. Information from preceding observations and simultaneous predictions from the second version of the Climate Forecast System (CFSv2) are considered. Consequently, the observed sea surface temperature (SST) over the mid-high latitude of the South Pacific in December, the observed sea ice concentration in the critical region of the Arctic in January, and the simultaneous SST from CFSv2 released in February are selected as predictors to develop the downscaling model. First, prediction models based on individual predictors are established to evaluate the prediction skills of different predictors, and subsequently, the singular value decomposition error correction method is employed to reduce errors in downscaling models. Additionally, the optimized ensemble scheme is utilized to synthesize hybrid downscaling models for summer precipitation over China on monthly scale with higher stability, and further seasonal prediction is conducted with results on monthly scale. The re-forecast results during the period 1983−2022 showed that the hybrid downscaling models derived from the optimized ensemble scheme exhibit comprehensive prediction skills compared with single-predictor models. The percentages of stations, at which the time anomaly correlation coefficients of re-forecast results are larger than the 90% confidence level, count for 90%, 88%, and 82% respectively for June, July, and August. The mean values of the spatial anomaly correlation coefficients are 0.39, 0.40, and 0.39, respectively, passing the 99% confidence level. In terms of real-time prediction, the hybrid downscaling models perform well on the monthly and seasonal scales during 2020–2022, when summer precipitation situations are anomalous and different from each other under similar La Niña events. The averaged PS scores of real-time predictions are 75, 75, and 70 for precipitation in June, July, and August, respectively. The PS scores for summer precipitation derived from monthly predictions are 72, 76, and 73 for 2020, 2021 and 2022, which are higher than the multiyear-averaged PS scores of real-time forecasts. These results reveal that seasonal predictions derived from effective monthly forecasts would improve the prediction skills of climate predictions on the monthly and seasonal scales.