Physics-Informed Machine Learning with Applications to Power Systems
- Zhang, Shaorong
- Advisor(s): Yu, Nanpeng
Abstract
In response to the demands of real-time decision-making and the challenges posed by insufficient and inaccurate physical model information, recent years have seen a significant surge of data-driven approaches aimed at addressing online control, optimization and monitoring problems within power systems. This thesis first summarizes a wide array of use cases in power distribution systems, including network reconfiguration and restoration, crew dispatch, Volt-Var control, demand response, and optimal power flow. The literature review reveals the versatility and potential of data-driven approaches to improve active distribution system operations. The existing data-driven algorithms are categorized into four main groups: mathematical optimization, end-to-end learning, learning-assisted optimization, and physics-informed learning. This categorization provides a structured overview of the current state of research in this field. Additionally, this thesis digs into enhanced algorithms such as non-centralized methods, robust and stochastic methods, and online learning, which represent significant advancements in addressing the unique challenges of active distribution systems. The discussion extends to the critical role of datasets and test systems in fostering an open and collaborative research environment, essential for the validation and benchmarking of novel data-driven solutions. In conclusion, this thesis outlines the primary challenges that must be navigated to bridge the gap between theoretical research and practical implementation, alongside the opportunities that lie ahead. These insights aim to pave the way for the development of more resilient, efficient, and adaptive active distribution networks, leveraging the full spectrum of data-driven algorithmic innovations.
By summarizing the existing works on data-driven modeling, monitoring, and control in active distribution networks, this thesis shows that physics-informed methods often have superior performance since more physical models and information can been fully utilized. Based on this observation, this thesis further develops physics-informed learning methods for learning power system dynamics and generating synthetic net load data..
First, this thesis proposes a Nearly-Hamiltonian neural network to predict transient trajectories and dynamic parameters of the power system by embedding energy conservation laws in the proposed neural network architecture. This inductive bias empowers the proposed model to learn the power system dynamics without explicitly using the exact functional form of the power system dynamic equations. The numerical study results on the single machine infinite bus system show that the proposed model produces accurate system trajectories and damping coefficient predictions. Furthermore, the proposed model significantly outperforms the baseline and Hamiltonian neural network.
Second, this thesis develops a novel neural ordinary differential equation (ODE) based algorithm to predict the transient trajectories of power systems. To handle noisy measurements, a noise-removal module is proposed, which is implemented before the neural ODE module. The proposed algorithm is validated using the IEEE 118-bus system. The numerical study results demonstrated the superior accuracy of the proposed model over the baseline neural network (NN) and its robustness against measurement noise. Furthermore, the analytics results verified the generalization performance across different fault durations and locations.
Third, this thesis presents a physics-informed diffusion model for generating synthetic net load data, addressing the challenges of data scarcity and privacy concerns. The proposed model embeds physical models within denoising networks, offering a versatile approach that can be readily generalized to unforeseen scenarios. Utilizing the real-world smart meter data from Pecan Street, this thesis validates the proposed method and conducts a thorough numerical study comparing its performance with state-of-the-art generative models, including generative adversarial networks, variational autoencoders, normalizing flows, and a well calibrated baseline diffusion model. A comprehensive set of evaluation metrics are used to assess the accuracy and diversity of the generated synthetic net load data. The numerical study results demonstrate that the proposed physics-informed diffusion model outperforms existing models across all quantitative metrics, showcasing at least a 20% improvement.
In conclusion, this thesis first provides a comprehensive review of the burgeoning field of data-driven algorithms and their application in solving increasingly complex decision-making, optimization, and control problems within active distribution networks. Based on the observation of the new trend of physics-informed learning to improve the performance of data-driven model, this thesis develops a Nearly-Hamiltonian Neural Network and Neural Ordinary Differential Equations to learn power system dynamics. Finally, a physics-informed diffusion model is proposed to generate synthetic net load data. The numerical study results demonstrate the superior performance of our proposed methods.