Search

Accurate subseasonal predictions of high surface air temperature (SAT) and heat wave events 10–30 days in advance are crucial for mitigating the risks of extreme weather; however, they pose a challenge for current operational models. In this study, we trained a convolutional neural network (CNN)‐based deep learning model to exploit the modulations in China's SAT by precursor signals across different timescales to improve predictions of future SAT and heat wave events. This CNN model demonstrated superior capability in capturing the evolution of SAT anomalies and the occurrence of heat wave events with forecast lead times beyond 20 days, compared with that of the operational models of the China Meteorological Administration and European Centre for Medium‐Range Weather Forecasts. Explainability analysis highlighted that subseasonal SAT predictability in China is driven primarily by large‐scale intraseasonal perturbations from both lower‐ and higher‐latitude regions of Eurasia and interannual variability. Plain Language Summary: Heat waves, marked by prolonged periods of extremely high surface air temperature (SAT), are among the most devastating natural disasters. Improving the predictive accuracy and extending the forecast lead times for heat wave events are vital for disaster prevention and mitigation. However, current operational models typically exhibit limited skill in predicting heat waves at subseasonal scales (2–6 weeks ahead). This is particularly true in relation to the densely populated regions of southern East China, for which the operational prediction systems of the China Meteorological Administration and European Centre for Medium‐Range Weather Forecasts are useful only up to approximately 2 weeks ahead. To address this limitation, we developed a deep learning model using a convolutional neural network approach that can reliably predict SAT anomalies and heat wave events in China with lead times of 10–30 days. Our analysis indicated that successful enhancement of the prediction skill of this model hinges on effectively leveraging the modulating effects of intraseasonal and interannual signals on China's SAT. Rather than focusing solely on specific timescale components, our findings suggest that considering interactions across multiple timescales could enhance subseasonal predictability. Key Points: Summertime surface air temperature (SAT) and heat waves in China are influenced by variabilities across different timescalesA deep learning model skillfully predicted SAT and heat waves for periods extending beyond 20 days by incorporating scale interactionsExplainability analysis linked model skill to modulation of local SAT by large‐scale intraseasonal and interannual signals [ABSTRACT FROM AUTHOR]

There are two main types of precipitation during the pre-flood season in South China, frontal precipitation in early period and monsoon precipitation in late period. It is related to not only the tropical system, but also the cold air in the middle and high latitudes. The extended range forecasting skills of the precipitation in the pre-flood season which depend on the atmosphere-ocean interaction and the internal changes of the atmosphere are still very low. There are biases in models compared to the observations, which makes it hard to directly use model in operational forecast. Therefore, in order to better apply model forecast data to extended period forecasts, the precipitation biases are corrected during the pre-flood season from 1983 to 2019 produced by the Dynamic Extended Range Forecast Operational System version 2.0 (DERF2.0) based on a non-parameter Quantile-Mapping (QM) correction method. Daily precipitation observation data from 261 stations in South China from 1983 to 2019 are selected for evaluation. On the basis of probability forecast of the original model outputs, the model biases are then corrected using monotone cubic spline interpolation combined with the observation. The models are established by cross samples and independent samples to validate the correction method's performance by absolute difference/percentage difference, root mean square error, temporal correlation coefficient and pattern correlation coefficient. It is found that the QM correction method can improve the model forecasting skills by effectively eliminating the systematic deviation of the model. It shows that the improvement of the method remains stable with different lead times and magnitudes of model biases. Further analysis shows that the main locations and average intensities of precipitation show better consistency with observation after correction. The QM correction method can generally capture the trend difference between the model and the observation, and effectively improve the inter-annual variability of model, but it has a poor ability for extreme events. On the other hand, the revised effect of the statistical scheme according to different percentile intervals is also significant. In addition, it shows that the correction performances of prediction are more consistent with the hindcast result.