How Storm Impact Analytics can Save Utilities; Part II: Machine Learning Methodologies
In last year’s Utility Products’ May issue, we discussed the importance of storm-impact analytics, noting how predictive information can have a positive effect on a utility’s ability to prepare for a storm.
By Don Leick
The 2017 hurricane season was the most expensive in U.S. history, causing an unprecedented amount of damage. In that year alone and for the first time in history, two Atlantic Category 4 hurricanes made landfall in the Continental United States. Current trends suggest this level of volatile weather will continue, meaning utilities, more than ever, need to be prepared for inclement weather to strike their territory.
A tree fell, taking wires with it, during a hurricane. Photo credit, iStock.com
With this trend comes the need for improved power restoration to meet the steadily growing customer and regulatory expectations. Better visibility on potential weather threats, and their impact, is key to improving restoration planning and deployment efforts.
In last year’s Utility Products’ May issue, we discussed the importance of storm-impact analytics, noting how predictive information can have a positive effect on a utility’s ability to prepare for a storm. One form of predictive analytics, machine-learning methodologies, is being increasingly deployed by utilities in various applications, from predictive maintenance to analyzing customer behavior. In this article, we’ll discuss exactly what machine-learning is and how it can be leveraged to advance utility storm preparation and restoration efforts.
Machine Learning Technology
Ultimately, it’s impossible to perfectly prepare for every storm. But improvements in preparation can make great strides in reducing operational costs, avoiding penalties from regulators, and improving the utility’s image with customers. To better prepare, utilities must be able to determine a forecasted storm’s damage potential and the resulting staffing needs.
Current approaches to storm preparation typically go like this: Someone at the utility looks at the weather forecast, then calls on their experience with a similar storm and how it impacted the utility. Many utilities also consult with professional meteorologists on a regular basis, which helps them better understand the risk. For example, the forecast may be given a numerical storm categorization—where a storm is rated by severity and expected impact on the utility’s service territory. While that’s a good approach, it’s certainly not foolproof. What utilities need is a better understanding of the real damage that will result to properly gauge staffing requirements for restoration.
A promising new approach to storm preparation is damage prediction with machine learning. Also known as artificial intelligence or neural network algorithms, machine learning is far more sophisticated and accurate for predicting damage and outages. Machine learning uses past history to identify patterns that enable future prediction of outages.
While it’s being used across many sectors, machine learning is a particularly good fit for the utility damage prediction problem. Every utility is different, whether it be by design, infrastructure age, maintenance practices, aggressiveness of vegetation management, etc. This means no two utilities will ever be impacted by weather in the same way. In fact, different areas of one utility may respond differently to the same storm. Machine learning can learn those differences and transform weather forecasts into actionable information. Quantitative predictions are provided to the operations team well before a storm’s arrival, such as an estimate of the outage incidents on a utility’s service territory, and within the various operational areas of the service territory. This information enables the right level of staffing, not too few and not too many resources. This avoids those problematic cases of excessive storm restoration times or procuring expensive third-party resources and then not using them—or at least reduces the frequency of their occurrence.
What Matters Most: Data
Because machine learning is based on training using historical data, the availability and quality of the data used as inputs is key to any machine learning process. Some of that data will be provided by the utility and others by the solution provider. Data needed for outage prediction can include:
• Historical outage incident information from the utility that is both time stamped and geo-located. Often two to five years of outage incidents, or trouble spots, are required to adequately train the models. Less frequent types of storm events, such as ice storms or tropical storms and hurricanes, will require more.
• Historical weather information that corresponds with a utility’s weather-related outages. This needs to be very fine resolution, gridded data.
• The utility’s overhead distribution system data in geospatial form. It’s important for the machine learning model to know where poles, lines and isolating (protective) devices are located. A simple example illustrates why this is important: the more lines a utility has in an area, the more likely there will be a problem.
• Tree trimming history to understand where the utility is in its four-year to five-year cycle. This is very useful, if available, because it improves the accuracy of predictions. For example, if trees were trimmed a year ago, there will be fewer outages than if trimmed four years ago.
• Factoring in trees to determine where they are in proximity to lines and when the spring/fall leaf changeover occurs for deciduous trees.
All of this data is then used to build predictive damage models specific to each type of weather event, whether it’s a thunderstorm, snow or high winds. The models can then be used to predict the impact of future storms based on forecasted weather. A number of predictive models will be used for each type of storm expected, because different machine learning algorithms are combined to get a range of potential incidents on a utility’s territory. Machine learning can determine where outages will be located throughout a utility’s entire network, which identifies areas that will be most affected, allowing operations teams to better plan for staging, pre-positioning or simply alerting local crews. It can also highlight specific areas of weakness in a utility’s infrastructure, as illustrated by outage incident history and the damage forecasts themselves.
Lineman doing repairs after a hurricane. Photo credit, iStock.com
If you are reading this and want to deploy machine learning but believe it will be a challenge to develop on your own, you’re absolutely right. Just one issue alone—the expense of a high-quality weather forecast at high resolution, including the needed historical data—is a showstopper for almost all utilities. Utilities, however, do not need to make this upgrade themselves. A few select vendors have built machine learning solutions specifically for the damage prediction problem, which can be applied to any utility. A utility simply provides their data to a vendor who will take care of the rest, working closely with the utility to show results. With the right solution, it doesn’t have to be a high-risk science project or a long open-ended consulting engagement. Meteorological insights can be applied as well, so it is not just a “black box” solution.
Before utilities begin to use machine learning operationally to gauge their critical staffing decisions, they need to understand the reliability of its predictions. A huge benefit of machine learning is the ability to see the accuracy of a model’s predictions before putting it into operation, using “cross-validation” techniques. For example, if a utility’s wind event model has been built with information from 100 previous storms, cross validation involves a rebuild of the model to include storm and outage information from 99 of the 100 storms. This intentionally leaves out the 100th storm. Then, the model is tested against that 100th storm. Because the model has no knowledge of the storm, it’s basically generating a prediction for forecasted weather. This process is then repeated to create a prediction for a second storm, rebuilding the model with the other 99 storms. This process continues until a prediction has been created for every one of the 100 storms, and accuracy metrics are compiled.
This process allows a utility to understand just how accurate the model is—as if it’s had years of experience with it. Cross validation also identifies the strengths and weaknesses of the prediction system. One weakness may be that utilities simply don’t have enough data for a particular type of storm, such as a hurricane. But as the utility is able to accumulate more storm data over time, the model can be re-trained with the latest storm data to improve accuracy, including rare events such as hurricanes.
2017 was a year unlike any other, and at this time the hurricane outlook for 2018 has strong similarities to last year. Utilities should no longer continue with the status quo, such as just looking to weather forecasts and meteorological consulting, alongside their own experience. Instead, the opportunity exists, with machine learning, to provide better predictions of a storm’s impact and the resulting damage prior to a storm’s arrival. With this additional guidance, the frequency of being under-prepared or over-prepared can be reduced, with great benefits to your restoration efforts.
About the author: Don Leick is senior product manager for DTN’s weather business. Leick leads the future direction and enhancement of online, mobile and alerting products. He has been the product manager for the WeatherSentry product for most of his 13 years with the company. Leick is currently focused on leading the improvement and application of DTN’s machine learning-based product for damage and outage prediction, Storm Impact Analytics.