Multivariate intermittent time series forecasting with fusion of structured information and temporal evolution information

After-sale accessory demand of complex equipment manufacturing enterprises tends to occur irregularly and fluctuates greatly, which leads to typical intermittent and small sample characteristics of demand data. When facing the series with high intermittence and large sudden demand, the existing prediction models are difficult to accurately capture the demand fluctuation rule and can not effectively predict the trend of parts demand. In order to improve the prediction accuracy and stability of multi-group parts, this paper proposes a new multivariate intermittent time series forecasting method by considering both structured information between sequences and time series evolution information. Firstly, a tensor-based light gradient boosting machine model is proposed, the original demand data is reconstructed through tensor decomposition to correct abnormal demand values in the sequences, and use the light gradient boosting machine to jointly predict multiple sequences. Then, a new linear decay correction model is constructed, and a correction factor is introduced into the linear decay exponential smoothing method to predict the demand and interval respectively for each sequence. Finally, the two prediction models are weighted and fused to obtain the final prediction results. Experimental validation is conducted on two complex equipment manufacturing enterprises' aftermarket parts demand datasets respectively, and the results show that, compared with several time series forecasting algorithms, the proposed method can effectively predict the demand fluctuation trend and improve the prediction accuracy and numerical stability. © 2024 Northeast University. All rights reserved.