Advancing Utility Adoption of Large Language Models: Practical applications in asset management

Presentation Description: Implementing asset management practices relies on large amounts of readily accessible trustworthy data about infrastructure and utility systems. Often, however, the data that utilities have available to support analyses is difficult to find, stored across multiple sources, and is time-consuming to process for analyses. As a result, many asset management programs advance to the stage where data collection efforts are not fully utilized to inform decision-making, and dashboards only report on a subset of important key performance indicators that do not provide all the necessary information to inform the best decisions. As a result, utilities are increasingly looking for solutions to address this common challenge, develop accurate and trustworthy planning models, and consistently get data in the hands of decision-makers.

The adoption of innovative technology represents an opportunity to enable these changes. Large language models (LLMs), like the technology powering ChatGPT, have recently emerged as a tool that can make utility data more accessible to decision-makers. These tools are now in widespread use for both personal and professional uses across many industries and offer new ways of analyzing, summarizing, and conducting complex reasoning about data. This presentation will demonstrate how utilities can adopt large language models and apply this technology across various tasks relevant to asset management. We will provide information about techniques for using large language models for complex tasks beyond chatbots, such as prompting and retrieval augmented generation. We will share several real-world examples of how utilities have used large language models in asset management programs to extract text from work orders; automate data entry; streamline regulatory reporting; and support information retrieval and question-answering about information contained in utility-specific documents, like operation and maintenance plans and engineering reports. The presentation will also include a forward-looking perspective on anticipated advances in generative artificial intelligence and a discussion on the risks and limitations of using LLMs on tasks related to critical infrastructure. We hope to inspire attendees to consider novel ways that LLMs can be tailored to meet the unique needs of utilities.