AI Large Model-Based Data Value Mining in Power Projects

Abstract

Power project usually has the characteristics of long construction period, many participants, complex supply chain and large number of project documents. Information such as schedule, earned value, cost, contract changes, quality, safety, equipment delivery, BIM / IoT, and meeting records are usually stored in separate systems. Therefore, although the data are available, they may not provide timely decision support for project management. Traditional statistical models are not good at interpreting narrative evidence. When the project conditions or supply conditions change, the prediction accuracy of the static model may also decrease. In order to solve these problems, this paper proposes Power-LLM ValueLoop, a closed-loop framework for data value mining supported by large models. The framework identifies entities, events, causal signals, urgency, conflict evidence, and confidence levels through a pattern-based large model layer, and associates relevant results with structured data such as schedule, cost, resources, quality, and supply chain. Subsequently, the semantic-time model is used to estimate and calibrate the probability of major delays or failure events in the next 4 weeks. Confirmed outcomes, human corrections, time-decayed learning, drift monitoring, and threshold adjustment form the feedback loop. The system can deliver traceable risk early warnings, identify key risk drivers, and provide action recommendations for review by project personnel. This paper uses a synthetic project portfolio for evaluation, including 720 power engineering projects, 10 start-up batches, and 15,653 project-week observation records. Controlled concept drift was introduced from the 6th batch. In the four batches after drift, the macro average ROC-AUC of the feedback model was 0.773, the macro average PR-AUC was 0.366, and the Brier score was 0.105. At the recall-oriented operating point, the model identified 78.1 % of the risks in the next four weeks, with an average warning advance of about 9.1 weeks. For 80.2 % of the event items, the system can issue an early warning at least 4 weeks before the event. Compared with the static AI fusion model, the feedback mechanism improves the recall rate by about 0.140, but the false alarm rate also increases from 0.275 to 0.391. This shows that the improvement of model performance still needs to rely on reasonable threshold management and manual review.  A supplementary Monte Carlo decision analysis examines prevalence and error-cost sensitivity at fixed operating rates. At the reported test prevalence, feedback has lower expected error loss only when the missed-event cost exceeds approximately 5.35 times the false-alarm cost. This conditional result does not establish savings in deployed projects.