Forecast-Driven Energy Management for a Smart Building Microgrid

Smart building microgrids contain multiple interacting resources that evolve on different time scales and under different sources of uncertainty. In the studied office-building setting, photovoltaic generation, tenant load, heat pump operation, battery energy storage, and electric vehicle charging together create a coordination problem for the building energy management system (EMS) rather than an isolated scheduling task for any single component.

This thesis aims to develop and evaluate a forecast-driven two-layer EMS for Magasin X. The study is based on real operational data from the building and combines external weather and electricity-price information with measured building and charging data. An autoregressive with exogenous input (ARX) thermal model is used to represent short-term building thermal dynamics for heat pump scheduling, while long short-term memory (LSTM) models are developed for forecasting photovoltaic generation, tenant load, and electric-vehicle-related demand. These models are integrated into a day-ahead EMS and a rolling EMS, and the resulting strategies are compared against an uncontrolled baseline and a perfect-foresight benchmark. The results show that the forecasting and thermal submodels, although imperfect, are sufficiently informative for downstream scheduling. In the annual comparison, the day-ahead and rolling strategies reduce grid-import cost by 27.01% and 29.95%, respectively, compared with 32.55% for the perfect-foresight benchmark. The rolling strategy also increases photovoltaic self-consumption to 84.99% and reduces monthly peak import in most months. For electric vehicle charging, 93.89% of sessions achieve at least 80% fulfillment and 82.00% reach at least 90%, indicating that the economic gains are accompanied by a measurable service-quality trade-off. The thesis therefore concludes that building EMS in this context should be understood as a multi-objective coordination problem rather than as a pure cost-minimization problem, and that forecast-driven two-layer control is practically valuable for the studied office-building case.