Driven by the high penetration of renewable energy sources and power electronic devices, modern power systems have become increasingly complex, intensifying the demand for accurate and intelligent fault detection. This paper analyzes a total of 81 references to explore the integrated application of artificial intelligence (AI) technologies across all stages of fault data processing, modeling, and optimization. The application potential of AI in fault data processing is firstly analyzed in terms of its performance in mitigating class imbalance, extracting feature information, handling data noise and classification. Then, the modeling of fault detection is classified into rule-driven, data-driven and hybrid-driven methods to evaluate their applicability in scenarios such as transmission lines and distribution networks. The accuracy of fault detection models is also investigated by studying the hyperparameter optimization (HPO) methods. The results indicate that the utilization of AI-driven imbalance handling enhances model accuracy by a range of 16.2% to 26.2%, while deep learning-based feature extraction techniques sustain accuracy levels exceeding 98.5% under a signal-to-noise ratio (SNR) of 10 dB. With a 99.96% detection accuracy, hybrid-driven models applied in fault detection perform the best. For the optimization of fault detection models, heuristic algorithms provide 6.92–19.375% improvement over the baseline models. The findings suggest that AI-driven methodologies in data processing demonstrate notable noise resilience and other benefits. For modeling fault detection, data-driven and hybrid-driven models are presently extensively employed for detecting short-circuit faults, predicting transformer gas trends, and identifying faults in complex and uncertain scenarios. Conversely, rule-driven models are better suited for scenarios possessing a comprehensive experience library and are utilized with less frequency. In the optimization of fault detection models, heuristic algorithms occupy a pivotal position, whereas hyperparameter optimization incorporating reinforcement learning (RL) is better suited for real-time fault detection. The discoveries presented in this paper facilitate the seamless integration of AI with fault detection in modern power systems, thereby advancing their intelligent evolution.
Maintaining a stable thermal environment is essential for ensuring data center thermal safety and reliable operation. However, existing studies lack a systematic synthesis of thermal environment measurement, modeling approaches, and their contributions to energy-efficient cooling. This review systematically analyzes 77 studies published since 2010 on experimental measurement, CFD simulation, and predictive modeling of data center thermal environments. The review identifies representative measurement strategies at rack and room scales, analyzes factors affecting CFD reliability, and evaluates the applicability of predictive models based on experimental and simulation data. Most experimental studies focus on temperature and airflow, while only about 30.7% of the 39 experimental studies include power measurements, limiting quantitative evidence linking thermal optimization with energy performance. Among the 58 CFD studies, Fluent was used in 41.4% of cases, while 70.6% of the 17 predictive-model studies relied on CFD data for model training. CFD reliability remained sensitive to boundary conditions, model simplifications, and validation strategies. Predictive modeling was dominated by machine learning and deep learning, while reduced-order and surrogate models were also applied, enabling rapid temperature-field and thermal-risk assessment but remaining constrained by training data quality and operating condition coverage. These findings establish an integrated framework in which measurements provide empirical benchmarks, CFD extends spatial and operational understanding, and predictive models enable rapid assessment for thermal environment management and cooling energy optimization.