Dissolved gas analysis (DGA) is the primary tool for incipient fault diagnosis in oil-immersed power transformers, yet most machine learning classifiers applied to DGA operate as black boxes whose predictions cannot be verified against established diagnostic standards. This paper introduces an Explainable Boosting Machine (EBM) for dissolved gas analysis (DGA)-based fault classification across seven IEC 60599 fault types. The EBM decomposes every prediction into per-feature score contributions that the engineer can audit at the level of individual gas ratios. On an independent test set of 70 samples drawn from a different transformer population, the EBM with default parameters achieves a macro-averaged F1-score (F1-macro) of 0.849 and a Matthews correlation coefficient (MCC) of 0.838, outperforming all black-box baselines with hyperparameter optimisation. A hybrid architecture pairing the EBM with a Light Gradient Boosting Machine (LightGBM) corrector raises F1-macro to 0.856 while preserving first-stage interpretability. The learned shape functions exhibit score transitions aligned with IEC 60599 diagnostic thresholds. This alignment provides independent evidence that the model recovers the underlying fault-gas chemistry. A controlled protocol sensitivity analysis shows that common evaluation choices inflate F1-macro by four to seven percentage points on average across models, with individual random partitions reaching eleven points. This inflation explains the gap with the 95–99% range reported in the literature.
This paper presents a new traveling wave (TW) detection method for transmission line (TL) fault location applications. The proposed solution defines an adaptive threshold from the statistical properties of the steady-state period of TW records, accounting for the effects of both Gaussian and non-Gaussian characteristics of real-world measurement noise. As a result, the user-defined setting process is no longer required, making the TW detection procedure more robust to the statistical variability of practical measurements and to errors associated with manual threshold adjustments. The validation of the proposed technique is carried out using a variety of real-world fault records taken from distinct manufacturers, whose distances were confirmed by the utilities' crew in the field. In each case, a classical double-ended TW-based fault location formulation is applied. The results demonstrate that the proposed adaptive strategy reliably detects the first fault-induced incident TWs, providing a settings-free and a manufacturer-independent methodology for post-fault analysis. Furthermore, statistical normality tests indicate that the steady-state noise cannot be fully described by a purely Gaussian model.
Although deep learning–based methods have achieved promising results in partial discharge (PD) diagnosis for gas-insulated switchgear (GIS), their interpretability and robustness under severe background noise remain limited. To address these issues, this paper proposes DPI-WaveNet, a dual-interpretable wavelet-driven network for GIS PD diagnosis in extreme noise environments. First, a learnable wavelet convolution is first embedded into the network input stage to replace randomly initialized kernels, enabling adaptive multi-scale time–frequency decomposition of raw PD signals. Subsequently, adaptive feature enhancement and denoising mechanism, integrating learnable band weighting and soft-threshold denoising, is then developed to strengthen PD-related discriminative frequency responses while suppressing noise-dominated components. To further improve interpretability, a Grad-CAM–based frequency-band interpretation method is introduced to reveal the contributions of key diagnostic frequency bands. In addition, a time–frequency prior-guided interpretable loss function is designed using frequency-focused and temporal sparsity constraints, encouraging the network to learn physically consistent discriminative response patterns for different PD types. Experimental results under multiple defect categories and various signal-to-noise ratio (SNR) conditions demonstrate that the proposed method achieves an overall classification accuracy of 99.05% ± 0.27% under ideal conditions while maintaining superior robustness and stable diagnostic performance under extremely low SNR conditions. These results validate the effectiveness and engineering potential of DPI-WaveNet for interpretable GIS PD diagnosis in complex noise environments.
Medium-voltage DC (MVDC) grids are key components in building modern power systems due to their high efficiency and flexibility. MVDC power systems are typically stiff systems, and their electromagnetic transient simulations often involve multi-time-scale dynamics, such as those of high-frequency DC transformers and low-frequency modular multilevel converters. To balance simulation accuracy and computational efficiency, the common strategy is to divide the network into subsystem partitions according to different time-scales and adopt corresponding simulation time-steps. At present, delay-based decoupling and multi-rate simulation approaches have been widely used, which may face the challenge of accuracy and stability for multi-time-scale MVDC grids. In response, this article proposes a multi-rate compensation method suitable for parallel electromagnetic transient simulation of MVDC power systems with multi-time-scale dynamics. Based on the network tearing approach, the method can realize delay-free decoupling and parallel simulation of different partitions without introducing artificial delay. It adopts different simulation time-steps for different partitions to further improve the simulation efficiency. The stability, accuracy, efficiency and computational complexity of the proposed multi-rate compensation method are analyzed through numerical stability analysis, error analysis and simulation of a medium-voltage flexible DC transmission system.
With the growing penetration of renewable energy, hybrid power system consisting of wind power, photovoltaics and thermal power are required to participate in primary frequency regulation (PFR) to satisfy grid codes. Owing to the nonlinearity of hybrid power system and the problem of switching among multiple operating conditions, traditional control methods face enormous challenges. Deep reinforcement learning such as double deep Q network (DDQN) can overcome these nonlinear drawbacks. However, its adaptability is extremely poor and requires training. In this work, the TL-DDQN algorithm using transfer learning is proposed to increase the adaptability of the DDQN algorithm. By optimizing the learning rate and target network assignment interval of the DDQN algorithm based on data correlation analysis in different scenarios, the practicality of the DDQN algorithm is improved, and the training time is reduced. Finally, an actual hybrid power system in China is adopted as the research object. In addition, the results of these simulations reveal that the TL-DDQN algorithm has significant advantages for increasing the PFR capability of hybrid power system, which proves that the TL-DDQN algorithm proposed in this paper can increase the adaptability of the DDQN and reduce training time.
