Enhancing the Robustness and Reliability of Smart Grids Using Machine Learning Approaches
Open AccessA smart grid is a complex system using power transmission and distribution networks to connect electric power generators to consumers across a large geographical area. Due to their heavy dependencies on information and communication technologies, smart grid infrastructure and applications, such as transmission lines and state estimation, respectively, are vulnerable to physical and cyber attacks and natural disasters.In this dissertation proposal, we target to effectively enhance the robustness and reliability of smart grids based on machine learning techniques including clustering, reinforcement learning, and neural networks. To be specific, we study this problem from the following three aspects. First, in smart grid systems, vulnerable lines may lead to cascading failures which can cause large-scale blackouts. Successfully detecting vulnerable lines can increase the stability of the smart grid systems and reduce the risk of cascading failures. By modeling a smart grid system into a directed graph, we investigate the problem of vulnerable line identification from a clustering perspective. By jointly considering the topological parameters and the electrical properties, we propose an affinity propagation based bus clustering algorithm to classify buses into clusters, where the center of each cluster represents the most influential bus in each partition. According to the clustering results, we design a vulnerable line identification scheme, which captures different types of potential critical lines in the smart grid system. Experiments over the IEEE-39 bus system demonstrate the effectiveness and correctness of our proposed algorithm. Second, operating IoT systems reliably and efficiently demands automatic and intelligent management and control schemes. However, existing IoT control systems are usually designed to operate in a limited predefined scenario (i.e., either the conceived “worst” case scenario or a group of frequent operation scenarios) and cannot be easily expanded to new applications. Such designs can hardly adapt to the smart gird systems with highly dynamic system variables, which include terrible weather conditions, electrical state estimation errors, and faulty human operations. Therefore, I propose to use an environment-adaptive emergency control scheme with reinforcement learning (RL). The proposed RL approach actively regulates the output power of the generators in real-time and thus prevents cascading failure and large-scale blackout. Our approach adopts the Deep Deterministic Policy Gradient (DDPG) to interactively train the smart grid system wherein it develops the optimal actions for assigning a proper output power value and accordingly guarantees the current output of transmission lines falling in a specific range. I will validate the proposed RL-based control through simulations of 50,000 contingency cases over the IEEE 118-bus test system. The experimental results thus far show that our approach can efficiently prevent cascading failure and blackout in non-predefined scenarios along with random system variables. Third, false data injection (FDI) attack is a classical attacking scenario in smart grids and leads to malicious anomalies. Specifically, FDI attacks tamper steaming data generated by grid sensors. Thus, the state variables (i.e., bus voltage angles and magnitudes represent two examples) will become anomalous and can lead to further system errors. To this end, my work presents a novel Wide & Recurrent Neural Networks (RNN) learning model to detect well-constructed FDI attacks that are not detectable by the existing smart grid bad data detection mechanism. In particular, the proposed Wide & RNN model combines the merits in its memory of the global knowledge of the system state information while capturing the temporal correlations among state variable measurement data. Extensive simulations based on real-world data show that the Wide & RNN model considerably increased FDI attack detection accuracy in comparison to Wide-only, RNN-only models and other existing works.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.