Short-term Quantitative Precipitation Prediction Based On Deep Learning

Precipitation is a meteorological characterization of a dynamic,non-linear,multi-time-scale circulation system.It is also the product of the combination of local circulation and thermal effects with terrain and landforms.China is located in the East Asian monsoon region.It is a region with frequent and frequent rainstorms,which may cause disasters such as flash floods and debris flows.Therefore,accurate precipitation estimation and short-term precipitation prediction are not only the basis for the rational development and scientific allocation of water resources,but also the key to ensuring social stability,the safety of people's lives and property,and maintaining the natural ecology and environment.At present,the mainstream precipitation prediction methods mainly include numerical weather forecasting and methods based on radar echo extrapolation.Numerical weather forecast(NWP)uses mathematical,physics,atmospheric dynamics and other methods to analyze the evolution of weather and forecast future weather.As an important tool for weather forecasting and climate forecasting,NWP products provide important decision-making basis for the daily operating forecasting.However,the numerical forecast itself has the disadvantages of uncertainty and parameterization errors.This uncertainty is mainly related to the fineness of the grid.The parameterization errors mainly refer to the initial value errors and the iterative errors in the calculation process.The numerical forecast takes into account various complex factors,which greatly increases the work cost of precipitation forecasting.Secondly,the NWP is better at a longer time range prediction,and the effect of predicting adjacent precipitation is not obvious.Radar echo extrapolation technology is based on the radar observation results at the current moment to retrieval the future position and intensity of radar echoes,which can quickly track and forecast the strong convection system.It is currently widely used in the rainfall forecasting.The mainstream method based on radar echo extrapolation is the optical flow method,and the limitations of the optical flow method are:(1)The optical flow method only considers the relative motion speed between two adjacent frames of images,and cannot calculate multiple frames of images(2)The steps of extrapolation and prediction using the optical flow method are separated,which is easy to cause errors,and it is difficult to further improve the accuracy of precipitation estimation.From the perspective of machine learning / deep learning,in the context of big data,this paper uses radar echo data and automatic weather station data to fully consider the temporal and spatial distribution characteristics of precipitation,and integrates the complex water involved in the precipitation process.Literature,physics,and other issues are reduced to the relationship between sample input,output,and error through a black box model.This thesis,has completed the following three aspects:(1)A dynamic radar quantitative precipitation estimation model combing wavelet transform with multi-time scale support vector machine(SVM)is proposed.Using the 2017 precipitation and radar observation data in East China,taking the estimated time as a reference,the radar reflectance and automatic station precipitation data for the first 30 minutes are selected and transformed into the wavelet domain as training data.Then SVM is used to establish a dynamic precipitation estimation model and performs precipitation estimation at intervals of 6 minutes.This proposed method uses the data from the previous period to dynamically perform precipitation estimation,which greatly improves the Threat Score(TS)value of precipitation estimation,effectively reduces the root mean square error(RMSE),and realizes refined precipitation estimation in East China.(2)A short-term radar quantitative precipitation prediction model based on Gaussian process regression is proposed.According to the radar echo sequence and precipitation data set provided by the Shenzhen Meteorological Bureau,the orientation gradient histogram and scale invariant feature transformation are used to extract the radar echo sequence image features,and then various existing machine learning methods such as SVM,random forest,Gaussian process regression and other methods are used to construct training models,compare the results of precipitation predictions of different models,and calculate the root mean square error based on the error between predicted rainfall and actual precipitation.Among them,the best performing Gaussian process regression model has a RMSE of 9.174 mm / h,which is about an improvement with RMSE of 16.5% over the first place in the CIMK AnalytiCup 2017 competition(RMSE is 10.99 mm / h).(3)A Tiny-RainNet network model for short-term approaching precipitation prediction is proposed.The short-term quantitative precipitation prediction is actually a spatio-temporal sequence prediction problem.It takes past radar data or radar chart sequences as input,and uses rainfall in the future as an output.General machine learning methods can only deal with spatial information,and are difficult for time series information.With reference to the Convolution Recurrent Neural Network(CRNN)network structure,a Tiny-RainNet model is proposed for short-term precipitation prediction based on the problems of short-term precipitation prediction and the advantages of convolutional neural networks(CNN)and bidirectional long-term and shortterm memory networks(BiLSTM).First,the Tiny-RainNet model uses the CNN to extract the context information of a single radar echo map,and then uses a two-way short-time log network to analyze and predict the context of the radar echo sequence.At the same time,adding a pooling layer after each convolutional layer and the BiLSTM layer effectively prevents the problem of poor test results caused by the fast fitting of the training set.The experimental results show that Tiny-RainNet achives a RMSE of 9.40mm/h.RMSE is respectively reduced by 34.01% and 16.88% compared with the linear regression result(RMSE is 14.69mm/h)and Conv-LSTM.