With the continuous expansion of power grid scale and the increasing complexity of operational structures (Yang et al., 2024), the available security margin of the system continues to shrink (Yi et al., 2023), leading to heightened vulnerability of the grid under disaster scenarios (Jufri et al., 2019). Extreme disasters such as earthquakes are characterized by sudden onset and high destructiveness (Yang et al., 2020). They can trigger multiple failures within a short period, including damage to transmission lines, failure of critical substation and control equipment, and abnormal load responses, posing significant threats to the safety and stable functioning of power systems (Nazemi and Dehghanian, 2020). Therefore, conducting research on power grid resilience assessment and enhancement under extreme disaster conditions is of great importance for improving system defense and restoration capacities.

The methods for evaluating power grid resilience in the context of extreme disasters are predominantly divided into two paradigms: post-disaster analysis (Ekisheva et al., 2022) and pre-disaster simulation (Panteli and Mancarella, 2015). The post-disaster analysis method is mainly based on historical data and post-disaster survey data to evaluate the resilience of the power grid in the face of disasters. Through real disaster situations and data, it reflects the performance of the power system in resisting disturbances and recovering from disturbances in actual disasters. For example, Liu et al. (2025) conducted a toughness assessment and risk study on the transmission tower line system under continuous earthquakes. It selected the actual mainshock-aftershock sequence from the international strong earth-quake database, used a series of scaled ground motions to conduct nonlinear dynamic analysis, and comprehensively captured the damage evolution law of the structure from elastic behavior to damage. However, this type of method can only be carried out after a disaster occurs and cannot be used for pre-disaster planning and early warning, and its scope of application is limited. The pre-disaster simulation method simulates the response and recovery process of the power system under disaster conditions by generating multiple fault scenarios before the disaster occurs, and can quickly identify the weaker parts of the power grid before the disaster, providing a basis for pre-disaster risk prevention (Tang et al., 2024). For example, Amani-Jouneghani et al. (2025) uses earthquakes as representative extreme events to construct a probability risk model for power facility damage under earthquake disaster conditions, generate a risk map that reflects the spatial distribution characteristics of infrastructure risks, and establish a multi-dimensional power grid resilience assessment system.

Furthermore, to mitigate the discrepancy between resilience assessment results and actual system performance during disasters, several studies have enhanced the modeling of seismic uncertainties and incorporated multiple earthquake scenarios for more comprehensive analysis. For example, Lagos et al. (2020) takes into account the uncertainty of system response, equipment vulnerability and earthquake disasters, and constructs a power grid resilience assessment framework under earthquake disasters from the perspective of load loss cost; Munikoti et al. (2021) proposed a modeling and evaluation method based on weighted functional graph theory for interconnected systems including electricity, water, and natural gas systems, and simulated and analyzed the resilience performance of the system under different earthquake scenarios.

However, with the continuous expansion of power grid scale, existing resilience assessment methods generally rely on large-scale scenario enumeration and complex physical-model-based computations, resulting in high computational complexity and low solution efficiency (Nazemi et al., 2020; Yang et al., 2025). In contrast, machine learning approaches represented by reinforcement learning (RL) can learn decision-making policies through interaction with the environment, enabling fast policy updates under uncertain disturbances. For example, Chen et al. (2023) applied a hierarchical reduction reinforcement learning method to emergency control of large-scale power systems, significantly reducing computational complexity by efficiently and accurately shrinking the action space. Zhou et al. (2020) employed a deep reinforcement learning (DRL) algorithm to obtain fast solutions for the optimal power flow problem under operational constraints, thereby providing power system operators with efficient and feasible real-time decision support. Moreover, most existing power grid resilience assessment methods focus on offline evaluation or post-disaster analysis of resilience performance under disaster scenarios. By contrast, DRL algorithms, leveraging policy iteration and experience accumulation mechanisms, are capable of dynamically adjusting control strategies. For instance, Huang et al. (2019) modeled the power system operation under disturbances as a sequential decision-making problem using a DRL-based approach, enabling dynamic adjustment and adaptive optimization of emergency control strategies.

In response, this paper proposes a resilience risk minimization strategy for power grids under earthquake disaster scenarios based on DRL. The proposed approach establishes a power grid functional loss and resilience risk assessment framework by constructing a potential seismic source zoning model and a power line failure probability model. By integrating a deep reinforcement learning algorithm, the method enables the development of a dynamic, earthquake-responsive strategy that minimizes the resilience risk of the power grid.

The structure of this paper is organized as follows: First, Section 2 presents the earthquake disaster scenario model. Then, Section 3 derives the resilience risk model for power grids under earthquake disasters. Next, Section 4 details the resilience risk minimization strategy for power grids based on DRL. Section 5 presents case studies to validate the effectiveness of the proposed method. Finally, Section 6 draws the re-search conclusion.

Traditional resilience assessment methods usually rely on static calculations of the power grid function loss expectation E L p s k under various earthquake scenarios to measure system risks. However, this type of method has low solution efficiency when faced with complex disaster evolution and multi-stage regulation, resulting in lagging optimization results and insufficient decision-making adaptability. In order to improve the solution efficiency and self-learning ability, this paper introduces a DRL framework and uses the Deep Q-Network (DQN) method to model the problem of minimizing the power grid resilience risk under earthquake disaster scenarios as a sequential decision-making optimization process. Through interactive learning of agents in disaster environments, DQN can quickly converge to an approximately optimal strategy in multiple earthquake scenarios and achieve adaptive optimization of resource scheduling.

In this method, the environment describes the dynamic evolution of the power grid under earthquake disturbances. The system state s t is defined by a set of observable variables that characterize the operating condition of the grid, including bus voltage magnitudes, line power flows, load supply levels, and the damage states of transmission lines under a given earthquake scenario. Based on the observed state s t , the intelligent agent selects a control action a t from a predefined action space to mitigate functional losses and resilience risk. Specifically, the action space consists of feasible grid regulation measures, such as load shedding at selected buses, generation re-dispatch, and network reconfiguration through line switching. After executing action a t , the system transitions to the next state s t + 1 according to the grid dynamics and earthquake impact, and receives an immediate reward r t , which is determined by the extent of power supply degradation and the resilience risk associated with the current earthquake scenario. Through continuous interaction with the environment and policy iteration, the agent learns an optimal control strategy that minimizes cumulative resilience risk over the decision horizon.

Based on the idea of DRL, a risk loss function J π is introduced, whose goal is to minimize the weighted expected resilience risk through strategy π a t | s t , as shown in Equation 24: J π = E π ∑ k = 1 N s e P o c c u r k ∑ t = 0 T γ t r t k (24)

Where, r t k represents the functional loss or risk cost at time t under earthquake scenario k. γ ∈ 0 , 1 is the discount factor, used to balance short-term and long-term resilience performance. Strategy π a t | s t represents the probability distribution of selecting action a t in state s t .

In order to achieve risk minimization, the cumulative loss expectation minimization can be transformed into a value function optimization problem in reinforcement learning (Wang et al., 2025). Define the state-action value function as Equation 25: Q π s t , a t = E π ∑ τ = t T γ τ − t r τ k | s t , a t (25)

Its optimal form satisfies the Bellman equation (Kamruzzaman et al., 2021), as shown in Equation 26: Q * s t , a t = E r t k + γ min a t + 1 Q * s t + 1 , a t + 1 (26)

This article uses the DQN algorithm and uses the deep neural network Q θ s , a to approximate the optimal value function Q * s , a . In each round of training, the agent collects samples s t , a t , r t k , s t + 1 from the experience replay pool, and calculates the temporal difference target based on the target network Q θ ¯ , as shown in Equation 27: y t = r t k + γ min a ′ Q θ ¯ s t + 1 , a ′ (27)

Update network parameters by minimizing the weighted mean square error, as shown in Equation 28: L θ = E P occur k Q θ s t , a t − y t 2 (28)

The network adopts a dual network structure to reduce training oscillations, uses an experience replay mechanism to break sample correlation, and adopts an ε greedy strategy to balance exploration and utilization (Tao et al., 2024), as shown in Equation 29: π ε a t ∣ s t = 1 − ε + ε A , a t = a t g , ε A , a t ≠ a t g . (29)

Where, ε represents the exploration probability, ε ∈ 0 , 1 , A represents the action set, and a t g denotes the greedy action in state a t .

After multiple rounds of iterations, the Q network approaches the optimal value function Q * , and the optimal control strategy can be given by the following Equation 30: π * s t = arg min a t Q * s t , a t = arg min a t E r t k + γ min a t + 1 Q * s t + 1 , a t + 1 (30)

The DRL agent adaptively learns the optimal power grid resource scheduling plan in the interaction of multiple earthquake scenarios, thereby minimizing the risk of power grid resilience assessment under earthquake disaster scenarios and improving the overall seismic resilience level of the system.

The IEEE 33-bus standard distribution system is adopted to verify the correctness and effectiveness of the proposed model and method. The spatial relationship between the seismic belt and the power system is illustrated in Figure 3, where the dashed lines represent elliptical isoseismal contours. The power system is located along the short-axis attenuation direction perpendicular to the seismic belt, with the nearest distance of 26.1 km and the farthest distance of 72.3 km from the seismic belt. It is assumed that the seismic belt contains three magnitude intervals: 6.6–7.0, 7.1–7.5, and 7.6–8.0.

To verify the effectiveness of the proposed method, the training parameters of the DRL model are set as follows: the learning rate α = 0.001 and the discount factor γ = 0.9. The number of training iterations is set to 5,000. Figure 4 illustrates the variation of the agent’s reward during the training process. As the training progresses, the agent continuously interacts with the environment, accumulates experience, and gradually learns an improved decision-making policy, resulting in higher cumulative rewards.

Taking the fault scenario induced by an earthquake within the potential seismic source zone as an example, the power system is located along the short-axis attenuation direction of the seismic belt, where the PGA attenuates most rapidly. Under this earthquake event, the PGA at the point closest to the epicenter within the power system has decreased to only 0.28 g. The seismic disaster causes line disconnections in the test system at lines 5, 8, 12, 13, 16, 18, and 30. To further evaluate the power supply capability and resilience of the proposed method under seismic disaster scenarios, a non-learning-based power grid resilience assessment and dispatch strategy is adopted as a benchmark for comparison (Lagos et al., 2020). The load demand and supply conditions for the benchmark method and the proposed method are shown in Figure 5.

As shown in the Figure 5, when multiple transmission lines are disconnected due to the earthquake, the overall power supply capability of the system decreases significantly. The traditional method exhibits a slower dispatch response to damaged lines and critical nodes during the early stage of the earthquake, resulting in a pronounced supply–demand gap during peak load periods and delayed recovery of power delivery in certain intervals. In contrast, the proposed method can rapidly identify critical power supply paths and allocate available resources in a more rational manner. As a result, the power supply capability is significantly enhanced, particularly during peak load periods, with an improvement of approximately 12% compared with the benchmark method. Based on this, a comparative analysis of the power grid functional performance under the two methods is presented in Figure 6.

As shown in the Figure 6, during the peak electricity demand period from 7:00 to 22:00, the proposed method is able to maintain a higher level of power supply functionality under the same conditions. Beyond the peak period, the overall system functionality under the proposed method also shows significant improvement compared with the benchmark method. In addition, the average decision-making time of the proposed method and the benchmark method was analyzed from the perspective of computational efficiency. The detailed comparison results are presented in Table 1.

As shown in Table 1, based on the proposed method for minimizing the power grid resilience risk index, the overall functional loss of the system is reduced by 18.1% compared with the traditional method. This demonstrates that the proposed method can guide more reasonable power allocation under earthquake disaster conditions, thereby effectively ensuring the functional integrity of the power system. Furthermore, the average decision-making time of the benchmark method is approximately 1.5 s, whereas the proposed method requires 0.18 s per decision, resulting in an improvement of 88.7% in decision-making efficiency for power grid resilience assessment and risk-minimizing dispatch under earthquake disaster conditions.

The resilience level not only captures the functional losses caused by line faults, but also accounts for the impacts of load supply–demand imbalances and power shortages. Figure 7 illustrates the variations of the power grid resilience level and the grid function expectation over time under the earthquake scenario.

During the evaluation of power grid resilience, it is observed that faults occurring on some transmission lines do not necessarily lead to a significant degradation of system functionality, and the impacts of line outages on the overall functional level of the power grid vary considerably. To further investigate the risk characteristics of individual transmission lines under earthquake conditions and their influence on the overall grid functionality, this study statistically analyzes the transmission lines with relatively high fault probabilities, as summarized in Table 2, and comparatively evaluates the impacts of their outages on system functional performance, as illustrated in Figure 8.

Based on the above case study results, when an earthquake occurs within the potential seismic source zone, the numbers of line disconnection events for Lines 2, 18, 3, 4, and 23 reach 6,183, 7,138, 6,136, 6,304, and 6,325, respectively. These five lines exhibit a strong correlation with the overall system functionality, and their outages result in pronounced functional losses of the power grid. Therefore, they should be identified as critical components for seismic disaster mitigation, where targeted pre-disaster reinforcement and strengthening measures are necessary to effectively reduce earthquake-induced line failure risks and enhance the power supply reliability and overall resilience of the grid under extreme events.

To address the power grid resilience risk problem under earthquakes and other extreme disasters, this paper proposes a resilience risk minimization strategy for power grids under earthquake disaster scenarios based on DRL. The main research conclusions are summarized as follows:

A multi-source earthquake fault scenario generation model is developed. By jointly considering the spatial heterogeneity of seismic belts and the attenuation characteristics of ground motion with distance, the proposed model more realistically captures the randomness and regional correlation of power grid component failures under seismic disasters. This provides high-credibility disaster scenario inputs for resilience assessment and dispatch optimization, effectively overcoming the overly idealized assumptions commonly adopted in traditional scenario construction methods.