Incremental Time Series Forecasting Based On Ensemble Learning

Time series prediction has an important role in real life and is a hot research area in the field of data mining.Time series data collected in nonstationary environments inevitably have variations in their data distribution,that is concept drift,which is the information that needs to be captured in focus for time series prediction tasks.Traditional offline time series analysis techniques based on statistics and deep learning,which are premised on the assumption that the predicted and historical data are independently and identically distributed,cannot effectively deal with the dynamic changes in data distribution.Although the periodic and quantitative incremental update model can solve the above problems to a certain extent,it cannot timely extract the distribution information of the latest data and adjust the model.Therefore,this paper uses ensemble learning technology to study how to capture concept drift in a timely and accurately,and incrementally update the forecasting model in order to improve the accuracy of time series forecasting.The main research content and contribution of this paper are as follows:(1)To outline and summarize existing time series forecasting techniques,incremental learning techniques,compare and analyze the advantages and disadvantages of each technique,and lay the necessary foundation for research on incremental time series forecasting based on ensemble learning.(2)Aiming at the problem that the periodic and quantitative incremental update model does not respond timely to the concept drifts existing in the time series,an adaptive incremental time series forecasting algorithm(Weight Local Adaptation And Adaptive Incremental Ensemble,WLAAIE)based on online ensemble learning is proposed.The algorithm first performs online prediction on the current time series data,and dynamically adjusts the weight of the model according to the prediction error to extract the evolution information of the data distribution near the concept drift point in time,then periodically train a new base prediction model with the incoming data,and finally,the ensemble prediction model is incrementally updated using the model weight adaptive strategy.Experiments on real datasets show that the WLAAIE algorithm has higher prediction accuracy than the existing classic algorithms such as DIL and IncLSTM.(3)Aiming at the problem that the periodic and quantitative incremental update model has a large number of unnecessary updates when predicting relatively stationary time series data,an incremental time series forecasting algorithm based on two-stage concept drift detection(Double Stages Concept Drift Detection And Adaptive Incremental Ensemble,DSCDAIE)is proposed.The algorithm first captures the concept drift accurately through two-stage concept drift detection.The first stage of detection uses two performance detection time windows to reflect the adaptability of the prediction model to the current data.The second stage of detection uses the results of statistical hypothesis testing in order to reduce the possibility of misjudgment caused by false concept drift,and finally,the results of concept drift detection are used to build incremental prediction models,reducing a large number of unnecessary model updates.The experiment on real datasets shows that the DSCDAIE algorithm has higher prediction accuracy and less time consumption than the existing IncLSTM,WLAAIE and other algorithms.