Smart metering data pattern recognition for identifying solar PV generation

The evolution of traditional electricity grids to smart grids requires analytical tools that can enable smart grid stakeholders to make decisions and effectively manage grid resources. Distribution network operators and other smart grid stakeholders need to use appropriate smart metering data analytic tools to address the smart grid problems they face. This thesis contributes to the knowledge base of algorithmic technologies suitable for smart grid analytics applications. Advanced Metering Infrastructure (AMI) is an integrated system of smart meters and data processing tools, and AMI is a core smart grid component. The implementation of AMI requires analytical tools for smart metering data. A focus is given to developing and evaluating tools for smart metering data analysis that enable AMI. Smart meters contain more information than simple meter readings and analytical tools to extract patterns from smart metering data are needed. More and more countries around the world have realised the importance of smart grids in recent decades and have developed smart grid projects; many households and premises worldwide have installed smart meters. The need for interpretable white box approaches to classification is addressed as much of the existing work on classification is focused on black box models. Black box methods are unsuitable for many smart grid applications due to the fact that the classifications they make cannot be explained. The methods applied in this thesis are suitable for application to real world smart metering problems, and their performance is evaluated using smart metering data provided by an industry partner to demonstrate this capacity. The work in this thesis contains three different smart metering classification approaches. They are a motif based approach, a knowledge based approach, and a tree based approach. All three approaches are demonstrated to be interpretable and effective at classification of smart metering data. The development and evaluation of these approaches contributes to the current knowledge base of smart grid data analytical tools. The first part uses a motif-based approach to identifying solar PV customers. The daily load pattern of smart metering data shows a strong periodicity, and these previously unknown but repeated patterns contain important information about electricity customers. Borrowing the idea of motifs from bioinformatics, these special patterns in the smart metering data can be seen as motifs. The motif-based approach applies Piecewise Aggregate Approximation (PAA) and Symbolic Aggregate Approximation (SAX) to segment and represent these patterns as motifs. The Nearest Neighbour method is used to classify these motifs and identify solar PV customers from their smart metering data. The second part takes a knowledge-based approach to generate pairs of templates that are used for classification. The understanding of the classification of smart metering data problem informs solving these kinds of classification problems. Template parameterisations are generated using knowledge of the classification problem and how pairs of templates can be optimised with evolutionary optimisation techniques to achieve good classification performance is shown. Applying knowledge-based approaches to define and address real industry problems can assist distribution network operators and stakeholders in building confidence in their applied analytical approaches. The third part explores an interpretable and effective approach to classifying daily load profiles that can provide insight into the significant features of the classification problem. The work in this part evaluates a tree-based approach with different preprocessing methods on the same smart metering dataset. This part shows that smart grid stakeholders have alternatives to black box classification methods. With the developed tree-based approach time can be spent finding distinguishing features in the data which can be used to construct effective classifiers, providing a very fast and inexpensive approach for addressing real industry issues.  When evaluating each of the three parts, this thesis uses the solar PV detection problem as an example of a real industry problem faced by distribution network operators. These three parts focus on developing and exploring methods to extract special patterns in the smart metering data of customers to address relevant industry issues and to expand academic knowledge of pattern recognition and smart grid analytics.