Intelligent fault diagnosis and security early warning method of new power system based on system network situation

Abstract

In the past few decades, China’s power demand has been increasing, and the power fiber plays a key role in ensuring the orderly dispatching of all links of the power system. The study used a wavelet decomposition and reconstruction method, which is a signal processing technique used to decompose complex optical power data into low-frequency and high-frequency signals with different frequency components. Through this decomposition, we can more clearly observe periodic fluctuations, trend changes, and noise components in optical power data. The study also examined different prediction models, including GRU, LSTM, ARMA), etc. The performance of these models in predicting optical power trends is then analyzed, taking into account their accuracy, stability, and computational efficiency. Finally, we carefully evaluated the GRU-ARMA combined prediction model and determined its superior performance in predicting optical power trends. The outcomes show that after adjusting the input data length of the gating cycle cell model and the relevant parameters of the autoregressive sliding mean model, the residual mean value was 0.0141. At the same time, the root mean square error calculated by the combined prediction model of the gating cycle unit-autoregressive moving mean model was 0.000618, which successfully improved the accuracy of predicting the optical power trend of power fiber. This research result provides an important reference for the aging state assessment of power fiber lines, and has an important practical application value for the maintenance of power fiber lines.

Keywords

Feature extraction network situation fault diagnosis security early warning power system

1. Introduction

With the continuous expansion of power systems and the rapid development of intelligent technologies, fault diagnosis and safety warning in power systems have become increasingly important. As an indispensable infrastructure in modern society, power systems play a crucial role in providing stable power supply, energy transmission, and supply. Traditional power system fault diagnosis and safety warning methods are usually based on offline data analysis and empirical judgment, and have problems such as low diagnostic accuracy and long response time. Given the complexity of power systems and the urgent need for real-time performance, traditional methods often cannot meet the requirements for fast and accurate diagnosis and early warning. Therefore, our research aims to propose an intelligent fault diagnosis and safety early warning system based on the system network situation, making full use of the correlation and information interaction between nodes and equipment in the power system to improve the accuracy of power system diagnosis and prediction. sex. This method has potential advantages because it can better capture the dynamic conditions within the power system and provide support for fast and accurate fault diagnosis and safety warnings [1]. Due to the ongoing development and intelligentization of power systems, power fiber optic sensing technology has been widely applied as an important sensing technology in power systems. Power fiber optics have the ability to real-time and accurately perceive parameters such as temperature, pressure, and vibrations in power systems, and convert them into optical signals for transmission and monitoring, offering new possibilities and potential for fault diagnosis and safety warning in power systems [2]. Therefore, the power optical fiber communication has become the main transmission mode of the power communication system, and the proportion in the system gradually increases. As the scale of power fiber communication network is expanding, there are serious problems such as the insufficient timeliness of fiber fault monitoring and the increasing difficulty of line maintenance [3]. To achieve online monitoring of fiber optic faults and early warning (EW) of power fiber optic line status, this study proposes to research power system fault diagnosis (FD) and security warning based on network situation. It aims to enhance the accuracy of power system diagnosis and prediction.

Research content contains four parts, the first part is for the network situation of power system FD and safety EW research results summary, the second part is based on the RBF network situation extract the corresponding security information, then based on power optical fiber FD and fault warning research, the third part of the power optical fiber FD and security EW verification and performance evaluation, the fourth part is a summary of the full text.

2. Related works

At present, the power industry is developing towards the direction of more intelligent and high quality. For enhancing the stability and reliability of power supply and ensure the smooth operation of smart grid, it is necessary to realize the fault prediction and intelligent EW through real-time acquisition and analysis of the status information of substation equipment. Furse et al. for addressing the problem of fault detection, positioning and diagnosis accuracy, the use of various reflection measurement methods, such as time domain reflector, sequence time domain reflector, the expansion frequency time domain reflector, etc., and using the genetic algorithm, neural network (NN), particle group optimization automatic analysis technology, for enhancing the accuracy of fault detection, positioning and diagnosis [4]. For enhancing the accuracy of existing transformer FD methods, Liao W et al. use the graph convolution layer as a classifier for finding the nonlinear relationship between dissolved gas and fault type. The back-propagation algorithm is applied to the training process of the graph convolutional networks [5]. Wang et al. achieved electrical power fiber insulation fault detection using a chaotic system. They created four distinct types of electrical power fiber insulation faults and employed a high-speed acquisition card to measure localized electrical signals. This enabled the establishment of dynamic error scatter diagrams as characteristic features for various electrical power fiber fault states. The research successfully identified these four different types of electrical power fiber faults, with a recognition rate of 97.5% [6]. Li et al. proposed a fault detection method based on adaptive adjustment S transform (ST) to solve the problem of how to achieve high-speed and robust short-circuit fault detection to ensure the reliable operation of DC power grid. This method is based on the frequency domain and can be used in A fault condition is detected within 0.3 milliseconds. This method effectively differentiates between faults and other transient conditions through high-frequency detection and low-frequency screening [7].

For effectively warning the risk of natural disasters in the power grid and reduce its impact on the power system, many scholars have studied the power disaster warning system. To detect high impedance faults (HIF) in solar photovoltaic integrated power systems, Veerasamy et al. employed three-phase current signals, encompassing signals under both normal operating and fault conditions, to extract features. They proposed an LSTM classifier for the recognition of HIF in photovoltaic integrated grids and compared its performance with other well-known classifiers. The robustness of the classifier was validated through the assessment of performance metrics such as Kappa statistics, precision, recall, and F-measure [8]. In order to address the limited research and lower accuracy of existing network security situational assessment methods in SDN environments, Du et al. proposed a network security situational assessment model based on an improved Backpropagation (BP) neural network. They utilized the Cuckoo Search algorithm to optimize the neural network’s weights and thresholds, aiming to achieve a global optimal solution. The results demonstrated that the situational assessment curve of this model closely approximates the actual situational values, with an accuracy rate reaching 97.61% [9]. Zhang et al. constructed an LSTM-DT model, applying the decision tree algorithm and long short-term memory network to address time series challenges in network security situational assessment. Simultaneously, they introduced the concept of attack probability, learned the features of the original dataset through stack sparse autoencoders, and predicted attack probabilities from the dataset processed by the LSTM network. The research results demonstrate that the network situational awareness method they proposed achieves a 95% accuracy in network security situational awareness [10]. Cheng et al. have developed a leakage detector and warning system, using high precision current transformer and high gain linear converter, can effectively detect the leakage current above 1 mA, can prevent the occurrence of serious problems such as fire caused by leakage [11]. In response to the problems existing in traditional power workers’ work clothes, Chen designed a ZigBee-based early warning system for power transmission and distribution workers. The system uses distributed fusion algorithms for signal transmission, which improves the accuracy and reliability of the transmitted information [12].

To sum up, many experts have studied FD and safety EW, and some progress has been made in power system FD and safety EW technology, but there are still some problems such as unidentification and insufficient identification accuracy. Various types of faults may occur in the power system, and some faults may be hidden and difficult to identify in time. Subtle grid anomalies or equipment failures may not be easily noticed initially, which may result in the failure causing some damage before the problem becomes obvious. At the same time, there is still a lack of comprehensive understanding of internal and external threats to the power system, and it is necessary to learn how to better integrate real-time data, machine learning algorithms, and network situation analysis to improve the effectiveness of security predictions. Therefore, the study proposes a new power system fault intelligent diagnosis and safety early warning method based on the system network situation. This method has many advantages such as comprehensive, intelligent, real-time, predictability, and adaptability, and helps to improve the reliability of the power system. and safety, reducing operating costs. It is hoped that the accuracy of power grid fault identification can be improved and the safe operation of the power system can be ensured.

3. Research on power system failure and EW based on network situation

With the increase of the scale and business form of the power information network, the traditional security technology is facing more and more complex security threats. To this end, the research puts forward the application of network security situation awareness to power information network security protection, based on the security protection equipment to extract network security elements for evaluation, and use the current and historical situation data to predict the future network security situation, to assist network security management personnel to prevent security risks [13].

3.1 Research on the algorithm based on RBF network

The feature vector extraction task is to reduce the dimensionality, optimise the data representation and improve the efficiency of neural network training [14]. The most representative and effective vectors are selected from a large number of feature vectors to eliminate redundant information and ensure an accurate and efficient training process. Radial Basis Function (RBF) neural network is selected for training [15]. The RBF neural network is a $m-n-h$ type neural network structure containing $m$ input nodes, $h$ hidden nodes and $n$ output nodes. In the figure, $X_{i}(i=0,1,\ldots m)$ is the input quantity, $\omega_{ij}$ represents the weight between the input layer and the hidden layer, $\nu_{ij}$ represents the weight between the hidden layer and the output layer, $\theta_{j}$ represents the threshold of the hidden layer, and $\theta_{k}$ represents the threshold of the output layer. $s_{j}$ The function expression is shown in Eq. (1). $g_{t}$ is the excitation function of the implicit layer and the output layer, respectively. The function expression is shown in Eq. (1).

$\displaystyle\Phi_{i}\left\|{x-c_{i}}\right\|=\exp\left[{-\frac{\left\|{x-c_{i% }}\right\|}{2\sigma_{i}^{2}}}\right]$ (1)

In RBF neural network, $\Phi_{i}$ in Eq. (1) is the basis function and $\sigma_{i}$ is used to perform normalisation. By appropriately setting the number of hidden nodes and the basis function, the nonlinear problem can be transformed into a linear problem. Let there are P training samples, then $X=\left\{{X_{1},X_{2},\ldots,\left.{X_{P}}\right\}}\right.$ in the output layer corresponds to the desired output $X p$ one by one; $\Phi_{i}\left\|{x-c_{i}}\right\|$ is the basis function of the $j$ th implied node; $c_{j}=\left\{{cj_{1},cj_{2}}\right.,\ldots,\left.{c_{jm}}\right\}^{T}$ denotes the centre vector of the basis function of the implied node; $y^{p}=\left\{{y_{1},y_{2}}\right.,\ldots,\left.{y_{m}}\right\}$ denotes the output value. The total error function of this network is expressed as Eq. (2).

$\displaystyle J=\sum\limits_{p=1}^{P}{J_{p}}=\frac{1}{2}\sum\limits_{p=1}^{P}% \sum\limits_{k=1}^{n}{\left({Y_{k}^{p}-y_{k}^{p}}\right)}^{2}=\frac{1}{2}\sum% \limits_{p=1}^{P}{\sum\limits_{k=1}^{n}{e_{k}^{p2}}}$ (2)

In Eq. (2), $c_{j}$ is the vector, $\nu$ is the network weights, $\delta_{j}$ is the width coefficient, and $\theta$ is the threshold value. Then the parameters are determined as shown in Eq. (3).

$\displaystyle d_{i}\left(t\right)=\left\|{x\left(t\right)-c_{i}\left({t-1}% \right)}\right\|,\left({i=1,2,\ldots,k}\right)$ (3)

Use Eq. (2) to calculate the Euclidean distance between the sample data and the centre value, then find the nearest point $r\left({1\leqslant r\leqslant k}\right)$ . Then find the cluster centre as shown in Eq. (4).

$\displaystyle c_{i}\left(t\right)=c_{i}\left({t-1}\right),\left({i=1,2,\ldots,% k,i\neq r}\right)$ (4)

Then determine the function limit value $\varepsilon$ and the learning rate $\beta\left(0\right)$ as shown in Eq. (5).

$\displaystyle c_{i}\left(t\right)=c_{i}\left({t-1}\right)+\beta\left(0\right)% \left[{x\left(t\right)-c_{r}\left({t-1}\right)}\right]$ (5)

Corrections are made to the learning rate and update samples as shown in Eq. (6).

$\displaystyle\beta\left({t+1}\right)=\frac{\beta(t)}{\sqrt{1+\text{int}(t/a)}}$ (6)

Determine the objective function as shown in Eq. (7).

$\displaystyle J=\sum\limits_{p=1}^{p}{J^{p}=\frac{1}{2}}\sum\limits_{p=1}^{p}{% \wedge\left(p\right)}\left[{Y^{p}-y^{p}(t)}\right]^{2}$ (7)

In Eq. (7), $\wedge\left(p\right)$ is the weighted factor matrix. The estimated value of weights is then calculated using Eq. (7) and the minimum value of $J$ is taken as shown in Eq. (8).

$\displaystyle\frac{\partial J(t)}{\partial V}=0^{r}$ (8)

The weights are finally solved for, using RSL for weight recursion, as shown in Eq. (9).

$\displaystyle\left\{{\begin{array}[]{l}\hat{v}_{jk}(t)=\hat{v}_{jk}(t-1)+K(t)% \left\{{Y^{p}-\left[{\Phi{}^{P}(t)^{T}}\right]}\right\}\hat{w}_{jk}(t-1)\\ k(t)=P(t-1)\cdot\Phi^{P}(t)\left\{{\left[{\Phi^{P}(t)^{T}}\right]P(t-1)}\right% .\cdot\Phi^{P}(t)+\\ P(t)=\left\{{1-k(t)\cdot\left[{\Phi^{P}(t)^{T}}\right]}\right\}P(t-1)\\ \end{array}}\right.\left.{\frac{1}{\wedge\left(p\right)}}\right\}^{-1}$ (9)

In Eq. (9), $vjk(t)$ is the weight matrix connecting the implicit layer and the output layer, $\Phi^{P}(t)=\left[{\Phi_{1}^{P}(t),\Phi_{2}^{P}(t),\ldots,\Phi_{\alpha}^{P}(t)% }\right]^{T}$ is the vector of the output of the implicit layer, and $\hat{y}_{k}(t)=\sum{vjk\cdot}\Phi_{j}^{P}(t)$ represents the output value of the kth node in the output layer. The study can extract the safety feature information of the power system by RBF neural network, which can monitor and predict the system operation status in real time, and discover the possible safety hazards in time, and provide powerful support to guarantee the operation safety of the power system, and the flowchart of the RBF neural network to extract the safety feature information of the power system is shown in Fig. 1.

Figure 1.

RBF NN to extract safety feature information of power system.

As shown in Fig. 1, the study of RBF NN was used to model and train the extracted security feature information, including determining the network parameters like the quantity of nodes and activation function types of the input, hidden and OL of the network, as well as the center and standard deviation of the HN. RBF NN has the advantages of robust, good interpretability, fast computing speed and strong scalability, and can better adapt to the complex and changeable security feature information in the power system.

3.2 Research on fault diagnosis in power systems based on power fiber optics

When it comes to algorithm research on network situation extraction and security feature information, the RBF (Radial Basis Function) network has been widely discussed and applied. This algorithm provides a powerful tool in the field of network security based on its success in the fields of pattern recognition and anomaly detection. The RBF network is well-known for its superiority in modeling nonlinear problems. Its core idea is to fit data through a set of radial basis functions to achieve modeling and analysis of network behavior. This method can not only identify known threats and attack patterns, but also detect unknown security threats, improving network security. In the process of extracting network security feature information, power optical fiber technology also plays an important role. Applying power optical fiber technology to fault diagnosis research can realize real-time monitoring and abnormality detection of power system status. By analyzing the data obtained by fiber optic sensors, potential faults and problems in the power system can be detected in real time and necessary measures can be taken in advance to ensure the reliability and stability of the power system. Therefore, fault diagnosis research based on power optical fiber not only helps to improve the operating efficiency of the power system, but also helps to ensure the safety of the power system.

Rayleigh scattering is a scattering phenomenon that occurs when light encounters particles much smaller than its wavelength. When the wavelength of the incident light is much larger than the size of the scattering particle, the incident light is scattered uniformly in all directions, usually centred on the scattering particle. In this scattering process, the intensity of the scattered light is proportional to the fourth power of the scattering angle and inversely proportional to the fourth power of the scattering wavelength. As shown in Eq. (10).

$\displaystyle I(\lambda)\infty\frac{1}{\lambda^{4}}$ (10)

Equation (10) The Rayleigh scattered intensity $I(\lambda)$ is inversely related to the wavelength of the incident light $\lambda$ . A longer wavelength of incident light results in a lower Rayleigh scattered light intensity. Assuming that the optical power intensity of the incident light is $p_{0}$ , the optical power intensity of the Rayleigh-scattered light can be calculated according to Eq. (11) when the light is transmitted to the point $z$ in the optical fibre and returned to the point of incidence.

$\displaystyle p_{0}=\frac{\tau\Delta c\alpha_{r}p_{t}}{4n_{1}}$ (11)

Fresnel reflection is a phenomenon of light reflection at an interface. $p_{t}$ is used to describe the optical power of incident light at the point of scattering. The calculation of Fresnel reflection involves the speed of light $c$ , the Fresnel reflection loss coefficient of the optical fibre $\alpha_{r}$ , and the relative refractive index of the optical fibre $\Delta$ and the refractive index of the fibre core $n_{1}$ . $n_{0},n_{1},n_{2}$ The refractive indices of different media are represented. The principle of OTDR fault detection is shown in Fig. 2.

Figure 2.

OTDR fault detection principle.

Optical Time Domain Reflectometer (OTDR) By analysing the reflection and scattering characteristics of optical signals, OTDR achieves the function of locating signals in optical fibres and detecting faults [16]. The ARIMA model is a classical time series forecasting algorithm model based on the ARMA model, and in applying the ARIMA model for When applying the ARIMA model for time series forecasting, the first step is to determine the smoothness of the series and perform appropriate difference operations on the non-smooth series to ensure the validity of the model [17]. In the $p$ th order autoregressive model $\textit{AR}(p)$ , the parameters to be estimated are the autoregressive coefficients $\varphi_{p}$ and the random noise term $\varepsilon_{t}$ represents irregular random information. Specifically, the current value $X_{t}$ is obtained by multiplying the most recent $p$ order lag term with the corresponding regression coefficient $\varphi_{p}$ and adding the random term $\varepsilon_{t}$ . This is shown in Eq. (12).

$\displaystyle X_{t}=\varphi_{1}X_{t-1}+\varphi_{2}X_{t-2}+\ldots+\varphi_{p}X_% {t-p}+\varepsilon_{t}$ (12)

The sliding average model $\textit{MA}(q)$ is usually assumed to have a zero mean. The process can be represented by Eq. (13), where the sliding average coefficient $\theta_{q}$ is the parameter to be estimated, which represents the weights of $\mu_{q-1}$ . By linearly combining the present and previous random error terms, the final time series data is obtained $X_{t}$ .

$\displaystyle X_{t}=\mu_{t}-\theta_{1}\mu_{t-1}-\theta_{2}\mu_{t-2}-\ldots-% \theta_{q}\mu_{q-1}$ (13)

The time series is affected by both the current and previous random error terms as well as the prior period values, then it can be modelled using the ARMA(p,q) model. The expression of the model is shown in Eq. (14), which combines an autoregressive process and a sliding average process and is expressed as an autoregressive moving average model of order $(p,q)$ , abbreviated as ARMA(p,q)

$\displaystyle X_{t}=\varphi_{1}X_{t-1}+\varphi_{2}X_{t-2}+\ldots+\varphi_{p}X_% {t-p}+\mu_{t}-\theta_{1}\mu_{t-1}-\theta_{2}\mu_{t-2}-\ldots-\theta_{q}\mu_{q-1}$ (14)

The design of the fibre optic fault detection device includes two parts, the lower computer hardware and the upper computer software, as shown in Fig. 3.

Figure 3.

Design scheme of power fiber optic fault monitoring equipment.

The lower position computer consists of STM 32 controller controller, optical power meter, OTDR module, optical circuit control, network communication and power supply circuit. The upper computer is mainly composed of optical fiber fault point type and distance display, working mode selection, optical power data storage, optical fiber line status warning and other functional modules [18].

3.3 Research on EW algorithm of power fiber line

Research on situation extraction and security feature information algorithms based on RBF networks provides us with a powerful tool that can be used to analyze and monitor the behavior of various complex networks, thereby achieving effective management of network security. However, the effectiveness of this cybersecurity depends not only on the security within the network, but also on the reliability of the infrastructure it supports, especially for critical areas such as power systems. Therefore, combining the feature extraction capabilities of the RBF network with the power fiber line status early warning algorithm can provide a higher level of guarantee for the operation and maintenance of the power system. The research on power optical fiber line status early warning algorithm aims to use advanced optical fiber sensing technology to monitor various parameters and line status in the power system to identify potential faults and problems in advance. By combining RBF Network’s experience in network situation extraction with the real-time monitoring capabilities of power optical fiber technology, we can achieve comprehensive monitoring of the power system, promptly identify network threats and power line problems, and take corresponding measures to ensure the power system continuous power supply and reliability.

With the increase of optical fiber use time, the optical power shows a trend of decrease. Therefore, by analyzing the network situation of optical power data, the working state of power fiber line is described, and the function of power fiber line state warning is realized. The schematic diagram of the optical power prediction method is shown in Fig. 4.

Figure 4.

Schematic diagram of optical power prediction method.

In Eq. (15), $\Psi\left(t\right)$ is a square-integrable function, $\Psi\left(t\right)\in L^{2}\left(R\right)$ , $L^{2}\left(R\right)$ are finite real number spaces, $\tau$ is a translation factor, and the scale factor is $a$ . The frequency domain data is reconstructed into time domain data as in Eq. (15).

$\displaystyle x\left(t\right)=\frac{1}{C_{\Psi}}\int\limits_{-\infty}^{\infty}% {\frac{1}{a^{2}}}W\left({a,\tau}\right)\Psi\left({\frac{t-\tau}{a}}\right)dad\tau$ (15)

During the discrete wavelet transform process, the signal first passes through a low-pass filter and a high-pass filter for filtering operations. The low-pass filter is used to extract the low-frequency part of the signal while the high-pass filter is used to extract the high-frequency part of the signal. The two sub-signals obtained after filtering are the low-frequency and high-frequency portions of the original signal. A downsampling operation is performed on the low-frequency part to obtain the next level of low-frequency signal. At the same time, downsampling is performed on the high frequency part to get the next level of high frequency signal. This process can be repeated many times and each time the obtained low frequency signal and high frequency signal can be filtered and down sampled again. In the discrete wavelet transform, it is necessary to amplify both the scaling factor $a$ and the translation factor $2^{j}$ , where $j$ is an integer greater than 0, so as to realise the processing of discrete data [19]. The principle of the three types of units is shown in Fig. 5.

Figure 5.

Three types of unit principles.

In view of the recurrent NN, Long Short-Term Memory (LSTM) is presented, which introduces three gating mechanisms: forgetting gate, incoming gate and output gate, which can effectively address the issue of gradient disappearance or gradient explosion in the recurrent NN. As an updated variant of the gating cycle unit (GRU), LSTM, combines the forgetting gate and the input gate into one update gate, making the operation process simpler and shorter operation time. For some data prediction, GRU has better results.

LSTM is a recurrent neural network suitable for processing time series data, which can capture long-term dependencies and temporal patterns in the data. In power systems, LSTM can be used to analyze historical data to identify potential abnormal patterns and trends, helping to detect signs of failure early. At the same time, LSTM can be used for fault prediction and diagnosis of power systems. By training the model using historical data, LSTM can learn patterns of normal operating conditions of the power system. When an abnormality occurs in the system, the model can detect patterns that do not match the normal patterns, thereby issuing early warnings or performing fault diagnosis. Various equipment in the power system, such as transformers, switchgear, etc., require regular monitoring of their status and health. LSTM can be used in conjunction with sensor data to monitor the working status and performance of equipment.

4. Verification of the power system FD and safety EW method

The study aims to confirm the speed of fault location (FL), the accuracy of FL, the duration of fault detection, and the accuracy of safety warning, so as to provide an objective evaluation of their processing effect. The study uses the wavelet function to verify the FD, and uses the 250 days of power grid fiber data provided by a power company to verify the security warning.

4.1 Validation of the FD method of power optical fiber based on feature information extraction

The study verifies the possible 6 kinds of HF signals in power fiber, and obtains the transient components of these signals at 10 MHz through simulation experiments. Now, the RBF NN is used for processing the data, and the wavelet energy of different frequency bands is calculated and extracted through the decomposition and reconstruction of 12-layer wavelets. Considering the complex NN and the large data processing capacity in the FD process of the wavelet NN, the wavelet energy is normalized to ensure the operation efficiency. After normalization, Fig. 6 intuitively shows the wavelet energy ratio of each frequency band, which more clearly shows the characteristics of different HF signals.

Table 1
OTDR fault detection test results

Test event type	Detection time (s)	Event point distance (m)	Error (m)
Fiber length	17	1011	9
B ending point	16	238	10
B reaking point	14	829	5
Joint point	17	581	8

Figure 6.

Three types of unit principles.

The energy of the transient failure is mainly distributed between the Ed 8 and Ed 12 frequency bands, with the Ed 12 and Ed 9 frequency bands having the most energy with a frequency of 10⁴–2 $\times$ 10⁴ Between Hz and 1250 to 2500 Hz. Specifically, the energy gradually decreases between Ed 10 to Ed 12 and Ed 1 to Ed 9, respectively. The energy of permanent failure is mainly concentrated in the LF band, and the HF component is less, and the Ed 12 frequency band (1200–2500 Hz) has the highest proportion. The energy distribution of the circuit breaker closing signal is relatively concentrated, mainly concentrated in 10⁴–2 $\times$ 10⁴Hz and 2 $\times$ 10⁴ 3.91 $\times$ 10⁴Hz within two frequency ranges. The OTDR fault test outcomes are indicated in Table 1.

The OTDR test involves the verification of the type and distance of the fiber fault point. First, a simulated power fiber with a length of 1 km is selected. A fiber bending point was artificially manufactured at 230 m from the power optical fiber fault monitoring equipment; a fiber connection point was manufactured at 570 m and a fiber fault point was manufactured at 820 m. By testing on multiple sets of data, the average results were obtained, and the specific data are recorded in Table 1. As shown in Table 1, the fiber monitoring equipment performed well in the detection of FD types with 100% accuracy. Meanwhile, the error was kept within 10 m in the event point distance measurement. The average detection time was 16 seconds.

4.2 Verification of power optical fiber security EW method based on wavelet algorithm

The study experimental data is based on conducting a 250-day experiment with a selected power company and analyzing the obtained optical fiber communication data results. During the experiment, 8 wavelet bases were used for 1 to 4 wavelet decomposition, and the data after each decomposition were replaced into the GRU-ARIMA optical power prediction algorithm model for prediction. The predicted RMSE values were finally compared with the original data, and the results are summarized in Fig. 7.

Figure 7.

Predicting RMSE values with different decomposition layers and wavelet basis functions.

According to the analysis results of Fig. 7, it is found that as the number of layers for wavelet decomposition increases, the loss of raw optical power data will increase, and thus the prediction error increases. In the layer 1 wavelet decomposition, the sym 4 wavelet basis function performs optimally and is therefore chosen as the wavelet basis function of the analysis method. Figure 7 presents the raw optical power data as graphs by performing the normalization process. Figure 8 shows the light power data sof LF and HF and their reconstructed data diagram.

Figure 8.

Comparison chart of LF and HF optical power data after wavelet decomposition and reconstruction.

Based on the observations in Fig. 8, the optical power data reconstructed after wavelet decomposition presents significantly different features. The data of the LF part showed an obvious downward trend when the monitoring days of the power fiber optical power data increased, while the data of the HF part was relatively stable, with no obvious upward or downward trend. This suggests that sun exposure, high temperature and humidity lead to aging fiber lines, which affects LF optical power data. The fluctuation of HF optical power data may be caused by the influence of the wind on the power fiber. Two LSTM and GRU algorithms were selected to train and predict the reconstructed LF partial optical power data, and to compare the prediction effects of the two algorithms. The comparison results are shown in Fig. 9.

Figure 9.

Comparison curve of LSTM and GRU prediction results.

In Fig. 9, the plots of the LSTM model prediction results, the GRU model prediction results, and the curves compared with the LF optical power test set data are shown. By observing the graph, the prediction results of the GRU model are closer to the original value of the test set. The GRU model is a modified model based on the LSTM model, which combines the forgetting gate and the input gate in the LSTM model into one update gate and reduces the processing of partial cell states. Thus, the GRU model is similar to the LSTM model in its algorithmic processing and analysis procedures. Given the overall stability of the HF optical power data, the ARIMA model was used to predict these data. Model parameters are shown in Table 2.

Table 2

Judgment criteria for model parameters

Model	AR( $p$ )	MA( $q$ )	ARIMA( $p$ , $d$ , $q$ )
ACF	Decay tends to 0 (geometric or oscillatory)	Censored after $q$ order	The attenuation tends to 0 after the $q^{\text{th}}$ order (geometric or oscillatory)
PACF	Censored after $P$ order	Decay tends to 0 (geometric or oscillatory)	Attenuation tends to 0 after $P$ order (geometric or oscillatory)

As shown in Table 2, time series stationarity needs to be ensured before applying the ARIMA model, and for non-stationary time series, differential processing is required. The parameter $d$ represents the number of differential treatments, using the ADF test and KPSS test [19] in MATLAB. Finally, the ADF test was 1 and the KPSS test is 0, so the value of parameter $d$ was determined to be 1. In addition, the parameter $p$ represents the number of lags of the time series itself in the AR model, while the parameter $q$ represents the number of lags in the prediction error used in the MA model. The values of the parameters $p$ and $q$ are determined mainly based on the autocorrelation coefficient (ACF) and the partial correlation coefficient (PACF). The normal distribution map of the test is shown in Fig. 10.

Figure 10.

Comparison curve of LSTM and GRU prediction results.

Figure 11.

Comparison curve of prediction results between LSTM-ARIMA and literature [20].

After determining the parameters $p$ , $d$ , and $q$ of the ARIMA model, it is essential for evaluating the stability of the model and judge the accuracy of parameter selection. Stability evaluation is mainly based on the statistical properties of the residuals, including the residual mean approaching zero and whether the distribution of the residuals conforms to a normal distribution. The test set residual data curves in Fig. 10 were analyzed and the mean residual value calculated was 0.0141, indicating that the prediction error is within the acceptable range. Moreover, in the test set residual normal distribution map in Fig. 10, it can be observed that the residual data basically agrees with the standard normal distribution curve after normalization. In conclusion, it is concluded that the parameter selection of the ARIMA model is accurate.

Based on the data in Fig. 11, it is observed that the ARIMA model predictions for the optical power HF data test set are very similar to the actual value. The calculated root mean square error (RMSE) of 0.000618 further confirms the high accuracy of the ARIMA (2,1,3) model in predicting optical power HF data. The comprehensive analysis of curve diagram and RMSE values shows that the model can accurately capture the changing trend of optical power HF data and show excellent prediction accuracy [20].

5. Conclusions

The research aims to solve the problems of low FD accuracy and power system safety warning. The RBF algorithm is proposed for feature extraction, and then OTDR is used to monitor optical fiber faults. In terms of safety warning, a combined optical power prediction algorithm is used for research. The RBF algorithm proposed in the study shows excellent performance in feature extraction and OTDR for optical fiber fault monitoring, and the fault detection accuracy reaches 100%, indicating that the model can accurately identify various fault types occurring in power systems. In addition, the event point distance measurement error always remains within 10 meters, proving the accuracy of the model in determining the fault location. The study also demonstrated the excellent performance of the combined optical power prediction model (GRU-ARIMA), with a root mean square error of only 0.000618, which means that the model’s predictions are very close to the actual observed values. The two algorithm models, GRU and ARIMA, perform well in low-frequency data prediction, and the ARIMA model also achieves good results in high-frequency data prediction. The energy of transient faults is mainly distributed in the Ed8 to Ed12 frequency bands, the energy of permanent faults is mainly concentrated in the low frequency band, and the energy of the circuit breaker closing signal is concentrated in 104–2104 Hz and 2104 $\sim$ 3.91*104 Hz. The optical fiber fault monitoring equipment performed well, with a fault diagnosis accuracy of 100%, event point distance measurement error within 10 meters, and an average detection time of 16 seconds. The experimental results also show that as the number of wavelet decomposition layers increases, the loss of original optical power data increases, leading to the problem of increased prediction errors, and the sym4 wavelet basis function was selected for analysis. Secondly, through normalization of data, the changing trends of low-frequency and high-frequency data on different days are graphically displayed. Next, the GRU model showed improvement in handling long-term dependencies in sequence data, while the ARIMA (2,1,3) model showed high accuracy in predicting optical power high-frequency data, verified by residual analysis and RMSE values. its superior prediction accuracy. A weakness of the study is the relatively small sample size covered by the data collection. Later studies will increase the number of subjects for data collection to obtain a broader data sample.

Footnotes

Acknowledgments

Supported by the 2023 Open Fund Project of the National Key Laboratory of Power Grid Safety (Research on intelligent fault diagnosis method of new power system with high proportion of new energy access, DZB51202301402).

References

Raoufi

Vahidinasab

Mehran

. Power systems resilience metrics: A comprehensive review of challenges and outlook. Sustainability. 2020; 12(22): 9698.

Sepulveda

Jenkins

Edington

Mallapragada

Lester

. The design space for long-duration energy storage in decarbonized power systems. Nat Energy. 2021; 6(5): 506-+.

Hatziargyriou

Milanovic

Rahmann

Ajjarapu

Canizares

Erlich

Hill

Hiskens

Kamwa

Pal

Pourbeik

Sanchez-Gasca

Stankovic

Van Cutsem

Vittal

Vournas

. Definition and classification of power system stability–revisited & extended. IEEE T Power Syst. 2020; 36(4): 3271-3281.

Furse

Kafal

Razzaghi

Shin

. Fault diagnosis for electrical systems and power networks: A review. IEEE Sens J. 2020; 21(2): 888-906.

Liao

Yang

Wang

Ren

. Fault diagnosis of power transformers using graph convolutional network. CSEE J Power Energy. 2020; 7(2): 241-249.

Wang

Liao

. Fault diagnosis for power cables based on convolutional neural network with chaotic system and discrete wavelet transform. IEEE T Power Deliver. 2021; 37(1): 582-590.

Ukil

Satpathi

Yeap

. Improved S transform-based fault detection method in voltage source converter interfaced DC system. IEEE T Ind Electron. 2020; 68(6): 5024-5035.

Veerasamy

Wahab

NIA

Othman

Padmanban

Sekar

Ramachandran

Hizam

Vinayagam

Islam

. LSTM recurrent neural network classifier for high impedance fault detection in solar PV integrated power system. IEEE Access. 2021; 9(22): 32672-32687.

Yao

Cao

Liang

Ren

. Network situation assessment method based on improved BP neural network. Electronics. 2023; 12(3): 483-501.

10.

Zhang

Kang

Xiao

. Research on network security situation awareness based on the LSTM-DT model. Sensors. 2021; 21(14): 4788.

11.

Cheng

Chen

Lin

Shen

Lin

Chen

. Leakage current detector and warning system integrated with electric meter. Electronics. 2023; 12(9): 2123.

12.

Chen

. Design of wearable intelligent monitoring and early warning clothing for power workers. IEEJ T Electr Electr. 2023; 18(5): 791-796.

13.

Lee

Chung

. The machine learning-based dropout early warning system for improving the performance of dropout prediction. Appl Sci-Basel. 2019; 9(15): 3093.

14.

Costa

Ambrósio

Frey

Andrade

. A multivariate statistical representation of railway track irregularities using ARMA models. Vehicle Syst Dyn. 2022; 60(7): 2494-2510.

15.

Ospina

Gondim

JAM

Leiva

Castro

. An overview of forecast analysis with ARIMA models during the COVID-19 pandemic: Methodology and case study in Brazil. Mathematics. 2023; 11(4): 3069.

16.

Zheng

Tang

Zhou

. A framework of adaptive multiscale wavelet decomposition for signals on undirected graphs. IEEE T Signal Proces. 2019; 67(7): 1696-1711.

17.

. A bearing fault diagnosis method based on wavelet packet transform and convolutional neural network optimized by simulated annealing algorithm. Sensors. 2022; 22(4): 1410.

18.

Higham

Mikaitis

. Anymatrix: an extensible MATLAB matrix collection. Numer Algorithms. 2022; 90(3): 1175-1196.

19.

Papp

Yildiz

. Alfonso: Matlab package for nonsymmetric conic optimization. Informs J Comput. 2022; 34(1): 11-19.

20.

Thomas

Chaudhari

Shihabudheen

Verma

. CNN-based transformer model for fault detection in power system networks. IEEE T Instrum Meas. 2023; 72: 2504210.

Intelligent fault diagnosis and security early warning method of new power system based on system network situation

Abstract

Keywords

1. Introduction

2. Related works

3. Research on power system failure and EW based on network situation

3.1 Research on the algorithm based on RBF network

4.1 Validation of the FD method of power optical fiber based on feature information extraction

Table 1 OTDR fault detection test results

Footnotes

Acknowledgments

References

Table 1
OTDR fault detection test results