Research On Short-Term Weather Forecast Based On Deep Learning

Nowcasting generally refers to the forecast of precipitation or strong convective weather in a certain area in a short period of time(0-2 h)in the future,and it plays an extremely important role in the prevention of daily meteorological disasters.my country is located in East Asia,on the west coast of the Pacific-Ocean,and the continental monsoon climate is significant.This makes China suffer from floods in summer and autumn.Therefore,it is urgent to establish an accurate nowcasting system.An accurate precipitation nowcasting system can not only effectively guide people's daily life,but also has important guiding significance for road traffic,air transportation and the issuance of meteorological disaster warnings.In recent years,as a novel machine learning algorithm,deep learning technology has begun to receive attention from researchers and business people.At present,deep learning has begun to be applied in the field of nowcasting.Compared with traditional algorithms,its effect has been significantly improved.Therefore,from the perspective of deep learning,this paper designs an end-to-end network model based on radar reflectivity image data to extract the spatiotemporal features contained in the input sequence,and then predict the prediction results we need based on these features.In general,this article mainly completed the following two aspects of work:(1)In practical applications,in order to obtain more accurate forecasts,meteorologists often need to grasp and analyze the changing trend of cloud clusters,which requires the externalization of the most accurate radar reflectance image sequence.Aiming at the extrapolation of radar reflectivity images in nowcasting,this paper proposes a spatio-temporal prediction model PPNet(Precipitation prediction network).This model converts the space-time problem of radar reflectivity image prediction into a video frame prediction problem.This effective angle conversion provides a lot of reference for the construction of the main framework of PPNet;Secondly,we introduced Generative adversarial networks(GAN),so that the generated radar images can retain more echo details;but because the pure GAN network structure lacks constraints on motion characteristics,the GAN network generates some detailed information that does not match the actual change trend of the cloud cluster causes the network to misreport the movement trend of some local echoes.To solve this problem,we introduced the optical flow network(Flow Net)to obtain and construct the optical flow loss of the echo image to constrain the generator to ensure the correct expression of the motion information.The effective combination of optical flow network and GAN not only ensures that the predicted cloud cluster has rich detailed features,but also better characterizes the movement law of the cloud cluster.Multiple experiments on the thunder-fax reflectance image data set provided by Shenzhen Meteorological Bureau and Anhui Meteorological Bureau show that the improvement of this paper is better than traditional optical flow method and other deep learning algorithms,and it has the advantage of predicting the trend of echo changes.great improvement,and better nowcasting accuracy has been achieved.(2)Aiming at the problem of precipitation prediction,this paper proposes a rainfall prediction model MAR-CNN(Multi residual attention-Convolutional network)based on convolution residual attention.The precipitation prediction problem is a regression problem.According to the radar reflectivity image data,the spatial characteristics of the cloud clusters are extracted,and the predicted precipitation is obtained by inversion.Different intensities of reflectance have different effects on precipitation.Strong reflectance areas have a greater impact on precipitation.In order to emphasize these strong reflectance areas,we introduces a multi-head attention mechanism that allows the model to distinguish which areas in the cloud cluster by itself is a strong reflectivity area;at the same time,in order to avoid the loss of global information caused by the attention mechanism,we introduce residual connections to ensure that the model can retain more global information while highlighting important features;In addition,the non-image information of the cloud,such as the cloud's moving speed,also has an important impact on precipitation.In response to this problem,we have introduced a second channel to extract the non-image features,based on the fused image features and Non-image features,predict the amount of precipitation.The final experimental results show that our improvement can make the characteristics of the strong echo area better,and the accuracy of precipitation prediction is also improved compared to other deep learning algorithms and machine learning algorithms.