[Paper Review] Revealing drivers and risks for power grid frequency stability with explainable AI

Johannes Kruse, Benjamin Schäfer+1 more|arXiv (Cornell University)|Jun 7, 2021

Energy Load and Power Forecasting70 references65 citations

TL;DR

This paper proposes an explainable AI (XAI) framework using gradient-boosted trees and SHAP values to predict and interpret frequency stability indicators—RoCoF, Nadir, MSD, and Integral—in three European power grids. It identifies load ramps, generation ramps, and forecast errors as key drivers of instability, demonstrating predictive performance 3.2 to 14.7 times better than a daily average baseline, with SHAP analysis revealing grid-specific risk factors and system-wide dependencies.

ABSTRACT

Stable operation of the electrical power system requires the power grid frequency to stay within strict operational limits. With millions of consumers and thousands of generators connected to a power grid, detailed human-build models can no longer capture the full dynamics of this complex system. Modern machine learning algorithms provide a powerful alternative for system modelling and prediction, but the intrinsic black-box character of many models impedes scientific insights and poses severe security risks. Here, we show how eXplainable AI (XAI) alleviates these problems by revealing critical dependencies and influences on the power grid frequency. We accurately predict frequency stability indicators (such as RoCoF and Nadir) for three major European synchronous areas and identify key features that determine the power grid stability. Load ramps, specific generation ramps but also prices and forecast errors are central to understand and stabilize the power grid.

Motivation & Objective

• To address the challenge of understanding and predicting power grid frequency stability in the face of increasing renewable energy penetration and complex system dynamics.
• To overcome the black-box nature of traditional machine learning models in critical infrastructure by applying explainable AI (XAI) techniques.
• To identify and quantify the most influential external features—such as load and generation ramps, electricity prices, and forecast errors—driving frequency deviations in synchronous European grids.
• To provide actionable insights for transmission system operators by linking model explainability to real-world operational risks and control strategies.
• To complement physics-based simulations with data-driven models that uncover hidden dependencies and emerging risks not captured by standard forecasts.

Proposed method

• Uses gradient-boosted tree models to predict four frequency stability indicators: RoCoF, Nadir, MSD, and Integral, using hourly external features.
• Applies SHapley Additive exPlanations (SHAP) to explain model predictions, enabling local and global interpretability of feature contributions.
• Incorporates both day-ahead forecasted features (e.g., load ramp day-ahead) and ex-post error features (e.g., forecast error generation ramp) as model inputs.
• Aggregates high-resolution frequency data into hourly indicators to align with operational time scales and improve model generalization.
• Validates model performance against a system-specific null model: the daily average profile of stability indicators.
• Generates dependency plots and interaction analyses using SHAP values to visualize feature effects and nonlinear relationships across the three synchronous areas (Continental Europe, Nordic, Great Britain).

Experimental results

Research questions

• RQ1 Which external features most strongly influence frequency stability indicators such as RoCoF, Nadir, MSD, and Integral in European power grids?
• RQ2 How do forecast errors in load and generation ramps affect frequency stability, and do these effects vary across synchronous areas?
• RQ3 To what extent can explainable AI models outperform simple baseline models (e.g., daily average profiles) in predicting frequency stability?
• RQ4 What are the key nonlinear and interactive dependencies between electricity market signals (e.g., prices) and frequency dynamics?
• RQ5 How do regional differences in grid operation and renewable integration shape the dominant drivers of frequency instability?

Key findings

• The XAI model outperforms the daily average baseline by 3.2× (Continental Europe), 6.9× (Nordic), and 14.7× (Great Britain) in predicting frequency stability indicators.
• Load ramps are consistently among the most influential features, with negative SHAP contributions indicating that larger ramps lead to lower (worse) Nadir values.
• Forecast errors in load and generation ramps are critical drivers of instability, with their impact varying significantly across synchronous areas.
• Electricity price ramps and day-ahead price signals also show strong, non-trivial effects on frequency stability, particularly in the Great Britain grid.
• SHAP interaction plots reveal that the combined effect of load ramps and forecast errors is more pronounced than individual contributions, indicating nonlinear system behavior.
• Synchronous generation levels show a complex, non-monotonic relationship with stability indicators, particularly in the Great Britain grid, where high levels correlate with increased instability under certain conditions.