A dual-stage approach for explainable time series forecasting in on-field cyber-physical systems

Time series forecasting (TSF) plays a key role in Cyber-Physical Systems (CPS), enabling critical tasks, such as predictive maintenance, resource optimization, and anomaly detection. Deep Learning (DL) models have recently gained popularity in TSF due to their strong ability to capture temporal dependencies. However, their substantial computational requirements often necessitate cloud-based deployment, which introduces privacy and energy concerns. Moreover, their inherent black-box nature limits interpretability, making it difficult for practitioners to understand and trust their predictions.To overcome these limitations, this paper proposes a dual-stage time series forecasting approach tailored for CPS applications. The approach is designed to meet the dual requirements of local execution at the edge and model interpretability. Specifically, in the first stage, clustering is applied to group time series data with similar patterns, while in the second stage, a Decision Tree Regressor (DTR) is trained for each cluster. By focusing on a narrower subset of data, each tree operates with reduced complexity, enabling the creation of simpler and more efficient models. Additionally, DTRs are inherently interpretable, offering a clear rationale for each prediction.The lightweight, interpretable design supports efficient on-device execution with low computational demands, making it ideal for CPS scenarios. Experimental results on 250 time series demonstrate a significant improvement in accuracy compared to classical DTRs, while substantially improving interpretability and reducing computational time against tree-based ensembles, as validated on a NUCLEO-STM32N657 device.