Prediction Of Short-term Rainfall Based On Convolutional LSTM And Random Forest

In all kinds of weather events,rainfall plays a vital role in human life.Accurate prediction of bad weather can reduce the occurrence of natural disasters.At the same time,forecasting rainfall information in advance can reduce the impact of rainfall on people's lives.However,rainfall forecasting is a very difficult task.Traditional rainfall forecasting schemes based on Z-I relation are vulnerable to the influence of detection conditions.On the other hand,different forecasting schemes need to be designed for different regions so that the generality is not high.Therefore,it is of great significance to design a rainfall prediction scheme with high accuracy and versatility.After studying the existing rainfall forecasting methods,this paper proposes a short-term rainfall forecasting method based on convolutional LSTM and random forest,and predicts the rainfall in the next 1-2 hours by using the Doppler radar information of multiple different altitudes in the past multiple moments.In this paper,the convolutional LSTM model is used to extrapolate radar maps to improve the accuracy of prediction,using random forest algorithm for ensemble learning can improve the generalization ability of the model and make it more versatile.The main methods are as follows:Firstly,the task of this paper is to forecast the rainfall in the next 1-2 hours,but it is impossible to predict the rainfall in the future by radar images of historical time series,so we use convolutional LSTM network to predict the radar maps of the next 1-2 hours.The full connection layer of each input layer on the basis of traditional LSTM network is replaced by convolution layer,which retains the temporal and spatial characteristics of radar maps and greatly improves the accuracy of extrapolation results.Secondly,the radar map extrapolated by convolutional LSTM has 400,000-dimensional characteristics,including many redundant features,and the feature dimension is too high,which causes memory explosion and can not be calculated.Therefore,a hierarchical feature extraction method is proposed in this paper.Firstly,the principal component analysis(PCA)method is used to reduce the dimension of each radar map,and secondly,the final results are obtained by splicing the radar map after dimension reduction at all times and altitudes.This design can not only solve the problem of memory explosion when PCA extracts features from high-dimensional radar maps,but also can retain the overall time sequence information and high information of radar maps;Finally,in the stage of rainfall prediction,in order to improve the generalization ability of the model,this paper uses the features after extrapolation and feature extraction of the radar map as input features,and randomly selects some features to construct a decision tree at each time,and integrates several decision trees to learn,and uses the model after learning to predict the rainfall in the next 1-2 hours.In this paper,the proposed model is compared with Z-I relation,support vector regression,and xgboost.The experimental results show that:(1)Radar map extrapolation using convolutional LSTM model can improve the performance of rainfall prediction task greatly;(2)The proposed model achieves the best results in hit rate,no alarm probability,critical success index,and root mean square error.As a result,especially in the stage of heavy rainfall,the effect has been improved more obviously,and the proportion of some indicators has been increased by more than 20%.