基于多要素的短临降水预报及可解释性分析

Precipitation nowcasting based on multiple factors and explainability analysis

当前的短临降水预报方法大多是基于雷达回波外推,没有充分考虑其他气象要素对降水生消演变的密切影响,从而限制了其预报的准确性。为解决此问题,基于风云四号B星数据,制作了包含四种背景气象要素、以定量降水估计为预报对象的短时临近降水预报数据集,提出了短临降水预报模型——MFPNM。以TransUNet为骨干,设计了并行双编码器分别提取预报对象和背景气象数据的高维时空特征;构造了内容编码模块将背景数据的空间特征作为预报对象高维特征向量的可学习位置编码;以已有的Transformer模块构建序列数据高维特征间的全局关系,以实现更准确的序列预测。MFPNM在风云-4B数据集和开源数据集上达到了最优水平,采用的指标包括临界成功指数、虚警率、均方根误差和结构相似性等。同时通过SHAP(shapley additive explanations)技术对模型进行了可解释性分析。实验结果及可解释性分析表明,该模型具有更好的预报准确度及可靠性。

The current methods for short-time precipitation nowcasting are based on radar echo extrapolation model,without fully considering the close influence of other meteorological factors on the evolution of precipitation generation and cancellation,thus limiting the accuracy of the forecasts.To address the above issues,this paper produced a short-time precipitation nowcasting dataset,and proposed the MFPNM(multiple factors precipitation nowcasting model).Based on data from the Fengyun-4B satellite,the dataset toke quantitative precipitation estimation as the forecast object and contained four background meteorological factors.Taking the TransUNet as the backbone of the model,this model proposed the parallel dual encoder to extract the high-dimensional spatio-temporal features of the forecast object and the background meteorological data,respectively.Besides,it constructed the content coding module to encode the spatial features of the background data as the learnable positional embedding of the high-dimensional feature vectors of the forecast object.It used a Transformer module to construct the global relationship between the high-dimensional features of the sequence data for better sequence prediction.The metrics used in this paper included critical success index,false alarm rate,root-mean-square error,and structural similarity,etc.The MPFNM was evaluated on two datasets(the proposed dataset and an open-source dataset)and outperformed the baseline models,and it was analyzed for explainability through the SHAP technique.The experimental results and explainability analysis show that the model has better forecasting accuracy and reliability.