AI can strengthen power grid resilience during disasters

AI’s appeal lies in its ability to handle high-impact, low-probability events. Extreme weather disasters fall squarely into this category, making them difficult to predict using historical averages or static models. AI systems, particularly those based on machine learning and deep learning, can detect subtle signals, integrate diverse data sources, and update predictions as conditions change.

Floods, heatwaves, hurricanes, wildfires, and earthquakes are increasingly disrupting electricity supply, leaving millions without power and inflicting heavy economic losses. New research published in the journal Energies now points to artificial intelligence as a decisive factor in whether modern power systems can withstand this new era of climate-driven risk.

Titled “Applications of AI for the Optimal Operations of Power Systems Under Extreme Weather Events: A Task-Driven and Methodological Review,” the review assesses how AI is being used to protect power systems before, during, and after natural disasters, while also warning of critical limitations that could undermine its effectiveness if left unaddressed.

Extreme weather is redefining power system risk

The study outlines the changing nature of risk facing power systems. Natural disasters are no longer isolated shocks but persistent, cascading threats. Floods can disable substations in low-lying areas, storms can topple transmission lines across wide regions, wildfires can destroy distribution networks, and heatwaves can simultaneously push demand to record levels while degrading generation capacity. These events often occur in combination, compounding damage and complicating recovery.

Traditional power system resilience strategies have relied heavily on physics-based and optimization-driven models. While these approaches are transparent and mathematically rigorous, the study finds they suffer from key limitations. They require precise modeling of system components and disaster scenarios, are computationally expensive, and often fail to generalize beyond the specific conditions for which they were designed. As grids grow more complex through the integration of renewable energy, distributed resources, electric vehicles, and smart devices, these methods become increasingly brittle.

The volume of data available to utilities has also exploded. Sensors, smart meters, satellite imagery, weather forecasts, and operational logs now generate streams of information that far exceed the capacity of traditional analytical tools. This data-rich environment is where artificial intelligence begins to show its value. The study positions AI not as a replacement for engineering knowledge, but as a complementary approach capable of learning patterns directly from data, adapting to uncertainty, and supporting faster decision-making under pressure.

AI’s appeal lies in its ability to handle high-impact, low-probability events. Extreme weather disasters fall squarely into this category, making them difficult to predict using historical averages or static models. AI systems, particularly those based on machine learning and deep learning, can detect subtle signals, integrate diverse data sources, and update predictions as conditions change.

How AI is being used before, during and after disasters

The research classifies AI use according to the operational problems utilities must solve when facing extreme weather. The first category is predictive tasks, which aim to anticipate future events and system states. These include forecasting natural disasters, predicting power generation and load under extreme conditions, identifying equipment likely to fail, and estimating the probability and location of outages. AI models have been shown to outperform traditional statistical approaches in many of these areas by capturing nonlinear relationships between weather variables, infrastructure conditions, and system behavior.

Accurate prediction enables proactive action. Utilities can pre-position crews, reinforce vulnerable assets, adjust generation schedules, and issue early warnings. The study highlights outage prediction and fault forecasting as particularly mature applications, where AI-driven models have achieved high accuracy and timely alerts, helping reduce the duration and severity of blackouts.

The second category covers descriptive tasks, which focus on situational awareness during and immediately after disasters. These tasks include fault detection and outage detection, allowing operators to quickly identify damaged components and affected areas. AI-based detection systems can analyze real-time data streams to locate faults even when parts of the network are inaccessible or communication is degraded.

The study notes that descriptive AI applications are critical for shortening restoration times. By rapidly classifying faults and outages, utilities can prioritize repairs, allocate resources more effectively, and avoid unnecessary delays. This capability becomes especially important when disasters disrupt both power and communication infrastructure.

The third category encompasses prescriptive tasks, which involve recommending or optimizing actions to restore service and maintain stability. These include load restoration strategies, dispatch of distributed energy resources, deployment of mobile energy resources such as battery systems and generators, and dynamic network reconfiguration. In these areas, reinforcement learning and nature-inspired algorithms have gained prominence due to their ability to learn optimal strategies through interaction with complex environments.

Prescriptive AI is where decision-making becomes most sensitive. The study underscores that these models must balance speed, reliability, and safety, as incorrect recommendations during emergencies can worsen outages or create new hazards. Despite promising results in simulations and pilot projects, the authors caution that prescriptive AI systems require rigorous validation before large-scale deployment.

Data, trust, and the limits of AI-driven resilience

While the study highlights significant progress, it is equally clear about the challenges that threaten AI’s real-world impact. Data quality emerges as a persistent obstacle. Power systems rely on heterogeneous data collected at different resolutions, formats, and sampling rates. Extreme events further exacerbate data gaps, as sensors may fail or communications may be disrupted precisely when information is most needed.

Another major concern is generalization. AI models trained on historical data may struggle to handle unprecedented events or rare combinations of hazards. The study warns that data-driven models can only predict what they have seen, raising questions about their reliability under novel conditions. This limitation is especially acute for extreme weather events, which are evolving rapidly due to climate change.

Interpretability also remains a critical issue. Many AI models operate as black boxes, producing accurate predictions without clear explanations. In power system operations, where decisions must be justified to regulators, operators, and the public, this lack of transparency can undermine trust. The study notes that engineers are often reluctant to rely on systems they cannot fully understand, particularly during emergencies.

The growing role of large language models introduces a new layer of complexity. While these models offer powerful data analytics capabilities, they also consume significant energy, creating a feedback loop between AI demand and power system capacity. The authors highlight warnings that rising AI workloads could strain electricity supply, making energy-efficient AI development an urgent priority.

Cybersecurity and data privacy are additional concerns. As grids become more data-driven, they also become more exposed to cyber threats. The study stresses the need for secure data-sharing platforms and privacy-preserving analytics, especially as smart devices proliferate across distribution networks.

To address these challenges, the authors outline several future research directions. These include decentralized sensing and edge computing to maintain situational awareness during outages, multi-modal decision-support systems that integrate weather, infrastructure, and social data, models that account for cascading disasters rather than isolated events, and secure platforms that protect sensitive information while enabling collaboration.