Voltage sag domain identification system based on multiple uncertainty scene generation

Abstract

The invention discloses a voltage sag domain identification system based on multiple uncertainty scene generation. The voltage sag domain identification system comprises a processor, a photovoltaic monitoring device, an electric power monitoring device and regulation and control hardware, the processor is configured to collect photovoltaic output power information of each photovoltaic station based on the photovoltaic monitoring device deployed on each photovoltaic station on the power distribution network, and train a deep convolutional generative adversarial network based on the photovoltaic output power information; generating a photovoltaic output scene by using a deep convolutional generative adversarial network; and determining a voltage sag domain identification result corresponding to the given parameter and the sensitive load threshold based on the photovoltaic output scene, the given parameter, the sensitive load threshold and the structure parameter of the power system. According to the description, the accuracy of scene generation is improved, the voltage sag domain of distributed photovoltaic access is described more accurately, and the accuracy of voltage sag domain identification is improved.

Description

Technical Field

[0001] The present specification relates to the field of voltage sag domain identification, and in particular to a voltage sag domain identification system generated based on multiple uncertainty scenarios. Background Art

[0002] The voltage sag domain refers to the set of grid fault points that cause concern and affect the production equipment of sensitive users of the bus. At present, a variety of methods have been used to identify the voltage sag domain, such as the critical distance method, the fault point method and the analytical method. The critical distance method is to calculate the critical fault point that causes the voltage sag of sensitive loads in the distribution network and the distance between the point and the bus where the sensitive load is located, so as to identify the voltage sag domain. The fault point method refers to selecting some virtual fault points on the line, and obtaining the voltage sag amplitude information of the bus where the sensitive load is located when a short-circuit fault occurs at each fault point through the short-circuit calculation method, so as to determine whether the virtual fault point is in the voltage sag domain of the sensitive load. The analytical method is a method for calculating the voltage sag domain based on the analytical formula of the sag amplitude, which is proposed by comprehensively considering the advantages and disadvantages of the critical distance method and the fault point method and using the power system short-circuit fault for analysis.

[0003] Nowadays, the widespread access of distributed photovoltaics will bring uncertainty on the source side, making the voltage sag propagation present a new law. Its output will be affected by multiple complex factors such as light intensity, temperature changes, load fluctuations, and is volatile. At the same time, the occurrence of voltage sag events in the distribution network is random. Therefore, in the context of distributed photovoltaic access to the distribution network, the output fluctuations of multiple units and the randomness of voltage sag events will bring multiple uncertainties to the distribution network, increasing the difficulty of accurately identifying the voltage sag domain. The existing voltage sag domain identification is often studied for deterministic power sources, assuming that the output power of distributed photovoltaics remains unchanged, ignoring the fact that the output of distributed photovoltaics is affected by the actual environment and has a certain degree of randomness and time-varying. In addition, the output power of photovoltaic units is mainly related to the light radiation, and the meteorological conditions in the local area are correlated. Therefore, the output of distributed photovoltaics in the same area has a certain temporal and spatial correlation.

[0004] Existing photovoltaic output generation often regards photovoltaic units as independent and only generates output for a single independent photovoltaic unit, ignoring the correlation between units. This results in low accuracy of scene generation and makes existing voltage sag domain identification methods inapplicable. Summary of the invention

[0005] One of the embodiments of the present specification provides a voltage sag domain recognition system based on multiple uncertainty scenarios, the system comprising a processor, a photovoltaic monitoring device, a power monitoring device and control hardware; the processor is configured to: based on the photovoltaic monitoring devices deployed at each photovoltaic site on the distribution network, collect photovoltaic output power information of each photovoltaic site, and train a deep convolutional generative adversarial network based on the photovoltaic output power information; generate photovoltaic output scenarios using the deep convolutional generative adversarial network; determine the voltage sag domain recognition result corresponding to the given parameters and the sensitive load threshold based on the photovoltaic output scenario, given parameters, sensitive load threshold and structural parameters of the power system; generate a voltage reduction response instruction based on the voltage sag domain recognition result; and reduce the voltage. The response instructions include indicating the monitoring frequency, monitoring accuracy, upload interval period and alarm threshold of the power monitoring device; sending the voltage reduction response instructions to the power monitoring device in the voltage sag domain; the power monitoring device is configured to: collect the power parameter sequence in the distribution network with the monitoring frequency and monitoring accuracy, and upload the power parameter sequence to the processor with the upload interval period; determine whether the voltage data in the power parameter sequence is lower than the alarm threshold; in response to being lower than the alarm threshold, send an alarm signal to the processor; the processor is also configured to: determine the estimated impact load corresponding to the node where the power monitoring device is located based on the alarm signal; generate control instructions and send them to the control hardware corresponding to the estimated impact load, so as to control the control hardware to perform compensation work. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0007] Figure 1 An exemplary scenario diagram of a voltage sag domain identification system generated based on multiple uncertainty scenarios as shown in this specification;

[0008] Figure 2 A schematic diagram of a flow chart of a processor determining a voltage sag domain identification result as shown in this specification;

[0009] Figure 3 A schematic diagram of the flow of compensation work for the processor and control hardware shown in this specification;

[0010] Figure 4 A schematic diagram of the data format of the input deep convolutional generative adversarial network shown in this specification;

[0011] Figure 5 This is a schematic diagram of the structure of the deep convolutional generative adversarial network generator model shown in this specification;

[0012] Figure 6This is a schematic diagram of the structure of the deep convolutional generative adversarial network discriminator model shown in this specification;

[0013] Figure 7 This is a schematic diagram of the Newton-Raphson method power flow calculation process shown in this manual;

[0014] Figure 8 This is a schematic diagram for calculating the voltage sag amplitude shown in this manual;

[0015] Fig. 9 This is a schematic diagram of the golden section search process shown in this manual. DETAILED DESCRIPTION

[0016] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of this specification. For ordinary technicians in this field, this specification can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.

[0017] It should be understood that the "system", "device", "unit" and / or "module" used herein are a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the words can be replaced by other expressions.

[0018] Flowcharts are used in this specification to illustrate the operations performed by the system according to the embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed precisely in order. Instead, the steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or one or more operations may be removed from these processes.

[0019] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other. The present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0020] refer to Figure 1 The present specification provides a voltage sag domain identification system based on multiple uncertainty scenarios, which may include a processor 10, a photovoltaic monitoring device 20, a power monitoring device 30 and control hardware 40.

[0021] The processor 10 is used to process data and / or information obtained from the photovoltaic monitoring device 20, the power monitoring device 30 and the control hardware 40. In some embodiments, the processor 10 may include one or more hardware processors 10, such as a central processing unit, a programmable logic device, etc.

[0022] The photovoltaic monitoring device 20 is a device for monitoring the working condition of each photovoltaic site (eg, photovoltaic unit). In some embodiments, the photovoltaic monitoring device 20 may include but is not limited to a smart meter, a voltage / power sensor, a photovoltaic array monitor, and the like.

[0023] In some embodiments, the photovoltaic monitoring device 20 can be used to collect photovoltaic output power information of each photovoltaic site.

[0024] The power monitoring device 30 is disposed in the power distribution network and is used to obtain the operation information of the power distribution network (eg, voltage, current, etc.). In some embodiments, the power monitoring device 30 may include but is not limited to a voltage transformer, a current transformer, and a smart meter.

[0025] The control hardware 40 is a device for compensating the load. In some embodiments, the control hardware 40 may include one or more of a backup power supply, a dynamic voltage regulator, and a load switch.

[0026] In some embodiments, the processor 10, the photovoltaic monitoring device 20, the power monitoring device 30 and the control hardware 40 can be connected to each other by wire or wireless communication.

[0027] like Figure 1 As shown, in some embodiments, the processor 10 is configured to perform one or more of the following steps:

[0028] Step 110, based on the photovoltaic monitoring device 20 deployed at each photovoltaic site on the distribution network, the photovoltaic output power information of each photovoltaic site is collected, and a deep convolutional generative adversarial network is trained based on the photovoltaic output power information.

[0029] In some embodiments, the photovoltaic output power information is the output power corresponding to each photovoltaic site. In some embodiments, the processor 10 can collect photovoltaic output power information of each photovoltaic site within a period of time (such as within a year) through the photovoltaic monitoring device 20.

[0030] In some embodiments, the processor 10 trains a deep convolutional generative adversarial network based on photovoltaic output power information, so that the deep convolutional generative adversarial network after training can generate photovoltaic output scenarios, thereby simulating multiple uncertainty scenarios in the power distribution network. The photovoltaic output scenario includes photovoltaic output power, and the adversarial network can generate a photovoltaic output scenario that is more comprehensive than the actual photovoltaic output scenario.

[0031] The specific process of generating photovoltaic output scenarios using Deep Convolutional GAN ​​(DCGAN) is as follows:

[0032] First, preprocess the data: In this manual, the input data used to generate the photovoltaic output scenario is the photovoltaic output power information of N sites in one year. Assuming that there are 365 days in a year, the data format of the photovoltaic output power information of N sites in one year is converted into a three-dimensional matrix, and the data size is expressed as N×T×D, where N is the total number of photovoltaic sites; T is 24, indicating 24 hours of data; D is 365, indicating that there are 365 days in a year. In addition, the historical photovoltaic data is divided into rainy season photovoltaic data and dry season photovoltaic data. The photovoltaic power information of all sites on each day can be recorded as {x n,t} in the form of a two-dimensional matrix, where n is the number of sites, t represents the time, and x n,t Represents the photovoltaic output power information of the nth site at time t, and the input matrix form is shown in formula (1):

[0033]

[0034] Before inputting data into the model, it is also necessary to perform feature scaling on the data, that is, mapping the original data to a specific range, removing the unit and unifying the dimension, and becoming a pure data sample, that is, normalizing and standardizing the data. Data normalization is to limit the data to a certain range, usually mapping the data to the range of -1 to 1 or 0 to 1. The former is also called Min-Max normalization. The commonly used Min-Max normalization calculation formula is shown in formula (2):

[0035]

[0036] The data scaling method used by the scene generation model is Min-Max normalization. The data format obtained after the above operation is shown in formula (2).

[0037] Generative adversarial network is a deep learning framework that represents a class of generative models. The structure of the generative adversarial network model is as follows Figure 3 As shown in the figure, unlike the structure of general neural networks, the generative adversarial network is an adversarial framework composed of two interconnected neural networks. These two networks are generally composed of deep networks, one network is called the generator network, which is used to learn the unknown distribution of data samples and expect the generated samples to deceive the true and false judgment of the discriminator network; the other network is called the discriminator network, which is used to distinguish the real input samples from the generated samples as much as possible.

[0038] The training goal of the generator is to make the probability distribution P of the generated sample G(z) G As close as possible to the probability distribution P of the real sample x r The training goal of the discriminator is to determine whether its input is a real sample or a generated sample as accurately as possible, and the discrimination result will be fed back in the form of a gradient function to optimize the network structure of the generator and the discriminator.

[0039] During the training process of these two networks, one of them is fixed to update the parameters of the other, and this process is repeated over and over again. The discriminator network feeds back the judgment results of the input samples to the generator network, so that the generator network continuously adjusts its own parameters to generate more realistic samples. The discriminator network adjusts its own parameters based on the continuously generated samples to better discover the differences in input samples. The generator network continuously improves the quality of generated samples, while the discriminator network continuously improves its discrimination ability. After the above-mentioned adversarial training, the generator network finally captures the potential distribution of real data and reaches a Nash equilibrium state, that is, the discriminator D cannot determine whether its input is a real sample or a generated sample. At this time, it can be considered that the generator G has learned the probability distribution of real samples, that is, the generated samples G(z) of the generator G and the real samples both obey the probability distribution P r .

[0040] The generator (G) and discriminator (D) of the generative adversarial network are not expressed in the form of explicit functions, and the choice of network structure is very flexible. For the convenience of expression, it is necessary to define the relevant symbols. The distribution law of historical data x is denoted as P r , and define a noise variable z. This noise variable is usually derived from an existing simple distribution P z The sampling is performed using, for example, a Gaussian distribution or a uniform distribution.

[0041] The generator performs upsampling through a fully connected layer network and transposed convolution, mapping the noise variable z to a generated data space G(z), and denoting its data distribution as P G . Because the noise input to the generator is a random variable, every time a new noise z is input, the output is also a new variable. The input to the discriminator is the real sample x or the generated sample G(z). The input sample is downsampled through the fully connected layer network, and the discriminator outputs a probability value to judge the input sample.

[0042] The goal of the generative adversarial network is to generate realistic samples that are difficult for the discriminator network to distinguish between true and false. To achieve this goal, it is necessary to define the loss function L for the generator network training. G And the loss function L of the discriminator network training GWhen the discriminator network is fixed and the generator network is trained, the probability of the generated sample G(z) should be larger, and the loss function L G Therefore, the generator network can use the loss function defined by formula (2):

[0043]

[0044] The training goal of the discriminator network is to distinguish P as much as possible. r and P G When the parameters θ of the generator network are fixed G , update the discriminator network parameters θ D When , the greater the difference in discrimination, the stronger the judgment ability of the discriminator network is, that is, maximizing the difference between E[D(·)] and E[D(G(·))]. According to the basic principle of generative adversarial networks, when the discriminator network can distinguish the input samples well, its loss function L D The smaller it is. Therefore, the loss function of the discriminator network can be defined as formula (3):

[0045]

[0046] The zero-sum game between the two networks is denoted as V(D,G). To optimize the loss function, both networks need to update the parameters of their own network based on the updates of the other network. The two networks are continuously trained alternately and their parameters are continuously updated until the adversarial network reaches Nash equilibrium. When the training converges, a pair of optimal parameters can be obtained. Because there are multiple extreme points in the high-dimensional data space, the parameters at this time and are two extreme points of V(D,G). The principle of generative adversarial network is a minimax optimization problem, which can be defined as formula (4):

[0047]

[0048] In the formula, V(D,G) represents a binary cross entropy function, the ultimate goal of which is to minimize the probability distribution P of the generated samples. G And the true sample probability distribution P r JS between ( Jensen-Shannon ) divergence.

[0049] DCGAN combines convolutional neural networks to process images. The network structures of the DCGAN generator model and discriminator model are as follows: Figure 4 and Figure 5 shown.

[0050] Compared with traditional generative adversarial networks (GAN), DCGAN has been modified in structure to make the combination of GAN and CNN in DCGAN more stable. The characteristics of deep convolutional generative adversarial networks include: using a fully convolutional network, canceling the fully connected layer, and using batch normalization.

[0051] Use a fully convolutional network: Replace the deterministic spatial pooling function (maxpooling) with a strided convolution, allowing the network to learn its own spatial downsampling. Use this approach in the generator to enable it to learn its own spatial upsampling and discriminator.

[0052] Eliminate fully connected layers: The strongest example is global average pooling, which has been used in state-of-the-art image classification models. Global average pooling improves model stability but hurts convergence speed. The middle ground of connecting the top convolutional features directly to the input and output of the generator and discriminator respectively works well. The first layer of the GAN takes uniform noise Z as input and can be called fully connected because it is just a matrix multiplication, but the result is reshaped into a 4D tensor and used as the start of a convolution stack. For the discriminator, the last convolutional layer is flattened and then fed into a single sigmoid output.

[0053] Use batch normalization: Transform the input of each unit to zero mean and unit variance to stabilize learning. This has proven to be a very important means of accelerating convergence and mitigating overfitting in deep learning. This helps deal with training problems that arise due to improper initialization and helps gradients flow in updated models. It turns out that this is crucial for getting deep generators to start learning, preventing the generator from collapsing all samples to a single point, a common failure mode observed in GANs. However, directly applying batch normalization to all layers can lead to sample oscillations and model instability, so batch normalization is only used on the output layer of the generator and the input layer of the discriminator.

[0054] Step 120, generating a photovoltaic output scenario using a deep convolutional generative adversarial network;

[0055] In some embodiments, after performing the data preprocessing in step 110, the processor can use random sampling noise as the input of the DCGAN generator and the historical output data of each photovoltaic site as the input of the discriminator. The deep convolutional generative adversarial network is trained using the training data until a Nash equilibrium is reached. When the DCGAN training is completed, a photovoltaic output scene similar to the real data generated by the generator is obtained.

[0056] Step 130 , based on the photovoltaic output scenario, given parameters, sensitive load thresholds and structural parameters of the power system, determine a voltage sag domain identification result corresponding to the given parameters and the sensitive load thresholds.

[0057] In some embodiments, the given parameter may be a power system distribution network parameter or a given distribution network parameter. The given parameter may include resistance, reactance, conductance, susceptance, reactive power, active power, generator parameters, transformer parameters, transmission line parameters, load parameters, etc.

[0058] For electrical equipment (load) in the power distribution network, if the voltage changes or changes suddenly, it will not work properly or its function will decrease. Such electrical equipment is called sensitive load. A short-term drop in voltage may affect the normal operation of sensitive loads and may even cause damage or production interruption. In some embodiments, the sensitive load threshold can be a value set according to manual experience or actual requirements of the power distribution network. When the working voltage of the sensitive load is lower than the sensitive load threshold, it means that the operation of the sensitive load may be affected.

[0059] In some embodiments, the structural parameter of the power system may be a topological parameter of the power system.

[0060] The voltage sag domain identification result is the voltage sag domain under the current photovoltaic output scenario. The voltage sag domain identification result may include one or more fault points corresponding to sensitive loads. Sensitive loads are electrical equipment that are affected by voltage and thus fail.

[0061] In some embodiments, the processor 10 can determine multiple voltage sag domains corresponding to given distribution network parameters and sensitive load thresholds based on photovoltaic output scenarios, given parameters, sensitive load thresholds, and structural parameters of the power system through a preset formula or algorithm, and determine a voltage sag domain identification result based on the multiple voltage sag domains. For more descriptions on determining the voltage sag domain identification result based on multiple voltage sag domains, please refer to the relevant description below.

[0062] Step 140: Generate a voltage reduction response instruction based on the voltage sag domain identification result.

[0063] In some embodiments, the voltage reduction response instruction includes indicating the monitoring frequency, monitoring accuracy, upload interval period, and alarm threshold of the power monitoring device 30. The processor 10 can generate a voltage reduction response instruction corresponding to the power monitoring device 30 at the position of the voltage sag domain corresponding to the voltage sag domain identification result. In some embodiments, the power monitoring device 30 can increase the monitoring intensity of the power distribution network based on the voltage reduction response instruction.

[0064] In some embodiments, the monitoring frequency refers to the number of times the power monitoring device 30 collects relevant power parameters within a preset time period (such as within 1 minute), the monitoring accuracy refers to the amount of data collected by the power monitoring device 30 during a single collection, and the upload interval period refers to the interval between two consecutive uploads of the collected relevant power parameters by the power collection device. For example, the voltage reduction response instruction may include instructing the power monitoring device 30 to collect relevant power parameters once per second and upload data to the processor 10 once every 30 seconds.

[0065] In some embodiments, the alarm threshold is a threshold used by the power collection device to determine whether an alarm needs to be issued, and can be determined by an expert. For example, when the power collection device is performing data collection, if the voltage value in the collected relevant data is less than the alarm threshold, there is no need to wait for the next upload cycle, and the power collection device immediately uploads the current data and can also send an alarm signal to the processor 10 according to the situation.

[0066] Step 150: Send the voltage reduction response instruction to the power monitoring device 30 in the voltage sag domain.

[0067] In some embodiments, the processor 10 may determine a plurality of power monitoring devices 30 within a voltage sag domain according to the location of the power distribution network, and send a voltage reduction response instruction to the power monitoring devices 30 .

[0068] refer to Figure 2 In some embodiments, after receiving the voltage reduction response instruction, the power monitoring device 30 is configured to perform one or more of the following steps:

[0069] Step 210 , collecting a power parameter sequence in the power distribution network at the monitoring frequency and the monitoring accuracy, and uploading the power parameter sequence to the processor 10 at the upload interval period.

[0070] In some embodiments, the relevant power parameters collected by the power monitoring device 30 can be used to construct a power parameter sequence in time order, and the elements in the sequence correspond to the collected parameter values.

[0071] Step 220: Determine whether the voltage data in the power parameter sequence is lower than the alarm threshold.

[0072] In some embodiments, when the voltage data in the power parameter sequence is lower than the alarm threshold, it indicates that sensitive loads may be affected and further processing is required.

[0073] Step 230 , in response to the temperature falling below the alarm threshold, sending an alarm signal to the processor 10 .

[0074] The alarm signal is used to remind the affected user (load) to adjust the power consumption characteristics of the sensitive load, for example, to remind the user to turn off the sensitive load, reduce the load of the sensitive load, etc. For more description of the alarm threshold, please refer to the relevant content of step 150 above.

[0075] In some embodiments, in order to reduce the impact on sensitive loads, after receiving the alarm signal, the processor 10 is further configured to perform one or more of the following steps:

[0076] Step 240: Determine the estimated impact load corresponding to the node where the power monitoring device 30 is located based on the alarm signal.

[0077] The estimated impact load is one or more sensitive loads that are expected to be affected in the voltage sag domain. In some embodiments, the processor 10 can determine the estimated impact load corresponding to the node where the power monitoring device 30 is located based on the alarm signal combined with the power monitoring device 30 and the position relationship of the sensitive load in the power distribution network. For example, all sensitive loads on a certain line in the voltage sag domain are taken as the estimated impact load.

[0078] Step 250 , generating a control instruction and sending it to the control hardware 40 corresponding to the estimated impact load, so as to control the control hardware 40 to perform compensation work.

[0079] In some embodiments, the control instructions may include working parameters or instructions for different devices according to the actual situation of the control hardware 40. The compensation work is to control the control hardware 40 to work based on the working parameters in the control instructions, thereby compensating for the estimated impact load to reduce the impact on the sensitive load.

[0080] In some embodiments, the control hardware 40 may include one or more of a backup power supply, a dynamic voltage regulator, and a load switch; the control instructions include one or more of a backup power supply instruction, a voltage self-regulation instruction, and a switch instruction.

[0081] The backup power supply is configured to: after receiving the backup power supply instruction, provide power supply corresponding to the estimated impact load; the dynamic voltage regulator is configured to: after receiving the voltage self-adjustment instruction, provide a compensation voltage corresponding to the estimated impact load; the load switch is configured to: after receiving the switch instruction, control the connection and disconnection of the estimated impact load and the distribution network.

[0082] The estimated impact loads in the voltage sag domain are processed accordingly by the control hardware 40, so that in some cases, the sensitive loads can continue to work normally or avoid damage to the sensitive loads caused by voltage fluctuations.

[0083] This specification proposes a voltage sag domain recognition system based on multiple uncertainty scenarios, using deep convolutional generative adversarial networks to track and simulate photovoltaic output characteristics, focusing on the inherent laws of distributed photovoltaic output, and mining the spatiotemporal correlation of the outputs of multiple distributed photovoltaic units. On the basis of taking into account the randomness and correlation of the output of distributed photovoltaic units, artificial data with the same dimension and similar probability distribution as the original photovoltaic output data is generated.

[0084] In some embodiments, the processor 10 is further configured to perform one or more of the following steps:

[0085] Based on the photovoltaic output scenario, the voltage of each node of the distribution network before the fault is determined using the Newton-Raphson method power flow calculation method;

[0086] The current bus is obtained from the lines of the distribution network. According to the given parameters, the impedance matrix of each node of the distribution network is calculated by adding branches or inverting the admittance matrix. Based on the voltage of each node before the fault and the impedance matrix of each node, the voltage sag amplitude of the node where the sensitive load is located is determined when different types of short-circuit faults occur at each node on the current bus.

[0087] Based on the voltage sag amplitude at the node where the sensitive load is located when different types of short-circuit faults occur at each node of the current bus, the voltage sag amplitude vector of the bus node is determined; the voltage sag amplitude vector of the bus node is compared with the sensitive load threshold to obtain a difference vector; and based on the difference vector, the line association vector is obtained;

[0088] Based on the line association vector, determine whether the current bus meets the analysis conditions;

[0089] In response to not being satisfied, the current bus is not within the voltage sag region, and the next line is selected from the power distribution network as the current bus to continue the determination;

[0090] In response to satisfying: the current bus is analyzed by Newton quadratic interpolation method to obtain an analytical expression for the voltage sag amplitude and a fault voltage equation of the current bus; and based on the fault voltage equation, the portion of the current bus within the voltage sag domain is obtained, and the next line is selected from the distribution network as the current bus to continue the judgment;

[0091] When the traversal and iteration of all lines in the distribution network are completed, multiple voltage sag domains corresponding to the given parameters and sensitive load thresholds are obtained, and the union of the multiple voltage sag domains is the voltage sag domain identification result.

[0092] In some embodiments, the voltage of each node of the power distribution network before a fault is determined using the Newton-Raphson method power flow calculation method, including:

[0093] After the deep convolutional generative adversarial network generates the photovoltaic output scenario, the voltage of each node before the fault is obtained using the Newton-Raphson method for power flow calculation. The Newton-Raphson method is a numerical method for efficiently solving nonlinear equations. When solving power flow calculation problems, the Newton-Raphson method is mainly used for solution calculation, and the function is gradually linearized through the Taylor series. This method has the characteristics of fast convergence speed and small error, and is an efficient method for solving complex nonlinear equations. Assume that the expression of the n-dimensional nonlinear equation system is:

[0094]

[0095] By linearly expanding the stepwise objective function through the first-order Taylor, the equation can be obtained:

[0096]

[0097] Will and Y ij Substituting the power equation of active power and reactive power into the equation, we can get the corresponding nonlinear equation expression of the power network:

[0098]

[0099] Further optimization calculations can obtain the modified equation based on the Newton-Raphson method:

[0100]

[0101] Where: is the node voltage; Y ij =G ij +jB ij is the node admittance.

[0102] like Figure 6 As shown, the calculation process is as follows:

[0103] Based on the clear grid structure parameters, the admittance matrix is ​​formed by substituting actual data, and the offset of each parameter (such as node voltage, reactive power, active power, etc.) is calculated according to the given initial value.

[0104] In response to the error meeting the error precision, the loop is jumped out and the result is output; in response to the error not meeting the error precision, the Jacobian matrix is ​​continued to be solved, and the correction voltage is solved by the correction equation to continue the calculation. Until the accuracy requirement is met, the loop is ended and the result is output, otherwise the loop process is still carried out.

[0105] In some embodiments, the processor 10 is further configured to:

[0106] First, according to the given parameters, the impedance matrix of each node in the network is calculated by adding branches or inverting the admittance matrix. Then, according to formulas (9) to (15), the voltage drop amplitude of the node where the sensitive load is located when different types of short-circuit faults occur at each node is obtained. Figure 7 As shown in FIG. 1 , a schematic diagram for calculating the voltage sag amplitude is shown.

[0107] In some embodiments, different types of short circuit faults include three-phase short circuit fault, single-phase ground short circuit fault, two-phase short circuit fault and two-phase ground short circuit fault. Assume that there is a fault point K:

[0108] When a three-phase short circuit fault occurs at the fault point K, since it is a symmetrical balanced fault, only the positive sequence is considered, and the three-phase voltage sag amplitudes at the sensitive load node S are the same, and the voltage sag amplitude of one phase is:

[0109]

[0110] Where: is the voltage sag amplitude at point S; is the positive sequence self impedance; is the positive sequence mutual impedance; is the voltage before the fault at point S; is the voltage before the fault at point K.

[0111] From formula (9), we can know that the voltage sag amplitude at point S is It is related to the system structure parameters and the system operation status before the fault. The operation status before the fault refers to whether the distribution network is operating normally, which can be obtained by the flow calculation before the fault. When an asymmetric short circuit fault occurs at the fault K with phase A as a special phase, the symmetrical component method needs to be used for analysis.

[0112] When a single-phase ground short circuit fault with phase A as the special phase occurs at the fault point K, the voltage sag amplitudes at point S are as follows:

[0113]

[0114] When a two-phase short circuit fault occurs in phases B and C, the three-phase voltage temporary drop amplitudes at point S are as follows:

[0115]

[0116] When a two-phase ground short circuit fault occurs between phases B and C, the three-phase voltage temporary drop amplitudes at point S are as follows:

[0117]

[0118] In formula (9) to formula (12): are the three-phase voltage sag amplitudes at the sensitive node S respectively; is the positive, negative and zero sequence self-impedance and mutual impedance; α=e j120° is the rotation factor.

[0119] When a short-circuit fault occurs on the system bus, its self-impedance and mutual impedance can be directly called from the node impedance matrix; however, when a short-circuit fault occurs at a certain point on the line, the calculation of its self-impedance and mutual impedance needs to be obtained by introducing the position variable p (0≤p≤1) and combining the node impedance matrix. Point S is the bus node where the sensitive load is located, and Z C is the sequence impedance of line FT. When the fault point K moves on line FT, its three-sequence self-impedance And the three-sequence mutual impedance between the fault point K and S All can be represented by the impedance matrix Z and the position variable p:

[0120]

[0121]

[0122] Where: are the self-impedances of each sequence at the system bus nodes F and T respectively; They are the mutual impedances of the system bus nodes F and T, the mutual impedances of the bus F and the sensitive load node S, and the mutual impedances of the bus T and the sensitive load node S, which can all be called from the system node impedance matrix; is the line sequence impedance between nodes F and T.

[0123] In addition, the voltage before the fault at the fault point K It can be represented by the fault location variable p:

[0124]

[0125] Therefore, the voltage sag amplitude at point S can be expressed by the pre-fault voltage and each sequence impedance. Substituting equations (13) to (15) into equations (9) to (12), we can obtain the function U(p) of the voltage sag amplitude at point S with respect to the position variable p when any short-circuit fault occurs at any fault point.

[0126] In some embodiments, based on the voltage sag amplitude at the node where the sensitive load is located when different types of short-circuit faults occur at each node of the current bus, determining the line association vector includes:

[0127] The voltage sag amplitude vector of the node where the sensitive load is located is obtained from the above, and the node determination vector B and line association vector L are calculated according to formula (11), and all lines are traversed one by one.

[0128] Calculate the voltage sag amplitude of sensitive load node S when each busbar short circuit fault occurs, and form an n-dimensional vector of busbar node voltage sag amplitude And compare it with the voltage sag threshold of sensitive load to get the difference vector ΔU S .

[0129]

[0130] By judging ΔU S The positive or negative value of the vector elements can tell whether each node is within the temporary drop region of the sensitive node. Therefore, the node judgment vector B is introduced:

[0131]

[0132] B i =1 means that busbar i is within the voltage sag range corresponding to the sensitive load; B i = 0 means that bus i is outside the voltage sag domain corresponding to the sensitive load. Therefore, in order to determine whether each line is included in the voltage sag domain, the line association vector L is introduced:

[0133]

[0134] Where: B i_F and B i_T It is the determination factor of temporary drop of the busbar of the connecting line.

[0135] In some embodiments, based on the line association vector, it is determined whether the current bus satisfies the analysis conditions. In response to not satisfying the conditions, the current bus is not within the voltage sag domain, and the next line is selected from the distribution network as the current bus. In response to satisfying the conditions: the current bus is analyzed by Newton's quadratic interpolation method to obtain an analytical expression for the voltage sag amplitude and a fault voltage equation of the current bus; and, based on the fault voltage equation, the portion of the current bus within the voltage sag domain is obtained, and the next line is selected from the distribution network as the current bus.

[0136] The analysis conditions include: if L i =0, indicating that the line is not in the voltage sag region, and the next line is calculated directly; if L i =1, indicating that one of the first and last nodes of this line is in the temporary drop region. It can be known that part of line i is in the temporary drop region and there is a unique critical point on the line. i =2, indicating that the first and last nodes of this line are both within the voltage sag region. The golden section search method is first used to solve the maximum value U of the voltage sag amplitude curve. max and compare it with the sensitive load voltage sag threshold U th For comparison. max th ​, indicating that this line is completely in the temporary drop region and there is no critical point, then the next line is calculated directly; if U max ≥U th , indicating that the line section i is located in the temporary drop region and there are two critical points on the line.

[0137] The whole process of the golden section search method is to find p a ≤p≤p b (p a =0,p b =1) the point p at which the voltage sag at the PCC point is the maximum value max . Use p max As interpolation points, it can be ensured that the interpolation curve passes through the maximum value of the voltage sag amplitude, providing a good initial value point for secant iteration. Figure 8 shown.

[0138] If the current busbar meets the analysis conditions, calculate the critical point.

[0139] The critical point refers to the fault location in the distribution network that causes the voltage sag amplitude at the bus where the sensitive load is located to be equal to the voltage sag threshold of the sensitive load. i =1, 2, the critical point is calculated. i =1, directly use (0,U F )、(0.5,U 0,5 ) and (1,U T )3 points are used as interpolation points, and the voltage sag amplitude analytical formula U(p) and the fault voltage equation are obtained by Newton quadratic interpolation method. L i =2, use (0,U F )、(p max ,U max ) and (1,U T )3 points are used as interpolation points of Newton's quadratic interpolation method to obtain the voltage sag amplitude analytical expression and the fault voltage equation.

[0140] Then the root p of the above fault voltage equation is ia As the initial iteration value of the secant iteration method, the accurate critical point position is obtained through iteration. The iterative expression of the critical fault distance based on the secant iteration method is as follows:

[0141]

[0142] Its convergence condition is:

[0143] ||U(p k+1 )|-U th |<ε (20)

[0144] The secant iteration method is an algorithm that approximates the exact root, and it does not need to calculate the differential of the voltage sag amplitude equation of the sensitive load, so it is more efficient than the Newton iteration method.

[0145] When the traversal and iteration of all lines in the distribution network are completed, multiple voltage sag domains corresponding to the given parameters and sensitive load thresholds are obtained, and the union of the multiple voltage sag domains is the voltage sag domain identification result corresponding to the photovoltaic output scenario.

[0146] The voltage sag domain identification system generated based on multiple uncertainty scenarios described in this specification is based on short-circuit calculation, uses quadratic interpolation method and numerical solution of nonlinear equations, optimizes the sag domain calculation process, classifies and calculates different lines, and reduces the calculation steps.

[0147] Compared with the existing voltage sag domain identification method, this method takes into account the impact of the dynamic output of distributed photovoltaic units on the voltage sag characteristics, as well as the randomness and spatiotemporal correlation of photovoltaic output itself, thereby improving the accuracy of scene generation, making the voltage sag domain characterization of distributed photovoltaic access more accurate, and improving the accuracy of voltage sag domain identification.

[0148] The processor 10 may also directly determine the voltage sag domain identification result through a machine learning model. In some embodiments, the processor 10 is further configured to:

[0149] According to given parameters, power parameter sequence, photovoltaic output power information, line basic data, future light data and sensitive load threshold, a power characteristic map is constructed; the power characteristic map is input into the voltage sag risk model to obtain the voltage sag risk area; and multiple voltage sag domains corresponding to the given parameters and sensitive load thresholds in the voltage sag risk area are calculated to determine the voltage sag domain identification result.

[0150] In some embodiments, the line basic data is the line characteristics of the distribution network, which can be determined by accessing the distribution network configuration information. The line basic data may include one or more of the line physical data, the types of each intersection, the intersection connection method, the load type, and the photovoltaic equipment class. In some embodiments, the future light data includes the light intensity at different times within a period of time (such as within a day), and the future light data can be obtained by accessing an external weather database. For more descriptions of given parameters, power parameter sequences, photovoltaic output power information, and sensitive load thresholds, please refer to the relevant content above.

[0151] The power characteristic graph is a graph consisting of nodes and edges between nodes. In some embodiments, the nodes of the power characteristic graph include sensitive load nodes, historical fault nodes, line crossing nodes, and photovoltaic access nodes. A historical fault node refers to a node where a fault has occurred at the location of the circuit in history, and is determined as a historical fault node; a line crossing node refers to a node at the intersection of a busbar and a branch line in a distribution network.

[0152] In some embodiments, the attributes of the nodes in the power characteristic graph may include power parameters, historical fault records, and regional environmental data.

[0153] The power parameters are the current and historical voltage, current, etc. of the node. The power parameters can be acquired based on the power parameter sequence corresponding to the node.

[0154] The historical fault records may include the fault cause, fault type, and fault time, etc. The historical fault records may be obtained by reading the running database; if the node has never failed, the historical fault records may be empty.

[0155] The regional environmental data refers to the environmental data of the region where the node is located. The regional environmental data may include temperature, precipitation, etc., and may be obtained through sensors or by accessing an external weather database.

[0156] In some embodiments, according to different node types, the nodes in the power characteristic graph may also include additional node attributes.

[0157] In some embodiments, when the node is a sensitive load node, the node attributes may further include: load type, historical voltage sag records, and sensitive load threshold.

[0158] The load type indicates the type of specific load connected to the node (such as whether the electrical equipment is a CNC machine tool, a server, or an X-ray machine, etc.). The load type can be obtained by accessing the database.

[0159] The historical voltage sag record is a record of the corresponding voltage sag when the performance of sensitive load equipment decreases or fails; the voltage sag record may include the cause of the sag (such as the type of fault), the magnitude of the sag, the duration of the sag, the time of the sag, etc. The historical voltage sag record can be obtained by accessing the operation database.

[0160] In some embodiments, when the node is a line intersection node, the node attributes may also include: intersection type (such as T-type, cross type), intersection connection method (such as direct connection, connection through a switching device, or connection through a transformer), and power distribution parameters; wherein the intersection type and intersection connection method may be obtained by accessing a database, and the power distribution parameters may include information such as voltage and current corresponding to the upstream and downstream lines of the node, which may be obtained through a power parameter sequence.

[0161] In some embodiments, when the node is a photovoltaic access node, the node attributes may also include: photovoltaic output power information, historical light data, photovoltaic equipment type, and future light data. Similar to future light data, historical light data may also be obtained by accessing an external weather database.

[0162] In some embodiments, the edge of the power characteristic graph corresponds to the power line between the nodes. In some embodiments, the edge of the power characteristic graph may have a direction consistent with the current in the power line. The attributes of the edge in the power characteristic graph may include line physical data (such as the length, cross section, material, etc. of the power line) and line power parameters in the power parameters. The line physical data and power parameters can be obtained by accessing a database. In some embodiments, the voltage sag risk model may be a GNN (Graph Neural Networks), which may be trained by training data. The input of the voltage sag risk model is a power characteristic map, and the output of the model is a voltage sag risk area.

[0163] In some embodiments, the voltage sag risk model can be obtained by training multiple labeled training data. Specifically, multiple labeled training data can be input into the voltage sag risk model, and a loss function can be constructed by using the labels and the results of the initial voltage sag risk model. The parameters of the initial voltage sag risk model are iteratively updated based on the loss function by gradient descent or other methods. When the preset conditions are met, the model training is completed and a trained voltage sag risk model is obtained. The preset conditions can be that the loss function converges, the number of iterations reaches a threshold, etc.

[0164] In some embodiments, the training sample includes a sample power characteristic graph constructed based on sample given parameters, sample power parameter sequence, sample photovoltaic output power information, sample line basic data, sample future light data, and sample sensitive load threshold. The sample power characteristic graph is similar to the power characteristic graph, which is a graph composed of nodes and edges between nodes. The nodes of the sample power characteristic graph include sample sensitive load nodes, sample historical fault nodes, sample line crossing nodes, and sample photovoltaic access nodes; the edges of the sample power characteristic graph are edges between sample nodes.

[0165] In some embodiments, the label corresponding to the training sample is the area of ​​the voltage sag domain corresponding to the sample sensitive load, and the label can be determined based on the actual scene corresponding to the sample power characteristic diagram or determined through software simulation.

[0166] In some embodiments, after the voltage sag risk model outputs a voltage sag risk area, the processor 10 may determine whether the bus in the voltage sag risk area meets the analysis conditions based on the method described above, obtain multiple voltage sag domains, and then obtain a voltage sag domain identification result.

[0167] The voltage sag risk model is used to determine the voltage sag risk area with a relatively high voltage sag risk, so that the processor 10 does not need to traverse all the lines in the distribution network, which greatly improves the speed of determining the voltage sag domain identification result and improves the response rate of the system.

[0168] In some embodiments, the processor 10 is further configured to determine a sensitive load threshold based on the load type, load aging degree, and load rated parameters of the sensitive load. In some embodiments, the load aging degree can be represented by the accumulated working hours of the sensitive load, and the load type and load rated parameters can be determined by accessing a database.

[0169] In some embodiments, the processor 10 can construct a sensitive load characteristic vector based on the load type, load aging degree, and load rated parameters of the sensitive load. The load type, load aging degree, and load rated parameters can respectively correspond to a dimension of the sensitive load characteristic vector. In some embodiments, the sensitive load characteristic vector can also include other dimensions, such as the dimension corresponding to the current operating temperature of the load.

[0170] In some embodiments, the processor 10 may determine the sensitive load threshold by performing vector matching in a vector database based on the sensitive load feature vector. The vector database is constructed by multiple groups of reference sensitive feature vectors and corresponding reference sensitive load thresholds. The processor 10 may match the reference sensitive feature vector matched with the sensitive load feature vector, and use the reference sensitive load threshold corresponding to the reference sensitive feature vector as the sensitive load threshold.

[0171] In some embodiments, the vector matching method can be to select a reference sensitive feature vector with the closest vector distance (such as Euclidean distance or cosine distance, etc.) as the matching reference sensitive feature vector. In some embodiments, other vector matching methods can also be used, such as fast similarity search, etc. In some embodiments, multiple groups of reference sensitive feature vectors and corresponding reference sensitive load thresholds in the vector database can be obtained in the following manner: first, experiment on the reference sensitive load and record its sensitive load type, load aging degree, and load rated parameters to construct a reference sensitive feature vector; then, during the experiment, experiment on the reference sensitive load with voltage drops of different magnitudes, and record the sensitive load threshold corresponding to the load as the reference sensitive load threshold. Exemplarily, assuming that the rated voltage of the reference sensitive load is 220V, its operating voltage can be gradually reduced, such as to 219V, 218V, etc. for experimentation. If it is found that the sensitive load cannot work normally (or is abnormal) when the voltage is lower than 215V, the reference sensitive load threshold is 215V.

[0172] By constructing a vector database, factors such as sensitive load type and load aging degree can be taken into account, and the sensitive load threshold can be determined more conveniently and accurately.

[0173] In order to obtain a voltage sag risk model with better performance and improve the learning effect of the training data, in some embodiments, the processor 10 may train the voltage sag risk model using a two-stage training method.

[0174] In some embodiments, the training process of the voltage sag risk model includes an initial training stage and an intensive training stage respectively performed using training data; the training data used in the initial training stage and the intensive training stage are consistent with those in the previous text, that is, the training data include training samples and labels corresponding to the training samples; the training samples include sample power characteristic maps constructed based on sample given parameters, sample power parameter sequences, sample photovoltaic output power information, sample line basic data, sample future illumination data and sample sensitive load thresholds, and the labels corresponding to the training samples are the areas of the voltage sag domain corresponding to the sample sensitive loads.

[0175] The difference between the training data used in the initial training phase and the intensive training phase is that the data sources for constructing the training data are different. In some embodiments, the initial training phase constructs training data based on a general data set obtained on a cloud platform for training; the intensive training phase constructs training data based on data actually collected from the distribution network for training.

[0176] The training data used in the initial training stage comes from the cloud platform, which may contain distribution network data sets in multiple scenarios, different scales, different weather conditions, etc., so that the processor 10 can quickly obtain a large amount of training data used in the initial training stage, and at the same time make the trained model have good generalization ability, and can better identify and deal with voltage sag risks.

[0177] The training data used in the enhanced training phase comes from data actually collected from the distribution network, making the data used in training closer to the actual application scenarios and improving the performance of the model in practical applications.

[0178] In some embodiments, in the training data of the initial training phase, the number of samples corresponding to power grids of different complexity is not less than a preset threshold, and the preset threshold is determined based on the complexity of the corresponding power grid. In some embodiments, the preset threshold may be positively correlated with the complexity of the power grid.

[0179] In some embodiments, a more complex power grid corresponds to more samples (such as data collected within 10 days of the power grid), and a less complex power grid corresponds to fewer samples (such as data collected within 3 days of the power grid), so that the model in the initial training stage has more samples to learn from the more complex power grids, thereby ensuring the generalization ability of the model.

[0180] In some embodiments, the complexity of the power grid can be determined based on the topology of the power grid itself, the complexity of the equipment connected to the power grid, and the weather complexity of the area. In some embodiments, the complexity of the power grid can be determined based on the weighted sum of the topology of the power grid itself, the complexity of the equipment connected to the power grid, and the weather complexity of the area. The weight can be determined according to the actual situation to reflect the importance attached to a certain factor.

[0181] In some embodiments, the more devices connected to the power grid and the more types of devices connected to the power grid, the higher the device complexity. In some embodiments, weather complexity can be represented by the mean of the variance of temperature, precipitation, light, etc. over a period of time (such as 1 day).

[0182] In some embodiments, the topological structure of the power grid itself can be obtained in the following way: first, the power grid structure is abstracted into a graph, which includes nodes and edge weights. The nodes correspond to the connection points and cross-connection points between the wires; the edge weights are the wires themselves. Then, the topological indicators such as the number of ring networks, the number of nodes and edges, the network density, and the node degree distribution in the graph are calculated to determine the topological complexity of the power grid.

[0183] In some embodiments, the number of ring networks can be calculated using a ring detection algorithm in graph theory, the network density can be calculated by calculating the ratio between the number of edges in the graph and the number of maximized edges (assuming that there are edges between every two nodes), and the node degree distribution can be calculated using a degree distribution algorithm in graph theory. In some embodiments, the more ring networks, the more nodes and edges, the greater the network density, and the greater the average clustering coefficient, the higher the complexity.

[0184] Using a two-stage training method to train the voltage sag risk model can quickly obtain a large amount of training data used in the initial training stage. At the same time, the trained model has good generalization ability and can better identify and handle voltage sag risks. In addition, using actual data for enhanced training can also improve the performance of the model in practical applications.

[0185] In order to make the photovoltaic output scenario generated by the deep convolutional generative adversarial network more robust, in some embodiments, the processor 10 is also configured to: generate multiple groups of randomly sampled noises with different noise characteristics and their training subsets based on the preprocessed photovoltaic output data, and the noise characteristics reflect the distribution information of the photovoltaic output; and train the deep convolutional generative adversarial network based on the multiple groups of randomly sampled noises with different noise characteristics and their corresponding training subsets.

[0186] In some embodiments, the processor 10 may divide the photovoltaic output power information after preprocessing in step 110 to generate random sampling noise with different noise characteristics. In some embodiments, the processor 10 may divide the photovoltaic output power information by season and / or by light intensity to better reflect the variation of photovoltaic output in different seasons or under different light conditions.

[0187] In some embodiments, the processor 10 can use non-parametric methods (such as kernel density estimation, empirical cumulative distribution function, Copula model) to generate multiple groups of random sampling noises with different noise characteristics based on the photovoltaic output power information divided by season and / or by light intensity.

[0188] In some embodiments, when using the above-mentioned multiple groups of randomly sampled noises with different noise characteristics and their corresponding training subsets to train a deep convolutional generative adversarial network, mixed training based on the training subsets or independent training based on the training subsets can be performed according to actual needs.

[0189] In subset-based hybrid training, a training subset (or a batch in a training subset) is randomly selected from multiple subsets for training in each iteration of the deep convolutional generative adversarial network to ensure that the generator and discriminator are exposed to data from different subsets during training, thereby learning a wider range of photovoltaic output patterns, allowing the generator to be exposed to a wider range of data distributions, thereby learning the diversity and uncertainty of photovoltaic output.

[0190] In the subset-based independent training, that is, after the training is completed, multiple sub-models corresponding to different seasons or light intensities can be obtained respectively. In some embodiments, the processor 10 can fuse multiple sub-models to obtain a comprehensive deep convolutional generative adversarial network. Through independent training, the generator of each subset can focus on learning the unique distribution characteristics of the subset. This targeted training method can generate more realistic photovoltaic output scenarios, especially for applications with high requirements for specific environmental conditions.

[0191] Although the embodiments of the present specification have been shown and described, it will be appreciated by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present specification, and that the scope of the present specification is defined by the appended claims and their equivalents.

Claims

1. A voltage sag domain identification system based on multiple uncertainty scenarios, characterized in that: The system includes a processor, a photovoltaic monitoring device, a power monitoring device and control hardware; The processor is configured to: Based on the photovoltaic monitoring device deployed at each photovoltaic site on the power distribution network, the photovoltaic output power information of each photovoltaic site is collected, and a deep convolutional generative adversarial network is trained based on the photovoltaic output power information; Generate photovoltaic output scenarios using the deep convolutional generative adversarial network; Based on the photovoltaic output scenario, given parameters, sensitive load thresholds and structural parameters of the power system, determining a voltage sag domain identification result corresponding to the given parameters and the sensitive load thresholds; Based on the voltage sag domain identification result, a voltage reduction response instruction is generated; the voltage reduction response instruction includes instructing the monitoring frequency, monitoring accuracy, upload interval period and alarm threshold of the power monitoring device; Sending the voltage reduction response instruction to the power monitoring device in the voltage sag domain; The power monitoring device is configured as follows: Collecting a power parameter sequence in the power distribution network at the monitoring frequency and the monitoring accuracy, and uploading the power parameter sequence to the processor at the upload interval period; Determining whether voltage data in the power parameter sequence is lower than the alarm threshold; in response to falling below the alarm threshold, sending an alarm signal to the processor; The processor is further configured to: Determine, based on the alarm signal, an estimated impact load corresponding to the node where the power monitoring device is located; Generate a control instruction and send it to the control hardware corresponding to the estimated impact load to control the control hardware to perform compensation work.

2. The system according to claim 1, characterized in that The processor is further configured to: Based on the photovoltaic output scenario, the voltage of each node of the distribution network before the fault is determined by using the Newton-Raphson method power flow calculation method; Obtaining a current bus from the lines of the power distribution network, calculating the impedance matrix of each node of the power distribution network by adding branches or inverting the admittance matrix according to the given parameters, and determining the voltage sag amplitude of the node where the sensitive load is located when different types of short-circuit faults occur at each node on the current bus based on the voltage of each node before the fault and the impedance matrix of each node; Determine the voltage sag amplitude vector of the bus node based on the voltage sag amplitude of the node where the sensitive load is located when different types of short-circuit faults occur at each node of the current bus; Determine the line association vector based on the voltage sag amplitude of the node where the sensitive load is located when different types of short-circuit faults occur at each node of the current bus; Based on the line association vector, determining whether the current bus meets the analysis condition; In response to not being satisfied, the current bus is not within the voltage sag region, and a next line is selected from the power distribution network as the current bus to continue the determination; In response to satisfying: analyzing the current bus by Newton quadratic interpolation method to obtain an analytical expression for the voltage sag amplitude and a fault voltage equation of the current bus; and, based on the fault voltage equation, obtaining the portion of the current bus within the voltage sag domain, and selecting a next line from the power distribution network as the current bus to continue the determination; When the traversal and iteration of all the lines in the power distribution network are completed, a plurality of the voltage sag domains corresponding to the given parameters and the sensitive load threshold are obtained, and the union of the plurality of the voltage sag domains is the voltage sag domain identification result.

3. The system according to claim 1, characterized in that The deep convolutional generative adversarial network includes: using a full convolutional network, canceling the fully connected layer and using batch normalization; After preprocessing the photovoltaic output power information, using random sampling noise as input of the generator of the deep convolutional generative adversarial network, using the historical output data of each photovoltaic site as input of the discriminator of the deep convolutional generative adversarial network, and using the training data to train the deep convolutional generative adversarial network until a Nash equilibrium is reached; After the deep convolutional generative adversarial network training is completed, the photovoltaic output scene generated by the generator that is similar to the real data is obtained.

4. The system according to claim 2, characterized in that The processor is further configured to: Based on clarifying the structural parameters of the power system, the admittance matrix is ​​formed after substituting actual data, and the offset of each parameter is calculated according to the given initial value; Perform error analysis, and in response to the error meeting the error precision, exit the loop and output the result; in response to the error not meeting the error precision, continue to solve the Jacobian matrix, and solve the correction voltage through the correction equation, and perform the error analysis again.

5. The system according to claim 2, characterized in that The different types of short circuit faults include three-phase short circuit fault, single-phase ground short circuit fault, two-phase short circuit fault and two-phase ground short circuit fault; When the short circuit fault is the three-phase short circuit fault, the processor is configured to: When the three-phase short circuit fault occurs at the fault point K, it is a symmetrical balanced fault, only the positive sequence is considered, and the three-phase voltage sag amplitudes at the sensitive load node S are the same, and the voltage sag amplitude of one phase is: in, is the voltage sag amplitude at point S, is the positive sequence self impedance, is the positive sequence mutual impedance, is the voltage before the fault at point S, is the voltage before the fault at point K; the voltage sag amplitude at point S It is related to the structural parameters of the power system and the operating state of the power system before the fault occurs. When an asymmetric short circuit fault occurs at the fault point K with phase A being a special phase, the symmetrical component method is used for analysis; When the short circuit fault is the single-phase to ground short circuit fault, the processor is configured to: When the single-phase grounding short circuit fault with phase A as the special phase occurs at the fault point K, the voltage sag amplitudes at point S are as follows: in, is the voltage drop amplitude of phase A at point S, is the voltage drop amplitude of phase B at point S, is the voltage sag amplitude of phase C at point S; is the voltage of phase A before the fault at point S, is the voltage of phase B before the fault at point S, is the voltage of phase C at point S before the fault; is the zero-sequence self-impedance at point K, is the positive sequence self impedance at point K, is the negative sequence self-impedance at point K; is the zero-sequence mutual impedance between point s and point K, is the positive sequence mutual impedance between point S and point K, is the negative sequence mutual impedance between point S and point K; is the voltage before the fault at point K; α=e j120° is the rotation factor; When the short circuit fault is the two-phase short circuit fault, the processor is configured to: When the two-phase short circuit fault occurs in phases B and C, the three-phase voltage sag amplitude expression at point S is: When the short circuit fault is the two-phase grounding short circuit fault, the processor is configured to: When the two-phase grounding short circuit fault occurs in phases B and C, the three-phase voltage temporary sag amplitudes at point S are as follows: When the two-phase ground short circuit fault occurs on the power system bus, the self-impedance and mutual impedance are directly called from the impedance matrix of each node of the distribution network; however, when the short circuit fault occurs at a certain point on the line, the calculation of the self-impedance and the mutual impedance needs to be obtained by introducing the position variable p (0≤p≤1) and the joint node impedance matrix, where point S is the bus node where the sensitive load is located, and point Z is the node of the bus where the sensitive load is located. C is the sequence impedance of line FT. When the fault point K moves on line FT, the three-sequence self-impedance And the three-sequence mutual impedance between the fault point K and S They are all represented by the impedance matrix Z and the position variable p: Where: They are the zero-sequence self-impedance, positive-sequence self-impedance and negative-sequence self-impedance of the system bus nodes F and T respectively; They are the mutual impedances of the system bus nodes F and T, the mutual impedances between the bus F and the sensitive load node S, and the mutual impedances between the bus T and the sensitive load node S, which are all called from the impedance matrix of each node in the distribution network. is the line sequence impedance between nodes F and T, p (0≤p≤1) is the position variable of the fault point K on line FT; The voltage before the fault occurs at the fault point K It is represented by the fault location variable p: in, is the voltage of bus node F before the fault occurs, is the voltage of bus node T before the fault occurs, p (0≤p≤1) is the position variable of fault point K on line FT; the voltage sag amplitude at point S is represented by the voltage before the fault and each sequence impedance, and the function formula U(p) of the voltage sag amplitude at point S with respect to the position variable p is obtained when any short-circuit fault occurs at any fault point.

6. The system according to claim 2, characterized in that The expression of the difference vector is: in, is the absolute value of the voltage amplitude at node S where the sensitive load is located when a short circuit fault occurs on bus 1, is the absolute value of the voltage amplitude at node S where the sensitive load is located when a short circuit fault occurs on bus 2, is the absolute value of the voltage amplitude at the node S where the sensitive load is located when a short circuit fault occurs on bus n; U th is the voltage sag threshold of sensitive load; ΔU S,1 It represents the difference between the absolute value of the voltage amplitude at the node S where the sensitive load is located and the voltage sag threshold of the sensitive load when a short circuit occurs on bus 1, ΔU S,2 It represents the difference between the absolute value of the voltage amplitude at the node S where the sensitive load is located and the voltage sag threshold of the sensitive load when a short circuit fault occurs on bus 2, ΔU S,n It represents the difference between the absolute value of the voltage amplitude at the node S where the sensitive load is located and the voltage sag threshold of the sensitive load when a short circuit occurs on bus n, ΔU S Indicated by ΔU S,1 , ΔU S,2 , ΔU S,n The difference vector composed of: By judging ΔU S The positive and negative values ​​of the elements in the difference vector are used to determine whether each node is within the temporary drop region of the sensitive node, and the node determination vector B is introduced: Among them, B1 represents the judgment result of bus 1, B2 represents the judgment result of bus 2, and B n Indicates the judgment result of busbar n, B i represents the judgment result of busbar i, B represents the node judgment vector; ΔU S，i It represents the difference between the absolute value of the voltage amplitude at the node S where the sensitive load is located and the voltage sag threshold of the sensitive load when a short circuit fault occurs on bus i; B i =1 means that busbar i is within the voltage sag range corresponding to the sensitive load; B i = 0 means that bus i is outside the voltage sag domain corresponding to the sensitive load. The inclusion of each line in the voltage sag domain is determined, and the line association vector L is introduced: Where: L represents the line association vector, L1 represents the result of the determination of the line 1 in the temporary drop domain, L2 represents the result of the determination of the line 2 in the temporary drop domain, L m Indicates the result of the determination of whether line m is included in the temporary drop domain; B1_ F is the temporary drop judgment result of bus node F of line 1, B 1_T is the temporary drop judgment result of bus node T of line 1, B 2_T is the temporary drop judgment result of bus node T of line 2, B m_F is the temporary drop judgment result of busbar node F of line m, B m_T is the temporary drop determination result of the bus node T of line m; The step of judging whether the current bus satisfies the analysis condition based on the line association vector includes: If L i =0, indicating that the line is not in the voltage sag region, and the next line is calculated directly; if L i =1, indicating that one of the first and last nodes of this line is in the temporary drop region, then the i part of the line is in the temporary drop region, and there is a unique critical point on the line; if L i =2, indicating that the first and last nodes of this line are both within the voltage sag region. The golden section search method is first used to solve the maximum value U of the voltage sag amplitude curve. max and the sensitive load voltage sag threshold U th For comparison; if U max <U th , indicating that this line is completely in the temporary drop region and there is no critical point, then the next line is calculated directly; if U max ≥U th , indicating that the line section i is located in the temporary drop region, and there are two critical points on the line; The whole process of the golden section search method is to find p a ≤p≤p b (p a =0, p b =1) the point p at which the voltage drop at the common connection point reaches the maximum value max , use p max As the interpolation point, it is ensured that the interpolation curve passes through the maximum value of the voltage sag amplitude, providing an initial value point for the secant iteration.

7. The system according to claim 2, characterized in that The processor is configured to: The root of the fault voltage equation p ia As the initial iteration value of the secant iteration method, the accurate critical point position is obtained through iteration. The critical fault distance iteration expression according to the secant iteration method is as follows: Among them, p k-1 represents the critical fault distance of the k-1th solution, p k represents the critical fault distance of the kth solution, p k+1 represents the critical fault distance of the k+1th solution, U(p k ) indicates that the position variable is p k The voltage sag amplitude at point S, U(p k-1 ) indicates that the position variable is p k-1 The voltage sag amplitude at point S, U th It is the voltage sag threshold of sensitive load; The convergence condition of the secant iteration method is: ||U(p k+1 )|-U th |<ε Among them, U(p k-1 ) indicates that the position variable is p k-1 The voltage sag amplitude at point S, U th is the voltage sag threshold of sensitive load, and ε represents the convergence threshold.

8. The system according to claim 1, characterized in that: The control hardware includes one or more of a backup power supply, a dynamic voltage regulator, and a load switch; the control instruction includes one or more of a backup power supply instruction, a voltage self-regulation instruction, and a switch instruction; The backup power supply is configured to: after receiving the backup power supply instruction, provide power supply corresponding to the estimated impact load; The dynamic voltage regulator is configured to: after receiving the voltage self-adjustment instruction, provide a compensation voltage corresponding to the estimated impact load: The load switch is configured to: after receiving the switch instruction, control the connection and disconnection of the estimated impact load and the power distribution network.

9. The system according to claim 2, characterized in that: The processor is further configured to: Constructing a power characteristic diagram according to the given parameters, the power parameter sequence, the photovoltaic output power information, line basic data, future light data, and sensitive load thresholds; Inputting the power characteristic diagram into a voltage sag risk model to obtain a voltage sag risk area; A plurality of the voltage sag domains corresponding to the given parameter and the sensitive load threshold in the voltage sag risk area are calculated to determine the voltage sag domain identification result.

10. The system according to claim 9, characterized in that: The training process of the voltage sag risk model includes an initial training phase and an enhanced training phase respectively performed using training data; the initial training phase constructs the training data for training based on a general data set obtained on a cloud platform; In the intensive training stage, the training data is constructed based on the data actually collected from the power distribution network for training; The training data includes training samples and labels corresponding to the training samples; The training sample includes a sample power characteristic diagram constructed based on sample given parameters, sample power parameter sequence, sample photovoltaic output power information, sample line basic data, sample future light data and sample sensitive load threshold; the label corresponding to the training sample is the area of ​​the voltage sag domain corresponding to the sample sensitive load; In the training data of the initial training stage, the number of samples corresponding to power grids of different complexity levels is not less than a preset threshold, and the preset threshold is determined based on the complexity of the corresponding power grid.

Images