A Granular Computing-Based Approach On Long-Term Prediction Of Time Series

Time series are widely presented in various fields such as finance,environment,and agriculture,and their research has gained continuous development driven by practical activities.In today’s time series research,it is an important task to provide accurate and interpretable long-term prediction.Traditional models focus on one-step-ahead forecasting,which usually require iterative computation when applied to long-term forecasting,resulting in the accumulation of errors that deteriorate the forecasting performance.On the other hand,most models tend to emphasize numerical accuracy while neglecting the interpretability and understandability of the model.Aiming at imitating human cognitive and reasoning mechanisms by employing granular computing(Gr C)as a vehicle,this paper present a novel approach to realize the long-term time series prediction with good numerical accuracy as well as semantic interpretation and comprehensibility,providing a new view for time series prediction.The proposed method first employs a sliding window strategy to smooth the raw time series and retain the main dynamic behavior.Subsequently,the smoothed time series is transformed into the corresponding granular time series consisting of information granules which are generated after the change patterns associated with semantic descriptions are extracted with the aid of clustering algorithm based on Dynamic Time Warping(DTW)distance.Finally,a Takagi-Sugeno(TS)architecture like granular model(Gr M)is formed by deriving the relations implying in the granular time series,and offers pattern inference and vector-level numerical prediction outputs.By simulating human intuition and inference,the method captures and portrays the dynamic behaviors and their evolutionary rules naturally existing in the time series with Gr C,enabling excellent interpretability and semantics in the information granulation,model construction and prediction processes,as well as easy understanding.The Gr M adopts a vector-based pattern-to-pattern reasoning mechanism that can effectively avoid cumulative errors,yielding more accurate long-term prediction of time series at the information granule level.Experiments on several datasets demonstrate that the proposed method achieves better numerical accuracy in long-term prediction along with better interpretability and semantics.