| The heavy precipitation caused by convective weather changes has the characteristics of rapid evolution and strong destructive power,which seriously threatens the safety of people’s lives and property.Therefore,a high-precision short-imminent heavy precipitation forecast method is proposed to provide relevant departments and departments that rely on weather decision-making in the real society.Provide strong support for people’s travel safety.With the rapid development of deep learning,it provides more cutting-edge technology and thought reference for all walks of life.However,in the field of meteorology,most forecasts are based on the fusion model of radar echo extrapolation and multiple regression,and its accuracy is not enough to meet the actual business needs.Therefore,by combining the relevant algorithm models of deep learning,the prediction accuracy,fitting speed and reaction rate of the prediction process can be further improved.The specific research contents are as follows:(1)Aiming at the problem that the dimensions of the radar echo map and the observation data set of the station are too large and do not match,a prefetcher is designed by using the self-encoding strategy,which compresses and fuses the dimensions of the radar echo map and the observation data of the station,and optimizes the Dataset,get a good feature distribution.The effectiveness of the prefetcher is verified by experiments.The model can optimize the data set,compress and fuse data dimensions,and provide a relatively good data distribution.(2)Aiming at the fact that the prediction accuracy of the Conv LSTM network model cannot meet the expected value,an encoder-attention feed-forward short-imminent heavy precipitation prediction model is proposed.The model combines Transformer’s encoding layer,attention and feed-forward neural network to encode and optimize the input features respectively,and refine the weight distribution of the features to help the model deeply analyze the features in the data,thereby optimizing the generalization ability of the model.The experimental results show that the encoder-attention feedforward network model can effectively improve the prediction accuracy of the model,and perform well on other evaluation indicators,proving the effectiveness of the model.(3)In view of the insufficient training and prediction rate of Conv LSTM and extrapolation estimation based on optical flow method,which may affect the timeliness of prediction,a double-gated prediction model based on attention is proposed to provide faster and more stable predictive performance.Accelerate model training through gated rectification attention,and use attention gating unit to subdivide feature weights to obtain a reasonable feature weight distribution,thereby improving the nonlinear learning ability of the baseline model.The experimental results show that the attention-based double-gated prediction model can guarantee a 2-6 times faster speed in the range of better prediction accuracy,and the evaluation indicators maintain a relatively good level,which proves the effectiveness of the model. |
