A Neural Network-Based Approach for Hierarchical Electricity Consumption Forecasting in Buildings

Energy forecasting in buildings is essential for efficient energy management and sustainability. Conventional forecasting methods often struggle to capture the complex behavior of energy usage in buildings, especially in hierarchical settings where energy consumption may vary across different energy uses or levels of aggregation (e.g., energy used for heating a room versus a floor or the complete building). This is because each load is typically forecasted separately, leading to incoherent forecasts across the hierarchy and preventing information exchange among the individual series it consists of. In light of these limitations, this paper explores the impact of hierarchical structure and granularity on forecasting accuracy. Unlike traditional methods that rely on a two-step reconciliation process, requiring the utilization of the bottom-up, top-down, or middle-out methods to ensure forecast coherency, the proposed method utilizes a global model to test different hierarchical configurations and make predictions directly without requiring additional reconciliation. Our findings reveal that models trained with deeper hierarchical understanding and finer granularity consistently outperform traditional hierarchical-forecasting methods, capturing unique consumption patterns at different levels of the hierarchy. However, they also reveal a trade-off between forecasting accuracy and computational efficiency, with finer granularity models incurring higher computational costs. The results underscore the importance of irregularities within the data set and how they affect forecasting accuracy. The study contributes valuable insights into the development of forecasting models for energy consumption in buildings, enabling better-informed decision-making and resource management.