Precipitation Forecast Spatiotemporal Sequence Prediction Research Based On The Fusion Of Deep Learning And Ensemble Learning

With the popularity of artificial intelligence in recent years,as one of the most popular machine learning methods,deep learning has applications far beyond traditional machine learning in the fields of computer vision and natural language processing.In this thesis,our research goal is to predict the future precipitation intensity distribution map based on the precipitation distribution map data,where both the input and prediction targets are image sequences.The problem of precipitation prediction is called Numerical Weather Forecasting(NWP)in the field of meteorology,and traditional methods require complex and detailed simulations of physical equations in atmospheric models.However,we notice that the input data is a spatiotemporal sequence of a certain length,and the output data is a spatiotemporal sequence of length 1.Therefore,we define this problem as the prediction problem of spatiotemporal series,and use deep learning to build an efficient and applicable model to achieve precipitation prediction.This thesis first combines a Convolutional Neural Network(CNN)with a Long Short-Term Memory(LSTM),and proposes a depth model Ext Conv LSTM to solve the problem of precipitation prediction in local areas..Then,based on the idea of integrated learning,using Ext Conv LSTM as the base model,a deep model integration system Ens Conv LSTM was further proposed.The CNN component in Ext Conv LSTM can efficiently capture spatial features,and the LSTM component can extract features in sequence time.Experiments show that Ext Conv LSTM can capture spatiotemporal features better than Fully Connected(FC)neural networks,CNN,and LSTM in local areas,and the prediction results are more accurate.At the same time,due to the complex geomorphology of the target area,the Ext Conv LSTM model with good performance in some areas may not be applicable in other areas with large differences.Therefore,based on the idea of ensemble learning,we train and build multiple Ext Conv LSTM-based models,and use the combination method of arithmetic average to obtain a deep model integration system Ens Conv LSTM suitable for complex landscapes.Experiments show that Ens Conv LSTM can effectively capture more spatiotemporal features,and its prediction performance is better than other deep learning algorithms.Although Ens Conv LSTM has achieved good prediction performance in areas with complex landforms,the actual precipitation data is incremental data recorded successively,and the model is not efficient for spatial feature extraction when the area area is large.Therefore,the idea of incremental learning paradigm is introduced in this article.By consolidating part of the previously learned convolutional layers,we first propose a novel Ext Conv LSTM-based incrementally constructed depth model Inc Conv LSTM(Incrementally constructed Ext Conv LSTM),and build on it Ens Inc Conv LSTM,a deep model integration system.Further,for the feature extraction of larger regions,considering that most regions may not have a great effect on the prediction results,we introduce an attention mechanism and propose an Ens Conv LSTM-based attention model depth system integration system Ens Att Conv LSTM(Ens Conv LSTM with Attention mechanism),greatly improving training efficiency.Experiments show that our proposed Ens Inc Conv LSTM and Ens Att Conv LSTM deep model integration system improves the algorithm training efficiency and obtains significantly better prediction performance than other deep learning algorithms.