Multiscale attention-guided LSTM for time series prediction

Abstract

With the advancement of the Industrial Internet of Things (IIoT) and smart manufacturing, multivariate time series forecasting has emerged as a critical pillar for equipment monitoring and production control. However, industrial time series often exhibit significant non-stationarity and multi-scale dynamic characteristics, and traditional time series prediction methods still have certain limitations in complex fluctuation intervals, especially in peak event prediction. This not only heightens operational uncertainty but also compromises the timeliness of anomaly detection and early warning systems. To address this issue, this paper proposes a time series prediction model based on multi-scale and attention-guided mechanisms, called MS-AG-LSTM (Multi-Scale Attention-Guided LSTM). This model extracts temporal features at different time scales through a multi-scale LSTM encoding structure and uses a temporal attention mechanism to adaptively weight key time segments, thereby enhancing the model's ability to express complex temporal dependencies. At the same time, a multi-task learning framework is introduced to jointly optimize the future statistical prediction task and peak risk classification task, improving the model's perception of key peak events. Experimental results show that compared to classical deep learning prediction methods, the proposed method achieves better performance in metrics such as RMSE and MAE, and demonstrates better stability in peak interval predictions. These findings indicate that the proposed method offers effective support for complex industrial time series forecasting and related condition monitoring tasks.