Generative Learning in Smart Grid

If a smart grid is to be described in one word, that word would be ’connectivity’. While electricity production and consumption still depend on a limited number of physical connections, exchanging data is growing enormously. Customers, utilities, sensors, and markets are all different sources of data that are exchanged in a ubiquitous digital setup. To deal with data complexity, many researchers recently focused on machine learning (ML) applications in smart grids. Much of the success in ML is attributed to discriminative learning where models define boundaries to categorize data. Generative learning, however, reveals how data is generated by learning the underlying distribution functions. In the past few years, generative models brought new dimensions to various domains. Computers became painters and composers. This thesis identifies three applications in the smart grid where generative learning has great potential. On the demand side, residential loads such as dishwashers and clothes driers are simulated using generative models. In specific, the latest developments in generative adversarial networks and kernel density estimators are levered to learn the underlying distributions of individual loads for both power consumption patterns and usage habits. Being data-driven, the learning process eliminates any biases introduced by rule-based models where predetermined fixed formulas describing each load are considered. The study demonstrates the flexibility, viability, and remarkable accuracy of the proposed framework. The resulting synthetic power consumption patterns and usage habits for individual loads are valuable sources for researchers to build or improve their data-driven models for demand-side studies. Non-intrusive load monitoring (NILM) is the second topic researched on the demand side. The goal in NILM is to identify the status of individual loads in a household by merely relying on a smart meter’s measurements without any hardware installations. The research focuses on identifying the operational condition of individual loads by developing a novel hybrid algorithm that combines the widely used generative technique, namely, hidden Markov model, with k-means clustering. The hybrid model is demonstrated to accurately identify the operation conditions of individual loads based on the ingested aggregate signal. Finally, for power transmission, a combination of generative models is proposed to estimate power states from a set of redundant measurements. Power state estimation is a fundamental technique in shedding light on the operational condition of the grid. A traditional state estimator is typically executed online and, in its non-linear formulation, involves a high level of computational complexity. Generative models shift that burden to the offline learning process. On the other hand, bad data detection and identification is a central feature in traditional estimators. As such, the developed framework integrated that feature in the data-driven state estimator by incorporating forward and backward generative adversarial networks. Simple domain knowledge is further incorporated in the model to improve its accuracy against the benchmark data-driven model. The proposed framework remarkably detected tampered measurements including false data injection.