Tag: Machine Learning

Categories
Article Research Products

Predicting The Impacts of Tropical Storms

My recently published paper, “A Data-Driven Decision Support Tool for Anticipating Tropical Storm Impacts to the United States Power Grid,” has the potential to be my most impactful yet. We developed a system that can accurately predict the impacts that tropical storms have on the power grid in the Continental United States. While not the first system of its kind, we demonstrate accurate predictions for a wide range of types and magnitude of storms. It has the potential to be a powerful decision-making support tool that helps managers in government and at utilities prepare more effectively for the worst types of natural hazards. For more details, please see our article: 10.1109/ACCESS.2024.3442768

Categories
Article Patents Research Products

Understanding the Resilience of the Power Grid

Photo by Jerry Wang via Unsplash

We’ve had a flurry of publications in advancing the understanding of the resilience of the power grid to extreme weather events in recent weeks! Hopefully these insights can be used to help improve the grid, so that we can all have reliable power even under climate change. The first paper describes a hybrid Machine Learning approach that integrates data from physical simulations to quantify the different levels of impact we can expect for different storms and infrastructural configurations [link]. This type of model has been used to evaluate proposed physical upgrades to the existing power grid and perform a cost-benefit analysis using real data from Connecticut [link]. This is a new type of planning for these types of projects, and can help make sure that the funding provided by the US Infrastructure Investment and Jobs Act is spent wisely. My co-authors and I firmly believe in the impact and novelty of this technology, and have even applied for a patent [link], as mentioned in a previous post. This work is just the beginning of the development of more sophisticated AI-powered digital twin models of the power grid that help us prepare for the future!

Categories
Patents Research Products

Framework For Evaluating Effectiveness of Adaptive Change

Just finished work on my second patent, one for a machine-learning based framework capable of quantifying the effectiveness of adaptive change in infrastructural systems. This is a tricky problem because as hazards get more severe, investments in infrastructural systems can appear to be less effective than they actually are.

Preliminary Results Applying this Approach to Infrastructure in Connecticut. Figure by William Hughes.

The trick is to build a weather-related impact model that is also trained on various technical aspects of the infrastructure. If the training data is good enough, the model will learn how these aspects influence the reliability of the system, and then you can use that model to test out various configurations of infrastructure.

We’re hoping that this new approach will be useful for power utilities, regulators, and others to optimize how they plan and adapt their systems to climate change.

Categories
Article Research Products

High Impact Storm Outage Prediction

Major storms are always much much worse than more typical events. In terms of power outages, they can be many orders of magnitude more impactful than typical events: sometimes cause about the same number of power outages in a day that a power utility would otherwise see over the course of a year.

Hurricane Isaias Over Florida
Hurricane Isaias Over Florida by NOAA

But extremes are hard to predict with machine learning for two major reasons.

  1. They’re rare events, and therefore you have few examples to learn from.
  2. Their impacts don’t fit the trends of more moderate events well and are much much larger.

But because these storms are so disruptive, they are very important to be able to predict accurately. In the journal Weather and Climate Extremes, we’ve just published a paper Improved Quantitative Prediction of Power Outages caused by Extreme Weather Events, which describes a data intensive approach of empirically predicting the impacts of storms like hurricanes on the power grid.

High Impact Storm Model Accuracy
Out-of-Sample High Impact Storm Model Accuracy with Select Storms Highlighted

Not only do the described methods have superior accuracy, because we consider a range meteorological factors not typically considered in impact analysis we’re able to describe a diversity of factors that can contribute to the most extreme impacts of weather.

Relative Contribution of Factors to Storm Impacts with Select Storms Highlighted
Categories
Article Research Products

Dynamic Prediction of the Effects of Tree Trimming

Tree trimming is a go-to solution for many power utilities to reduce the number of power outages caused by storms. The effects of this are intuitive: because so many outages are caused by trees in storms, cutting back the trees will reduce outages.

A Trimmed Tree
A Tree trimmed around the Power Lines

Quantify or predicting this reduction is much more difficult, and yet, so critical for effective cost-benefit analysis. In Dynamic Modeling of the Effect of Vegetation Management on Weather-Related Power Outages published in Electric Power Systems Research we do just that by building a dynamic machine learning power outage prediction model, using the amount of aggressive tree trimming as one of the predictor variables. With this dynamic model we were able to run a counter-factual analysis to determine that historical tree trimming has reduced outages in Connecticut somewhere between 25 and 42%, and that additional tree trimming could have further reductions in outages in the future.

Reduction in Outages due to Tree Trimming Over Time by William O’Taylor
Categories
Article Research Products

The Effects of Tree Trimming on Power Reliability in Storms of Different Strengths

We just got a new article published in a special issue of Sustainability. It uses statistical methods to quantify the effects of Enhanced Tree Trimming (a rigorous tree trimming standard being used by Eversource), differentiating between storms of different intensities. We show that this type of tree trimming is able to reduce power outages during all types of storms, but the effects diminish somewhat for strong storms: which can cause much more severe damage like downed trees.

Tree Canopy Cover in Connecticut (Image by Will O’Taylor)

We’re hoping that with these results, utilities and regulators will be better able to prioritize measures to make the power grid more reliable, and help prepare for an increasingly uncertain climate. Check it out here. Check out the full text of the article here.

Categories
Conferences Research Products

AGU 2021 Fall Meeting

The AGU 2021 Fall meeting just wrapped up in New Orleans, and it was quite interesting to be there. It was a ‘hybrid’ conference with both in-person and online elements. It was a great chance to present our work in-person and online, and connect with other people in the field.

Jackson Square, New Orleans

Check out the video below of what we presented. It describes the process of developing a machine learning-based weather-related power outage prediction model specifically for high impact storms. These storms are generally difficult to get right because they’re so rare and diverse, but our model seems to be able to understand the pertinent processes well enough to accurately predict the impacts of the biggest storms that have affected the Northeastern US in the past decade.

Extreme Weather Event Outage Prediction Presentation