Concept drift monitoring for industrial load forecasting with artificial neural networks

Abstract

Long Short-Term Memory (LSTM) models are frequently applied for industrial energy load forecasting. However, real-world production systems are highly dynamic which poses major challenges. Concept drifts potentially lead to performance degradation that affects systems optimization for the worse. In this work, Concept Drift Detection (CDD) for industrial energy load forecasting with LSTM models is researched. For this purpose, five CDD algorithms are evaluated using the active power of a machine tool. Drift Detection Method (DDM) and Kolmogorov-Smirnov Windowing (KSWIN) proved to be particularly effective with easily interpretable and reasonable hyperparameters.