An ac substation reactive voltage control method based on deep neural network

Abstract

The application discloses an AC substation reactive power and voltage control method based on a deep neural network, comprising the following steps: acquiring real-time SCADA measurement information of a power grid in a controlled area; judging whether there is a voltage power over-limit problem in the AC substation in the controlled area according to the measurement information; if there is, determining all feasible control strategies according to the existing problem; inputting substation operation data in the measurement information into a deep neural network model which is constructed and trained in advance as input values corresponding to each feasible control strategy; outputting the AC substation bus voltage amplitude corresponding to each feasible control strategy, determining the AC substation bus voltage and power factor corresponding to each feasible control strategy; and determining the optimal control strategy according to the safety evaluation result. The method effectively solves the problems of high dimension, high dynamics, high nonlinearity and high randomness of a large power grid, and can be applied to unknown working conditions such as unanticipated faults and non-pre-set operation modes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid dispatching, in particular to an alternating current substation reactive voltage control method based on a deep neural network. BACKGROUND

[0002] With the continuous access of high-voltage AC / DC hybrid, high proportion of renewable energy, the gradual application of energy storage devices, and the change of power market rules and market participants' behavior, the power electronics characteristics of the power system are increasingly evident, and the uncertainty, dynamics and diversity of power grid operation have significantly increased. In addition, network security and frequent natural disasters and other external uncertainties increase the potential risk of safe operation of the power grid, bringing unprecedented challenges to the formulation and optimization of real-time control decisions of the power system. Under normal circumstances, the design and operation philosophy of the power grid is to ensure the safe and stable operation under N-1 (or part of N-k) conditions, and to develop relevant standards to examine a number of safety indicators including substation bus voltage, power factor, and line flow before and after the fault.

[0003] Therefore, real-time monitoring of substation abnormalities and taking quick and effective dispatching control measures are crucial to the safe and economic operation of the power grid. The current voltage and power factor control measures of the substation are mostly based on historical experience analysis to develop control measures. In complex scenarios such as prominent network structure contradictions, AC / DC interaction, and interaction between sending and receiving end power grids, the control strategies given are difficult to cope with the rapid changes in the state of the system, directly leading to overly conservative or optimistic control measures. Sensitivity analysis based on power grid models and other algorithms are usually difficult to calculate the most effective control strategy in real-time in the actual operation environment of the power grid due to the influence and constraints of the time-varying, highly nonlinear, dynamic and random nature of the large power grid, insufficient model accuracy, and limited online computing resources, and cannot achieve the best results. SUMMARY

[0004] The present application provides an alternating current substation reactive voltage control method based on a deep neural network to effectively control the voltage and power factor level of the alternating current substation, aiming at the problems existing in the prior art.

[0005] The technical solution of the present application is as follows:

[0006] An alternating current substation reactive voltage control method based on a deep neural network, comprising:

[0007] Obtaining real-time SCADA measurement information of the power grid in the controlled area;

[0008] According to the real-time SCADA measurement information, it is judged whether there is a voltage and power overrunning problem in an alternating current substation in a controlled area; if there is, all feasible control strategies are determined according to the existing voltage and power overrunning problem, then for each feasible control strategy, substation operation data in the real-time SCADA measurement information is input into a deep neural network model which is constructed and trained in advance as an input value, and the amplitude of the bus voltage of the alternating current substation corresponding to each feasible control strategy is output, and then the bus voltage and power factor of the alternating current substation corresponding to each feasible control strategy are determined; then the safety of the bus voltage and power factor of the alternating current substation of each feasible control strategy is evaluated, and the optimal control strategy is determined according to the evaluation result, and the control instruction is determined according to the optimal control strategy.

[0009] Further, the real-time SCADA measurement information includes the amplitude of the bus voltage of the substation, the active and reactive power values of the transmission line, the capacitor measurement value, the main transformer tap position, and the high and low side power measurement value of the main transformer.

[0010] Further, the substation operation data includes the photovoltaic field voltage setting value, the main transformer reactive power measurement value of the alternating current substation, the main transformer active power measurement value of the alternating current substation, the main transformer tap position of the alternating current substation, the line state of the alternating current substation, and the capacitor state of the substation.

[0011] Further, the specific steps of performing safety evaluation and determining the optimal control strategy according to the evaluation result are: calculating the Euclidean distance between the bus voltage and power factor corresponding to each feasible control strategy and its preset safety target, and selecting the control strategy corresponding to the minimum Euclidean distance as the optimal control strategy.

[0012] Further, the control strategy includes the switching of the capacitor in the alternating current substation, the change of the transformer tap, and the adjustment of the photovoltaic field voltage setting value.

[0013] Further, the construction and training method of the deep neural network model includes:

[0014] According to the set control target, a deep neural network model is constructed, and the input value of the deep neural network model is composed of the photovoltaic field voltage setting value, the main transformer reactive power measurement value of the alternating current substation, the main transformer active power measurement value of the alternating current substation, the main transformer tap position of the alternating current substation, the line state of the alternating current substation, and the capacitor state of the substation, and the output value is the amplitude of the bus voltage of the substation;

[0015] The historical SCADA data set is obtained, and the data in the historical SCADA data set is divided into input values and output values;

[0016] The input values and output values are normalized;

[0017] The historical SCADA data set is divided into a training set and a test set in a ratio of 80% and 20%, the training set is input into the constructed deep neural network model for training, and the test set is used to verify and improve the performance of the deep neural network model.

[0018] Further, the deep neural network model comprises an input layer, a set number of hidden layers and an output layer, each layer has a set number of neurons, and full connection layers are used to connect different layers of the neural network.

[0019] Compared with the prior art, the beneficial technical effects of the present application are:

[0020] The present application proposes a data-driven AC substation reactive voltage control method, which uses a deep neural network model to deeply mine the correlation between substation SCADA data, can effectively extract the state information of power grid operation and control, quickly output the optimal control strategy, and can ensure the voltage safety and power factor compliance of the power grid under various operating conditions. The most effective regulation and control strategy is calculated in the actual operation environment of the power grid.

[0021] The AC substation reactive voltage control method of the present application aims at the problems existing in the obtained SCADA data, obtains the AC substation bus voltage amplitude under all feasible control strategies, and then converts the AC substation bus voltage amplitude into the substation bus voltage and power factor. According to the safety evaluation result of the substation bus voltage and power factor, the optimal control strategy is determined. This method effectively solves the problems of high dimension, high dynamic, high nonlinearity, high randomness and the like of large power grids, and can be applied to unknown working conditions such as unanticipated faults and non-pre-set operating modes. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The control flowchart of the present application in the embodiment is shown in the figure;

[0023] Fig. 2(a) is a distribution of voltage amplitude prediction error in the training set;

[0024] Fig. 2(b) is a distribution of voltage amplitude prediction error in the verification set;

[0025] Fig. 2(c) is a distribution of voltage amplitude prediction error in the test set;

[0026] Figure 3 The power factor evaluation result in the embodiment is shown in the figure;

[0027] Figure 4 The regulation and control system flowchart provided in the embodiment is shown in the figure. DETAILED DESCRIPTION

[0028] Embodiment one:

[0029] The deep neural network-based alternating current substation reactive voltage control method of the embodiment comprises:

[0030] Real-time SCADA measurement information of the power grid in the controlled area is acquired; the real-time SCADA measurement information comprises substation bus voltage amplitude, transmission line active and reactive power values, capacitor measurement values, main transformer tap position, and main transformer high and low side power measurement values.

[0031] According to the real-time SCADA measurement information, it is determined whether there is a voltage and power over-limit problem in the alternating current substation in the controlled area; if there is, all feasible control strategies are determined according to the existing voltage and power over-limit problem, and then for each feasible control strategy, the substation operation data in the real-time SCADA measurement information is input into the deep neural network model that is pre-constructed and trained as an input value, and the alternating current substation bus voltage amplitude corresponding to each feasible control strategy is output. The substation operation data comprises photovoltaic field voltage set value, alternating current substation main transformer reactive power measurement value, alternating current substation main transformer active power measurement value, alternating current substation main transformer tap position, alternating current substation line state, and substation capacitor state. The control strategy comprises switching of the capacitor in the alternating current substation, change of the transformer tap, and adjustment of the photovoltaic field voltage set value.

[0032] Further, the alternating current substation bus voltage and power factor corresponding to each feasible control strategy are determined; then the alternating current substation bus voltage and power factor of each feasible control strategy are subjected to safety evaluation, and the optimal control strategy is determined according to the evaluation result, and the regulation and control instruction is determined according to the optimal control strategy. The specific steps of safety evaluation and determination of the optimal control strategy according to the evaluation result are as follows: the Euclidean distance between the substation bus voltage and power factor corresponding to each feasible control strategy and the preset safety target of the substation bus voltage and power factor is calculated, and the control strategy corresponding to the minimum Euclidean distance is selected as the optimal control strategy.

[0033] Embodiment two:

[0034] A further optional design of the embodiment is that the construction and training method of the deep neural network model are designed, and the construction and training method of the deep neural network model comprises:

[0035] Based on the set control objectives, a deep neural network model is constructed. This model includes an input layer, a predetermined number of hidden layers, and an output layer. Each layer has a predetermined number of neurons, and the different layers are connected by fully connected layers. The input values ​​of this deep neural network model consist of feature data comprising the photovoltaic power station voltage setpoint, reactive power measurement values ​​of the AC substation main transformer, active power measurement values ​​of the AC substation main transformer, tap position of the AC substation main transformer, AC substation line status, and substation capacitor status. The output value is the substation bus voltage amplitude.

[0036] Obtain the historical SCADA dataset and divide the data in the historical SCADA dataset into input and output values;

[0037] Standardize the input and output values;

[0038] The historical SCADA dataset was divided into training and test sets at a ratio of 80% and 20%, respectively. The training set was used to train the constructed deep neural network model, and the test set was used to verify and improve the performance of the deep neural network model.

[0039] Example 3:

[0040] To ensure voltage safety and power factor compliance of the power grid under various operating conditions, such as Figure 1 As shown, this embodiment provides a rapid control strategy for substation bus voltage and power factor exceeding limits. Within a designated power grid area, it adjusts capacitor switching, transformer tap changes, and photovoltaic power station voltage setpoints within the area to meet various safety requirements of substations in that area, including both base state and fault conditions. The method includes the following steps:

[0041] Step 1: Collect real-time SCADA measurement information of the power grid within the controlled area online and analyze the data quality (including missing and erroneous data). The collected SCADA measurement information includes substation bus voltage amplitude, active and reactive power values ​​of transmission lines, capacitor measurements, main transformer tap changer locations, and main transformer high, medium, and low voltage power measurements. If the bus voltage and power factor of all substations within the controlled area are within safe limits, no control operations will be performed in this step; the system will continue to wait for new SCADA sampling data. The sampling period for real-time SCADA measurement information can be selected by the user: 4 seconds, 30 seconds, 1 minute, 10 minutes, or 15 minutes.

[0042] Step 2: By analyzing real-time SCADA data, potential voltage and power factor exceedance issues in AC substations are detected. The safe voltage range and power factor range differ for substations of different voltage levels. Voltage and power factor exceedances are well-known information and will not be elaborated upon in this invention.

[0043] Step three, when the AC substation bus voltage amplitude or power factor exceeds the normal range, for each AC substation with voltage or power factor out of limits, the following steps are taken to regulate:

[0044] Extract the available regulation resources in the AC substation, including capacitors, photovoltaic station voltage setting, main transformer tap, etc.

[0045] For voltage or power factor out of limits, all feasible control strategies are traversed (including switching of capacitors in the AC substation, change of transformer tap, adjustment of photovoltaic station voltage setting) and the neural network model evaluation control effect is obtained using the deep neural network model, that is, the change of AC substation bus voltage and power factor after implementation of different control strategies is predicted. The evaluation of control effect needs to consider the bus voltage level and power factor level at the same time, and the corresponding safety range is quantitatively evaluated. The evaluation index is defined as the Euclidean distance between the current operating point and the control target, as shown in formula (1). The reaction in the data feature vector after the controller action is called the neural network model, which is used to estimate the voltage value and power factor after control. Figure 3

[0046] The control strategy with the optimal regulation performance (i.e., the shortest distance from the center point of the safety area in the AC substation bus voltage-power factor diagram after regulation) is screened out, as shown in formula (2), and the control strategy and the expected regulation effect are output. Figure 3

[0047] Further, the embodiment can also perform unified verification after outputting all substation regulation instructions, and send the regulation instructions to the equipment for closed-loop control.

[0048] The input of the above deep neural network is the "input data feature" (feature data), and the output is the predicted bus voltage amplitude. The feasible control strategy is actually part of the input data feature. By changing the control strategy, the input information of the neural network model is changed to predict the impact on the bus voltage after the change. That is, to predict the control effect. For example, the input information of the deep neural network model includes capacitor state and main transformer tap information, so the expected effect (change of bus voltage amplitude) of the control action can be predicted by switching capacitors and changing the main transformer tap position.

[0049] Embodiment four:

[0050] ​​The embodiment is further designed on the basis of Embodiment Three, and further includes a training step of a deep neural network, specifically including: collecting massive historical SCADA data to form a feature data vector, and training a deep neural network model for the controlled area using a deep learning algorithm. The selected data features include the photovoltaic power station voltage set value, the transformer reactive power measurement value, the transformer active power measurement value, the transformer tap position, the transformer line state, and the transformer capacitor state in the collectable SCADA data. The SCADA original data adopts a true value, and after collection, needs to be normalized, that is, for all transformer power measurement values, SBASE=100MVA is adopted; for the photovoltaic power station voltage set value, the base voltage is taken for normalization. The input feature data is shown in Table 1.

[0051] Table 1 Input data features

[0052]

[0053] The output data (as shown in Table 2) is the transformer bus voltage amplitude in the historical SCADA data, and the original value needs to be converted to a normalized value (normalized value=true value / reference value) according to the bus voltage reference.

[0054] Table 2 Input data

[0055] Output data Substation bus voltage magnitude [Vm1, Vm2,...]

[0056] The collected historical SCADA data needs to be uniformly stored in a data file according to the data timestamp. Different types of data with the same timestamp are stored in two data files according to the above input data feature table and output data format. Secondly, the following steps are used to train the deep neural network model:

[0057] Step one: read the “input data features” file and the “output data” file, and check the timestamp synchronization. Eliminate data items with different timestamps.

[0058] Step two: divide the data into a training set and a test set according to a certain proportion, wherein a subset in the training set can also constitute a validation set to verify the performance of the neural network model.

[0059] Step three: construct a multi-layer neural network, including an input layer, a hidden layer 1, a hidden layer 2… a hidden layer N, and an output layer, and specify the number of neurons in each layer. The different layers of the neural network are connected by full connection layers.

[0060] Step four: use the Adam optimization method, specify the sample training times (epoch), use the training set data to train and obtain the deep neural network model.

[0061] Step five: verify and improve the performance of the neural network using the test set. Performance indicators can be measured using prediction error indicators such as MAE, RMSE or MRE. The model training effect is shown in Figure 2, where Figure 2(a) shows the distribution of voltage amplitude prediction error in the training set; Figure 2(b) shows the distribution of voltage amplitude prediction error in the validation set; Figure 2(c) shows the distribution of voltage amplitude prediction error in the test set

[0062] The embodiment innovatively proposes a substation voltage and power factor automatic control method based on a deep neural network model, which can provide fast and accurate online autonomous control measures for regional substations to solve power grid safety problems. Its main technical characteristics include: deeply mining the correlation between substation SCADA data, effectively extracting the state information of power grid operation and control; with sub-second reaction speed; can provide online verification of control strategy effectiveness; effectively solve the problems of high dimension, high dynamic, high nonlinearity, high randomness, etc. of large power grids, and can be applied to unknown conditions such as uncovered expected faults and non-pre-set operation modes.

[0063] The application proposes an alternating current substation reactive power and voltage control method based on a deep neural network. The method provides a general platform for solving power grid dispatching problems. First, collect, process and analyze massive historical SCADA data to form data features for the deep learning model. Second, train the deep learning model to establish a matching model between the feature data and the substation bus voltage amplitude. By adjusting the training method and parameters, the deep learning model is obtained through offline training, so that the voltage amplitude prediction error is controlled within 1%. Third, use the deep learning model to predict the expected effect of different control instructions, so as to select the most effective control instruction, and finally form a substation real-time control instruction to control the capacitor switching, adjust the transformer tap position and the photovoltaic field station voltage setting value. In the real-time operation stage, the SCADA data is driven to collect the substation information in real time, detect the abnormality, and the deep learning model gives the optimal control strategy, which can optionally include auditing the control decision, that is, the control instruction, and finally forms an effective closed-loop control after auditing.

[0064] Embodiment five:

[0065] The embodiment provides an alternating current substation reactive power and voltage control system based on a deep neural network, as shown in Figure 4 , which includes a SCADA real-time data acquisition and analysis module, a deep neural network model, a real-time decision module and a dispatching instruction auditing module.

[0066] The SCADA real-time data acquisition and analysis module is used to obtain the SCADA measurement information of the controlled regional power grid;

[0067] determining whether the AC substation in the controlled area exists voltage and power factor out-of-limit problems; if the result is yes, traversing all feasible control strategies for the substation existing voltage and power factor out-of-limit problems, obtaining the SCADA specific data and parameter values of the substation and inputting them into the pre-trained deep neural network model;

[0068] The deep neural network model is used to output the substation bus voltage amplitude based on the input values;

[0069] The real-time decision module is used to determine the substation bus voltage and power factor according to the output substation bus voltage amplitude, evaluate the substation bus voltage and power factor based on the preset safety target, and finally determine the regulation and control instruction according to the evaluation result.

[0070] The regulation and control instruction auditing module is used to audit the regulation and control instruction generated by the decision module.

[0071] In the real-time operation stage, the SCADA data is driven, the substation information is collected in real time, the abnormality is detected, the optimal control strategy is given by the deep learning model, the regulation and control instruction is issued to the substation equipment after the audit, and finally the effective closed-loop control is formed.

[0072] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative but not restrictive. Those skilled in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, which are all within the protection of the present application.

Claims

1. A reactive power and voltage control method for AC substations based on deep neural networks, characterized in that: include: Acquire real-time SCADA measurement information of the power grid within the controlled area; Based on the real-time SCADA measurement information, it is determined whether there is a voltage or power overrun problem in the AC substation within the controlled area. If so, all feasible control strategies are determined based on the existing voltage or power overrun problem. For each feasible control strategy, the substation operation data in the real-time SCADA measurement information is input as the input value to a pre-built and trained deep neural network model, which outputs the AC substation bus voltage amplitude corresponding to each feasible control strategy, thereby determining the AC substation bus voltage and power factor corresponding to each feasible control strategy. Then, a safety assessment is performed on the AC substation bus voltage and power factor of each feasible control strategy, and the optimal control strategy is determined based on the assessment results. The control command is then determined according to the optimal control strategy. The substation operation data includes the photovoltaic power station voltage setpoint, the reactive power measurement value of the AC substation main transformer, the active power measurement value of the AC substation main transformer, the tap position of the AC substation main transformer, the line status of the AC substation, and the capacitor status of the substation. The specific steps for conducting a safety assessment and determining the optimal control strategy based on the assessment results are as follows: calculate the Euclidean distance between the substation bus voltage and power factor corresponding to each feasible control strategy and their preset safety target, and select the control strategy with the smallest Euclidean distance as the optimal control strategy. The control strategies include switching capacitors in AC substations, changing transformer taps, and adjusting the voltage setpoint of photovoltaic power plants.

2. The reactive power and voltage control method for AC substations based on deep neural networks according to claim 1, characterized in that: The real-time SCADA measurement information includes the substation bus voltage amplitude, active and reactive power values ​​of transmission lines, capacitor measurement values, main transformer tap position, and main transformer high, medium and low power measurement values.

3. The reactive power and voltage control method for AC substations based on deep neural networks according to claim 2, characterized in that: The construction and training methods of the deep neural network model include: Based on the set control objectives, a deep neural network model is constructed. The input value of the deep neural network model consists of feature data composed of photovoltaic power station voltage setpoint, AC substation main transformer reactive power measurement value, AC substation main transformer active power measurement value, AC substation main transformer tap position, AC substation line status, and substation capacitor status. The output value is the substation bus voltage amplitude. Obtain the historical SCADA dataset and divide the data in the historical SCADA dataset into input and output values; Standardize the input and output values; The historical SCADA dataset was divided into training and test sets at a ratio of 80% and 20%, respectively. The training set was used to train the constructed deep neural network model, and the test set was used to verify and improve the performance of the deep neural network model.

4. The reactive power and voltage control method for AC substations based on deep neural networks according to claim 1, characterized in that, The deep neural network model includes an input layer, a set number of hidden layers, and an output layer. Each layer has a set number of neurons, and the different layers of the neural network are connected by fully connected layers.

Images