Energy Forecasting and Decision Making using Data Analytics in Smart Grid

Abstract

In a world driven by a massive demand for energy, energy forecasting and intelli gent decision making are crucial. Load forecasting remains challenging due to the non-linearity, multiple seasonality, varying variance, and random fluctuations in time series data. This thesis explores several statistical methods for short-term load fore casting using multiple hourly load time series to address the challenge of multiple sea sonality. We drive the motivation to propose an extended version of a state-of-the-art short-term load forecasting method combining statistical method and deep learning model from the literature. Our proposed forecasting method combines context vector based dilated Recurrent Neural Network (RNN) with Holt Winter’s inspired exponen tial smoothing method. The time series are preprocessed for deseasonalization and normalization using the exponential smoothing method. The preprocessed time series are fed to the RNN, adopting the sliding window size to capture the weekly seasonal pattern. The context vector based dilated RNN further helps capture the seasonality and long-term dependencies. Analysis of simulation results and study of error metrics demonstrate the potential of our proposed short-term load forecasting methodology. Furthermore, this thesis also studies the effect of environmental variables and building design for improved energy management. Simulations of our proposed con troller for heating, ventilation, and air conditioning system are performed with differ ent building energy designs. The result analysis shows that good thermal insulation and room seal status can help reduce almost half of the total building energy con sumption without disturbing the thermal comfort of occupants. Independent simu lations for the working of our proposed lighting controller are also conducted. The illuminance setpoints are intelligently decided by our proposed fuzzy logic based en ergy controller considering different outdoor and indoor variables. Simulation results show that considering a visual comfort standard with a higher luminance value still minimizes energy consumption without jeopardizing visual comfort. The combined result analysis confirms that our model helps improve user comfort without paying any additional cost.