On the path to climate-neutral production, industrial energy systems require greater flexibility in electricity consumption. Accurate short-term load forecasting using artificial neural network (ANN) models enables industrial entities to adapt energy consumption in response to fluctuations in energy supply. Load forecasting plays a key role in the operation of industrial energy systems, allowing for optimized energy management and significant reductions in energy consumption and costs. While maintenance of physical assets is considered state-of-the-art, the continual operation of ANNs poses a long-standing challenge. This is because ANNs are prone to catastrophic forgetting (CF), which describes the process by which learning new patterns erases the existing knowledge of the network. To overcome the obstacle of CF, we propose a framework that combines concept drift detection (CDD) with continual learning (CL) methods. For CDD, drift detection method (DDM) from statistical process control is applied. The retraining of the ANN model with CL is based on two rehearsal methods with k-means clustering and random sampling, in which a suitable core dataset is constructed from the data stream. The framework is validated using a long short-term memory (LSTM) model for short-term load forecasting of the main active power connection of an industrial throughput parts cleaning machine (TPCM) and benchmarked against a state-of-the-art learning in isolation (LI) approach. The results show that DDM with k-means clustering improves forecasting performance by 9.1 % compared to LI while introducing minimal computational overhead. Consequently, the methods presented contribute to adaptive, efficient, and robust load forecasting in industrial production systems.
Zink et al. (Thu,) studied this question.