CNN-XGBOOST Short-term Heavy Rainfall Forecast Model And Its Application Researc

Short-duration heavy precipitation is an important weather phenomenon with the characteristics of small scale and rapid development,which is one of the important factors triggering natural disasters and is also the focus and difficulty of weather forecasting.The current short-range forecast for short-duration heavy precipitation mainly uses radar echo data,and the short-duration heavy precipitation forecast model based on radar echo extrapolation is easily blocked by high mountains in mountainous areas,which makes the echo data missing and distorted and generates false echoes,resulting in lower forecast accuracy.Therefore,we need a short-duration heavy precipitation forecast model that is not affected by the topography,suitable for mountainous and hilly areas,and meets the demand for early warning,so as to provide support for refined short-duration heavy precipitation disaster warning and disaster prevention and mitigation decisions.The causes of short-duration heavy precipitation are related to physical parameters such as air humidity,moisture in the atmosphere and temperature and humidity.In this paper,we use reanalysis data,combined with deep learning and machine learning methods,to construct a model for short-duration heavy precipitation forecasting using atmospheric physical parameters,and apply it to short-duration heavy precipitation disaster risk warning in the chemical industry,with good results.The specific research of the paper is as follows.(1)In view of the poor applicability of radar echo-based extrapolation short-duration heavy precipitation forecasting models in hilly areas,a short-duration heavy precipitation forecasting model based CNN-XGBOOST is constructed to forecast whether short-duration heavy precipitation will be monitored in the next 1 hour at weather stations using atmospheric physical quantity parameters.The model uses the XGBOOST classifier instead of the output layer of the traditional CNN model;the CNN set becomes a trainable feature extractor to automatically obtain features from the input;and the XGBOOST set becomes a recognizer at the top layer of the network to generate results and provide more accurate outputs.The experimental results show that the CNN-XGBOOST-based short-duration intense precipitation forecast model forecasts higher TS scores than the control model by 0.133-0.564.(2)For the situation that the traditional CNN model cannot distinguish the importance of feature channels,the attention mechanism is introduced to improve the original model.After extracting the features,the attention mechanism is used to apply weights to different channels separately to consciously focus on the features and improve the forecasting accuracy of the model.The experimental results show that the forecasting accuracy of the model improves with the introduction of the SE and CBAM modules respectively compared with that without the attention mechanism,with the maximum improvement of 0.20 in the forecast TS score with the introduction of the SE module and 0.92 with the addition of the CBAM module.(3)In response to the lack of short-duration heavy precipitation disaster-causing forecasting models for chemical enterprises in southern Guangxi,a short-duration heavy precipitation disaster-causing risk assessment and prediction model for chemical enterprises in southern Guangxi is constructed based on the above technical method.Through the concept of representative meteorological stations,the representative meteorological stations of each chemical enterprise are calculated,and the corresponding hourly precipitation intensity disaster thresholds are constructed for different short-duration heavy precipitation disaster warning levels.The model was validated by using individual cases of short-duration heavy precipitation in mountainous hilly areas and plain coastal areas,and the forecasting accuracy of the model was 0.375 and 0.419 for the two cases of short-duration heavy precipitation,and 0.4 and 0.429 for chemical enterprises,which can meet the warning requirements of chemical enterprises and provide data support for short-duration heavy precipitation disaster prevention and mitigation warning for chemical enterprises.