1. Introduction Dual power supply for 110kV transmission lines offers significant advantages in improving power supply reliability and system stability. Statistics show that 70%–80% of all line faults are single-phase grounding faults. When both ends of the protected line have power sources, the influence of the transition resistance will produce a capacitive or inductive offset component in the measured impedance, causing the protection range to expand or shrink. This offset component is related not only to the inherent characteristic parameters of the line but also to the potential angle between the equivalent power sources at both ends of the line [1, 2, 3]. Utilizing a BP neural network to construct an impedance relay, enabling distance protection to adapt reliably to changes in system and fault conditions, is a worthwhile area of ​​exploration [4, 5]. This paper uses the fundamental amplitudes of the three-phase voltage and current after the fault, along with their phase difference, as inputs to the relay fault phase selection sub-neural network to determine the phase of a single-phase grounding fault. Based on the fault phase determination results, the voltage and current fundamental amplitudes of the fault phase, their phase difference, zero-sequence voltage and zero-sequence current, and the potential angle between the power supplies at both ends of the line are used as inputs to the relay fault location and ranging sub-neural network to achieve accurate fault location and ranging. Transient simulation of a given power network using EMTP confirms the feasibility of the proposed design. 2. Neural Network-Based Single-Phase Grounding Impedance Relay 2.1 Structure of the Neural Network-Based Single-Phase Grounding Impedance Relay The proposed design does not perform impedance calculation; instead, it uses a neural network to identify fault states from combinations of various system parameters. Considering that the proposed design uses the voltage and current during the fault and their phase difference as inputs, and can output the distance from the fault point to the protection installation location, the term "impedance relay" in conventional distance protection is still used. The neural network-based single-phase grounding impedance relay consists of two parts: a fault phase selection sub-network (NN1) and a fault location and ranging sub-network (NN2), as shown in Figure 1. ● Fault Selection Phase Subnetwork (NN1) [img=348,199]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/11-3.jpg[/img] [img=348,198]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/11-1.jpg[/img] The neural network model used in this paper is a multi-layer feedforward neural network, employing an improved BP method with variable learning rate and momentum term addition. The transfer function of each layer in the network is the logarithmic sigmoid function. The selection of the number of hidden layers and the number of neurons in each hidden layer is crucial to the successful training and generalization performance of the neural network. Through extensive experiments, it was determined that the number of neurons in the first and second hidden layers of the fault phase selection subnetwork is 42, and the number of neurons in the first and second hidden layers of the fault location and ranging subnetwork is 36. The experiments showed that choosing two hidden layers makes the neural network easier to train and its generalization ability is also very good. 2.2 Ideal Output Characteristics of a Single-Phase Grounding Impedance Relay Based on a Neural Network The ideal output characteristics of the fault phase selection are: output 1 when t1 (t2, t3) is the fault phase, and output 0 when it is not a fault phase. When 0 ≤ t1 (t2, t3) < 0.5, it is a non-fault phase; when 0.5 ≤ t1 (t2, t3) ≤ 1, it is a fault phase. For fault location, the setting range is designed to be 90% of the total length of the protected line. When 0 ≤ a1 < 0.5, it is an external or reverse fault; when 0.5 ≤ a1 ≤ 1, it is an internal fault. For fault location, when 0 ≤ a2 < 0.09, it is a reverse fault; when 0.09 ≤ a2 ≤ 0.79, the fault distance is (lBC·a2－0.09)/0.70km; when 0.79 ≤ a2 ≤ 1.00, the fault distance is (lBC＋lCD·a2－0.79)/0.70km. The ideal output characteristics of fault location and fault distance are shown in Figure 2. [img=376,275]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/12-1.jpg[/img] Notice: Fault location x refers to the percentage of the total line length between the fault point F and the protection installation point B. When -30≤x＜0, the total length of the line refers to line AB; when 0≤x≤100, the total length of the line refers to line BC; when 100＜x≤130, the total length of the line refers to line CD. [b]3 Training and Testing of Single-Phase Grounding Impedance Relay Based on Neural Network[/b] The fault samples in this paper are from the EMTP simulation of single-phase grounding fault in the system shown in Figure 3. The sampling frequency is 1kHz, and the fault characteristic quantities are calculated using the full-wave Fourier algorithm. A total of 1440 training samples and two sets of test samples (840 samples in the first set and 360 samples in the second set) were made. Before training with the neural network, the original data of the samples were normalized to facilitate the training of the neural network. Referring to reference [4], the most severe fault current value at the protection installation location was taken as the reference value of the current, and the rated voltage was taken as the reference value of the voltage. The phase difference φ was divided by π. The potential angle Δδ is set to 0 when it is -5 degrees and 1 when it is 5 degrees. For values ​​between -5 and +5 degrees, it is linearly assigned, such as 0.25 when Δδ is -2.5 degrees and 0.85 when Δδ is +3.5 degrees. 3.1 Training and Testing of the Fault Phase Selection Sub-Network (NN1) The error characteristics between the actual output and ideal output of the fault phase selection sub-network for training samples are shown in Figure 4(a), the error characteristics between the actual output and ideal output for the first set of test samples are shown in Figure 4(b), and the error characteristics between the actual output and ideal output for the second set of test samples are shown in Figure 4(c). [img=393,135]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/12-2.jpg[/img] From Figure 4, it can be seen that due to the significant differences in voltage, current amplitude, and phase difference between the faulty and non-faulty phases, the correct operation rate of the neural network relay fault phase selection is 100% for 1440 training samples and 1200 test samples, and the operation performance is good. 3.2 Training and Testing of Fault Location and Fault Tolerance Sub-network (NN2) 3.2.1 Output Characteristics of Fault Location and Fault Tolerance Sub-network for Training Samples For the trained neural network, Figures 5(a), (b), and (c) are generated for the operation characteristics of 1440 training samples under three system operating conditions (potential angles Δδ of 5, 0, and -5 degrees) and eight transition resistances to reflect the training effect of the neural network. As can be seen from Figures 5(a), (b), and (c), the fault location and fault ranging output characteristics of the neural network relay for the training samples are very close to the ideal operating characteristics. This indicates that the neural network has correctly distinguished the training samples and obtained the ideal output characteristics through learning from the samples. The operating characteristic diagrams of the training samples also show that, under various system operating conditions and transition resistance conditions, the neural network relay can reliably not operate for reverse faults and faults outside the protection zone. For faults on the local line, its protection range can reach 90% of the total length of the line, thus achieving both the designed protection range and meeting the selectivity requirements. [img=325,336]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/12-3.jpg[/img] 3.2.2 Test Results of the Fault Location and Fault Ranging Subnetwork on the First Set of Test Samples Based on two system operating conditions different from the training samples (potential angle Δδ of 2.5 and -2.5 degrees respectively) and seven transition resistances (Rg of 2.5, 7.5, 15, 27.5, 42.5, 60, and 85 Ω respectively), a second set of 840 test samples was generated at the same fault location as the training samples (see Figure 2) to test the generalization performance of the neural network. Figures 6(a) and (b) show the test results of the fault location and ranging subnetwork on the first set of test samples. [img=310,255]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/13-1.jpg[/img][img=310,258]http://zszl.cepee.com/cepee_k jlw_pic/files/wx/jdq/2002-1/13-2.jpg[/img][img=310,252]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/13-3.jpg[/img] Figure 6 shows the output characteristics of the fault location and ranging subnetwork for the first set of test samples. From Figures 6(a) and (b), it can be seen that, with the same fault point locations as the training samples, although the test samples were made under different system operating conditions and transition resistance conditions than the training samples, the fault location and fault ranging output characteristics of the neural network relay for the test samples are still close to the ideal operating characteristics, indicating that the neural network has good generalization performance. From the operating characteristic diagram of the test samples, it can be seen that under various system operating conditions and transition resistance conditions different from the training samples, the neural network relay can reliably not operate for reverse faults and faults outside the protection zone. For faults on the local line, its protection range can reach about 90% of the total length of the line, exhibiting good selectivity and directionality. [img=365,261]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/13-4.jpg[/img] [img=365,381]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/13-5.jpg[/img] [img=329,129]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/14-1.jpg[/img][img=329,253]http://zszl.cepee.com/cepee_kjlw_pic/files/wx/jdq/2002-1/14-2.jpg[/img] 3.2.3 Fault Location and Fault Tolerance Sub-network's Verification Results for the Second Set of Test Samples A second set of 360 test samples were created based on four system operating conditions (potential angles Δδ of 3.5, 1.5, -1.5, and -3.5 degrees respectively), five transition resistances Rg of 0.5, 25, 45, 65, and 90Ω respectively), and the location of the fault point, which differed from the training samples and the first set of test samples. Among them, samples 1-60# are reverse direction faults, samples 61-300# are in-zone faults, and samples 301-360# are out-of-zone faults. Figure 7 shows the test results of the fault location and ranging sub-network for the second set of test samples. As can be seen from Figure 7, this neural network relay can adapt to changes in system operating conditions and transition resistance, has good directionality and operating performance, and achieves 100% accuracy in locating various faults such as reverse direction faults (at -03), faults near the protection device (at 03), faults at the end of the protection range (at 85 and 88), and out-of-zone faults (at 101). Its protection range can reach 88% of the total length of the line. The error of the ranging output is within ±0.1, that is, within 15% of the total line length. 4. Analysis and Conclusion The well-trained neural network impedance relay can not only adapt to changes in the 0-100Ω transition resistance, but also to changes in system operating conditions (the angle Δδ between the two system potentials varies between -5 degrees and +5 degrees). Among them, the accuracy of fault phase selection is 100%, the protection range of fault location can reach 88% of the total line length, and the directionality and selectivity are good. In addition, the neural network impedance relay also has a certain fault location function. [b]References[/b] ... IEEE Transaction on Power Delivery, 1994, 9(1): 480-491. 〔4〕 Duan Yuqian, He Jiali. Distance protection based on artificial neural network 〔J〕. Proceedings of the CSEE, 1999, 19(5): 67-70. 〔5〕 Wang Xiaoru, Qian Qingquan, Wu Sitao. Research on adaptive distance protection based on neural network 〔J〕. Automation of Electric Power Systems, 1998, 22(12): 27-30.