Deep learning-based postprocessing for hourly temperature forecasting

In this article, a prediction model based on spatiotemporal stacked ResNet (Res-STS) for hourly temperature prediction is designed. On the timescale, the Res-STS removes the gate structure of the long short-term memory (LSTM) model, and the data of multiple consecutive time nodes are stacked together to preserve all temporal characteristics of the data. A point-to-point data mapping relationship is developed on the spatial scale to maximize the impact of large-scale environmental background field characteristics on a single grid point. Based on the historical gridded data from the China Meteorological Administration land data assimilation system (CLDAS) and the optimal factor dataset of the European Centre for Medium-Range Weather Forecasts Integrated Forecasting System (ECMWF-IFS) from 2017 to 2020, hourly temperature prediction models based on convolutional long short-term memory (ConvLSTM) and Res-STS model are developed, respectively. Furthermore, the prediction results of the two models in 2021 are compared with the ECMWF-IFS. The results show that the root mean square error (RMSE) of the prediction results by ConvLSTM and Res-STS models are both smaller than that of ECMWF-IFS. Specially, the Res-STS model performs best: it reduces the RMSE by 20.8% (24.5%) compared with the ConvLSTM (ECMWF-IFS). Specifically, the RMSE peaks in the afternoon when the daily maximum temperature occurs, while it is relatively smaller at night. Res-STS demonstrates a significant improvement in forecast performance compared with ECMWF-IFS, while ConvLSTM's correction during the period of maximum temperature occurrence has been enhanced. Moreover, the forecast performance of the Res-STS model is least affected by terrain compared with those of the ConvLSTM and ECMWF-IFS. For the regions with terrain height greater than 1 km, the model Res-STS evidently improves the RMSE.