Learning Optimal Crew Dispatch for Grid Restoration Following an Earthquake

The evolution of power system resilience research has been driven by increasing natural disaster impacts and the need for rapid recovery. Early efforts focused on preventive measures like seismic retrofitting and infrastructure hardening, guided by standards such as IEEE 693 and IEC 61850. Subsequently, adaptive control strategies, including network reconfiguration and automated load shedding, emerged to mitigate damage effects. Optimization techniques like MILP and heuristics (genetic algorithms, simulated annealing) have been used for post-disaster scheduling, but their computational demands limit real-time applicability. Recent advances in AI, especially deep learning, have shown promise in modeling complex system behaviors, yet their integration into disaster response remains nascent. This paper builds upon these developments, introducing a scalable, AI-based framework that leverages Transformer and DRL to address the core challenge of rapid, high-quality dispatch in large-scale, uncertain environments.

Post-earthquake restoration involves complex, multi-objective decision-making under high uncertainty. Traditional methods like MILP, while optimal, are computationally intensive and unsuitable for real-time applications. The dynamic nature of damage evolution, resource constraints, and communication disruptions further complicate the problem. Existing heuristic approaches offer speed but sacrifice solution optimality. The core challenge is to develop a decision-making system that can quickly generate reliable dispatch strategies, adapt to evolving damage states, and operate under severe computational constraints, thereby enabling effective emergency response and minimizing societal impacts.

The key innovations include: 1) employing Transformer models to encode high-dimensional, temporally dependent system states, capturing long-range interactions that traditional models overlook; 2) integrating these with deep reinforcement learning to learn adaptive dispatch policies from simulated damage scenarios, enabling rapid inference; 3) developing a damage scenario generation pipeline based on seismic standards and Monte Carlo methods, reducing computational complexity while maintaining scenario diversity. These innovations collectively enable the system to produce near-instantaneous, high-quality dispatch plans, overcoming the limitations of classical optimization methods and setting a new standard for disaster response in power systems.

• �� Damage modeling: Utilize seismic standards and GMPEs to estimate ground motion intensities (PGA), then apply fragility curves to compute failure probabilities for network components.
• �� Scenario generation: Conduct Monte Carlo simulations to produce diverse damage scenarios, each representing different failure patterns.
• �� Scenario aggregation: Use dimensionality reduction techniques to synthesize damage patterns into a geospatial impact map, simplifying complex data.
• �� System impact assessment: Perform power flow calculations for each scenario, evaluating system loss via combined failed component count and ENS, weighted appropriately.
• �� Scenario selection: Identify representative scenarios based on return periods, ensuring coverage of typical and extreme events.
• �� Model training: Develop a Transformer-based policy network, trained offline with simulated data using reinforcement learning algorithms such as policy gradient methods.
• �� Online deployment: Use the trained model to generate dispatch strategies in real time, guiding repair crew routing and resource allocation.

The experimental setup involves applying the proposed framework to a 2869-node European gas and power network, simulating damage scenarios across three seismic intensities (low, medium, high). The model's hyperparameters, including Transformer depth, attention heads, and reinforcement learning parameters, are tuned via grid search. Baseline comparisons include traditional MILP and heuristic algorithms like genetic algorithms. Metrics evaluated encompass restoration time, dispatch solution accuracy, and computational time. Cross-scenario robustness tests assess model generalization, while ablation studies isolate the contributions of Transformer and DRL components. Results consistently show that the proposed method reduces average recovery time by 40%, maintains solution errors below 2%, and achieves inference within seconds, demonstrating its practical viability.

The framework achieved a 40% reduction in average restoration time compared to MILP-based approaches, with solutions generated in less than 5 seconds—an improvement of over 3000 times in response speed. The dispatch strategies closely matched optimal solutions, with less than 2% deviation in total cost or time. Across different damage scenarios, the model maintained high robustness, with performance degradation under extreme damage conditions remaining below 5%. The results confirmed the model’s ability to adapt to diverse damage patterns and resource constraints, outperforming heuristic methods in both speed and solution quality. These findings validate the framework’s potential for real-world deployment, especially in time-critical disaster scenarios.

This AI-driven dispatch framework can be integrated into emergency management systems for rapid response to earthquakes and other natural disasters. Utilities can deploy it to generate real-time repair schedules, guiding crews efficiently and reducing downtime. The approach requires pre-established damage models and scenario libraries, making it suitable for regions with detailed seismic data. Additionally, the framework can be extended to multi-resource coordination, supporting broader infrastructure resilience efforts. Long-term, it can serve as the backbone of autonomous disaster response platforms, enabling fully automated, scalable, and adaptive recovery operations that minimize societal and economic impacts of natural calamities.

The model’s reliance on accurate seismic damage modeling poses a challenge, as uncertainties in earthquake parameters can affect decision quality. The training process demands significant computational resources, limiting rapid updates or adaptation to new data. Its current scope focuses primarily on distribution network dispatch, with limited consideration for multi-infrastructure coordination or multi-hazard scenarios. Real-world deployment also requires robust communication infrastructure, which may be compromised during disasters. Future work should address these limitations by incorporating uncertainty quantification, distributed computing, and multi-system integration to enhance robustness and applicability.