Artificial Intelligence and Machine Learning is an Important Tool to Improve Cost Estimation for Natural Disasters in Electric Utilities

The normal procedure to approve and allocate funding requires an extensive cost estimation process — one that is particularly difficult for electric infrastructure work that can require specialized workers and equipment. In the context of natural disaster recovery, this estimating work is rife with uncertainties (and, of late, made even more difficult because of  Covid-19). For instance, in Puerto Rico after Hurricane Maria, significant electric repairs needed to be performed in hard-to-reach mountainous areas, demanding helicopters with highly trained work crews. At the same time, electric infrastructure repair is critical and urgent — other repair work can’t go on without it. 

Access to data is a key factor for success in any cost estimation process. Over the past several years, several entities have collected cost data from electric utility repair and replacement efforts following natural disasters. Examples of these include commercial cost databases such as RSMeans or EOS, privately supported research efforts such as Facebook Data for Good, academic exercises such as Arizona State University’s SHELDUS program, and government-supported databases such as NOAA’s Nighttime Lights, and FEMA’s Grants Manager. A collection of these alternative data sources, together with data locally collected, could be aggregated into large databases that could facilitate the cost estimation process for electric utilities in the aftermath of natural disasters. However, such data is not completely structured and includes a mix of information that does not follow a standard format.  

We propose that Artificial Intelligence and Machine Learning (ML) methods could be suitable techniques for extracting useful cost estimation information from the type of non-structured databases described above. AI uses techniques to train computers to acquire and apply knowledge. ML is a subset of AI which focuses on identifying previously unseen sources of value in data, often patterns across variables that identify of previously unseen correlations and improve predictive capabilities. ML techniques are effective when working with data sets that are too large or diverse (e.g, text, numeric, qualitative) for easy processing.  

AI/ML has seen pioneering application in diverse areas such as finance, engineering and medicine in recent years. In the context of construction, several commentators and academics have pointed out the value of these methods for cost estimation. Artificial neural networks (ANN), random forest algorithms, and self-organizing maps are a few examples of techniques that have been applied to construction projects.  

AI/ML, if applied in a disaster recovery context for electrical utilities, might significantly improve cost estimating capability and responsiveness. The following are potential AI/ML applications:

• Quick benchmarking of proposed costs for new projects by mapping them to similar projects from previous disasters in order to judge reasonableness;
• Generation of visualization maps of large disaster cost databases to identify patterns across similar repairs from other disasters;
• Early prediction of repair/replace costs based on historical data; and
• Generation of robust sensitivity analyses that identify the impact of new resilience codes and standards.  

Barriers to progress in this field have often been driven by unreasonable expectations that AI/ML can solve any data problem. Additionally, it must be recognized that the curation of appropriate and useable data is equally important as the analytical techniques used to extract knowledge from it. The curation of data is a challenging task in the disaster recovery context, given the unstructured format of many of the databases. Therefore, investments in cleansing, normalization, transformation and labeling of the data may be required before AI/ML algorithms can be applied.  

In spite of the caveats mentioned above, the ever-increasing amount of data and the availability of these advanced data mining techniques can improve the efficiency and accuracy of cost estimation in a disaster recovery context. This can be achieved by:

• Investments in acquisition of disaster cost databases;
• Training of cost estimators on AI/ML concepts;
• Research on AI/ML applications to disaster related cost estimations; and
• Socializing of AI/ML approaches to cost estimation with utility- and disaster-related stakeholders. 

Dr. Ismael Arciniegas Rueda (Ph.D., Economics, State University of New York at Albany) is a senior economist at the nonprofit, nonpartisan RAND Corporation, where he leads teams supporting FEMA on issues such as validation of electric utilities cost estimates and electric utilities codes and standards. In addition, Arciniegas leads research on energy markets and energy risk management. Arciniegas has held leadership positions at international energy companies including AES, PSEG, Constellation, and TransAlta, where he led teams that successfully participated in wholesale, retail, and renewable energy markets. He serves as adjunct faculty on energy economics and time series at Catholic University of America in Washington, D.C. 

Dr. Parousia Rockstroh is an Associate Mathematician at RAND. His areas of expertise include modeling, simulation, and cost estimation. He has applied these methods across projects in the Army, Air Force, and Homeland Security. Rockstroh has led and co-led several projects to develop conceptual models for electrical utility infrastructure projects. These projects included cost models for high-voltage transmission segments, distribution networks, and electrical substations. The methods and models in these projects were developed for cost estimation of the electrical grid in Puerto Rico, following damage from Hurricanes Irma and Maria. Rockstroh holds a PhD in Mathematics from the University of Cambridge.