Transformer-enhanced spatiotemporal neural network for post-processing of precipitation forecasts

Numerical Weather Prediction (NWP) models are extensively utilized worldwide and have played a pivotal role in weather forecasting. Precipitation is subject to various intricate factors, rendering it one of the most challenging factors to predict. Additional post-processing steps are required to reduce biases and achieve reliable predictions for precipitation-related decision-making. In this work, we propose a transformer-enhanced spatiotemporal neural network called TransLSTMUNet for short- and medium-range precipitation post-processing. Firstly, TransLSTMUNet employs convolutional operators to extract localized meteorological features. Secondly, it capitalizes on transformer architecture to enrich these extracted features with a broader, global perspective of spatial information. Thirdly, TransLSTMUNet leverages ConvLSTM to further enhance the features with temporal information. Furthermore, to address challenges posed by imbalanced distribution of precipitation intensity, we design a novel loss function called quantile weighted mean squared error (QWMSE). QWMSE simultaneously considers both normal and intense precipitation during the model' s training phase. In the experiments, the THORPEX Interactive Grand Global Ensemble (TIGGE) dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) is used as input for post-processing the precipitation forecasts. The experiments show that the precipitation forecasts post-processed by TransLSTMUNet exhibit the best overall performance compare with the eight post-processing baselines. It significantly improves the forecast performance of TIGGE forecasts. Specifically, TransLSTMUNet enhances the accuracy (ACC) metric by 12.14 % and increases the threat scores (TS) for 24-hour accumulated precipitation of 0.1 mm, 10.0 mm, 25.0 mm, and 50.0 mm by 8.30 %, 9.77 %, 31.60 %, and 51.25 % respectively. By effectively integrating the strengths of convolutional and transformer methodologies, the proposed TransLSTMUNet model offers a novel approach for post-processing precipitation forecasting. This model design has the potential to inspire various other research avenues within the hydrological domain and beyond.