Efficient Load Forecasting Model Assessment for Embedded Building Energy Management Systems

Growing urbanization tendencies have increased the role that buildings play in future energy systems. We present a comparative study of load forecasting methods for large commercial buildings’ energy traces. Forecasts can be used to improve control algorithms at the local and grid level and to devise practical anomaly detection algorithms. The main approaches for classical modeling, machine learning, and deep learning methods are illustrated with considerations for statistical relevance and contextual information. Model parameter design is achieved using heuristics or systematic parameter space searches for robustness. For the available building consumption datasets, two short-term prediction methods are deployed: Seasonal AutoRegressive Integrated Moving Average (SARIMA) and Recurrent Neural Networks (RNN). An extensive evaluation of prediction performance was conducted based on several metrics. The daily load energy forecast for the 7 considered buildings showed higher precision when using RNN methods compared to the SARIMA auto-regressive algorithm. Implementation of the proposed embedded online building management system can be replicated for other buildings.