Machine learning applications for building energy analytics

Abstract

As smart meters in buildings continue to be implemented worldwide, an unprecedented volume of time series data sets related to building energy consumption and performance have been collected. However, traditional analytics models, and physics based approaches used in the building sector face challenges in effectively handling and managing time series data from smart meters. This is mainly because of its inherent variation, rapid velocity, and the need for advanced data preprocessing methods to handle it. This thesis proposes a range of data-driven frameworks and models to address several key challenges in the domain of building energy consumption and performance. The primary focus revolves around three key components: 1) Predicting the presence of occupants in buildings through the application of pre trained data-driven models, this thesis proposes a novel transfer learning framework. This framework leverages past knowledge from similar domains, enabling efficient adaptation to new occupancy prediction tasks and boosting accuracy, especially in scenarios with scarce training data. 2) Examining the effectiveness of employing global forecasting models under the context of building energy consumption, rather than relying on a singular forecasting model. Unlike single forecasting methods, the proposed global forecasting models can simultaneously learn from multiple time series associated with building energy consumption. This helps uncover hidden connections in smart metering data, improving transfer performance and knowledge across various forecasting tasks. 3) Unravelling the underlying properties of energy consumption through the analysis of time series features, this thesis presents a forecastability framework tailored to meet this specific need. The framework relies on a feature matrix extracted through interpretable time series feature extraction techniques. Through a supervised learning, the framework is trained to establish a mapping between the extracted features and the target label of interest. This enables the prediction of how forecastable a given time series is within the context of energy consumption. Keywords: Machine Learning, Forecasting Building Energy Consumption, Smart metering.