Energy Analytics: Application of Machine Learning to Electricity Data Analytics

Fossil fuel usage results in consequences that harm the environment, economy, and human life, among which climate change is the most important one. Renewable energy resources and electrification of conventional systems are potential ways to replace fossil fuel usage. A smart grid is an infrastructure that facilitates the transition to more environmentally friendly energy sources. Smart meters provide large-scale data that raises the need for analysis. Such analysis is crucial to maintaining and planning to build more efficient smart grids. Traditional methods for analysing time series data recorded by smart meters are not flexible and do not generalise well to unseen cases. Deep learning techniques have proved to be useful in analysing such complex data and providing insights for smart grid planners and operators. The aim of this thesis is to explore the effectiveness of deep learning techniques in analysing energy data. The thesis develops a new tool to forecast energy generation from photovoltaic (PV) cells. Solar energy is one of the main forms of renewable energy resources which is captured by PV cells. Since PV generation heavily depends on heavily uncertain variables, like weather conditions, the PV power generation problem is a challenging problem. The stochastic nature of the input features introduces uncertainty that makes power generated from PV cells less dispatchable compared to conventional generation methods. Therefore, to make PV generation better dispatchable, an accurate forecast is required. We propose a deep learning-based forecaster, that combines the power of convolutional and recurrent neural networks, that outperforms state-of-the-art methods. The second research addressed here is to identify electric vehicles (EVs) from smart meter recordings. The electrification of transport systems will help decarbonize the transport systems. However, charging EVs imposes overhead on the grid that can cause voltage fluctuations. Therefore, EV identification has a significant application in the management of the local grid by energy distributors. The major issue is that the number of EV users is much lower than non-EV users which raise the imbalance issue for the dataset. Inspired by anomaly detection ideas, we propose a novel EV identification technique based on deep learning that takes an unsupervised approach and outperforms baselines. By doing so, we do not need to manipulate the input data. Instead, we rely on the abundance of the non-EV part of the input data. Finally, the thesis develops deep learning techniques for fault/anomaly detection in multi-sensor systems. The major issue associated with this problem is the lack of labeled time series which raises the need for unsupervised learning. The other challenge is the multivariate nature of such data and the correlation information that exists between several sensors. Often, energy data come in the form of multivariate time series with significant inter correlations. Taking this correlation into account is a major challenge that impacts the performance of anomaly detectors. We propose a deep learning that considers the correlation between different sensor time series explicitly. Our method models both temporal and spatial correlations between different time series. We prove empirically the effectiveness of our model.