Global vs. standalone forecasting for load-level prediction in distribution service areas

Load-level prediction in distribution areas is vital for ensuring grid reliability and resource optimization. This paper compares two modeling strategies — global and standalone — for this task. In the global approach, a shared model is trained using historical data from multiple areas, while the standalone method builds a separate model for each area. Two architectures, CNN and DLinear, are adopted. Load levels are derived from predicted load ratios and classified into three categories: light, normal, and high. Results show that global modeling consistently outperforms standalone modeling in classification accuracy under the same architecture, especially for DLinear. This study offers practical insights into collaborative forecasting across service areas.