Breakthrough for power distribution cable diagnosis using Artificial Intelligence

By Dr Ke ZHU, Dr Genghong LU, Mr HO Nam Yim, Mr Chao LEI, Mr WONG Ming Kwong, Mr IP Chi Lam, Mr PUI Yu Lin, and Dr Siqi BU

Smart power distribution cable diagnosis with AI

The Hongkong Electric Co Ltd (HK Electric) has been providing Hong Kong with one of the most reliable supplies of electricity in the world (viz., reliability rating higher than 99.999% since 1997). Such success is inseparable from its development and application of innovative technologies to enhance operational performance. At HK Electric, we encourage and support all kinds of innovations, either incremental or genuinely transformative. Innovation is essential to a sustainable business, and thus, promoting innovation and developing innovative capabilities are integral parts of our corporate strategy and management principles. HK Electric has pooled these different experiences into the HK Electric Innovation Report 2023 and uploaded it to the company website for reference and knowledge transfer (see Figure 1).

In what follows, we will introduce one example of the innovation projects on power cable diagnosis at HK Electric. In 2022, HK Electric began its collaboration with the Centre for Advances in Reliability and Safety Ltd (CAiRS), an InnoHK research centre affiliated with The Hong Kong Polytechnic University, to explore the feasibility of diagnosing power cables using Artificial Intelligence (AI) method. This represents a significant breakthrough in the condition assessment of power distribution cables in electrical engineering.

Diagnosing power cables with AI method represents a significant breakthrough in the condition assessment of power distribution cables in electrical engineering.

Figure 1: HK Electric Innovation Report 2023, available for reference at https://www.hkelectric.com/zh/sustainability/running-a-sustainable-business/innovation

Power distribution cable diagnosis

HK Electric has been in continual operation since 1889 to provide electricity in Hong Kong. Electricity is generated at Lamma Power Station, then transmitted through transmission (275 kV and 132 kV) and distribution networks (22 kV and 11 kV) to various customers. The transmission and distribution network of HK Electric mainly uses underground power cables, providing system protection against extreme weather conditions such as typhoons and proving ideal for a densely loaded city like Hong Kong (see Figure 2). Though the design life of these cables can exceed 50 years, the actual performance of the cables will be seriously affected if defective components, such as cable sections or joints with water-tree growth, are not removed promptly⁶. Hence, continuous condition assessment is significant and conducted regularly for power cables to ensure network reliability and safety.

In order to evaluate the robustness of the power distribution cables of HK Electric, the integrity of the cable insulation is measured by Very Low Frequency (VLF) testing⁵. This measurement method, supported by IEEE 400.2, was introduced in around 2010. It treats the cable circuit as a simplified equivalent circuit consisting of a resistor and a capacitor, with the tan-delta (TD) value defined as the ratio between the resistive current and the capacitive current (see I_R and I_C in Figure 3). This value represents the dielectric loss, which increases if the cable degrades over prolonged service life or deteriorates due to water tree(s). In HK Electric’s practice, the mean TD values of three phases (L1, L2, and L3) are tested at 0.5, 1.0, and 1.5U₀ (see Figure 3), and then classified under a 5-level Health Index (HI) system. An HI of 1 indicates that the cable is relatively safe for continual operation, while an HI of 5 indicates that the cable insulation is severely degraded and requires prompt attention. The corresponding maintenance plan is scheduled based on different HI (see Table I). About 5,000 cables have been tested, either proactively or after cable work in the past ten years. Among them, about 5% of 11-kV power distribution cables were identified with severe conditions (namely, HI = 5), and they have been preemptively rescued from in-service failure.

Table I: Evaluation on cable circuit comprised of XLPE cable section (≦1,000m)

Figure 2: Three-phase underground power distribution cables deployed at HK Electric

The merits of this successful method are summarised as follows. Firstly, the measurement results are presented as simple numerical values, to which onsite engineers can refer immediately. Secondly, the measurements from a given phase can be compared with those from adjacent phases to evaluate the overall cable condition. Meanwhile, through periodic testing, a comparison can be made with past records, aiding in the identification of degradation trends in the cable condition over time. Thirdly, there is a low risk of test failure, since the over-rated voltage measurement at 1.5U₀ is merely transient. Lastly, the measurements remain stable and are not influenced by external electromagnetic fields at substations. However, this method has been ascertained to still be inadequate by considerations involving the following aspects⁷. Firstly, the category and corresponding threshold of HI, as established by IEEE 400.2, do not adequately reflect the nature of the network at HK Electric. The cable failure risk is roughly classified into five levels instead of being accurately quantified based on the specifics of HK Electric’s network. Secondly, some cables were found to have failed shortly after the cable test, despite being identified with a low HI index in the practical operation. This might be because the condition evaluation process is only interpreted through mean TD values and absolute partial discharge values. According to the latest publication entitled Cable Diagnostic Focused Initiative (CDFI) released by Georgia Institute of Technology, an expansion of criteria (namely, using differential TD under excitation voltage and the change of TD with time) is introduced to provide more information about cable systems⁴. Lastly, while the HI can only evaluate the overall cable insulation level, it cannot pinpoint which specific components are at risk, even if the cable is identified with a poor HI. Therefore, HK Electric has collaborated with CAiRS to improve this methodology over the past two years.

The collaboration commenced in January 2022 and was completed in June 2023. During this period, the following work was accomplished: (i) Inspection of the sample of faulty power distribution cables at HK Electric (see Figure 4); (ii) Providing CAiRS with data of power distribution cables for analysis; (iii) Theoretical study and application of simulation software (e.g., COMSOL) for analysis of water-tree growth; (iv) Development of AI-based software for power cable diagnosis; and (v) Conducting a blind test to verify the effectiveness of the model.

Figure 3: Equivalent model for TD testing and an example of measurement at 0.5/1.0/1.5 U₀.

Figure 4: Workshop visit from CAiRS to inspect faulty power distribution cables at HK Electric

Cable diagnosis aided by Artificial Intelligence (ensemble learning method)

Existing power distribution cable diagnosis relies mainly on the formation and physical appearance of the water tree (e.g., vented or bow-tie tree, and its length), and on diagnosis experiences accumulated by engineers. While this methodology yields reliable diagnostic results, it requires considerable engineering resources. To address this challenge, an AI-based cable diagnosis method has been developed, leveraging insight gained from comprehensive field data. As shown in Figure 5, AI development has undergone three surges, culminating in the creation and application of various AI technologies across different industries, improving human productivity and quality of life³. With powerful computing power and exceptional learning capabilities, AI can perform predictions simply by learning from vast amounts of data without the need to understand complex physical principles. The AI model is trained and validated using the training data. Once the AI model is trained, the corresponding AI algorithm will use new testing data as inputs to evaluate cable condition.

With powerful computing power and exceptional learning capabilities, AI can perform predictions simply by learning from vast amounts of data without the need to understand complex physical principles.

Figure 5: Breakthrough of AI development and its application for power distribution cable diagnosis

The following section introduces a novel power cable condition assessment model based on AI, namely, ensemble learning method. Ensemble learning is the technique that combines multiple individual models (also called base or weak learners) to improve the overall performance and reduce the variance and bias of the model. By combining multiple models trained on different subsets of the data or using different algorithms, ensemble learning-based cable diagnosis can produce more accurate and robust prediction results than any individual model (e.g., support vector machine and naïve Bayes). The objective of developing this AI-based cable diagnosis method is to assist system operator in further evaluating the probability of cable failure and predicting the cause of such failure.

A. Model input

In the construction of AI model, the input parameters play a critical role. In addition to mean TD values as stated in the HI, the change of TD with time or voltage has been revealed by CDFI to be able to provide additional hints about the evaluation of cable condition. Therefore, the following three parameters are investigated and used in the AI model (see Figure 6):

(i) Tan-delta stability (STD): This parameter refers to the time dependence of TD. A series of TD values are measured at U₀ (a set of 8 measurement is adopted in the practice of HK Electric). The standard deviation (STD) of these measurements is then calculated. Generally, the measured TD should remain stable for a healthy cable unless there is a vibration of water molecules, which can alter the conduction properties of the insulation layer over time. This variability is why STD can serve as a reference for evaluating water-tree growth in the cable. According to IEEE Std 400.2™-2013 and CDFI, a cable is considered defective if a large STD is observed.

(ii) Difference of tan-delta (DTD): This parameter represents the difference between the mean TD values measured at 0.5 and 1.5U₀. Similarly, a cable is deemed weak if the DTD is large, which indicates that the insulation losses increase more dramatically as the test voltage goes up.

(iii) R value: This parameter is calculated as a ratio of the highest mean TD at 1 U₀ (1.5 U₀) of one phase to the lowest mean TD at 1U₀ (1.5U₀) of another one among three phases. This ratio denotes the insulation discrepancy among the three phases. Generally, the mean TD of the three phases should be roughly similar when tested under the same applied voltage. If a large R-value is observed, the insulation degree of one phase is remarkably worse than the others, which is uncommon for a three-phase power cable. Therefore, this value is also adopted as the input of the AI model.

In conclusion, these three metrics (namely, STD, DTD, and R), obtained from very-low frequency TD testing, are compiled as the model inputs for further AI model construction.

Figure 6: Input parameters for AI model for evaluating cable condition

B. Model construction

The developed AI model, consisting of two sub-modules, is presented in Figure 7 (blue part). One sub-module can provide a probability of cable failure (PoF), while the other can predict the risky component (namely, cable body or joint) if the cable is identified with a high PoF. Further details are illustrated below.

Figure 7: Model structure of the ensemble learning-based cable condition monitoring model (with the blue part indicating the model training process and the orange part being the testing)

(i) Probability of failure (PoF) evaluation: the goal of PoF evaluation model is to estimate the PoF of cables based on the measurement of mean TD, STD, and DTD.

In practice, there is only a small part of cables with a failure history. Therefore, the number of normal samples outweighs that of failed ones, which results in data imbalance during AI model training. This imbalance can lead to degraded performance of the trained model. To solve this problem, the Synthetic Minority Over-sampling Technique (SMOTE) has been used for data pre-processing¹. SMOTE operates by selecting examples that are close to the feature space, drawing a line between them and the examples in the feature space, and then drawing a new sample at a point along that line. The objective of SMOTE is to create as many synthetic samples as possible for the minority category (namely, samples with a history of cable failure), thus balancing the number of normal and faulty samples.

Extreme gradient boosting (XGBoost), a supervised learning algorithm, has been adopted in the PoF model². In this algorithm, decision trees are created in sequential form. Weights play an essential role in XGBoost. Weights are assigned to all the independent variables, which are then fed into the decision tree that predicts results. The weights of variables predicted incorrectly by the tree are increased, and these variables are then fed to the second decision tree. These individual classifiers/predictors then ensemble to give a strong and more precise model. Consequently, by combining the estimates of a set of simpler, weaker models, the XGBoost has a greater generalisability to different cable conditions and has been proven more accurate compared to other traditional machine learning methods.

The model output of XGBoost is the PoF. Two probabilities are given to indicate the probability of the cable being normal P_norm and the probability of the cable failing P_fail. The summation of these two probabilities is equal to 1. The higher P_fail is, the more likely the cable can fail during the operation.

(ii) Probability of fault component (PoFC) prediction: The PoFC prediction model is developed to estimate the component most likely to lead to a high PoF of the cable. In this sub-model, the R-value is used as an input for the AI model. To improve the efficiency of the model, data standardisation and principal component analysis (PCA) methods are used in the data pre-processing. Data standardisation is the process of converting data to a common format, while PCA is a statistical technique used for analysing datasets and increasing the interpretability of data. It linearly transforms the data into a new coordinate system where the data can be described with fewer dimensions.

For the PoFC model, gradient-boosted decision trees (GBDT) have been trained to learn the nonlinear relationship between the TD features and the fault component. GBDT is an iterative decision tree model that utilises the additive algorithm for prediction, continuously reducing residuals generated in the training process. GBDT uses the forward distribution algorithm and selects the classification and regression tree (CART) learner as a weak base learner. During training, each iteration of the decision tree involves adjusting the values of the coefficients and weights of the input variables to minimise the loss function (the measurement of the difference between the predicted and actual target values). GBDT generates numerous weak learners through multiple iterations, with each learner trained on the basis of the residual of the previous learner. Finally, it integrates these multiple weak learners into a single strong learner by weighing the summation of each tree.

The model output of GBDT is the PoFC. Two probabilities are given to indicate the probability of the fault component being cable body P_body and the probability of the fault component being cable joint P_joint. The summation of these two probabilities is equal to 1. The higher P_body is, the more likely the cable body is the defect location.

C. Model validation

For convenience, a power distribution cable diagnosis system has been developed, as shown in Figure 8. The system integrates the trained AI models and can provide PoF prediction and PoFC prediction functions. By selecting the specific cable type (namely, XLPE-XLPE or XLPE-PILC cable), the AI model task (namely, model training or testing), and the diagnosis task (namely, cable risk evaluation or fault component prediction), the system can conduct cable diagnosis based on the input data. The final prediction results are visualised with two pie charts, with different colors in them indicating different cable conditions and fault components.

In order to evaluate the efficiency of the developed AI model, a blind test was conducted in which CAiRS was provided with the latest testing records of cables from HK Electric. The data was processed according to the steps outlined in the orange part of Figure 7 to calculate the PoF and PoFC of each cable based on VLF testing results. The findings and merits of this developed AI model are summarized below:

(i) Quantified risk of cable failure: With this method, each cable now has a specific failure probability instead of a rough range classified by 5-level HI. This will assist HK Electric in effectively prioritising cables maintenance.

(ii) Enhanced accuracy of cable assessment: By incorporating DTD, STD, and R as AI model inputs in addition to the mean TD value, this model is revealed to be able to identify some cables that are at risks, even if they are marked as HI ≠ 5. A list of examples is shown in Table II (on the left side), where these cables failed in service despite not showing severe conditions during the previous cable inspection. These cables were identified with low failure risks by HI but were flagged by the AI model (P_fail > 0.50). This result will assist HK Electric in identifying and investigating risky cables for further action if necessary.

(iii) Capability of fault component prediction: In addition to issuing alarms (viz., high P_fail) during cable condition assessment, the system can also predict the risky component leading to high P_fail. On the right side of Table II, a list of examples is provided where the risky components were identified for cables flagged with high risks. Furthermore, it was found that these predicted components matched the faulty ones recorded by HK Electric.

Figure 8: Illustration of the AI-based power distribution cable diagnosis system

Table II: Application of the developed AI Model in power cable diagnosis

Conclusion

In this article, we have introduced a new methodology for the condition assessment of power distribution cables using an AI model. The merits of this method were presented, and some validated results have been presented to indicate its feasibility. The improved inspection capability can further enhance the reliability and security of power delivery in HK Electric.

Last but not least, this is also an excellent showcase for the collaboration between industry and academia (see Figures 9 and 10). HK Electric identifies the pain point in its business operation and offers domain knowledge of cable diagnosis to CAiRS while working on the project. Meanwhile, CAiRS contributes its expertise in AI to mitigate safety and reliability risks for industrial partners.

Figure 9: Video production and promotion of collaboration, available at https://www.cairs.hk/en/media_detail/index/39

Figure 10: Management-level visit and project discussion between HK Electric and CAiRS

Acknowledgements

The authors would like to acknowledge, with gratitude, the permission granted by the Management to publish this article. This article is also supported by Centre for Advances in Reliability and Safety (CAiRS), an InnoHK Research Cluster of the HKSAR Government.

About the authors

Dr Ke Zhu, Mr Wong Ming Kwong, and Mr Ip Chi Lam are affiliated with The Hongkong Electric Co Ltd. Dr Genghong Lu, Mr Ho Nam Yim, Mr Chao Lei, and Mr Pui Yu Lin are affiliated with the Centre for Advances in Reliability and Safety. Dr Siqi Bu is affiliated with the Department of Electrical and Electronic Engineering, The Hong Kong Polytechnic University, and also with the Centre for Advances in Reliability and Safety. All authors have contributed equally.

References

1. Chawla N V, Bowyer K W, Hall L O, and Kegelmeyer W P (2002). ‘SMOTE: Synthetic Minority Over-sampling Technique’. Journal of Artificial Intelligence Research, 16, pp.321-357. doi: https://doi.org/10.1613/jair.953.
2. Chen T, and Guestrin C (2016). XGBoost: A Scalable Tree Boosting System, Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. San Francisco, California, USA. 13 - 17 August. New York: Association for Computing Machinery. pp.785-794.
3. Furse C M, Kafal M, Razzaghi R, and Shin Y J (2020). ‘Fault Diagnosis for Electrical Systems and Power Networks: A Review’. IEEE Sensors Journal, 21(2), pp.888-906. doi: https://doi.org/10.1109/JSEN.2020.2987321.
4. Hernandez-Mejia J C (2016). Dissipation Factor (Tan δ). Report by National Electric Energy Testing, Research and Applications Centre of Georgia Tech Research Corporation.
5. Institute of Electrical and Electronics Engineers (2013). IEEE Std 400.2- 2013: IEEE Guide for Field Testing of Shielded Power Cable Systems Using Very Low Frequency (VLF)(less than 1 Hz). doi: https://doi.org/10.1109/IEEESTD.2013.6517854.
6. Zhu K, Lee W K and Pong W T, (2017). ‘Energization-Status Identification of Three-Phase Three-Core Shielded Distribution Power Cables Based on Non-Destructive Magnetic Field Sensing’. IEEE Sensors Journal, 17(22), pp. 7405-7417. doi: http://dx.doi.org/10.1109/JSEN.2017.2756905.
7. Zhu K and Ng T F (2022). ‘Improvement of cable defect identification for power distribution networks’. HKIE Transactions, 29 (1), pp. 37-45. doi: https://doi.org/10.33430/V29N1THIE-2021-0016.