Research On Minute-level Precipitation Forecast In Guizhou Based On Deep Learnin

Precipitation nowcasting within 0 to 2 hours is of great relevance in government departments taking measures to damage avoidance from weather disasters,disaster relief and weather-dependant social activity guarantees and production of operation in various industries.Up to now,However,literature shows that global coarse-resolution numerical weather prediction(NWP)models have challenges in generating accurate precipitation with a lead time of 0–2 h,affected by the spin-up issue and the difficulties in non-Gaussian data assimilation.,also the extrapolation nowcasting based on radar data and products has great limitations.Aiming to obtain high-quality precipitation nowcasting more efficiently and directly,this paper constructed various precipitation nowcasting models within 0 to 2 hours over Guizhou,China base on minute-level multi-element observation data.The Convolutional Long Short-Term Memory(ConvLSTM)and Predictive Recurrent Neural Network(PredRNN)models were used as comparative DL models,and the Lucas–Kanade(LK)Optical Flow method was selected as a traditional extrapolation baseline.The forecasting skills and training efficiency of deep learning models are improved by means of data quality controlling,increasing the time series,the number of training samples,training with heavy precipitation data set(Hea-P data set),and training with heavy precipitation datasets contains physical quantity(HCPQ data set).Comprehensive evaluations of the precipitation nowcasting is carried out base on traditional pointto-point,including RMSE,ETS and POD.Also,a rainstorm case on the Method for Object-Based Diagnostic Evaluation is included.The following main conclusions are obtained:(1)Based on the multi-element observation data in Guizhou,several high-quality training datasets were established,deep learning models for precipitation nowcasting that can directly obtain precipitation nowcasting was constructed.By means of supplementation of missing values and normalization can improve training efficiency and forecasting skills of deep learning models.Increasing the time series length and sample number of training dataset and training with heavy precipitation dataset were both efficient ways to improve forecasting skills exceeding the precipitation threshold of 2.5 mm.At the same time,training with Hea-P data set could improve the Probability Of Detection(POD)and stability of nowcasting.On this basis,the HCPQ data set further improved the forecasting skills exceeding the threshold of 2.5 mm,and also greatly improved the ability of precipitation nowcasting with microscale precipitation exceeding the threshold of 8.0 mm.The improvement of forecasting skills is more obvious at the lead time of 2h.The PredRNN model trained with the 8-year HCPQ data set performed the highest forecasting skills among all models.All the areas with high forecasting skills were mainly associated with the southern,southeastern and northern regions of Guizhou.(2)The Improvement Skill Score(ISS)of the comprehensive verification index shows the HCQP data set played a more important role in improving the forecasting skills of deep learning models than Hea-P dataset.The HCQP data set obtained a higher POD while reducing the deep learning model’s FAR,POFD,and RMSE.The location and shape of precipitation nowcasting became more accurate,and the ConvLSTM model outperformed the LK Optical Flow model at the lead time of 2h.The deep learning models trained with data sets of 8-year time series length performed the best improvements.(3)A rainstorm case on the Method for Object-Based Diagnostic Evaluation shows that the inconsistencies between the precipitation nowcasting of the LK Optical Flow model and observation are increasing over time.The optical flow vector calculated based on the precipitation that was about to experience significant fluctuations will obtain nowcasting that deviated significantly from the observation.Different from the characteristics of slight movement and single linear change of LK Optical Flow model,the deep learning model breaks through the limitations of traditional methods by training with the HCPQ dataset,could well capture the complex nonlinear changes in the precipitation,and then better precipitation nowcasting which exceeds significantly that of the traditional LK Optical Flow model was obtained.