Comparative Analysis of Hierarchical Electric Load Forecasting Techniques in the Manufacturing Sector

Abstract:
Electric load forecasting plays a crucial role in reducing energy-related costs, integrating renewable energy sources, and enhancing the system's energy flexibility. Simultaneously, increasing digitalization enables high-resolution energy consumption monitoring, such as at the machine or manufacturing hall level. Leveraging these granular measurements can improve forecasting accuracy and provide detailed predictions at multiple levels, facilitating more effective load management. This study evaluates two hierarchical electric load forecasting approaches based on machine learning techniques and compares them with a non-hierarchical baseline using the same models. The evaluation is conducted on five real-world datasets from the manufacturing sector. The results indicate that while incorporating hierarchical structure does not consistently improve prediction accuracy for datasets with shorter timeframes, it demonstrates benefits for those with extended temporal coverage. Furthermore, Temporal Convolutional Networks (TCN) and Extreme Gradient Boosting (XGB) significantly outperform linear regression (TCN 15.3% and XGB 14.4% average SMAPE improvement), with the former achieving the most consistent performance, particularly in non-hierarchical settings on temporally limited datasets and in hierarchical settings on longer datasets, while the latter slightly outperformed it in a few cases.