Demand-side analytics for sustainable energy

Abstract

The rapidly growing demand for energy poses one of the biggest challenges in our society. This challenge is critical not just for the power utilities but also for the environment as it leads to increased carbon footprint and climate change. There have been concerted efforts towards sustainable energy. One of the main contributions towards effective energy management is the deployment of smart grid technology, in particular including demand-side management (DSM) techniques. DSM involves tasks like non-intrusive load monitoring or energy disaggregation, demand forecasting, anomaly detection, outage management, etc. These tasks empower customers to make more informed decisions about their energy consumption, adjusting both the quantity and timing of their energy usage. The goal is to reduce the overall energy consumption and also to save the cost of building an additional generation capacity to meet the critical peak demands. In the first DSM task, the problem of energy disaggregation is explored. Energy disaggregation is a single channel blind source separation problem where the task is to estimate the consumption of each electrical appliance given the total meter reading. In this work, two issues that are often overlooked - the problem of missing data, and, the problem of outliers are addressed. The first problem arises when the smart-meter cannot transmit the readings to the server owing to the failure of wireless communication interfaces. The second problem arises from transients, surges and other non-linear effects. A modified dictionary learning-based disaggregation framework is used to address these problems. A recent work [1] in this area showed that instead of employing the usual Euclidean norm cost function for dictionary learning, better results can be achieved by learning the dictionaries in a robust fashion by employing an l1- norm cost function. This is because energy data is corrupted by large but sparse outliers. In the second work, an approach to improve robust dictionary learning is proposed. This is done by imposing a low-rank penalty on the learned I coefficients. The ensuing formulation is solved using a combination of Split Bregman and the Majorization Minimization approach. In the third work, the existing work of robust dictionary learning is extended by modeling non-linear perturbations as sparse error and applying robust versions of dictionary learning for disaggregation. On top of the basic (unsupervised) robust dictionary learning formulation, two supervised variants are proposed. In the first supervision, dictionaries are learned such that they are incoherent; this ensured that the dictionaries from different appliances look different from each other. In the second formulation, discriminating sparse codes are introduced, such that the codes generated for each appliance would not look alike. In the final work in the area of energy disaggregation, a new method based on the transform learning formulation is introduced. Several recent techniques, such as discriminative sparse coding, power let disaggregation, and deep sparse coding, are based on the synthesis dictionary learning/sparse coding approach. The proposed method is based on its analysis equivalent. The theoretical advantage of the analysis dictionary compared to its synthesis counterpart is that the former can learn from fewer training samples - this has implications in reducing the cost of energy disaggregation. The next contribution to DSM in this thesis is load forecasting. It is a technique used by power utilities to predict the power needed to meet the demand and supply equilibrium. In the first contribution towards this task, the problem of one-day-ahead short-term load forecasting is considered. As the effects of weather, as well as prior consumptions, are nonlinear functions, the formulation is based on non-linear Kalman filtering algorithms. In the second work, the focus is to improve the accuracy of building-level demand forecasting. For the said purpose, a regressing deep dictionary learning approach is proposed. There are two versions of the algorithm - synthesis and analysis. In this work, point forecasting, as well as profile forecasting, is performed. The last contribution to DSM is to detect abnormal energy consumption behavior in residential buildings. Understanding, identifying, and addressing abnormal energy consumption in buildings can lead to energy savings and the detection of faulty appliances. This work investigates two key challenges found in energy anomaly detection research: (1) the lack of labeled ground truth, and (2) the lack of consistent performance accuracy metrics. In the first challenge, labeled ground truth is imperative for training and benchmarking algorithms to detect anomalies. In the second challenge, consistent performance accuracy ii metrics are crucial to quantifying how well, algorithms perform against each other. Two approaches that help in the automatic annotation of the ground truth data from publicly available datasets are proposed: a statistical approach for short-term data and a piecewise linear regression method for long-term data. Finally, we aim to detect anomalies that we define as power consumption during a power outage (negative anomaly) and power theft (positive anomaly). A robust principal component analysis (RPCA) technique for separating anomalies from the normal component is employed.