Continual learning for very short-term load forecasting: A case study on parts cleaning

The transition to climate-neutral energy systems demands increased flexibility in electricity consumption by industrial entities. Accurate very short-term load forecasting enables adapting energy use in response to supply fluctuations. However, real-world production systems are dynamic and subject to concept drift, which causes model performance to degrade over time due to new data patterns. Consequently, the models must incrementally acquire and consolidate knowledge which is referred to as continual learning (CL). In CL, catastrophic forgetting (CF) poses major challenges by causing performance on earlier tasks to degrade after learning new ones. In our study, we developed a novel framework based on concept drift detection (CDD) and CL for adaptive modeling that proves to effectively mitigate CF and offers high knowledge retention. The validation was conducted using data from a throughput parts cleaning machine (TPCM) which is part of a representative production chain of the metalworking industry. Five regularization-based CL methods were compared providing insights into the relative strengths and weaknesses of the algorithms under identical conditions. The experimental results show that synaptic intelligence (SI) and memory aware synapses (MAS) improve forecasting performance by 21 % compared to traditional offline learning approaches. While learning without forgetting (LWF) and online elastic weight consolidation (OEWC) provide increased robustness, LWF is additionally characterized by its ease of use. Furthermore, these methods introduce minimal computational and memory overhead. The findings confirm that the proposed CDD-CL framework enables efficient and robust load forecasting in dynamic industrial environments.