Furqan Amjad

This paper presents Power-Q, a hybrid multi-agent framework that combines rule-based anomaly screening with Large Language Model (LLM)-driven analytical agents to deliver explainable, near real-time supervisory insights into power quality disturbances. The architecture integrates decentralized screening agents dedicated to solar, wind, and load domains with LLM-based Analyst Agents that perform structured root cause interpretation, generate live event summaries, and support natural language interaction with operators. The framework was evaluated on a simulated five-bus microgrid using parametric fault injection and real-time data streaming. Results show that Power-Q achieves sub-200 ms anomaly detection latency, produces live summaries within approximately 2 seconds, and generates complete diagnostic reports in under 4 seconds. The Supervisor Agent demonstrated alignment with key reporting elements defined in IEEE standards, producing interpretable and structured diagnostic outputs across the evaluated scenarios. The Ticker Agent achieved complete event coverage with clear, operator-readable summaries, while the Query Agent enabled effective human-in-the-loop interaction. Overall, Power-Q provides a software-centric, operator-focused framework designed to support scalable deployment in principle, enhancing situational awareness and decision support in smart grid environments.

This paper presents a comprehensive review of the applications, challenges, and future directions of Large Language Models (LLMs) in the Electrical Power Domain (EPD). Leveraging transformer-based architectures such as GPT, BERT, and LLaMA, LLMs have shown transformative potential across power system applications including load forecasting, fault diagnosis, regulatory compliance, question answering, risk assessment, and intelligent data analysis. Through a systematic analysis of over 45 studies, the review highlights measurable benefits such as up to 20% improvement in load forecasting accuracy, 30% reduction in operational response time, and 40% decrease in manual workload for regulatory tasks. LLMs have demonstrated strong adaptability through zero-shot and few-shot learning and are capable of processing multimodal inputs for real-time decision-making. However, limitations such as high computational costs, lack of domain-specific datasets, limited explainability, and concerns around regulatory alignment hinder widespread deployment. To address these gaps, the paper outlines research opportunities including domain-specific fine-tuning, scalable deployment strategies, multimodal integration, and the development of unified benchmarks such as ElecBench. Overall, the integration of LLMs in power systems represents a significant step toward more intelligent, reliable, and sustainable energy management.