Application of Artificial Intelligence in Power System Monitoring and Fault Diagnosis

1. Introduction

Emerging technologies such as artificial intelligence (AI), big data analytics, and deep learning have gained widespread attention in recent years and have demonstrated great potential for application in many industrial fields. In power systems, AI and other technologies are also being used as new and powerful tools to replace traditional techniques for feature modeling, performance control, and fault diagnosis in order to obtain superior results. This Special Issue, “Application of Artificial Intelligence in Power System Monitoring and Fault Diagnosis”, aims to introduce the latest advances in this field and discusses the application of AI technology in power system modeling and control, state estimation, performance diagnosis, and prognosis, among other fields.

The scope of this Special Issue includes, but is not limited to, the following:

• Data-based abnormalities analysis of thermal power systems and nuclear power systems;
• Fault diagnosis and prediction of wind turbines based on SCADA data;
• Modeling, monitoring, and diagnosis of waste-to-energy, biomass power, and tidal power systems;
• Data-based fault characteristics analysis of power generation equipment;
• Power equipment health monitoring based on vibration signals, sound signals, image signals, thermal infrared signals, etc.
• Control and performance monitoring of photovoltaic power generation systems;
• Modeling, scheduling, control, and monitoring of microgrid systems;
• SOC estimation, SOH estimation, fault detection, isolation, and localization of lithium battery systems;
• State estimation and performance evaluation of large-scale energy storage systems.

From a total of 24 submissions, 10 research papers were published in this Special Issue, with 14 rejected.

2. Highlights of Published Papers

This section provides a summary of this Special Issue of Energies, covering published articles [1,2,3,4,5,6,7,8,9,10] which address several topics related to AI technologies in power system performance monitoring.

In [1], Barnabei et al. designed a Supervisory Control and Data Acquisition (SCADA)-based framework for the unsupervised anomaly detection of district heating (DH) network generating units. The framework relies on a multivariate machine learning regression model and then uses a sliding threshold approach for the subsequent processing of the model residuals generated during the testing phase. The system was tested against major failures occurring in gas-fired generating units at the DH plant in Aosta, Italy, and the results showed that the framework can detect anomalies successfully.

In [2], Lin et al. proposed a new method for shunt capacitor monitoring. The method monitors the shunt capacitor bank via the synchronous voltage and branch current of the shunt capacitor bank, calculates the capacitance parameters of the ungrounded star-connected capacitor bank using the parameter symmetry of the capacitor parameter calculation method, and identifies the abnormal state of the capacitor according to the statistical method. The simulation established by PSCAD verified that the relay protection device could effectively monitor the early abnormal condition of the capacitor bank.

In [3], Jawad et al. proposed a fault diagnosis method based on probabilistic generative models to remedy the shortcomings of existing fault detection methods for high-voltage direct current (HVDC) transmission systems. The method uses wavelet transform based on ant colony optimization and artificial neural network to detect different types of faults in HVDC transmission lines. The experimental results showed that the proposed method has higher accuracy and stronger robustness in the fault diagnosis of HVDC transmission systems compared with existing methods, such as support vector machines and decision trees.

In [4], Pujana proposed a hybrid model-based method for developing a digital twin (DT) model for wind power conversion systems. The method combines the advantages of physical models with advanced data analysis techniques to obtain knowledge from actual operational data while preserving physical relationships, thereby generating synthetic data from non-occurring events to detect and classify faults. Compared with existing DT methods, the method proposed in this paper has significant advantages in accuracy and interpretability.

In [5], Xia et al. proposed a multi-model fusion ensemble learning algorithm based on stacked structures to detect power theft. To solve the problem of existing methods being unable to further improve the accuracy of electricity theft detection, a heterogeneous ensemble learning method is used to construct a heterogeneous integrated learning model for stacked structure electricity theft detection using different powerful individual learning superposition integration structures to achieve the accurate detection and identification of electricity theft.

For identifying different types of partial discharges (PDs) in gas-insulated switchgear (GIS), Zheng et al. proposed an improved feature fusion convolutional neural network (IFCNN) method in [6], which solves the problem of traditional methods requiring a large quantity of statistical discharge data. By fusing time-frequency features, the method can uncover more local features of potential discharge pulses and increase the recognition accuracy to 95.8%.

In [7], Luo et al. designed an automatic machine learning-based lifetime prediction model (AutoML) for accurately estimating and predicting the capacity and lifetime of Li-ion batteries. The features of CC and CV phases are extracted using optimized incremental capacity (IC) curves, and the noise is removed using the Kalman filtering algorithm. They then built AutoML, which can automatically generate the appropriate processing flow, addressing the issues of information redundancy and high computational cost. By validating the NASA dataset, they demonstrated a significant improvement in the model’s ability to predict battery life on small-scale datasets.

In [8], Bai et al. proposed an HOG-SVM-based power system equipment identification method. First, wavelet transform is performed on the sound signals of power system equipment collected from the field to obtain wavelet coefficient-time maps. Then, the HOG features of the images are selected, and the selected features are classified using an SVM classifier. Moreover, the method also combines sound signal and image processing to effectively take advantage of image processing and avoid the limitations of sound signal processing. Finally, simulation experiments demonstrated that the proposed method can accurately identify and classify power system equipment.

In [9], Chen et al. proposed a deep-learning-based method for the intelligent modeling of the incineration process in waste-to-energy plants. The output variables are selected regarding safety, stability, and economy. The input variables are determined by eliminating invalid redundant variables using the Lasso (Least absolute shrinkage and selection operator) algorithm and a multi-input multi-output model based on feature selection, and CNN-BiLSTM is established. The results showed that the model can fully exploit the data features under multi-dimensional input feature parameters, and that it has higher accuracy and applicability than the traditional model.

Finally, in [10], Zhang et al. constructed a short-term wind speed prediction model based on variable support segments (VSS). At first, the method decomposes the historical wind speed series into several components using the variational mode decomposition method. Then, an improved transformer model is used to predict the predicted values of each element, and these predicted values are summed to obtain the future wind speed prediction. Experimental results showed that the prediction accuracy of the improved transformer model is significantly higher than that of other prediction models.