Research On COVID-19 Propagation Trend Prediction Algorithm Based On Deep Learning

In the real world,non-smooth time series data are more prevalent,and their data distributions become more complex over time.Traditional research methods usually employ smoothing techniques to eliminate temporal distribution differences to improve the predictability of the data.However,in COVID-19 time series data,as a kind of time series data with non-smoothness,the predicted value of the model often lags behind the true value when the smoothing technique is used to predict its propagation trend.In addition,the smoothing technique may lead to overfitting of the model and fail to accurately reflect the characteristics of the real data,both of which can seriously affect the accuracy of the prediction.Therefore,this study aims to improve the accuracy of COVID-19 propagation trend prediction by deeply mining the features of COVID-19 time-series data and optimizing the deep learning model structure to better accommodate non-stationary time series modeling.Higher requirements are placed on the modeling capability and generalization of the deep learning model,and the specific research work includes the following aspects:(1)The data set used for this study was constructed using global and U.S.daily confirmed case data,and these data were preprocessed and characterized.The COVID-19 time series decomposition and Augmented Dickey-Fuller test(ADF)hypothesis test showed that the timeseries data of COVID-19 were non-stationary.(2)To solve the problem that the predicted value lags behind the true value in COVID-19 prediction,an improved Multi-Layer Deep Time Convolutional Neural Network(MDTCNet)prediction algorithm is proposed in this paper.The method has the ability to capture the deep features and long time dependence of time series,and uses a multilayer perceptron based feature fusion network for prediction output to improve the nonlinear representation of temporal features.Multiple benchmark models are compared in the experiments,and it is shown that the method significantly reduces the prediction lag and improves the prediction accuracy in the global and U.S.COVID-19 propagation trend prediction tasks,and achieves both short-and long-term prediction of COVID-19 propagation trends.(3)A Transformer prediction model based on time series decomposition(Sequence Decomposition Transformer,SDTransformer)is proposed for the non-smoothness of COVID-19 time series data.The model is internally constructed with a time series decomposition module,which aims to discover more potential complex patterns in the data and learn the relationship between these patterns through an attention mechanism,thus improving the feature extraction ability of the model for smooth time series.The experimental results show that the time series decomposition Transformer model proposed in this paper achieves higher prediction accuracy in COVID-19 propagation trend prediction compared with the multilayer deep time convolutional neural network and Transformer prediction model,and verifies the effectiveness of the model in predicting non-stationary time series from several perspectives.