Fault TroubleShooter (FaultTS)

There are currently several fault monitoring systems in the LV network. However, their cost is often high and their installation/operation is not always straightforward, making it difficult to obtain information in a short time span, impacting the decision time. Thus, to alleviate this issue, we developed a sensor that gathers the necessary data and sends it to a server so that our web application is able to calculate the impedance and distance to the fault.

The purpose of this work is to take advantage of the data given by the sensor to classify a fault into 5 categories: LG, LL, LLG, Line-Line-Line (LLL), Line-Line-Line-Ground (LLLG) and, with this information, obtain the impedance and distance to the fault, thus decreasing the time needed to pinpoint a fault and correct the issue. This is paramount to Distribution System Operator (DSO)’s as they may be penalized by the regulator if their Key Performance Indicator (KPI)’s deteriorate, being the time the service is down the most important one.

There are currently several methods to detect and locate faults. In the present work we focus on impedance based methods, which require the voltage and current phasors before and during the fault. These methods can be broadly divided in two groups: single-ended and double-ended methods. In the former, the currents and voltages are measured at one end of the cable and, in the latter, these variables are measured at both ends.

In order to assess our method’s performance, we carried out experiments at the PNDC, in which controlled short-circuits were induced in PNDC’s LV network. The experiments consisted in inducing LG, LL, LLG, LLL and LLLG faults in two different positions for 5 different resistances (0 , 0.75 , 5 , 10 ,20 ) inatotal of 50 experiments.

Impedance methods assume a clear distinction between the current before and after the fault. Asthefault resistance increases, both become increasingly similar, which prevents any impedance based algorithm from giving an accurate distance estimation. Thus, a total of 20 experiments were considered: 10 corresponding to a fault resistance of 0 and 10 corresponding to a fault resistance of 0.75 .

We used the relative error to assess the algorithm’s performance. Wewanted to assess the following metrics:

• M1: Percentage of fault events detected;
• M2: Percentage of fuse blown events detected;
• M3: Percentage of faults correctly classified;
• M4: Impedance relative error;
• M5: Distance relative error.

Considering metrics M1 and M2, the algorithm was able to detect 100% of the fault events (M1) and 100% of the fuse blown events (M2). This includes all the fault resistances mentioned above, from 0 to 20 .

Concerning metrics M3-M5, these were only defined for fault resistances of 0 and 0.75 and do not apply to the 5 , 10 and 20 fault tests. This is due to the fact that the increase in the fault resistance leads to current values in the same order of magnitude before and during the fault, which is highly detrimental to the performance of impedance-based methods. However, these events were detected by the Eneida DTVI and the waveform data was gathered for future analysis, as we plan to cover these higher impedance faults with a different approach.

For the fault resistances considered (0 and 0.75 ), the algorithm correctly classified 100% of the faults (M3).

Considering M4, we obtained 7/10 (70%) events with a relative error lower than the maximum admissible relative error defined (15%) for a fault resistance of 0 , the median relative error was 11.37% and the maximum relative error was 27.51%. For a fault resistance of 0.75 all the events were below the maximum admissible relative error defined (50%), the median of the relative error was 2.49% and the maximum relative error was 6.14%.

As for M5, we obtained 6/10 (60%) events with a relative error lower than the maximum ad missible relative error defined (15%) for a fault resistance of 0 , the median relative error was 11.35% and the maximum relative error was 21.81%. For a fault resistance of 0.75 9/10 (90%) events presented a relative error lower than the maximum admissible relative er ror defined (50%), the median relative error was 20.88%, and the maximum relative error was 53.62%.

It was also within the scope of this work to assess whether the defined acceptance criteria were appropriate to evaluate the performance. The results indicate that while metrics M1-M3 were well defined, metrics M4 and M5 were not adequate to obtain a clear understanding of the performance of the algorithms. The success criteria defined for distance (M5) were set as a relative error, this implies that the absolute error increases with the distance to the fault, which was not confirmed by the test results. For 0 faults, 9 out of 10 tests present an absolute error of less than 100m. For 0.75 faults, 8 out of 10 tests have an absolute error under 200m. In both cases, an outlier was registered for a distance of 897.5m to the fault. For these outliers, at 0 fault impedance, the absolute error was in the order of 200m, while for 0.75 it increased to around 300m.