Analysis And Processing Of Spatiotemporal Precipitation Forecasting By Considering Data And Model Uncertainties

Precipitation is an important meteorological and hydrological element in geospatial information science,and is the main driving factor for floods and drought disasters.In2019,the national direct economic losses caused by floods reached 192.27 billion yuan,and droughts caused 23.601 billion kilograms of food losses.Accurate spatiotemporal prediction of precipitation is of great significance for flood and drought disaster prediction and early warning,smart city management,and agricultural water resource allocation.Due to the measurement errors in geographic spatiotemporal data,the imperfection of prediction methods,and the existence of random factors,the spatiotemporal prediction of precipitation is uncertain.The current rainfall prediction methods can be divided into two categories: numerical prediction methods based on numerical weather models and data-driven prediction methods based on statistical machine learning.Due to certain errors in the input data of the numerical model and the imperfections of the numerical model,the results of numerical precipitation prediction often have predictive errors,which require further post-processing.Statistical machine learning methods have uncertainties in the input data,model parameters and variable selection,leading to uncertainty in precipitation prediction results.This paper focuses on the data and model uncertainty of precipitation prediction and related issues,including nonlinear post-processing of precipitation numerical prediction,Bayesian ensemble precipitation prediction that considers model uncertainty,and precipitation forecasting using deep learning as well as considering data and model uncertainty.The main contents of this study are as follows:(1)Due to the errors of geographic spatiotemporal data and prediction methods,the precipitation prediction results based on numerical models have systematic deviations and noise,which makes it difficult to carry out accurate applications at the regional scale.Quantile mapping is a commonly used statistical post-processing method.The quantile mapping method considers the difference in the distribution of the data,but does not consider the nonlinearity and correlation of the prediction error.This paper proposes a nonlinear post-processing method for precipitation numerical prediction based on wavelet coupled machine learning,which combines the advantages of wavelet multiscale decomposition and nonlinear modeling of machine learning to model the nonlinear prediction error of numerical prediction.Among them,the wavelet transform is used for multiscale decomposition of the results of numerical rainfall prediction,and decomposed into a series of low-frequency and high-frequency signals.The lowfrequency and high-frequency data after wavelet multiscale decomposition are used for regression modeling by support vector machine and random forest algorithm,and the post-processed precipitation prediction results are obtained.This paper takes the monthly rainfall prediction in China as an example,collects the monthly rainfall prediction data of the global numerical model from 1982 to 2016,applies the quantile mapping method and the wavelet coupled machine learning method proposed in this paper to perform the precipitation prediction post-processing,and obtains the ground rainfall prediction at 518 precipitation stations.The experimental results show that the post-processing method of precipitation numerical prediction based on wavelet-coupled machine learning proposed in this paper is superior to the prediction results of the original numerical prediction and quantile mapping method,and the predictive root mean square error is reduced by 18-40 mm(21-33%)compared with the quantile mapping method,which improves the accuracy of post-processing predictions.(2)Due to the measurement errors in the data and the imperfect model structure and parameters,there is a certain degree of uncertainty in the prediction.The prediction of a single model is difficult to give a confidence interval for the prediction.Predictive uncertainty is an important indicator that characterizes the predictive ability of a model,and represents the degree of divergence of the forecasting.For statistical machine learning methods,the existing prediction uncertainty research mainly obtains the prediction uncertainty by randomly regularizing the parameters of the model,ignoring the uncertainty of the model structure and variable selection.In this paper,by considering the diversity and uncertainty of model structure and variable selection,a Bayesian ensemble rainfall prediction method that considers model uncertainty is proposed.This method is based on Bayesian theory to estimate the posterior probability of the prediction results of different model structures and different predictor variables.The posterior probability is used as the weight to fuse the prediction results of multiple models to obtain the predicted posterior mean and variance.The experiment took monthly-scale rainfall prediction in China as an example,designed a series of rainfall prediction experiments based on statistical and numerical methods,and applied the proposed precipitation prediction method to fuse these prediction models in a Bayesian way.Experimental results show that the Bayesian ensemble precipitation prediction method proposed in this paper is superior to the single best prediction method in terms of precipitation prediction accuracy.In the 1(2)month precipitation prediction,the correlation coefficient is increased from 0.94(0.85)to 0.95(0.88)and the root mean square error skill score increased from 0.48(0.20)to 0.54(0.26).(3)The prediction uncertainty comes from the noise of the data and the imperfection of the model.The existing precipitation prediction uncertainty analysis method mainly considers the model uncertainty and ignores the influence of data uncertainty on rainfall prediction.Based on variational inference and Monte Carlo sampling theory,this paper proposes a predictive uncertainty probability estimation method that comprehensively considers the uncertainty of data and model.This method first obtains the random error of the input and output data based on the three-cornered hat algorithm.It is assumed that the data error obeys the Gaussian distribution and is propagated during the model training.The model error randomly selects the parameters of the deep learning network through the random sampling algorithm.In the training process,the data and model uncertainty are simultaneously modeled,and the data uncertainty is propagated and model uncertainty is randomly sampled in the prediction stage to obtain a set of prediction results to quantify the uncertainty of the prediction.The experiment takes the weekly rainfall prediction in southern China as an example,designs a deep learning encoding-decoding structure as the prediction method,and compares the data and model uncertainties that only consider data uncertainty,only consider model uncertainty,and comprehensively consider data and model uncertainty.The experimental results show that the proposed method by this study that comprehensively considers the uncertainty of the data and the model has higher prediction accuracy,and the prediction uncertainty is smaller than the existing two data and model uncertainty coupling modeling methods,suggesting higher forecasting reliability.