[Paper Review] Just-In-Time Learning for Operational Risk Assessment in Power Grids

Oliver Stover, Pranav Karve+5 more|arXiv (Cornell University)|Sep 26, 2022

Electric Power System Optimization4 citations

TL;DR

This paper proposes JITRALF, a just-in-time machine learning framework that trains hourly risk surrogates to accelerate real-time operational risk assessment in power grids with high renewable penetration. By using asymmetric loss functions to prioritize accurate prediction in unsafe operating regions, JITRALF reduces computational burden while maintaining high accuracy in estimating risks of insufficient reserves, load shedding, and operating costs—enabling cost-effective, timely risk mitigation decisions.

ABSTRACT

In a grid with a significant share of renewable generation, operators will need additional tools to evaluate the operational risk due to the increased volatility in load and generation. The computational requirements of the forward uncertainty propagation problem, which must solve numerous security-constrained economic dispatch (SCED) optimizations, is a major barrier for such real-time risk assessment. This paper proposes a Just-In-Time Risk Assessment Learning Framework (JITRALF) as an alternative. JITRALF trains risk surrogates, one for each hour in the day, using Machine Learning (ML) to predict the quantities needed to estimate risk, without explicitly solving the SCED problem. This significantly reduces the computational burden of the forward uncertainty propagation and allows for fast, real-time risk estimation. The paper also proposes a novel, asymmetric loss function and shows that models trained using the asymmetric loss perform better than those using symmetric loss functions. JITRALF is evaluated on the French transmission system for assessing the risk of insufficient operating reserves, the risk of load shedding, and the expected operating cost.

Motivation & Objective

• Address the computational infeasibility of real-time Monte Carlo-based risk assessment in power systems with high renewable variability.
• Overcome the challenge of changing grid topology and generator commitments across hours in day-ahead and real-time operations.
• Improve prediction accuracy in high-risk operating regions where underestimation poses greater system reliability risks.
• Enable grid operators to delay and minimize risk-mitigating interventions using up-to-date, accurate risk estimates.
• Develop a scalable, practical framework for real-time risk assessment using probabilistic forecasts and digital twin simulations.

Proposed method

• Train dedicated machine learning surrogates for each hour of the day to account for changing generator commitments and system topology.
• Use just-in-time learning: train models after the day-ahead security-constrained economic dispatch (DA-FRAC) using simulation data from the DA risk assessment.
• Employ a novel hazard-aware, asymmetric loss function that penalizes underestimation of risk more heavily than overestimation, especially near or beyond operating limits.
• Train models to predict key quantities of interest (QoIs) such as risk of insufficient reserves, load shedding, and expected operating cost without solving SCED problems.
• Integrate probabilistic forecasts of wind, solar, and load as inputs, and use digital twin simulations to generate high-fidelity training data.
• Validate model performance using real-time risk estimation on the French transmission system under high-, medium-, and low-risk scenarios.

Experimental results

Research questions

• RQ1 Can machine learning surrogates trained just-in-time significantly reduce the computational cost of real-time operational risk assessment in power systems?
• RQ2 How does using an asymmetric loss function improve model performance in critical unsafe operating regions compared to symmetric loss functions?
• RQ3 To what extent can JITRALF accurately predict risk of insufficient reserves, load shedding, and operating costs in real-time using up-to-date forecasts?
• RQ4 Can real-time risk estimates enable grid operators to delay and minimize costly risk-mitigation actions while maintaining system reliability?
• RQ5 How does the hourly training of surrogates account for the dynamic changes in grid topology and generator commitment patterns?

Key findings

• JITRALF achieves high prediction accuracy for risk of insufficient operating reserves, load shedding, and expected operating cost, with mean absolute error (MAE) as low as 0.02×10⁴$ for low-risk scenarios.
• The model trained with the asymmetric hazard-aware loss function outperforms symmetric loss models, especially in high-risk scenarios, reducing error by up to 50% in critical regions.
• For high-risk scenarios, JITRALF correctly identifies the need for large-scale reserve commitment, while for low-risk cases, it accurately predicts minimal required interventions.
• Real-time risk estimation enables operators to delay commitment of additional generators until the risk is credible, potentially reducing operational costs and emissions.
• The framework successfully identifies risky conditions in real time, with load shedding risk predictions showing MAE of 18.1×10⁴$ (HAL: 21.6×10⁴$) in high-risk cases.
• The mean risk values for load shedding are 83 (high), 14.9 (medium), and 0 (low), and JITRALF's predictions align closely with these benchmarks, demonstrating strong fidelity.