Online assessment and control for voltage stability with machine learning-based approaches

Abstract

Voltage stability has been attracting significant attention in the research community since many large blackouts around the world since the 1960's. To prevent catastrophic blackouts attributed to voltage instability, theories and techniques for voltage stability assessment and control were proposed. However, with the increasing penetration of renewables and dynamic loads, the power system model is more complex, posing huge challenges to traditional model-based approaches for voltage stability due to their high dependency on models, high computational burden and low online efficiency. Therefore, this thesis focuses on data-driven approaches that can overcome these shortcomings by employing machine learning techniques. We make contributions in four aspects.

Firstly, a novel data-driven learning and control method is proposed for long-term voltage stability (LTVS). With the knowledge accumulated off-line and feature extractions, we establish the relationship directly between the system dynamics and optimal control actions through principal component analysis and artificial neural networks, which helps to quickly find an optimal control action to restore bus voltages online with a good system performance.

Secondly, a real-time continuous monitoring system with sliding 3D convolutional neural network (3D-CNN) is developed for LTVS assessment. The dynamic responses and topology information of a power system are strategically fused and visualized with sequential state images. The bus indices are reordered to strengthen the topology information. With the localized weight-shared convolution operations, 3D-CNN can be conducted more efficiently, and extract features independently of spatial locations. The translation invariance characteristic makes it highly generic to unknown scenarios, which is further enhanced after bus indices reordering. The sliding window enables 3D-CNN to handle varying degrees of disturbance and performs continuous online monitoring.

Thirdly, short-term voltage stability (STVS) is explored, and a data-driven approach based on spatial-temporal graph convolutional networks (ST-GCN) is proposed for STVS assessment. Instead of converting system information during dynamic processes into regular images, ST-GCN employs graph convolutions on the power network directly to preserve the spatial correlations between buses, which is more natural and reasonable. By generalizing the convolutional operations to irregular graphs, it is also weight-shared and translation invariant. Besides, several traditional convolutional layers on the time dimension are applied to capture temporal correlations.

Finally, a distributed and transferable model for STVS assessment is developed by learning the relationship between system dynamics during faults and the corresponding transient voltage security index. The model is named gated recurrent graph attention network (GRGAT), where the attention operations of bus information are performed directly on the system topology and the system dynamics are captured with gated recurrent units. Particularly, as the attention operations are independent between buses, GRGAT is distributed during online applications and can adapt to the change of topological structures including connection status and the number of neighbor buses, which makes transfer learning possible. Then, adversarial adaptation is proposed to transfer learned knowledge for another modified network.

The effectiveness and potential of these data-driven approaches are verified in various case studies. In summary, this thesis provides new insights into the applications of machine learning techniques in power systems.