Xiaosheng Peng

Application of K-Means method to pattern recognition in on-line cable partial discharge monitoring

On-line Partial Discharge (PD) monitoring is being increasingly adopted in an effort to improve asset management of the vast network of MV and HV power cables. This paper presents a novel method for autonomous recognition of PD patterns recorded under conditions in which a phase-reference voltage waveform from the HV conductors is not available, as is often the case in on-line PD based insulation condition monitoring. The paper begins with an analysis of two significant challenges for automatic PD pattern recognition. A methodology is then proposed for applying the K-Means method to the task of recognizing PD patterns without phase reference information. Results are presented to show that the proposed methodology is capable of recognising patterns of PD activity in on-line monitoring applications for both single-phase and three-phase cables and is also effective technique for rejecting interference signals.

Denoising and feature extraction in PD based cable condition monitoring systems

There has been increasing application of on-line partial discharge (PD) based cable insulation condition monitoring among utilities worldwide due to the ability of on-line PD monitoring to allow incipient insulation faults to be detected and aged cable replacement program to be prioritised. However the application is also accompanied with a number of challenges. Data from on-line PD monitoring systems shows presence of higher levels of interference, including sinusoidal RF noise, switching pulses, PD from local plant, radio and power line carrier communication systems, etc. The biggest challenge associated with on-line cable PD monitoring is to distinguish PD generated in cable insulation from noisy raw data, which requires not only application of data denoising techniques but also feature extraction techniques to differentiate signals coming from different sources based on their characteristics. This paper aims to overcome the above-mentioned challenge. Following a brief introduction the paper introduces an effective denoising technique involving the adaptive second generation wavelet transform (ASGWT). To describe the various PD pulses, which the authors have observed from on-line cable PD monitoring systems, methods for PD feature extraction are discussed. These include analysis of raw PD signal, phase-resolved PD pattern, etc. Finally, based on data denoising and feature extraction, signal classification for an on-site PD testing experiment is introduced.

Rough set theory applied to pattern recognition of Partial Discharge in noise affected cable data

This paper presents an effective, Rough Set (RS) based, pattern recognition method for rejecting interference signals and recognising Partial Discharge (PD) signals from different sources. Firstly, RS theory is presented in terms of Information System, Lower and Upper Approximation, Signal Discretisation, Attribute Reduction and a flowchart of the RS based pattern recognition method. Secondly, PD testing of five types of artificial defect in ethylene-propylene rubber (EPR) cable is carried out and data pre-processing and feature extraction are employed to separate PD and interference signals. Thirdly, the RS based PD signal recognition method is applied to 4000 samples and is proven to have 99% accuracy. Fourthly, the RS based PD recognition method is applied to signals from five different sources and an accuracy of more than 93% is attained when a combination of signal discretisation and attribute reduction methods are applied. Finally, Back-propagation Neural Network (BPNN) and Support Vector Machine (SVM) methods are studied and compared with the developed method. The proposed RS method is proven to have higher accuracy than SVM and BPNN and can be applied for on-line PD monitoring of cable systems after training with valid sample data.

SDMF based interference rejection and PD interpretation for simulated defects in HV cable diagnostics

Partial Discharge (PD) in cable systems causes deterioration and failure, identifying the presence of PD is crucial to Asset Management. This paper presents methods for interference signals rejection and for PD interpretation for five types of artificial defect in 11 kV ethylene-propylene rubber (EPR) cable. Firstly, the physical parameters of the artificial defects used for PD signal generation are introduced. Thereafter, the sample stress regime, PD testing and detection systems, including IEC 60270 measurement system and High Frequency Current Transformer (HFCT), are outlined. Following on, a novel Synchronous Detection and Multi-information Fusion (SDMF) based signal identification method is developed, to separate PD and interference signals within raw data. Finally, a comparative PD analysis of two detection systems is carried out and several characteristics of insulation related PD signals of EPR cables are reported. The SDMF based data pre-processing and the comparative PD activity analysis contribute to improvement of PD pattern recognition and assist in on-line PD monitoring systems.

Transfer function characterization for HFCTs used in partial discharge detection

High frequency current transformers (HFCTs) are widely employed to detect partial discharge (PD) induced currents in high voltage equipment. This paper describes measurements of the wideband transfer functions of HFCTs so that their influence on the detected pulse shape in advanced PD measurement applications can be characterized. The time-domain method based on the pulse response is a useful way to represent HFCT transfer functions as it allows numerical determination of the forward and reverse transfer functions of the sensor. However, while the method is accurate at high frequencies it can have limited resolution at low frequencies. In this paper, a composite time-domain method is presented to allow accurate characterization of the HFCT transfer functions at both low and high frequencies. The composite method was tested on two different HFCTs and the results indicate that the method can characterize their transfer functions ranging from several kHz to tens of MHz. Results are found to be in good agreement with frequency-domain measurements up to 50 MHz. Measurement procedures for using the method are summarized to facilitate further applications.

A successful on-site PD testing experience of 11kV EPR cable insulation systems

A power generating station in the United Kingdom has reported a number of in service failures in its 11kV single core Ethylene-Propylene Rubber (EPR) insulated cables. Several industrial companies have carried out on-site condition assessment to determine whether other insulation defects were present but no conclusive results have been found due to presence of strong background electrical noise. The present authors were invited to carry out on-site testing to demonstrate the effectiveness of their denoising techniques. This paper presents the processes of the on-site cable partial discharge signal detection experience and the signal processing of the raw data. Following a brief introduction to the tests, equipments and connections, the paper analyses sources of different types of interference signals. These are found to originate mainly from UPS Inverter Supplies or 11kV Motor drive circuits. Thereafter, second generation wavelet transform (SGWT) data denoising algorithm is introduced. SGWT is proved to be an effective denoising technique for the detected data. Also presented in the paper are PD pattern identification and PD source localization methods which are used to identify the source of the PD signal. Finally the diagnosis results, with indication of potential insulation defect and cable joint problems, are provided.

A very short term wind power prediction approach based on Multilayer Restricted Boltzmann Machine

The wind power prediction (WPP) is challenging as a large amount of data with complex nonlinear relationship should be fitted by the prediction method. To improve the accuracy, WPP based on the Multilayer Restricted Boltzmann Machine (MRBM), which is a deep learning neural network with strong feature interpretation ability, is presented in the paper. To explore the influencing factors of prediction accuracy, the number of hidden layers and the number of nodes in each layer of MRBM are studied. Furthermore, the classic Back Propagation Neural Network (BPNN) based WPP, as a reference, is compared with the MRBM method. The results show that the accuracy of MRBM based WPP is higher than that of BPNN based WPP. The Root Mean Square Error (RMSE) of the MRBM based prediction is 4.5% lower than that of BPNN in some period, and the error distribution of MRBM based WPP is with better concentration ability than that of BPNN based WPP.

A very short term wind power forecasting approach based on numerical weather prediction and error correction method

The rapid development of wind power brings great stability challenge to power system due to the randomicity and uncertainty of wind power output. One effective way to overcome the challenge is wind power forecasting. A novel very short-term wind power forecasting approach based on numerical weather prediction and error correction method is presented in the paper. From the statistical analysis of the error probability density distribution of Back Propagation Neural Network (BPNN) and Support Vector Machine (SVM) based wind power prediction, an error estimation model is presented and an error correction method is proposed. The key step of correction method is calculating the corresponding prediction uncertainties according to the predict value. This method combines the advantages of both error estimation model and error probability density distribution, which is evaluated with the data from 7 wind farms and numerical weather prediction, and is proved to be an effective way to improve the accuracy of very short term wind power forecasting.

Optimal feature selection for partial discharge recognition of cable systems based on the random forest method

Optimal feature selection is one of the most significant challenges of Partial Discharge (PD) pattern recognition of cable system as the number of PD features is with large quantity, which reduces the efficiency of PD pattern recognition methods and restricts the effectiveness of PD based condition monitoring and diagnostics. To overcome the challenge a random forest method is presented in the paper to select the most effective PD features from 18 different kinds of features. Firstly, PD data from five types of artificial defects based on IEC 60270 system in High Voltage (HV) lab is introduced. Secondly, PD data pre-processing and feature extraction are carried out and 18 kinds of PD features are extracted from the raw data. Thirdly, the random forest method based optimal feature selection is presented in details and compared with Lasso method. Finally, the feature selection methods are evaluated with Random Forest and Logistic Regression based PD pattern recognition method. The top 6 features are recommended for PD pattern recognition based on the experimental data and the random forest method.

Knowledge discovery from on-line cable condition monitoring systems

Detection and diagnosis of partial discharge (PD) activity has been widely adopted in electrical plant condition monitoring. For many years incipient partial discharge faults in power cables have been identified through off-line investigation techniques. With the development of measurement technology, more recently, continuous on-line monitoring systems are being installed, because in comparison with off-line measurement, it owns more advantages such as low cost, easy set-up etc. This has been instigated with the aim of reducing unexpected failures. Unfortunately, due to a lack of knowledge rules which can be applied to the data detected from on-line PD condition monitoring, this technology has not shown its full potential so far. This paper presents work on the analysis and development of a knowledge acquisition system based on rough set (RS) theory. Results prove that the proposed algorithm can successfully discover the hidden correlations between cable faults and PD measurement data and improve the effectiveness of on-line condition monitoring systems.