Bridging the Development–Production Gap in Industrial Energy Management Systems

An Experimental Study for Data Drifts

Abstract

Industrial Energy Management Systems (EMSs) combine data-driven forecasting with optimisation to schedule Behind-the-Meter (BTM) batteries, Photovoltaic (PV) generation, grid exchange, and flexible loads. However, when industrial operating conditions change after deployment, offline-trained models degrade, affecting the quality of operational decisions. This paper studies the impact of data drift on an industrial warehouse EMS with a BTM battery, PV generation, and mandatory forklift charging. Drift regimes are detected using a multivariate Maximum Mean Discrepancy (MMD) permutation test and grouped into non-drift, forklift-drift, and demand-drift periods. Six prescriptive configurations are compared: static Predict-then-Optimize (PTO) and Predict-andOptimize (PAO) baselines, their continual-learning (CL) variants with replay memory, and two robust optimisation (RO) formulations based on box and ellipsoidal uncertainty sets. Results show that forecasting accuracy is an insufficient proxy for decision quality, as lower MAE does not necessarily yield better schedules. CL improves performance under critical-load drift, reducing the optimality gap by up to 83%, while ellipsoidal RO achieves the best aggregate performance, reducing the optimality gap by 87% relative to PTO. These findings motivate drift-aware, decision-centric evaluation for deployed industrial EMSs.