• A physics-informed multimodal LLM (DSM-EQA) for energy question answering is introduced. • Unified fusion of time-series, weather data, and grid topology enables accurate reasoning. • Physics-constrained learning ensures 94.6% compliance with power, voltage, and thermal limits. • Achieves 89.4% numerical accuracy and reduces hallucinations to only 3.2%. • Validated with real grid events, showing 91% agreement with historical operator decisions. The rapid integration of renewable energy has dramatically increased power system complexity, requiring decision-support tools that deliver accurate, physically consistent, and human-interpretable responses to operational queries. This work proposes DSM-EQA, a physics-informed multimodal large language model that seamlessly fuses operational time-series, meteorological data, and grid topology into a unified latent representation, adapts a pretrained language model via low-rank adaptation, injects physics knowledge through continuous prefix conditioning, and simultaneously generates fluent natural-language explanations and quantitative system indicators under strict power-balance, voltage security, and thermal-limit constraints enforced by differentiable regularization. Evaluated on a physics-verified benchmark of over 10,000 question-answer pairs across six core energy management tasks, DSM-EQA achieves 87.3% Decision Support Accuracy, 94.6% physics compliance, and 89.4% numerical accuracy surpassing GPT-3.5-Turbo by 29.2 percentage-points and GPT-4o by 16.7 percentage-points while reducing hallucinations to 3.2% ( C o h e n ′ s d = 0.82 , p < 0.001 ), with benchmark case studies confirming 91% alignment with operator-validated reference decisions. DSM-EQA provides a strong, physically grounded foundation for trustworthy AI decision support in safety-critical energy infrastructure during the global energy transition.
Abbas et al. (Fri,) studied this question.