Integrated audit intelligence: A multi-task learning model for classification and opinion generation in power systems

Auditing in electric power systems involves complex decision-making based on large volumes of unstructured and domain-specific records. Under the digitalization and decarbonization of integrated energy operations, such audit records also reflect cyber-physical and operational risks, so Artificial Intelligence (AI)-driven audit intelligence can support resilience enhancement by improving risk visibility and accelerating standardized corrective decision-making. Automating this process requires models that can understand context, identify issues, and generate relevant textual outputs. This research presents an integrated Multi-Task Learning framework for power-system auditing that jointly performs audit issue classification and audit opinion generation within a unified model. The system employs multiple encoders and decoders anchored on a pre-trained transformer-based language model to process heterogeneous audit content and learn both common and task-specific patterns. Experiments on real-world datasets from electric utilities demonstrate that the proposed multi-task learning framework outperforms conventional single-task baselines in both classification and generation tasks. The model delivers consistent improvements in classification metrics such as F1-score (harmonic mean of precision and recall), along with enhanced generation quality across Bilingual Evaluation Understudy (BLEU), Recall-Oriented Understudy for Gisting Evaluation (ROUGE), Distinct-n (lexical diversity), and Bidirectional Encoder Representations from Transformers Score (BERTScore), which measures semantic similarity. Additional ablation studies are conducted to assess the contribution of different input types. The results confirm the effectiveness of the proposed approach and highlight its potential as a scalable and intelligent solution for advancing audit automation in the power sector. • Novel integrated MTL model for audit issue classification and opinion generation. • Dual-encoder Transformer fuses audit narratives and cited provisions via attention. • Real utility audits: classification F1 improves from 0.8042 to 0.8309 vs baselines. • Generation quality improves on BLEU, ROUGE, Distinct-n, and BERTScore metrics. • Ablations show both inputs are essential; removing one drops F1 up to 7.3 pts.