Toward self evolving quality assurance frameworks for AI driven intelligent energy management software

The subject matter of the article is the processes of designing and validating a self-evolving quality assurance (SEQA) framework for artificial intelligence (AI)-driven intelligent energy management software (IEMS). The goal is to develop a scalable and adaptive SEQA framework that enables continuous optimization and reliability- and trustworthiness-oriented quality assurance in dynamic, heterogeneous operational environments. The tasks to be solved are: to formalize a unified QA architecture integrating reinforcement learning for adaptive control and federated learning for distributed calibration; to develop a robust cross-domain adaptation mechanism ensuring energy-aware trust calibration; and to empirically validate the framework's performance against baseline models across multiple real-world energy datasets. The methods used are: reinforcement learning for policy-driven optimization, federated learning for privacy-preserving model aggregation, trust calibration techniques for reliability assessment, and experimental benchmarking on NASA, UCI, and OPSD datasets. The following results were obtained: the proposed SEQA framework successfully integrates reinforcement learning-based local adaptation with federated policy aggregation, achieving continuous self-evolution of QA performance across heterogeneous energy management scenarios; the cross-domain adaptation mechanism ensures robust generalization capability, with F1-scores exceeding 0.86 and reliability remaining above 0.91 under diverse operational conditions; experimental validation demonstrates consistent improvements in reliability (F1-score increases by 6–8%), calibration accuracy (Expected Calibration Error reduced to 0.024), and energy efficiency (up to 13%) compared to baseline QA models; the framework maintains stable performance under dynamic data distributions, with ablation studies confirming that each component—reinforcement learning, federated evolution, and continual replay—plays a critical role in enabling robust self-evolving quality assurance. Conclusions. The scientific novelty of the results obtained is as follows: 1) the proposed SEQA framework introduces a unified adaptive paradigm that synergistically combines reinforcement learning, federated calibration, and cross-domain adaptation, enabling autonomous, continuous quality evolution in IEMS; 2) the developed cross-domain mechanism achieves robust generalization and energy-aware performance balancing, addressing key limitations of static and single-domain QA approaches; 3) the extensive experimental validation demonstrates consistent improvements in reliability, calibration accuracy, and energy efficiency, confirming the framework's practical applicability for long-lifecycle industrial deployments; 4) the integration of adaptive aggregation intervals and policy pruning mechanisms minimizes redundancy during synchronization while maintaining near-linear scalability on distributed federated nodes, validating the framework's feasibility for deployment in real-world smart grid and industrial IoT environments