WeCARE - Weather-Outage-Modeling

Weather-Outage Modeling

Welcome to the power outage modeling section! In this section, we dive deep into power outage and weather data and link them to find grid reliability matrix to see how reliability changes both temporally and spatially over Florida. We go even deeper, and based on the power outage data, weather data, grid information, and social vulnerability, we build power outage models to predict different types of power outages (no/minor, moderate, and major) for each county in Florida.

Subteam

Prof. Thomas Wahl
Mohammad Siddiqur Rahman
Meghana Nagaraj
Alejandra Rodriguez Enriquez

Introduction to weather-outage modeling

The power grid is essential for functioning communities and societal prosperity. However, the U.S. grid is increasingly fragile, largely due to extreme weather and climate events. Weather-related power outages have already doubled in the last decade compared to the one prior. This trend is expected to worsen as extreme weather becomes more frequent and the majority of the U.S. grid infrastructure surpasses its life expectancy. By analyzing the risk and reliability of the grid, we can better identify vulnerable communities, improve emergency response strategies, and build resilience through the installation of solar photovoltaic systems.

Weather-outage modeling approach

Conventionally, weather-outage modeling focuses on short-term forecasting to predict daily disruptions based on immediate conditions. This study, however, adopts a broader scope, developing a wide-area, long-term model to analyze the evolution of outage risks from 2015 to the present, and plans to reconstruct from 1979, as well as projecting trends for the next 30 to 50 years. To achieve this, we first characterize the fundamental relationship between extreme weather and grid instability. We then utilize Florida as a case study to assess historical risk and reliability, subsequently demonstrating the predictability of power outages through a machine learning model.

Acquire, analyze, and process outage data

Power outage models rely on historical weather information, power outage records, and environmental factors. To ensure model reliability, we require training data with long temporal coverage and high spatial resolution. We retrieve power outage data from various sources, such as governmental agencies, private providers. We assess data quality and select the appropriate dataset for model development. Appropriate preprocessing of power outage and weather data is required to identify power outage risk and the precondition of building a weather-outage model.

Outage risk maps from EAGLE-I

We developed weather-related outage risk maps for Florida using historical weather and outage data. The Environment for the Analysis of Geo-Located Energy Information (EAGLE-I) power outage data with 15-minute temporal resolution is linked to weather data to build the final version of the risk maps. Maps can be derived for various metrics; at the moment, we derive commonly used reliability metrics such as the system average interruption duration index (SAIDI). The framework used to derive the metrics and maps is described in more detail in the following paragraph.

Risk maps framework

In the first step, we want to filter EAGLE-I power outage data to remove minor power outages that were unlikely to be weather-related and could be due to daily routine system operations. Those events are short in duration and only affect a small number of customers. The challenge is to identify an appropriate threshold to remove all those minor events initially. After testing different values, we decided to use the 85th percentile value derived from the raw outage time series and only retain events that exceeded that threshold in terms of customers affected. Next, using a declustering window of 60 minutes (where no outages occur) between events, we identify individual power outage events and calculate duration, maximum number of customers out, and customer hours of interruption (duration multiplied by maximum number of customers). Typically, reliability metrics like SAIDI are derived by using all power outage events. However, our goal is to create risk maps for weather-related power outage events. To identify weather-related outages, we link the outage data to various weather variables that can potentially trigger or contribute to outage events. We found that precipitation and wind speed were the most dominant drivers causing outages in our study region. Hence, we identify weather-related power outages as events when either precipitation or wind speed exceeds their respective 90th percentile values or both when both exceed their 85th percentiles simultaneously. All other outage events are removed (i.e., values set to zero).

Grid reliability

Results indicate significantly lower total outage counts in northwest Florida compared to the rest of the state, with the exception of the Lake Okeechobee region and Monroe County (Fig. 1a). Conversely, northwest Florida experiences a notably higher frequency of weather-related outages. In terms of reliability metrics, non-weather and non-TC SAIDI exhibit similar spatial patterns (fig. 1c, e) while weather-related SAIDI closely aligns with TC SAIDI trends (1d, f).

Fig. 1: (a) Total number of power outages (>85th percentile, >1 hour duration, independent events >1 hour apart from each other) between 2015 and 2022 across 67 Florida counties; (b) percentage of weather-related outages; (c) SAIDI from non-weather events; (d) SAIDI from weather events; (e) SAIDI from non-tropical weather events; (f) SAIDI from tropical weather events.

Weather outge modeling results

The Power outage prediction model was finalized for both binary and multi-class classification scenarios. We see significant improvements in model performance for both binary and multi-class classification results compared to initial model testing. The binary model has a comparatively lower precision score compared to the recall score, mostly in South Florida for 2021. Liberty County in the Northwest has the lowest precision score. This is because there was only one recorded power outage, and the model fails to accurately predict that. The model achieves satisfactory F1 and F2 scores.

Fig. 2. Maps show evaluation metrics of binary model for different counties of Florida. (a) precision, (b) recall, (c) F1, and (d) F2 score.

The multi-class classification model also shows good results in terms of evaluation metrics. A few counties in southwest Florida have comparatively lower values, likely due to limited outage data.

Fig. 3. Maps show evaluation metrics for different counties of Florida. (a) precision, (b) recall, (c) F1, and (d) F2 score.