Load Forecasting for an Industrial Building with Aggregated and Disaggregated Load

Short-term load forecasting in multipurpose research buildings is challenging because diverse load categoriesex hibit heterogeneous and often unpredictable usage patterns, complicating the design of reliable energy-management strategies. This work provides a systematic empirical benchmarkth at compares direct total load prediction with a disaggregated approach that forecasts each load category separately and aggregates the results. Using Long Short-Term Memory networks applied to different load categories across multiple training durations and seasons, we deliver practical guidance on when to use each aggregation strategy. Our analysis reveals that direct forecasting achieves lower mean absolute error in most scenarios, with disaggregation providing benefits only for loads with independent, sporadic usage patterns. We explore conditionsfavoring disaggregated predictions, including behavioral diversity between categories, predictability of individual components, and representative training data