Enhancing Smart Grid Resilience Using AI Models

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Jun 14 2024

In an article published in the journal Nature, researchers presented a graph reinforcement learning (GRL) model for managing power outages in smart grids, enhancing resilience by optimizing control policies for power restoration.

The model, utilizing a capsule-based graph neural network, accounted for network topology variations and interdependencies. Validated on modified Institute of Electrical and Electronics Engineers (IEEE) test networks, it showed near-optimal, real-time performance, significantly reducing energy loss and demonstrating generalizability across diverse outage scenarios.

Background

Resilience enhancement in power distribution networks (DN) has become crucial due to the increasing frequency of extreme events causing power disruptions. Traditionally, DNs were seen merely as links between transmission networks and consumers. However, with the rise of distributed energy resources (DERs) and smart grids, DNs are now considered capable of autonomous operation.

Self-healing capabilities in smart grids, involving fault location, isolation, and service restoration (FLISR), are essential for minimizing disruptions. Previous work has explored reconfiguration and load shedding for DN restoration, employing methods such as heuristic, meta-heuristic, and mixed-integer programming. However, these approaches face scalability issues, and computational inefficiency, and are not well-suited for real-time decision-making in complex, unbalanced networks.

This paper addressed these gaps by presenting a GRL model using a capsule-based graph neural network. The model optimized real-time control policies for power restoration, accounting for network topology variations and interdependencies, and demonstrates superior performance in test networks.

Graph-Based Scenarios and Optimization for Outage Management

The researchers presented a GRL model for managing power outages in DNs. The model's training scenarios were generated by simulating outages through randomized edge removal in the DN's graph representation. This approach captured localized failures and cascading effects, gradually increasing the fraction of edge failures from zero to 50%.

The outage management problem was formulated as an optimal power flow problem, leveraging mixed-integer programming. The model included variables for switching and load shedding, while control variables managed power flow. The DN's power consumption was balanced by generation constraints, ensuring stability. Voltage and current constraints were applied to all buses, with specific conditions for slack buses and outage lines. The model employed a three-phase branch flow formulation with conic relaxations for enhanced accuracy. Kirchhoff’s voltage law and second-order conic inequality constraints were incorporated to ensure realistic power flow.

The training was conducted over 36 hours with up to two million steps. The GRL model, using a capsule-based neural network, showed effective training completion and performance compared to a multi-layer perceptron (MLP). The model’s simulations were run on a system with an Intel Core i7 processor, using Python, OpenDSSDirect application programming interface (API), Networkx, and Gurobi optimizer.

Key Findings and Performance Evaluation

The researchers evaluated a graph convolutional actor-critic policy (GCAPS) model for outage management in power DNs, comparing it with traditional methods and an MLP-based approach. Key findings included superior convergence and performance of GCAPS across various network sizes.

Case studies demonstrated GCAPS' effectiveness in maintaining voltage stability and optimizing energy supply during outages in 13, 34, and 123-bus networks. Comparison with baselines highlighted GCAPS' near-optimal decision-making and real-time performance advantages.

The architecture leveraged graph-based representations and RL for scalable and efficient emergency response, validated through simulations and comparative experiments. Overall, GCAPS showed promise for enhancing the resilience of distribution networks in real-world scenarios, offering a robust framework for managing outages and ensuring reliable power supply.

Enhancing DN Resilience

The model integrated grid-forming and feeding DER modes, enabling both grid-connected and islanding reconfiguration for outage management. Load shedding ensured operational feasibility and voltage stability. Utilizing on-policy RL and GCAPS neural networks, the model effectively integrated nodal properties and topology information.

Evaluated on IEEE test networks, it outperformed traditional methods, demonstrating near-optimal decision-making and constraint adherence. Its real-time response and scalability made it suitable for large DNs. However, training on larger networks requires advanced computational resources. Future work should explore the model's applicability to heterogeneous networks and address communication breakdowns' impact on outage resolution. This extension could involve modeling interconnected power and communication networks as multi-layered graphs.

Conclusion

In conclusion, the GRL model utilizing a capsule-based graph neural network presented a promising solution for managing power outages in distribution networks. Validated on IEEE test networks, it demonstrated superior performance in real-time outage management, offering near-optimal decision-making and scalability.

The integration of grid-forming DER modes and load shedding enhanced network resilience, ensuring reliable power supply during emergencies. However, addressing computational challenges associated with training on larger networks is crucial for practical implementation. Further research is needed to address computational challenges and explore the model's applicability to heterogeneous networks and communication breakdown scenarios.