Battery Energy Storage Deployment: Transforming Planning to the Radical Average

• In Event: Innovation and Risk in the Power Grid

Abstract

Battery energy storage (BES) is being rapidly deployed across the United States. This technology is both providing traditional energy, capacity, and ancillary services to the grid and is increasingly being linked to data center development. To date, 53.4 GW of BES has been deployed. However, an additional 469.3 GW of planned capacity across 2,535 projects is making its way through state regulatory dockets and local siting processes. Strategically analyzing how this technology is being conceptualized, managed, and integrated into the energy system has the potential to transform the planning and operation of electricity systems. However, analyzing state public utility commission dockets is notoriously time-consuming and detailed work, difficult to do at scale, and a strain on the capacity of state regulators. In this paper, we explore the use of large language models (LLMs) and AI agents to structure and probe BES-related public utility dockets. In doing so, we explore whether large-scale document analysis can provide reliable and replicable analytical support for PUC document analysis, paired with human experts in the loop. The scope of our research presents a case study of the approval of 332 MW of storage, out of a total of 1.4 GW of storage needed to support a proposed data center in Saline Township, Michigan. The storage is part of the utility’s integrated resource plan and aligns with recent state statutory requirements and consumer protections. With these additions, DTE Electric’s total storage capacity will be 2.6 GW, making it among the largest utility energy storage portfolios in the country. The BES systems match the data center’s contracted demand. Data collection includes public documents from two interrelated PUC dockets with 12 orders, 590 filings, and 6,551 public comments, along with state statutes, media releases, and other official documents.
Our analysis proceeds in two streams. From the official filings, AI agents extract project facts and timeline as well as stakeholders, validated through multi-agent cross-checks. From the public comments, we inductively identify concerns through unsupervised clustering. We then apply expert judgment to retain and consolidate those most relevant to our interest in battery storage and develop a well-defined codebook of topics. Finally, we deductively classify every concern against this codebook, with human validators reviewing the raw comments, identifying systematic errors, and iteratively revising the classification prompts. We then combine the streams to examine the sentiment attached to each concern, how concerns and positions shift around key decisions, whether stakeholders align into coalitions, and whether these patterns differ between the data-center-linked storage projects and the other projects in the docket. The analysis provides insight into how this BES-data center deployment is being framed, debated, and supported.BES integration presents a fruitful avenue for methodological development and testing because it requires fundamental rethinking of how utilities, grid operators, and regulators assess resource adequacy, plan energy infrastructure investments, and engage in new business models. Our initial findings suggest that LLM- and AI-assisted analysis requires well-defined research questions and iterative human-AI inter-coder reliability checks.

Authors

• Yifan Liu, Harvard University
  Presenting Author

• Stephanie Lenhart, Boise State University
  Non-Presenting Co-Author

• Elizabeth J. Wilson, Aalto University FINLAND
  Non-Presenting Co-Author