Research On Short-term Rainfall Forecast Model Based On Multi-task Deep Learning

Short-term rainfall prediction is one of the most important research topics in meteorological prediction.Accurate and efficient rainfall prediction can provide effective scientific basis for human activities and national security.In recent years,the results of deep learning research in weather prediction are significant.Although the short-term rainfall prediction model based on deep learning has surpassed most of the traditional rainfall prediction models,there are still many areas for improvement in deep learning itself.At present,most of the meteorological data obtained by many meteorological centers come from multiple observation sites,which have high-dimensional and chaotic characteristics.To effectively model these data,we need to simulate the internal spatial-temporal structure of the data.Therefore,our research work is as follows:(1)In view of the long time dependence of recurrent neural network short-term rainfall prediction model in single-task,a short-term rainfall prediction model based on recurrent gated network is proposed.In the construction of recurrent neural network,we choose the gated recurrent network with gating algorithm.The gated recurrent network is a variant of the long short-term memory network.Compared with the long short-term memory network,it is simpler,and the performance of the model will not decline when the calculation cost is reduced.Experiments show that the problem of time dependence in the process of rainfall prediction has been effectively solved,and the prediction cost of the model has been significantly reduced.In view of the fact that different weather data sets are tested in the short-term rainfall prediction model of single-task convolution neural network,the model is easily affected by the data types and produces instability,a rainfall prediction model based on dynamic convolution neural network is proposed.In the traditional convolution neural network,the convolution layer is static,and the parameters of the filter as the convolution layer are fixed,which results in the low compatibility of the model.In this paper,the filter size in the convolution layer is dynamically assigned by a function corresponding to the input,so that the filter can adjust its size according to the input data of the network,and the influence of different data on the model is significantly reduced.Compared with other models,the dynamic convolution neural network model is more stable and accurate,and the accuracy is improved by 9.5%.(2)In view of the lack of the ability to model the inherent spatial-temporal relationship of multi-site meteorological data when the existing deep learning method is used to build the rainfall prediction model for multi-site data,a short-term rainfall prediction model based on multi-task convolution long short-term memory network is proposed.By combining the spatial characteristics of convolutional neural network with the temporal characteristics of recurrent neural network,we propose a long short-term memory network with convolutional structure,which not only solves the problem of redundancy of long short-term memory network in processing spatial data,but also encodes spatial information,so that the temporal and spatial characteristics of Meteorological data can be effectively used.Further research,we will combine the meteorological data of multiple sites for learning,allowing sites to transfer the learned knowledge from one site to other related sites,and modeling the correlation between sites,so that the hidden key clues between sites can be fully utilized.Compared with the existing excellent models,the error rate of the multi-task convolution long short-term memory network model is reduced by 3.2%,and the success rate is increased by 10%.