Research On Time Series Data Prediction Method Based On Multi-model Fusion

Time series is an orderly and interrelated data series with the time scale,and the data characteristics have a certain change law with time.Time series data forecasting has received extensive attention in various fields such as business,environment,medicine,and industry.Accurate forecasting plays an important role in saving resources,improving efficiency,reducing costs,and optimizing configuration.Time series prediction method is a means to calculate future observations according to historical time series data and relevant characteristics.At present,the prediction of time series mostly depends on deep learning methods.Although many achievements have been achieved,there are still many deficiencies in the preprocessing of time series data,periodic feature extraction and fuzzy prediction of high-dimensional data.Aiming at the problems of data loss and abnormal data in daily life,a time-series data preprocessing method(BE-GAIN)based on the fusion of Boundary Balanced Generative Adversarial Network(BEGAN)and Generative Adversarial Imputation Network(GAIN)is proposed,which uses the proportional control coefficient to ensure the interpolation equilibrium in the interpolation process,At the same time,the loss reconstruction of the Generative Adversarial Imputation Networks(GAIN)is carried out,and the global convergence measure is used to stabilize the convergence,so as to finally produce the optimal estimation of the missing and abnormal data.The datasets from five different domains are employed in the experiments.The results show that BE-GAIN has higher data interpolation performance than common interpolation algorithms,which establishes a reliable data set for subsequent prediction.Complex periodic feature extraction has always been the difficulty of time series prediction.In response to this problem,A fusion network model(TSANet)based on temporal convolutional network(TCN)and self-attention(Self-Attention)mechanism is proposed.The TSANet model first uses TCN as the core prediction model,and then uses the parallel global and local convolution modules and the fusion structure of Self-Attention to enhance the feature correlation,and uses the Long Short-Term Memory(LSTM)network as the unit component of the local convolution module.Finally,Autoregressive(AR)models are then used as a means to capture linear features.Experimental results present that the TSANet model has higher prediction performance compared to popular deep learning models.The prediction cost of high-dimensional time series data is high on TSANet and most complicated depth model.To decrease the sophistication of prediction,A network model based on matrix factorization(MF)model and TSANet fusion(TSANet-MF)is proposed for the application scenario of high-dimensional time series data ambiguity.The TSANet-MF model uses the TSANet model to perform the temporal regularized MF algorithm,which further improves the conversion performance of high-dimensional time series data to low dimensional time series data training,and finally uses TSANet to achieve fast fuzzy prediction of the original data.The experimental results present that the TSANet-MF model has higher prediction efficiency compared with TSANet,and also has excellent prediction performance compared with the benchmark model.