Studying Precipitation Nowcasting Approach Based On Online Tensor Learning

Precipitation forecast has always been one of the hot issues studied by meteorologists.It is of great significance to forecast the precipitation accurately,especially the nowcasting for guiding the daily travel and the enterprise management.Machine learning and deep learning are used in most precipitation forecasts.The shortcoming is as follows: offline learning(batch learning)is adopted and the datasets are trained repeatedly.Therefore,the efficiency of model training is low.As a high order extension of matrix,tensor is a powerful tool to study spatio-temporal data.Currently,the research of nowcasting using tensor model is in the initial stage.Based on the tensor structure,an online learning model of nowcasting is proposed in this thesis.According to the mobility of meteorological data,the model can realize hourly forecast at multiple stations and online calibration.First,the online tensor regression model is used to forecast the hourly data of multiple stations.However,there may still be systematic deviation between the predicted and the actual value.Then,on the basis of the tensor regression model,the frequency matching method is used to calculate the dynamic calibration coefficients,and the predicted value is calibrated online to reduce the deviation.Finally,the forecast value is obtained by online tensor regression and frequency matching method.In the two actual datasets,the algorithm reduce the error between the predicted and the actual value on the datasets,which has a good performance in the evaluation indicators of the sunny and rainy.