Research On Time Series Prediction Based On Hybrid Neural Networks

With the development of big data technology,the application of time series prediction in many fields has been greatly improved.At the same time,due to the ability of independently mining data feature information,deep learning methods have gradually been introduced into time series forecasting,which has received widespread attention and attention.Compared with traditional methods,deep learning methods can combine spatiotemporal multi-dimensional information to extract time series features,so as to complete time series analysis,and have the advantages of not requiring manual extraction of features,predicting multiple sets of series at the same time,and capturing the correlation between sequences.However,there are still disadvantages such as unclear judgment of the importance of time series features and incomplete extraction of feature information.This dissertation takes hybrid neural network as the core model structure,studies the improvement of time series prediction methods from the perspectives of feature importance,data information nature and model parameters,and the specific research content is as follows.(1)To address the problem that conventional time series forecasting methods based on deep learning neural networks do not consider the different importance of time series features and lack segmentation processing for different feature information,this dissertation mainly proposes a method for time series prediction based on hybrid neural network with attention mechanism AT-CNN-LSTM.This method mixes the convolutional neural network and the long short-term memory network,uses the hybrid neural network to extract the differences of different importance features in time series,uses parallel attention as a branch,expands the receptive field of CNN,improves the performance of short-order feature extraction,and connects LSTM with the attention mechanism in the form of a pathway,so that it can more accurately mine long-order information.The results of ablation and comparative experiments show that this method can obtain more accurate prediction results when processing datasets with complex data types and large feature numbers.(2)To address the problem that the existing hybrid neural network structures cannot accurately predict the short-term mutation of time series data,and only use the time domain information,and cannot accurately model different frequency information from a single angle,this dissertation proposes a hybrid neural network structure with wavelet transform MDWT-AT-CNN-GTLSTM,establishes a deep learning model for time series prediction.Firstly,the time series data is decomposed into high-frequency and low-frequency information through multi-level discrete wavelet transformation,and then the short-order features are extracted by a convolutional neural network with an attention mechanism,and the long-order information mining by the fusion gated long-term short-term memory network pays attention to the short-term mutation information,so as to process the time series from both time and space aspects,and realize the feature classification and prediction of the time series through feature extraction and construction.Experiments on UCR dataset and feature significant difference dataset show that this method can obtain better accuracy and prediction accuracy.(3)To address the problem that CNN structure does not extract sufficient feature information during time series prediction,this dissertation proposes a method of hyper-parameter optimization in time series prediction based on hybrid neural networks.This method selects hyper-parameters according to the structural characteristics of CNN,and then adopts the algorithm of mixing particle swarm and gradient descent to optimize the hyper-parameters through strong global search performance and fast convergence,which avoids the situation that the hyper-parameter optimization value falls into the local optimization,and solves the problem of poor feature information extraction effect when processing the time series prediction task caused by incomplete optimization.Experimental results on typical datasets show that the proposed method improves the accuracy of feature information extraction without changing the structure of CNN,gives full play to the potential of CNN in time series prediction tasks,and saves hardware resources and computing costs occupied by processing.