Abstract

Energy, as an essential aspect of socioeconomic growth, has remained an intriguing issue for many researchers worldwide. The rising need for energy drives academics and researchers to develop novel solutions for improved energy utilization.
The main contribution of this thesis is two-fold. First, we introduce an occupancy prediction framework to improve the efficiency of energy management systems in smart buildings. Occupancy prediction heavily depends on historical occupancy-related data collected from various sensor sources. Unfortunately, a major problem in that context is the privacy constraints of collecting training data. This situation inspired us to rethink the occupancy prediction problem, proposing the use of an original principled approach based on occupancy estimation via interactive learning to collect the needed training data. Following that, the collected data, along with various features, was fed into several algorithms to predict future occupancy. This fold mainly proposes a weakly supervised occupancy prediction framework based on office sensor readings and occupancy estimations derived from an interactive learning approach. Two studies are the main emphasis of this part. The first is the prediction of three occupancy states, referred to as discrete states: absence, presence of one occupant, and presence of more than one occupant. The purpose of the second study is to anticipate the future number of occupants, i.e, continuous states. Extensive simulations were run to demonstrate the merits of the proposed prediction framework's performance and to validate the interactive learning-based approach's ability to contribute to the achievement of effective occupancy prediction.

Second, given that an accurate electric power load forecasting framework is critical for power utility companies as it increases control over the relevant infrastructure, resulting in significant improvements in energy management and scheduling, we propose an encoder-decoder model that takes advantage of the expressiveness of transformer-based encoders to produce probabilistic forecasts. Two real-world datasets are utilized to incorporate the performance of the proposed framework on two different types of data: hourly load data from the power supply company of the city of Johor in Malaysia and hourly load consumption data from one of Grenoble Institute of Technology's buildings. The former represents aggregated data, which makes identifying patterns and trends easier, but the latter was taken from a single building (non-aggregated), which increases the difficulty of forecasts. The model's performance is discussed across multiple time horizons, including 24-hour, 1-week, and 1-month predictions. The framework achieved notable improvements compared to the used baseline, Amazon DeepAr, where accuracy was improved from 87.2 percent to 96.2 percent for Malaysian data and from 52.3 percent to 68.2 percent for Grenoble data for 24 hours ahead forecasting, from 84.7 percent to 89.7 percent for Malaysian data, and from 45.5 percent to 57.2 percent for Grenoble data for 1 month ahead forecasting.