This work focuses on cloud-to-ground (CG) lightning, the type of atmospheric discharge most prone to affect power distribution systems in the Brazilian South-Southeastern region. This region was selected to house important industrial activities and have a Lightning Detection and Location Network (LDLN) with expressive coverage. The LDLN detects lightning events in real-time by using a Time of Arrival (TOA) technique, inferring the characteristics of the event, such as its location and peak current.
The developed CG lightning forecast model leverages Machine Learning (ML) to predict the incidence of events. The region of interest is divided into a grid, and predictions are made for each grid point in five-minute increments, extending up to one hour into the future. Unlike traditional binary forecasts that only indicate the occurrence of lightning, our model predicts the number of lightning events per grid point. This quantitative approach enables power companies to assess the severity of potential lightning activity, allowing for more informed decision-making and proactive measures to safeguard assets. After experimentation with several algorithms, the selected method was a gradient-boosting regressor. The model was trained solely with past lightning occurrences from 2018 until 2023, thus requiring only one data source to generate predictions. This, in turn, facilitates the emission of forecasts in short time steps.
The foremost model's performance is verified for a lead time of five min., with a Median Absolute Error (MAE) of 0.0038 lightning per element, and the lowest performance is associated with a lead time of one hour, with an MAE of 0.0064. The model proved capable of predicting the location and quantity of CG lightning for up to one hour in the future. The main drawback of the model is its inability to predict the first lightning occurrence in a thunderstorm, known as lightning initiation, for its predictions are made based on past lightning events.
The second product, the correlation tool, is designed to evaluate whether specific power outages were caused by lightning. This tool considers two main factors: the spatial distance between the lightning event and the affected power asset and the temporal proximity between the lightning event and the outage. By analyzing these factors, the correlation tool determines the likelihood that a given lightning event was responsible for an outage.
To develop this tool, we extensively analyzed historical power outage data from 2020 to 2024. We correlated these outage events with recorded lightning strikes, identifying the critical spatiotemporal parameters influencing the likelihood of a lightning-induced outage. This historical dataset allowed us to develop a robust probability function that can assign a likelihood score to each lightning event, indicating the probability that it caused an outage.
The correlation tool's primary utility is its ability to provide post-event analysis. By accurately attributing outages to specific lightning events, power companies can better understand the vulnerabilities in their infrastructure and refine their mitigation strategies.
This understanding can lead to targeted improvements in lightning protection systems, more effective resource allocation for repairs, and enhanced operational planning to prevent future outages.
A significant finding of our analysis of power outages associated with lightning discharges is that the discharge's peak current is not the main factor in causing an outage. We verified that the power distribution system assets are more vulnerable to large numbers of lightning events in a short time span (i.e., tens of minutes) than isolated lightning with an intense peak current.
Together, these tools provide a comprehensive solution for anticipating, mitigating, and managing lightning's impacts on power distribution systems. The forecasting tool allows for early warnings and preparedness, while the correlation tool offers post-event analysis to determine causality and refine mitigation strategies. By integrating these capabilities, power distribution companies can enhance their operational resilience, reduce downtime, and minimize the economic impact of lightning-induced outages.
This project demonstrates the potential of machine learning and data analytics to transform the management of natural hazards in power distribution systems. Our findings highlight the importance of precise forecasting and accurate event correlation in mitigating lightning effects, ultimately contributing to more reliable and efficient power delivery.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner