Building Energy Management: A Data-Driven Approach Using Clustering and Load Forecasting

The increase of smart meters in the grid has led to the generation of a vast amount of high dimensional energy data with improving temporal resolution. During analysis, relying on short samples like a day or week of data, could lead to wrong conclusion due to seasonal dynamics and customer behavior variations. To effectively utilize the vast amount of information, it must be compressed into a low-dimensional representation. This thesis explores the state-of-the-art dimensionality reduction techniques for a high-dimensional, non-linear energy dataset and proposes a novel deep learning based method to address the limitation of existing approaches. The proposed method exploits LSTM based Variational Autoencoders to generate an encoded representation. By utilizing this method, a 8760-dimensional dataset can be condensed into 10 dimensions, which can be further reduced by PCA to a 2D representation for visualization purposes. The effectiveness of this dimensionality reduction technique is demonstrated through its application to a large number of residential buildings, followed by the implementation of clustering algorithms on the reduced dataset. With this foundation, the thesis explores the application of clustering algorithms for energy consumption forecasting at the individual building level. Accurate load forecasting is crucial for both utilities and consumers in smart grid environments, enabling users choose a more appropriate electricity consumption scheme and resource optimization for utilities. However, forecasting load for individual building is challenging compared to the aggregated load because of high volatility and uncertainty in the load profile patterns. Several machine learning and deep learning models have been developed in the past but such exploration either produced high error or required individual model for each building, which is not feasible for large number of buildings. To address this challenge, a novel deep learning ar chitecture is developed, utilizing the aforementioned clustering algorithm to group buildings with similar consumption patterns. The forecasting model is based on a Sequence-to-Sequence architecture, with separate models developed for each cluster. Other experiments are conducted using baseline LSTM, non-clustered Seq2Seq, and Seq2Seq+KMeans models. These are performed for comparison with the proposed model. For evaluation, Mean Absolute Percentage Error is used. In summary, this thesis contributes to the field of smart grid analysis by providing innovative solutions for dimensionality reduction and load forecasting, enabling more efficient and accuracy energy management strategies in context of complex power system.