Probabilistic forecasting of weather-driven faults in electricity networks: a flexible approach for extreme and non-extreme events

Electricity networks are vulnerable to weather damage, with severe events often leading to faults and power outages. Timely forecasts of fault occurrences, ranging from nowcasts to several days ahead, can enhance preparedness, support faster response, and reduce outage durations. To be operationally useful, such forecasts must quantify uncertainty, enabling risk-informed resource allocation. We present a novel probabilistic framework for forecasting fault counts that captures typical and extreme events. Non-extreme faults are modeled linearly interpolating estimates from multiple additive quantile regressions, while extreme events are described through a discrete generalized Pareto distribution. To incorporate the impact of weather fluctuations, we use ensemble numerical weather predictions, which helps to quantify uncertainty in the forecasts. This approach is designed to provide reliable fault predictions up to four days ahead. We evaluate the model through numerical experiments and apply it to historical fault data from two electricity distribution networks in Great Britain. The resulting forecasts demonstrate substantial improvements over business-as-usual and alternative modeling approaches. A practitioner trial conducted with Scottish Power Energy Networks from October 2024 to March 2025 further demonstrates the operational value of the forecasts. Engineers found them sufficiently reliable to inform decision-making, offering benefits to both network operators and electricity consumers.

💡 Research Summary

This paper introduces a novel probabilistic forecasting framework, named X‑flexForecast, designed to predict daily fault counts in electricity distribution networks under varying weather conditions. The authors address a critical gap in the literature: most existing short‑term fault prediction tools either provide deterministic point forecasts, focus on a single day horizon, or lack a rigorous treatment of uncertainty, especially in the tails of the distribution where extreme events occur.

The methodology separates the problem into two regimes. For non‑extreme fault counts, a set of additive quantile regression (AQR) models is fitted at several probability levels (e.g., 0.1, 0.5, 0.9). Each quantile model combines linear terms with smooth spline functions of weather covariates (wind speed, precipitation, lightning, temperature, etc.). Linear interpolation between the estimated quantiles reconstructs the full conditional cumulative distribution function (CDF), allowing the generation of coherent predictive intervals and full predictive distributions for any lead time up to four days.

For extreme fault counts—defined as days when the observed number of faults exceeds a high integer threshold u—the authors employ a discrete Generalized Pareto (DGP) distribution to model the exceedances Y−u. The DGP’s scale (σ) and shape (ξ) parameters are themselves modeled as smooth functions of the same weather covariates within a Generalized Additive Model for Location, Scale, and Shape (GAMLSS) framework. This mixture approach ensures that the heavy‑tailed behavior of the fault count distribution is captured without imposing unrealistic continuity assumptions on inherently discrete data.

Weather forecast uncertainty is incorporated through an ensemble of Numerical Weather Prediction (NWP) outputs (e.g., ECMWF ensemble). Each ensemble member is fed into the AQR‑DGP pipeline, producing a member‑specific predictive distribution. Bayesian Model Averaging (BMA) then combines these member distributions, assigning weights based on historical predictive skill. This step naturally propagates meteorological uncertainty into the final fault count forecasts.

Model performance is evaluated in two stages. First, extensive simulation studies using synthetic data generated from Poisson, Negative Binomial, and over‑dispersed processes demonstrate that X‑flexForecast consistently outperforms standard Poisson regression, single‑quantile models, and naïve tail‑only approaches across proper scoring rules such as CRPS, Brier score, and log‑likelihood. Second, the framework is applied to real‑world data from two British distribution networks spanning ten years and eleven districts, as well as a pilot deployment with Scottish Power Energy Networks (SPEN) from October 2024 to March 2025. In the real‑data experiments, the method yields substantially lower CRPS and Brier scores for both “amber” (moderate) and “red” (extreme) fault bands. Notably, the proposed model avoids the over‑prediction of red‑event probabilities that plagues a naïve baseline, delivering calibrated probabilities that align with observed frequencies.

Operationally, engineers at SPEN used the probabilistic forecasts to adjust crew scheduling, pre‑position maintenance resources, and prioritize inspection routes. Qualitative feedback reported reductions in average outage duration and emergency response costs, confirming the practical value of the uncertainty‑aware predictions.

The paper acknowledges several limitations. The choice of the exceedance threshold u is data‑driven and may need adaptation for different networks or climate regimes. Extremely rare “super‑extreme” events (e.g., hurricane‑scale storms) remain challenging due to limited training examples. Spatial dependence between neighboring districts is modeled only implicitly through separate district‑level fits, potentially missing cross‑district contagion effects.

Future research directions include extending the framework to a hierarchical Bayesian model that jointly captures spatial correlation, implementing sequential Bayesian updating to incorporate real‑time NWP revisions, developing adaptive threshold selection mechanisms, and exploring transfer‑learning or synthetic‑data augmentation to bolster tail estimation for ultra‑rare events. The authors argue that, with these extensions, X‑flexForecast could become a generic tool for weather‑driven risk forecasting across other critical infrastructures such as gas, water, and transport networks.