Precipitation Nowcasting Based On Machine Learning And Research On Its Interpretability

As the main component of weather forecast,Quantitative Precipitation Nowcasting(QPN)has a great impact on transportation,social production,deployment of major events,and military missions.The traditional Numerical Weather Prediction(NWP)model based on physical equations consumes a lot of computing resources with the requirements of increasingly high resolution.In addition,it tends to perform poorly on nowcasting problems due to the spin-up issue.At the same time,with the widespread use of space-based,land-based,and sea-based observation equipment and the continuous improvement of observation resolution,massive amounts of meteorological observation data are generated every day around the world.These data are often just stored and not used efficiently.Therefore,it is necessary to develop a set of data-driven method to efficiently use multi-source data in response to the high temporal and spatial resolution requirements of short-term precipitation nowcasting tasks.This paper uses machine learning and deep learning techniques,combined with multi-source data such as satellite,radar and automatic weather station observations,to construct a precipitation nowcasting model and study the interpretability of machine learning models.The contributions are as follows:(1)Due to the high degrees of freedom and nonlinearity of machine learning algorithms,it is hard to add physical constraints.Therefore,this paper collects multisource observation data such as satellites,radars,and rain gauges data in East China during the flood season(May to September)from 2017 to 2018 are collected to study the precipitation nowcasting problems in this region,which is under the influence of the complex weather system.Subsequently,it uses multi-source data as the proxy of physical constraints and matches the corresponding algorithms for the physical properties of different data.The optical flow method is used to predict the advection of satellite clouds.According to the rapid changes of the radar echo within 0-2 hours,the convolutional neural network is used for feature extraction and prediction;the random forest algorithm is used to replace the traditional empirical formula(Z-R relationship)to simulate the nonlinear relationship between precipitation,radar echo reflectivity and cloud top brightness temperature.Finally,a precipitation nowcasting model MSDM with accurate prediction results and physical significance was constructed.(2)The interpretability of machine learning is often criticized by the meteorological community.In this paper,the interpretable machine learning algorithm Rulefit is used for quantitative precipitation estimation tasks to explore the interpretability of the model in the prediction process with the data introduced in previous section.A tree model is taken as an example,the decision-making mechanism of each step of Rulefit prediction and the resulting rules are analyzed in detail.The results show that its prediction of the large value of radar echo and the low value area of satellite cloud top brightness meets the physical understanding.Rulefit focuses on areas with large precipitation values in the predicting process,which is consistent with the concerns of human experts in the forecasting process.(3)Using model-agnostic methods,Partial Dependence Plot(PDP),this paper analyzes the main features that affect the model’s prediction results,and study the impact of feature interaction,sudden changes in predicted values,and influence of sample distribution on the final results.The study found that radar data is the main feature that affects the prediction results,and when satellite data and radar data are input into the model at the same time,the sample distribution is more uniform,and the prediction is more accurate.(4)In this section,SEVIR storm event dataset is used and the attentional feature fusion module(AFF)with inception is constructed,which is used for multi-scale feature extraction and deep fusion of radar,satellite and lightning data.It helps to maintain the shape of the radar echo and improves the prediction accuracy of large value areas.