Time Series Forecasting Based On Contrastive Learning

In recent years,time series forecasting has received much attention due to the continuous development and popularization of data science technology.Time series forecasting is a method of forecasting future data based on the trends and periodicity of historical data.Time series forecasting is widely used in finance,meteorology,transportation,and other fields,and is an important tool for decision-making and resource allocation.Contrastive learning is an effective method for time series forecasting,which is based on the model’s learning of feature representation,and can better handle the complexity and non-stationarity of time series data,improving the accuracy of prediction results.The research work of this paper includes the following aspects:(1)A contrastive learning model based on hierarchical contrast and compound consistency is proposed to address the problem of learning a universal and generalized representation of time series data.Based on the representations learned by the contrastive learning model,a forecasting model is constructed.Hierarchical contrast allows the model to learn fine-grained,multi-scale time stamp representations,while compound consistency enables the model to effectively learn the general representation of time series.As a result,the contrastive learning-based forecasting model has better forecasting performance and transferability.(2)A novel contrastive learning model is proposed in this study based on multivariate adaptive wavelet decomposition and non-stationary attention blocks.The objective is to overcome the challenges of the severe degradation of model prediction performance and the ineffective modeling of important frequency information in non-stationary time series data.Multivariate adaptive wavelet decomposition can capture time correlation that is closely related to frequency,and promote time-frequency consistency contrast.Non-stationary attention blocks can learn the non-stationary characteristics of the original sequence that are eliminated by the stationary processing.Therefore,by combining multivariate adaptive wavelet decomposition and non-stationary attention blocks to improve the existing contrastive learning model,it can achieve superior performance in non-stationary time series prediction tasks.(3)A contrastive learning model based on frequency-improved Legendre Memory Units is proposed in this study to tackle the challenge of retaining historical information in neural networks without overfitting to the noise present in the history.The Legendre Memory Unit,which has stronger projection capabilities for historical information,replaces the original Embedding in the non-stationary attention block of the existing model.The Fourier transformation in the frequency-enhancing layer effectively filters out the noise in the historical information to prevent deviation from the prediction,thus improving the long-term forecasting performance of the existing model.