Research On Prediction Methods For Correlation Characteristics Of Time Series Data

Time series prediction is hot topic research in the field of deep learning applications.Capturing complex correlation characteristics between time series data is the key problem to accurate prediction.The current researches fail to provide solutions to the following problems.Specifically,how to deal with the different degrees of impact of multiple non-predictive time series on the target series at different time stages.The mutation phenomenon contained in the time series data will significantly affect the change rule of the target series,how to learn this information in the historical data through the prediction method.In most previous studies,the prediction method is usually designed only from the perspective of the neural network model,while the objective function and optimization algorithm completely ignore the temporal change information.In this paper,we study the prediction methods for the correlation characteristics of time series data from three perspectives: the objective function,the neural network model and the optimization algorithm to help the model converge in iterative training.(1)In aspects of objective function,this paper proposes a novel time series prediction evaluation mechanism to learn characteristics of temporal correlation.Specifically,this paper proposes to include the mean absolute error(MAE)and root mean square error(RMSE)loss in the objective function to evaluate different temporal change amplitude errors.Meanwhile,the second-order difference technique is used in the penalty term of the objective function to adaptively capture the impact information of the abrupt and slow change information of each sequence on the target series.(2)In aspects of the neural network models,this paper proposes a series of prediction models that capture short-term abrupt change and temporal dependencies and complex temporal changes.Specifically,this paper first designed a variant of the long-short-term memory(LSTM)network,introducing transformation gating based on the deformation of the hyperbolic tangent function to map the temporal information flow to the most obvious interval to better capture short-term temporal change information.Subsequently,in order to capture the different impact information of different non-predictive time series on the target series,this paper designs a new multi-stage attention network composed of influence attention mechanism and temporal attention mechanism.In the influence attention mechanism,the same and different time stage attention mechanisms are used to capture this influence information.In addition,an attention score readjustment mechanism is designed in this model to comprehensively capture the influence weights of multiple non-predictive time series.Aiming at the problem of capturing mutation information in historical observation data,this paper proposes a new deep learning model to solve.In the attention mechanism of the encoder,a new function mapping mechanism is designed to fuse historical hidden state and cell state information.Moreover,the LSTM with a transformation mechanism is used in the encoder to capture the mutation information existing in the input information flow.Besides,an adaptive self-paced curriculum learning mechanism is designed to obtain mutation information that may be ignored between small batches of samples.(3)In aspects of optimization algorithms,this paper proposes two adaptive stochastic optimization algorithms that consider the temporal correlation information in the gradient error flow.Specifically,this paper proposes an optimization algorithm(called Adaptive Hybrid Multidimensional Gradient,Ada HMG)based on the popular adaptive learning rate optimization algorithm framework for the correlation characteristics of time series data.The algorithm uses a mixed high-order multi-dimensional update strategy for the second-order moment estimation history and current information to process the temporal information in the error gradient flow.Subsequently,this paper also solves the problem that the Ada HMG algorithm introduces too many hyperparameters to increase the difficulty of parameter adjustment.In this paper,a new optimization algorithm(called Adaptive Moment Forget Gradient,Ada MFG)is proposed to adaptively memorize and forget the first-and second-order moment estimation information,so that the proposed optimization algorithm greatly improves the ability to capture the temporal change information in the error gradient flow.Finally,this paper provides comprehensive and in-depth experimental research on multiple open source time series datasets with different scales and different application fields.The experiment results not only verified that the proposed method in this paper is more effective and advanced than the state-of-the-art(SOTA)model,but also through the ablation experiment to clarify the role of each component in the solution.