Research on real-time data processing and evaluation of new power system wide-area digital metering equipment based on deep learning algorithm

The typical technical realization scheme of real-time data management system is shown in Figure 1. The real-time data management system obtains its data from the distributed control system (DCS), network control system, water system, coal system, and ash system (PLC) of each power plant. In order to keep the production control system free from interference from the outside world and to ensure the safe and stable operation of the production control system, the real-time data are sent from the control system to the enterprise management network in the form of unidirectional transmission through the isolation equipment. After receiving the data, the real-time data management system displays and analyzes the data, and presents dynamic information about the equipment to the managers of the enterprise.

Figure 1.

Realtime data management system typical technology implementation scheme

The topology of the real-time data management system is shown in Figure 2. The real-time data management system collects data information from real-time data sources of the power plant, attaches the information to the graphical elements, takes the graph as the information indexing interface, and presents the information it covers in an intuitive form in front of the manager. In addition, it analyzes and compares real-time data and historical data to achieve the purpose of understanding the change of equipment status at any time, so that equipment failures can be monitored at the beginning of the occurrence. The topology diagram is as follows: 1)

The real-time data source (usually the production control system of the power plant) sends the real-time data in the form of unidirectional transmission to the digital mining engineers’ station through the gateway machine, which can isolate the real-time data source from the real-time monitoring system, so that the production control system is free from the interference from the outside world, and ensure the safe and stable operation of the production control system.

Figure 2.

Extension structure of real-time data management system

The real-time data management system is divided into four parts: general power system graphic editing system, real-time data analysis system, data acquisition system and equipment health status evaluation system, and the functional structure of the system is shown in Figure 3.

Figure 3.

Functional structure diagram of the system

The basic principle of artificial neural network is to continuously acquire knowledge through the external environment and optimize its own parameters (e.g., weights) according to certain learning rules, so that the actual output gradually converges to the ideal output over time [19-22]. According to the amount of information about the external environment, there are three types of artificial neural network learning methods: supervised learning (with teacher learning). Unsupervised learning (learning without a teacher). Re-inspired learning (reinforcement learning). The artificial neural network continuously learns the knowledge acquired from the external environment through these three learning methods. Its learning criteria include: error correction learning. Hebb Learning. Competitive Learning. When solving the problem, it optimizes its performance according to the learning criterion, and finally achieves the ideal output. From the above analysis, it can be seen that the most important thing in the learning process of artificial neural network is to have enough samples to learn and optimize its performance. Generally speaking, we need to apply computer programming software to build an artificial neural network model and set its specific parameters. According to the data provided by the outside world and the structure of the artificial neural network to formulate a specific standard, in the artificial neural network in each layer of each node, according to the formulated specific standards to deal with the data, so that the artificial neural network in accordance with the standards of each input will eventually get a corresponding output. At the same time, the artificial neural network compares the actual output with the ideal output, and if there is a deviation between the two and the size of the deviation exceeds the accuracy requirements, the artificial neural network effectively corrects the standard in real time, which is what we call the training process. This process also reflects the characteristics of artificial neural networks in terms of their ability to effectively learn. Through training, the artificial neural network effectively and accurately corrects deviations from specific criteria. This will make the difference between the actual output and the ideal output smaller in the next training process.

In power systems, historical fault information is often stored in automated devices such as fault recorders. Artificial neural networks can be trained to learn and explore criteria based on the large amount of historical fault information available in the power system. In order to realize the real-time data transmission performance and steady-state performance evaluation of wide-area digital metering equipment, two relatively independent sub-networks are constructed in this paper. 1)

Transmission characteristic discriminative network ANN1

As the name suggests, this sub-network is able to accurately judge the real-time data transmission performance of power system wide-area digital metering equipment, so that the protection can be accurately acted under different data transmission modes of the power system as well as in the face of different types of faults inside or outside the area. In order to achieve this purpose, and because judging the location and type of faults requires us to know the three-phase current magnitude of the power system before and after the faults, the input layer of the sub-network ANN1 has six nodes, which are the current magnitude before the faults, I_a, I_b, I_c, and the current magnitude after the faults, i_a, i_b, and i_c, respectively; and then we need to determine the type of the faults and the number of nodes in the output layer of the fault and location discriminating sub-network ANN1. Number of nodes. The real-time data transmission method of the power system, the type of fault, and whether the fault occurs inside or outside the zone are all needed to be accurately discriminated by ANN1. According to the above analysis, these different output information can be expressed by five output layer nodes: node one output value is 1 when the power system operation mode is the maximum operation mode, and the value is 0 when the power system operates in the minimum operation mode. Node two outputs a value of 1 for single-phase faults and a value of 0 for no faults. Node three output value is 1 for two-phase fault, value 0 for no fault. Node four output value is 1 when three-phase fault, the value is 0 when there is no fault. Node five output value of 1 when the area fault, the circuit breaker trips, the value of 0 when the area outside the fault, the circuit breaker does not trip. In this way, the number of input layer nodes of the fault type and location discriminating sub-network ANN1 is taken as 6 and the number of output layer nodes is taken as 5.

The real-time data from the device is utilized to train the two sub-networks for continuous learning, in order to correct the connection weights and thresholds of all neurons in the sub-networks. For each sub-network, the BP algorithm is used to continuously back-propagate the error value, so that all neurons in the artificial neural network correct their own weights through their own error signals, which also makes the two sub-networks converge to the global optimal solution at the fastest speed. And to meet the requirements of the artificial neural network discrimination accuracy. In this paper, the fault samples used to train the two sub-networks come from the data set collected above, through which different fault states in different operation modes of the power system are simulated. The neurons in the two sub-networks are made to form their respective connection weight matrices and threshold matrices, and finally the two sub-networks are examined in conjunction with the evaluation indexes and datasets to see whether the actual outputs of the networks are the same as the ideal outputs.

The power information integrated transmission WRED parameters are configured as shown in Table 1, and the DSCP-based weighted fair queuing mechanism is configured for the access routers of the substation, in which the emergency control real-time data service is configured as low latency queue, and its latency requirements are guaranteed on a priority basis. At the same time, the fixed access rate (CAR) mechanism is configured on the interface between the access router and the core network router to limit the input and output rates of the bursty management information service, fault information service and high-bandwidth multimedia data service to the server side. In terms of congestion avoidance, the weighted random early detection (WRED) mechanism is set up in the border router and core router, and the threshold threshold and maximum loss probability of each message can be clearly seen that the maximum probability of discarding the device management information in the multimedia information is 0.992 and 0.997, i.e., the probability of discarding is the largest in comparison with other data types.

The end-to-end delay of packets with and without artificial neural network is shown in Fig. 5, where (a) to (f) are emergency control and protection information, wide-area surveillance information, SCADA information, fault information, management information, and multimedia information, respectively. Under ANN1 (Artificial Neural Network) condition, real-time data service is guaranteed with better quality of service, although the delay of fault information, management information and multimedia information is slightly increased in the ANN case, but the fault information and management information generally use TCP protocol, which is guaranteed by retransmission mechanism in both transport and application layers, and will not affect their reliability. Multimedia messages generally use the UDP protocol, and this service allows a certain delay and packet loss rate, so its quality of service is also not greatly affected. In the case of ANN, the end-to-end delay of emergency control and protection is about 6.08 ms. the end-to-end delay of wide-area monitoring information is about 8.19 ms. the end-to-end delay of SCADA information is about 12.27 ms. the end-to-end delay of fault information is about 18.08 ms. the end-to-end delay of management information is about 23.19 ms. the end-to-end delay of multimedia information is about 29.39 ms. Data exchange between control centers and substations and control centers are similar and can yield similar delay results, which are not discussed in this paper.

Figure 5.

End-to-end delay analysis

Under ANN1 network conditions, the setup link LSR (switching router) 2-LSR9 fails at the moment of 200 seconds, and the failure is recovered at the moment of 500 seconds, which causes one main LSP (LSP2) of the substation to fail. Network simulation is performed to obtain the traffic of two main LSPs of the substation as shown in Fig. 6, where (a)~(b) are LSP1 and LSP2, respectively, and it can be seen that when the LSR2-LSR9 working link fails, it automatically switches the data flow to the other main LSP (LSP1) to ensure that the communication will not be interrupted. When the link failure is fixed, the traffic on the two main LSPs returns to normal.

Figure 6.

An LSP is faulty. Procedure

The data flow received by the emergency control and protection server at the control center in the case of a communication link failure, in both the no-ANN1 network and ANN1 network scenarios, is shown in Figure 7. As seen in Fig. 7(a), in the case of no ANN1 network, when the communication link fails, the traffic received at the server side is unstable, and the data flow at the server side recovers after about 10 seconds after the link fails. The reason for this is that the switching of data flow cannot be realized automatically, and traffic recovery under link failure can only be achieved by re-routing and searching for reachable paths. Under this mechanism, the time required for re-routing is related to the routing algorithm, the size of the network, and the time required for its convergence when the network is re-routed is within tens of seconds. And as seen in Fig. 7(b), under the ANN1 network, when the communication link fails, the server side can still receive a stable data stream. The reason for this is that ANN1 has a self-healing fault recovery mechanism, which can realize the switching of traffic between multiple LSPs, and MPLS can quickly detect the faulty failed components, and its fault recovery time is in the millisecond level. It can better meet the needs of real-time control of power systems.

Figure 7.

Emergency control and protection of data received by the server

To ensure that the substation maintains communication reliability in case of failure of the two main LSPs, a backup LSP is set up for each substation. Take substation one as an example. In order to verify the self-healing recovery mechanism of the ANN network in this situation, the following scenario is set up: the working links LSR2-LSR3 and LSR2-LSR9 fail at the same time at the moment of 200 seconds and recover at the moment of 500 seconds, which will lead to the failure of two main LSPs of the substation. In this paper, we simulate this and get the traffic on the two main LSPs and the backup LSP of substation 1 as shown in Fig. 8, which shows that when the two main LSPs fail, the data traffic of substation 1 is automatically switched to the backup LSP, and the main LSP returns to the normal working state after the failure is recovered. Meanwhile, similar traffic characteristics to Fig. 8(b) can be obtained at the server side, indicating that the server can still receive stable data traffic. This ANN network can ensure the reliability of real-time data in the power system.

Figure 8.

Two LSPS are faulty.

The limiting cut-off time (CCT) is used as an index to construct the degree of transient stabilization, however, the calculation of the CCT needs to be repeatedly performed by the time-domain simulation method, which is a heavy computational task and takes a long time. Therefore, the disturbed degree based on the cluster envelope integral of the generator power angle trajectory is constructed S: (1)S=∫t=0l=Ti(maxδi−minδi)dt

Where: δ_i is the power angle of the ind generator and T_s is the simulation time. According to Eq. (1), the disturbance degree S is calculated for all the samples labeled as stable in the temporary stability sample set, and then S is subjected to the maximum-minimum normalization process and its value is converted to the interval [0, 1] to obtain the stability degree index M_s: (2)Ms=Smax−SSmax−Smin

Where: S_max and S_min are the maximum and minimum values of S for all stable samples in the temporary stability sample set, respectively. When S is closer to S_max, the value of M_s is smaller and closer to 0, which represents that the system has a smaller degree of stability: on the contrary, the value of M_s is larger and closer to 1, which represents that the system has a larger degree of stability.

When the system loses transient stability, the time experienced by the sample to experience destabilization after fault removal is used as a quantitative index of the destabilization degree of the destabilized sample. The time between fault resection and the occurrence of instability is defined as: (3)T=tus−tct

Where: t_us is the moment when the system is destabilized (i.e., the power angle difference between any two generators is greater than 360°), and t_ct is the fault removal time. According to equation (3), T of all samples labeled as unstable in the temporary stability sample set are calculated, and then T is subjected to maximum-minimum normalization and its value is converted to the interval [0, 1] to obtain the instability degree index M_us: (4)Mus=Tmax−TTmax−Tmin

Where: T_max and T_min are the maximum and minimum values of T for all the destabilization samples in the temporary stabilization sample set, respectively. When T is closer to T_max, the value of M_us is smaller and closer to 0, which means that the system has a smaller degree of instability; on the contrary, the value of M_us is larger and closer to 1, which means that the system has a larger degree of instability.

First, the proposed ANN1 model-based power system wide-area real-time data is used to obtain the deterministic equipment stability samples and the deterministic equipment instability samples, and the deterministic stability samples of the power system equipment are inputted into the ANN1 model for the assessment of the degree of stability. Similarly, the determined instability samples of the power system equipment are input to the ANN1 model for training and evaluation. The construction method of the assessment model for the degree of stability and destabilization is to use the ANN to learn the mapping relationship between the sliding time window input features corresponding to the corresponding response time and the degree of transient stability, with the feature set d as input, and the equipment parameter information selected by ANN1 and ANN1. The stabilization and power system equipment instability degree assessment models are trained with power system equipment stabilization samples and equipment instability samples, respectively, so that the root mean square error loss function between the predicted and true values is minimized. The performance assessment index of the system equipment temporary stability degree prediction model uses the root mean square error to measure the prediction accuracy of the stability degree.

The changes in the root mean square error of the assessment accuracy indicators of the stability assessment model and the instability assessment model with response time are shown in Table 2. The root-mean-square error of the ANN2-based power system equipment stability assessment model and the ANN2-based instability assessment model decreases with the increase of the response time, and it can be seen that the assessment accuracy improves with the increase of the response time.

In order to visualize the prediction performance of the stabilization/destabilization degree assessment model under different response times. Figures 9(a) and (b) show the comparison of the real and predicted values of the stability degree of the test samples at response time t=0.0138s and response time t=0.1s for the ANN2-based power system equipment stability degree assessment model, respectively. Figures 10(a) and (b) show the comparison plots of the true and predicted values of the destabilization degree of the test samples of the ANN2-based power system equipment destabilization degree assessment model at response time t = 0.0138s and response time t = 1s, respectively. It is worth noting that when the response time is only 0.0138s, the root-mean-square error between the predicted value and the true value of the stability assessment model proposed in this paper is only 0.0056, and the root-mean-square error between the predicted value and the true value of the destabilization assessment model is slightly larger at 0.0138, but the error is still in an acceptable range. In actual operation, the low degree of stability (the smaller the value of Ms, the lower the degree of stability) given by the ANN2-based power system equipment stability assessment model should be taken to take relevant preventive and control measures to improve the stability of the power system. The more serious degree of instability (the larger the value of Mus, the more serious the degree of instability) given by the ANN2-based power system equipment stability assessment model should be prioritized to take emergency control measures to prevent the occurrence of system instability. It can be seen that based on the ANN network method proposed in this paper, the degree of stability/destabilization of power system equipment can be accurately assessed, which provides a useful reference for the implementation of subsequent preventive and emergency control measures for power systems.

Figure 9.

Testing results of stability degree at different response time

Figure 10.

Testing results of instability degree at different response time