Peer-Reviewed Research

The digital transformation of energy asset performance management (APM) is accelerating; however, most existing systems remain model-centric, fragmented, and heavily dependent on manual expert interpretation. Although advanced machine learning and physics-based assessment models are widely deployed across energy generation assets and electrical grid infrastructures, they are typically executed in isolation, limiting their ability to provide coherent, real-time, and context-aware decision support. This paper addresses this structural limitation by introducing an agentic artificial intelligence (AI) framework that formalizes analytical orchestration as a goal-driven, modular multi-agent system for large-scale energy portfolios.

We propose an agentic AI framework in which analytical tasks are decomposed into structured reasoning steps, dynamically linked to heterogeneous domain tools. The contribution lies in designing an intelligent orchestration layer capable of selecting, sequencing, and synthesizing existing analytical models according to user intent and operational context. The framework is implemented as a prototype multi-agent architecture in Python, using graph-based workflow orchestration to model execution dependencies and structured reasoning traces to ensure transparency and explainability. A retrieval-augmented generation (RAG) module integrates contextual technical knowledge and produces structured narrative syntheses of results, enabling consistent interpretation across heterogeneous analyses.

The proposed architecture is evaluated through a large-scale operational assessment use case integrating established analytical methodologies from OpenOA, including Monte Carlo annual energy production estimation, long-term gross energy modeling, availability and electrical loss quantification, wake and resource loss assessment, condition-based diagnostics, and energy yield gap analysis. The agentic layer coordinates these modules through dynamic tool selection and graph-based execution workflows, generating auditable reasoning traces for each analytical request. Validation is conducted on a portfolio of 42 wind farms, representing approximately 5 GW of installed capacity across North America, using three years of historical SCADA data. More than 400 analytical queries were executed. Preliminary results indicate efficient end-to-end execution times of 5–10 seconds for lightweight queries, diagnostic narrative inconsistency rates below 5% in expert-reviewed evaluations, and a structural reduction in workflow coordination overhead through integrated orchestration.

By formalizing analytical orchestration as an agentic control layer, this work advances digital transformation in energy APM from model-centric automation toward goal-oriented analytical intelligence. The proposed architecture reduces dependency on fragmented expert workflows, enhances cross-technology consistency, and provides a scalable foundation for hierarchical multi-agent coordination. The results demonstrate the practical viability of agentic AI as a structural innovation for next-generation, AI-driven energy infrastructure asset performance management systems.