Wanqi Yuan

The large-scale integration of renewable energy has led to complex topologies in 110 kV distribution networks, posing significant challenges for accurate fault section location. Meanwhile, high-resistance grounding faults complicate fault type identification. To address these problems, this paper first proposes a novel fault diagnosis method based on mapping time series measurements to RGB images. Specifically, three-phase currents and voltages are measured at the terminals of 110 kV lines and mapped to three channels of RGB images. The fault diagnosis problem, which includes fault type identification and fault section location, is reformulated as two image classification tasks. Then, a lightweight ResNet (LW-ResNet) is proposed to solve the fault type identification problem by refining the residual blocks in traditional ResNet. Finally, a Squeeze-and-Excitation (SE) attention module is combined with LW-ResNet (LW-SE-ResNet) to improve the accuracy of fault section location, especially for faults near T-junctions. Simulation results demonstrate that under simulation conditions with synchronization and measurement errors, LW-ResNet achieves 97.8% accuracy in fault type identification, and LW-SE-ResNet achieves 95.6% accuracy in fault section location. Compared with traditional methods, the proposed fault diagnosis method offers superior accuracy and robustness. The real-time capability is verified, as the proposed methods achieve high fault diagnosis accuracy using data from the initial 0.25 cycle after a fault occurs, and the inference latencies are less than 5 ms.

Faulty phase selection (FPS) methods are used to solve the problem of single phase to ground fault (SPGF) in the distribution network. When a false FPS (FFPS) issue occurs, the short-circuit current in the faulty line suddenly increases. It is crucial to diminish damages caused by the FFPS issue and rapidly re-select the faulty phase. Therefore, this paper proposes a fault-tolerant method based on an active-intervention-type arc suppression device with a soft switch mode (AASD-SSM) and a fast FPS method based on wavelet packet transform (WPT). To strengthen the fault-tolerant performance of FPS, the AASD-SSM is first developed with an intermediate resistance and a soft switch. Then the value of intermediate resistance is determined by the short-circuit current in AASD-SSM. To realize the accuracy and rapidity of FPS, a modified WPT (MWPT) method based on frequency band energy is presented. In different fault scenarios, the number of wavelet decomposition layer is updated by MWPT and the SPGF issues in the faulty lines can be monitored in real-time. Simulation results demonstrate that AASD-SSM improves the fault-tolerant performance of FPS. The phase selection accuracy of MWPT is better than that of WPT in all fault scenarios and the computational complexity of MWPT is reduced for more than 10%.

With the continuous development of the smart grid, details of residential users’ electricity consumption have received lots of attention. In recent years, a new concept of nonintrusive load monitoring (NILM) based on graph signal processing (NILM-GSP) has been proposed. NILM allows researchers to access the electricity consumption profile of every household appliance (AP) online, which greatly reduced the cost of grid operation. This article proposes a new NILM based on alternating optimization (NILM-AO) to solve the load disaggregation problem. Since the power consumption of residential users is basically stable in its operation cycle, a power consumption constraint is constructed to make NILM realistic. In order to improve real-time processing capability, we propose a statistical downsampling method (NILM-AODM) to find the optimal sampling rate. Simulation results show that NILM-AO is more accurate than NILM-GSP, and the sampling rate of NILM-AODM is also smaller than that of NILM-AO, which leads to less communication load and time delay.

Topology identification (TI) is an essential problem in the distribution network due to exponentially growing power grid size in recent years. In this study, it is reformulated as a regularised alternating convex optimisation problem. Then an application based on the current injection model is proposed. Compared to the traditional algorithm optimising ‐norm, which may lead to overfitting, a new ‐norm minimisation problem with ‐norm regularisation is proposed to solve the trade‐off problem with non‐convex constraints. The proposed method reduces the size of the training data set compared with the traditional TI method. Simulation results show that the recovery performance of the proposed algorithm is superior to the traditional one in additive white Gaussian noise scenario.