A power grid reliability assessment method based on fine-grained data-driven

Abstract

The present invention belongs to the field of power systems, and specifically relates to a power grid reliability assessment method based on fine-grained data drive; the method includes randomly extracting power grid system status data using the Monte Carlo method; calculating the power difference of the system DC power flow before and after the power grid system is disconnected; inputting the power difference and other power data into the power flow module as the input feature vector of the sample to be tested; calculating the regression result and classification result of the sample to be tested, and determining the problem sample; inputting the input feature vector of the problem sample into the optimal power flow module, calculating the classification result of the problem sample, and determining the fault sample; performing minimum load shedding analysis on the fault sample, and according to the analysis result, combining the variance coefficient to calculate the uncertainty of the fault sample, and outputting the reliability assessment result of the fault sample. The present invention can be widely used in the online calculation of the operation reliability of large power grids, and is particularly suitable for the scene after large-scale access of high-proportion new energy.

Description

Technical Field

[0001] The present invention belongs to the field of power systems, and aims to realize efficient calculation of large power grid reliability assessment by using deep neural networks, and specifically relates to a power grid reliability assessment method based on fine-grained data drive. Background Art

[0002] For power systems with a high proportion of renewable energy and strong uncertainty, large-scale power grid reliability assessment is an important tool for preventing major blackouts. The reliability assessment process includes power flow calculation and optimal power flow calculation. In order to obtain high-precision assessment results, large-scale power grid reliability assessment needs to analyze a large number of system states (i.e., repeatedly perform a large number of power flow and optimal power flow calculations). Therefore, large-scale power grid reliability assessment faces the problem of heavy computational burden. With the access of a high proportion of renewable energy, the complexity of reliability assessment models and the surge in the frequency of computational requirements have further exacerbated the above problems.

[0003] In recent years, with the continuous development of information technology and deep learning technology, data-driven methods have shown great potential in reducing the computational burden of reliability assessment, and such methods are increasingly favored by scholars. The current data-driven methods for reliability assessment usually establish a classifier to reduce the frequency of optimal power flow calculation by identifying whether the samples are marked as "successful" or faulty. After classification, only the samples marked as faulty are subjected to optimal power flow calculation with the goal of minimizing load reduction, and the samples marked as "successful" do not need to be analyzed. In reliability assessment, since most system states belong to the "successful" category, the above method can save a lot of computational costs.

[0004] However, it is difficult for existing methods to accurately obtain high-precision reliability assessment results. The main reasons are: the nonlinearity of the power flow equation, the discreteness of branch disconnection, and the imbalance between success / failure samples make the learning task of reliability assessment very complicated. In order to reduce the learning complexity and realize data-driven reliability assessment, existing methods ignore the imbalance problem, the uncertainty of random branch failures, or / and assume that the power flow is linear, which will cause the reliability assessment to face a significant loss of accuracy and is difficult to promote and apply in actual industry. Therefore, there is an urgent need for more effective data-driven methods to ensure the accuracy of large power grid reliability assessment results. Summary of the invention

[0005] Based on the problems existing in the prior art, the purpose of the present invention is to address the deficiencies of the data-driven power system operation reliability calculation method and propose a power grid reliability assessment method based on fine-grained data drive. The technical solution to implement the present invention includes:

[0006] Obtaining power grid system status data, and randomly extracting a portion of the power grid system status data according to the variance coefficient using the Monte Carlo method;

[0007] Input the power difference of the sample to be tested, the active power vector and reactive power vector of the load, the maximum active power and reactive power that the unit can generate, and the minimum active output and reactive output of the unit into the data-driven power flow module;

[0008] The power flow regression module calculates the regression result of the sample to be tested, and the power flow classification module predicts the classification result of the sample to be tested, and outputs the regression result and the classification result as the output feature vector of the sample to be tested;

[0009] When the regression result exceeds the corresponding preset threshold, or / and the mark of the classification result is a problem mark, the sample to be tested is regarded as a problem sample;

[0010] Input the input feature vector of the problem sample into the data-driven optimal power flow module, predict the classification result of the problem sample, and output the classification result as the output feature vector of the problem sample to determine the faulty sample;

[0011] Performing minimum load shedding analysis on the fault sample to calculate the active power of the inflow branch and the outflow branch, as well as the voltage amplitude and the voltage phase angle;

[0012] According to the calculation results of the minimum load shedding analysis of the fault samples, the uncertainty of the fault samples is calculated in combination with the variance coefficient, and the reliability evaluation results of the fault samples are output.

[0013] After the present invention adopts the above technical solution, it mainly has the following effects:

[0014] 1. Compared with the existing data-driven operation reliability assessment method, the data-driven power flow module introduced in the present invention can be used as a buffer to effectively identify problem samples, i.e., the system state where the power flow exceeds the limit, and reduce the learning complexity of identifying fault samples. In addition, the buffer can be used to pre-identify a portion of non-problem samples as successful samples, which can also effectively alleviate the problem of sample imbalance.

[0015] 2. The data-driven fine-grained reliability assessment framework proposed in the present invention includes a power flow module and a minimum load shedding calculation model, so the framework can calculate any reliability index. At the same time, through the feature vector and learning strategy designed by the present invention, the framework can effectively cope with the random fluctuations of load / new energy and predict the random failures of power grid equipment without assuming linear power flow or unchanged power grid structure, and has high calculation accuracy.

[0016] 3. The present invention can be widely used in online calculation of operational reliability of large power grids, and is particularly suitable for situations where high-proportion new energy sources such as wind and electricity are connected on a large scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a framework diagram of fine-grained operation reliability evaluation based on a deep neural network in an embodiment of the present invention;

[0018] Figure 2 The figure is a flow chart of a power grid reliability assessment method driven by fine-grained data in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] Figure 1 is a power grid reliability assessment framework diagram used in the embodiment of the present invention, which is based on a deep learning network, such as Figure 1 As shown in the figure, the framework diagram mainly includes four parts. First, some system states are obtained by sampling and input into the data-driven problem sample identification model. In this identification model, a data-driven power flow module including a data-driven power flow classification module and a power flow regression module is required to find problem samples, and then find fault samples from the problem samples. Next, only the minimum load shedding analysis is required for the fault samples, and the reliability index is statistically analyzed according to the analysis results. Among them, the samples to be tested are divided into successful samples, fault samples and ordinary samples; the ordinary samples here will be screened out by the data-driven power flow module, and the remaining problem samples will be determined. The problem samples include successful samples and fault samples.

[0021] Compared with the traditional data-driven reliability assessment framework, the framework of the present invention introduces a buffer (i.e., a data-driven power flow module) to identify problem samples, i.e., system states including line overload, voltage over-limit, or system decoupling, rather than directly identifying fault samples. The buffer divides the recognition of all system states into two stages. The first stage is to distinguish whether it is a problem sample, and the second stage is to distinguish whether the problem sample is a fault sample. These two stages can effectively reduce the learning complexity and alleviate the problem of sample imbalance. Considering that the incorrect identification of problem samples may lead to a decrease in the accuracy of large power grid reliability assessment indicators, the framework combines data-driven power flow regression and classification results to make the identification results of problem samples more accurate. In particular, if a sample is identified as a problem sample by a data-driven power flow regression model or a data-driven power flow classification model, it is regarded as a problem sample.

[0022] It can be understood that the above framework alleviates the overfitting problem to a certain extent by establishing a buffer to first identify the problem samples, and further improves the identification accuracy of the problem samples by combining the power flow regression results with the power flow classification results.

[0023] In addition, the present invention also considers that how to construct a feature vector that effectively characterizes the nonlinearity and discreteness of operational reliability and design a deep learning strategy that is suitable for each data-driven module in the above framework is the key to realizing the functions of the above framework. Based on this, Figure 2 A flow chart of a power grid reliability assessment method based on fine-grained data driving in an embodiment of the present invention is given, such as Figure 2 As shown, the evaluation method includes:

[0024] 101. Obtain power grid system status data, and randomly extract a portion of the power grid system status data according to the variance coefficient using the Monte Carlo method;

[0025] In the embodiment of the present invention, since the power grid system status data is very massive and huge, it is realistic to directly process this data. Based on this, the Monte Carlo method is used in the embodiment of the present invention to randomly sample an appropriate amount of system states, and part of the system states can be extracted according to the variance coefficient as the extraction ratio.

[0026] In the embodiment of the present invention, the variance coefficient is expressed as:

[0027]

[0028] Among them, α represents the variance coefficient; represents the estimated value of the reliability index; Indicates estimated value The reliability index here can include the probability of power shortage (Loss of Load Probability, LOLP), and the expected power shortage (Expected Energy Not Supplied, EENS), etc. The variance coefficient can be calculated according to the estimated value of the corresponding reliability index, and the corresponding extraction ratio can be guided according to the variance coefficient. The variance coefficient is used to represent the estimated value. The uncertainty of the Monte Carlo method can also be used as one of the convergence conditions of the Monte Carlo method.

[0029] 102. Preprocess the extracted power grid system state data to calculate the power difference of the system DC power flow before and after the power grid system is disconnected;

[0030] In the embodiment of the present invention, it is necessary to preprocess the extracted power grid system state samples, and the preprocessing method adopts the z-score method as shown in the following formula, and divides the samples into a training set, a validation set and a test set (samples to be tested);

[0031]

[0032] Among them, v mean and v std are the mean and standard deviation of the vector V respectively.

[0033] In the embodiment of the present invention, the accuracy of sample classification is considered to be an important factor in how to effectively characterize the random failure of branches. Traditional technology usually uses available transmission capacity as a feature vector to characterize the uncertainty of power grid branches. However, this method is difficult to effectively reflect the connection relationship between branches and the importance of disconnection of a branch. In this regard, the present invention proposes to use the power difference calculated by the system DC flow before and after disconnection as a feature vector to characterize the uncertainty of branches.

[0034] First, the present invention calculates the power of each branch DC power flow according to the admittance matrix and the voltage phase angle, which is expressed as P k =B k δ k ;

[0035] Among them, P k represents the power of the DC power flow on branch k, B k represents the admittance matrix on branch k, δ k represents the voltage phase difference on branch k, where k∈[1,2,…,L], and L represents the total number of branches in the power grid system.

[0036] Secondly, the present invention calculates the power difference of the DC power flow before and after the disconnection of different branches through the transfer distribution factor DF, which is expressed as

[0037] Among them, DF k-l represents the change in DC power flow on branch k caused by the interruption of branch l, P l represents the DC power flow on branch l, l≠k, k,l∈[1,2,...,L].

[0038] Finally, the present invention performs matrix splicing on the power difference of the DC power flow of each different branch before and after the disconnection, and obtains the power difference vector of the DC power flow of all branches before and after the branch is disconnected, which is expressed as ΔP DC =P a -P b ;

[0039] Among them, ΔP DCThe power difference vector of the DC power flow of all branches before and after the branch is disconnected, P a With P b They represent the DC power flow vectors of all branches before and after the branch is disconnected.

[0040] It is understandable that under certain system conditions, the change in the flow of each branch can reflect the change in the admittance matrix B, and thus to a certain extent reflect the topological connection relationship of the power grid. In addition, the change in the power value of the DC flow also reflects the relative importance of disconnecting different branches.

[0041] It is worth noting that the present invention does not need to store the branch DC power flow change values ​​before and after the different branches are disconnected, which is very helpful for practical applications and can reduce computing power and storage space. This is because the feature vectors before and after the different branches are disconnected can be updated online through the transfer distribution factor DF, which can avoid processing the global state data of the power grid system. This method can effectively reduce the global burden, and the calculation overhead and storage overhead are smaller than global operations, which is more suitable for large-scale power grid reliability assessment.

[0042] 103. Obtain the power difference of the sample to be tested, the active power vector and reactive power vector of the load, the maximum active power and reactive power that can be generated by the unit, and the minimum active output and reactive output of the unit, and input them into the data-driven power flow module as the input feature vector of the sample to be tested;

[0043] After preprocessing the system status data, the power flow equation and the minimum load shedding optimization model can be solved based on the Newton method and the interior point method respectively to solve the load active vector and reactive vector, the maximum active power and reactive power that the unit can generate, the minimum active output and reactive output of the unit, and the power difference of the system DC power flow before and after the power grid system is disconnected.

[0044] The power flow equation is expressed as:

[0045]

[0046]

[0047] Among them, P i (V,δ) represents the injected active power at node i, Q i (V,δ) represents the injected reactive power at node i; V i is the voltage amplitude at node i, V j is the voltage amplitude at node j, V represents the voltage amplitude vector; N represents the total number of nodes in the power grid system; G ij is the real part of the node admittance matrix from node i to node j, B ij is the imaginary part of the node admittance matrix from node i to node j; δ ijrepresents the voltage phase difference from node i to node j, δ ij =δ i -δ j , δ i represents the voltage phase angle at node i, δ j represents the voltage phase angle of node j, and δ represents the voltage phase angle matrix.

[0048] The minimum load shedding model is expressed as:

[0049]

[0050] st

[0051] P i (V,δ)-PD i +C i =0,i∈ND

[0052] Q i (V,δ)-QD i =0,i∈ND

[0053]

[0054]

[0055] 0≤C i ≤PD i ,i∈ND

[0056]

[0057] V i min ≤V i ≤V i max ,(i∈N)

[0058] Among them, C i represents the amount of load loss at node i; ND represents the number of load nodes in the power grid system, PD i with QD i Respectively represent the active load and reactive load at node i; PG i With QG i are the active power and reactive power generated by node i respectively; is the minimum active output of the unit at node i; is the maximum active output of the unit at node i; is the minimum reactive power output of the unit at node i, is the maximum reactive power output of the unit at node i, NG represents the number of all generator nodes in the system, T k (V,δ) represents the transmission power on branch k, represents the maximum transmission power on branch k, L represents the number of all branches in the power grid system, V i min is the minimum voltage amplitude at node i, V i max is the maximum voltage amplitude at node i.

[0059] In the embodiment of the present invention, after extracting a part of the power grid system status data, each piece of data can be used as a sample to be tested. Since the input feature vector of the deep neural network needs to include information such as the load, new energy, and topological structure of the large power grid, its input information is relatively complex. In order to simplify the processing, the present invention solves the power flow equation and the minimum load shedding optimization model based on the Newton method and the interior point method respectively, and obtains the power difference of the system DC power flow before and after the power grid system is disconnected, the load active vector and reactive vector, the maximum active power and reactive power that can be generated by the unit, and the minimum active output and reactive output of the unit as input, and constitutes the input feature vector input into the data-driven power flow module, wherein the input feature vector X of the reliability assessment is expressed as:

[0060] X=[P D ,Q D ,PG max ,QG max ,PG min ,QG min ,ΔP DC ]

[0061] The input feature vector X is also the input feature vector of the sample to be tested in the present invention, P D and Q D are load active vector and load reactive vector respectively; PG max and QG max are the maximum active power and the maximum reactive power that can be generated by the generator set respectively; PG min and QG min They are the minimum active output and minimum reactive output of the generator set respectively.

[0062] It should be noted that the new energy is regarded as load in the present invention. The system load and available power generation capacity are constructed as input feature vectors to characterize the uncertainty of new energy, load demand and generator status.

[0063] It can be understood that in the embodiment of the present invention, under the premise of inputting the characteristic vector X, the power flow regression equation is used to solve it in combination with the input data of the input characteristic vector X, and the regression results of the bus node voltage amplitude vector, the branch power vector, the unit output at the balancing node, i.e., the active power vector, and the reactive power vector output by the unit are obtained through the power flow regression module.

[0064] 104. The power flow regression module calculates the regression result of the sample to be tested, and the power flow classification module predicts the classification result of the sample to be tested, and outputs the regression result and the classification result as the output feature vector of the sample to be tested;

[0065] In an embodiment of the present invention, the data-driven trend module includes a trend regression module and a trend classification module. The data-driven trend module is a deep learning network, wherein the trend regression module and the trend classification module can also be sub-modules of the deep learning network, which are mainly used to calculate the regression results and predict the classification results.

[0066] In an embodiment of the present invention, the regression result calculated by the power flow regression module is used to characterize the node voltages of all buses, all line powers, and generator output powers, while the classification result predicted by the power flow classification module is used to distinguish whether the sample to be tested is a problem sample.

[0067] In the embodiment of the present invention, the construction of the output feature vector is mainly determined by each data driving module. Therefore, the output feature vector of the reliability evaluation is expressed as:

[0068] Y PF =[V,T,PG ref ,Q G ,F PF ]

[0069] Among them, Y PF represents the output characteristic vector of the sample to be tested; V is the bus node voltage amplitude vector; T is the branch power vector; PG ref is the active power vector of the unit output at the balance node; Q G is the reactive power vector output by the unit; F PF It is the sign vector of the problem sample, which takes 0 or 1.

[0070] It is understandable that the vector V, T, PG ref ,Q G is output by the power flow regression module, and F PF Whether it is the output of the power flow classification module, the two together constitute the output feature vector of the data-driven power flow module.

[0071] 105. When the regression result exceeds the corresponding preset threshold, or / and the mark of the classification result is a problem mark, the sample to be tested is regarded as a problem sample;

[0072] In an embodiment of the present invention, considering that the incorrect identification of problem samples may lead to a decrease in the accuracy of large power grid reliability assessment indicators, this embodiment combines the data-driven power flow regression results and power flow classification results to make the identification results of problem samples more accurate. Specifically, if a sample to be tested is identified as a problem sample by the power flow regression module or the power flow classification module, the sample to be tested is regarded as a problem sample. For example, assuming that the regression result of the voltage amplitude is 1.8pu, and the preset threshold of the voltage amplitude is 1.05pu, then obviously, the regression result of the voltage amplitude has exceeded its corresponding preset threshold, then the sample to be tested at this time will be treated as a problem sample for subsequent secondary classification.

[0073] In the embodiment of the present invention, the preset threshold is a preset value of the project, and those skilled in the art can complete the setting of the corresponding threshold according to actual conditions, and the present invention does not make any specific limitation on this.

[0074] 106. Input the input feature vector of the problem sample into the data-driven optimal power flow module, predict the classification result of the problem sample, and output the classification result as the output feature vector of the problem sample to determine the faulty sample therein;

[0075] Since only the problem samples are screened in the above process, but whether the problem samples are fault samples is unknown, the present invention also adopts a data-driven optimal power flow module to further process the classification and identification problem of the problem samples. Therefore, the present invention can use the corresponding input feature vector as the input feature vector in the data-driven optimal power flow module based on the classification result of the problem samples, which can be expressed as:

[0076] X1=[P D ,Q D ,PG max ,QG max ,PG min ,QG min ,ΔP DC ]

[0077] in, The input feature vector X1 here is also the input feature vector of the problem sample of the present invention.

[0078] In the embodiment of the present invention, in order to quickly identify faulty samples, the power flow module is no longer set in the data-driven optimal power flow module, and only the classification module is set. In order to enable the classification module to effectively identify faulty samples from problem samples, the output feature vector is set to Y OPF =[F OPF ], where F OPF Is a flag vector indicating whether the system is in a fault state, which can be 0 or 1.

[0079] It can be understood that in the embodiment of the present invention, based on the input feature vector X of the problem sample, the minimum load shedding model is used to solve the problem in combination with the input data of the input feature vector X, and the data-driven optimal power flow module is used to determine whether the current node is load shedding, and the result of the flag vector of the fault system state is obtained.

[0080] Since both the data-driven power flow module and the data-driven optimal power flow module are composed of deep neural networks, the present invention provides a unified description of the deep neural network. The input layer and the middle layer of the deep neural network use one of the most commonly used activation functions in the field of deep learning, the ReLU activation function. For the activation function of the last layer of the neural network, a linear activation function is selected to enable the neural network to capture a wider range of outputs. Therefore, for the data-driven module of reliability assessment, the mapping relationship between its input and output can be expressed by the following formula.

[0081]

[0082]

[0083] Among them, Y PF Represents the output feature vector of the sample to be tested; Represents the weight matrix between the i-th layer neural network and the i+1-th layer neural network of the sample to be tested; Represents the activation function of the test sample X in the i-th layer of the neural network, Represents the bias vector between the i-th layer of the neural network and the i+1-th layer of the neural network of the sample to be tested; i=1,2,...,n, n represents the number of neural network layers; θ PF = {W PF ,b PF}; Y OPF The output feature vector representing the problem sample; Represents the weight matrix between the i-th layer neural network and the i+1-th layer neural network of the problem sample; Represents the activation function of the problem sample X in the i-th layer of the neural network, Represents the bias vector between the i-th layer neural network and the i+1-th layer neural network of the problem sample; i=1,2,...,n, n represents the number of neural network layers; θ OPF = {W OPF ,b OPF}.

[0084] In order to obtain the optimal deep neural network parameter θ, that is, θ PF and θ OPF , the mean square error function is usually used as the loss function L, expressed as:

[0085]

[0086] Where N represents the number of samples; and i are the actual and predicted values ​​of the i-th sample.

[0087] Then, the corresponding optimization goal is to minimize L. Correspondingly, the Rmsprop or Adam algorithm can be used to solve the above unconstrained optimization problem.

[0088] In a preferred embodiment of the present invention, a load shedding sample sensitive loss function is proposed for the intelligent classification module of success samples / fault samples, that is, success / fault system status, to improve the learning accuracy of the deep neural network for misclassified fault samples. Specifically, the load shedding sensitive loss function proposed by the present invention is as follows:

[0089] L OPF =βL1+L2

[0090]

[0091]

[0092] Among them, L1 represents the loss function of the successful sample, that is, the mean square error of the classification results of the successful sample; L2 represents the loss function of the fault sample, that is, the mean square error of the classification results of the fault sample; Ω s represents the set of successful samples, Ω f represents the set of fault samples. The parameter β is a user-defined weight, ranging from (0, 1). Through this loss function, we can minimize L OPF To ensure the overall accuracy of classification; at the same time, by adjusting the weight β, it makes the system state more sensitive to incorrect prediction of faults.

[0093] It should also be understood that in an embodiment of the present invention, the power flow equation and the minimum load shedding optimization model can be solved based on the Newton method and the interior point method respectively to obtain the power difference, the load active vector and reactive vector, the maximum active power and reactive power that can be generated by the unit, the minimum active output and reactive output of the unit, and the power difference of the system DC power flow before and after the power grid system is disconnected as the input feature vector; at the same time, the branch power, voltage amplitude, unit output, power flow over-limit sign, load shedding sign, etc. are retained as output feature vectors, the input feature vector is input into the power flow module, and the problem sample is predicted by combining the result of the output feature vector, and the input feature vector is input into the optimal power flow module, and the fault sample is predicted by combining the result of the output feature vector. This dual network prediction method can realize high-precision online calculation of large power grid reliability assessment.

[0094] 107. Perform minimum load shedding analysis on the fault sample to calculate the active power of the inflow branch and the outflow branch, as well as the voltage amplitude and voltage phase angle;

[0095] In the embodiment of the present invention, since the fault sample has been determined, it is necessary to perform minimum load shedding analysis on the fault sample, and the calculation of the active power of the inflow branch and the outflow branch, as well as the voltage amplitude and the voltage phase angle includes:

[0096]

[0097] st

[0098] P i (V,δ)-PD i +C i =0,i∈ND

[0099] Q i (V,δ)-QD i =0,i∈ND

[0100]

[0101]

[0102] 0≤C i ≤PD i ,i∈ND

[0103]

[0104] V i min ≤V i ≤V i max ,(i∈N)

[0105]

[0106]

[0107] Among them, C i represents the amount of load loss at node i; ND represents the number of load nodes in the power grid system, P i (V,δ) represents the injected active power at node i, Q i (V,δ) represents the injected reactive power V at node i i is the voltage amplitude at node i, δ i represents the voltage phase angle at node i, δ j represents the voltage phase angle at node j, δ ij represents the voltage phase difference from node i to node j, δ ij =δ i-δ j ;PD i with QD i Respectively represent the active load and reactive load at node i; PG i With QG i are the active power and reactive power of node i respectively; G ij is the real part of the node admittance matrix from node i to node j, B ij is the imaginary part of the node admittance matrix from node i to node j; is the minimum active output of the unit at node i; is the maximum active output of the unit at node i; is the minimum reactive power output of the unit at node i, is the maximum reactive power output of the unit at node i, NG represents the number of all generator nodes in the system, T k (V,δ) represents the transmission power on branch k, represents the maximum transmission power on branch k, L represents the number of all branches in the power grid system, V i min is the minimum voltage amplitude at node i, V i max is the maximum voltage amplitude at node i, and N represents the total number of nodes in the system.

[0108] 108. According to the calculation results of the minimum load shedding analysis of the fault sample, the uncertainty of the fault sample is calculated in combination with the variance coefficient, and the reliability evaluation result of the fault sample is output.

[0109] In the embodiment of the present invention, the reliability index can be statistically analyzed based on the calculation results of the minimum load shedding analysis of the fault samples. The reliability indexes mainly considered by the present invention include: Loss of Load Probability (LOLP) and Expected Energy Not Supplied (EENS). At the same time, the variance coefficient α and the number of sampling samples can be used as the convergence conditions of the Monte Carlo method in the present invention.

[0110] In order to better illustrate the evaluation scheme of the present invention, the following is evaluated in combination with specific power grid system status data. In this embodiment, the IEEE RTS-79 system, the IEEE RTS-96 system and their modified examples are used to verify the effectiveness of the present invention. The example information is as follows:

[0111] Example 1: IEEE RTS-79 system. Assume that the load curve follows a normal distribution with a standard deviation of 10% of the expected value and the historical annual peak load of the IEEE RTS-79 system is the expected value.

[0112] Example 2: IEEE RTS-96 system In this system, three IEEE RTS 79 systems are connected through six transmission lines.

[0113] Example 3: Modified IEEE RTS-79 system with a 20% penetration rate of renewable energy.

[0114] Example 4: Modified IEEE RTS-96 system with a 20% penetration rate of renewable energy.

[0115] The convergence condition of the Monte Carlo method is when the variance coefficient α is less than 5% or the number of samples reaches 100,000. The reliability result of the Monte Carlo method combined with the Newton-Raphson method and the interior point method is an exact solution. The hyperparameters of the deep neural network are detailed in Table 1. The following indicators are used to quantify the accuracy of the power flow calculation. P vm It refers to the proportion of voltage amplitude absolute error exceeding 0.001pu. va It refers to the proportion of voltage phase angle absolute error exceeding 0.01rad. pf / P qt Refers to the proportion of the absolute error of the active power flowing into / out of the branch exceeding 5MW. M_VM, M_VA, M_PF and M_PT correspond to the root mean square error of the above variables.

[0116] Table 1 Hyperparameters of different deep neural network intelligent modules under different examples

[0117]

[0118] The accuracy of a classifier is usually measured by the following metrics: ACC, SEN, SPE, g-mean.

[0119]

[0120] in

[0121] in

[0122]

[0123] Table 2 shows the data-driven power flow regression results under different examples. From Table 2, we can see that the root mean square error M_VM of the voltage amplitude does not exceed 5.00×10 in Examples 1 to 4. -4 pu; the RMS error of the voltage phase angle M_VA does not exceed 5.00×10 in Examples 1 to 4. -3rad; the root mean square error of branch power including M_PF and M_PT does not exceed 2.22MW. In addition, the probability indicators of voltage and power are less than 1.5% in Examples 1 to 4. Therefore, due to the strong approximation ability of deep neural networks, the calculation accuracy of all variables of interest is high.

[0124] Table 2 Data-driven power flow regression results under different examples

[0125]

[0126] Based on the high-precision power flow regression results, the problem samples can be identified by comparing them with the power system operation limit values. Table 3 lists the classification results of the problem samples by different methods under different examples. As shown in Table 3, under all examples, the overall accuracy (ACC) of the problem sample classification by the data-driven power flow regression and classification models exceeds 99.45% and 98.85%, respectively. It is worth noting that SEN quantifies the proportion of correctly classified problem samples to the total number of problem samples. The larger the SEN value, the fewer the problem samples that are misclassified. Using data-driven regression or classification results alone, the minimum values ​​of SEN in Examples 1-4 are 99.45% and 99.17%, respectively. By combining the data-driven regression results and classification results, that is, taking the union of the problem samples finally judged by the data-driven regression model and the classification model, the minimum value of the above SEN can be increased to 99.43% and 99.82% while maintaining high accuracy (ACC), thereby obtaining a more reliable classification result, which verifies the effectiveness of the present invention.

[0127] Table 3 Results of different methods for classification problem samples under different examples

[0128]

[0129] Table 4 lists the classification results of successful / faulty samples under different conditions. It can be seen from the number of problem samples and the number of faulty samples in the table that the sample imbalance problem is very serious in all cases. In addition, it can be intuitively seen that the imbalance ratio of successful samples to faulty samples can be reduced by first identifying the problem samples instead of directly classifying all samples as successful / faulty. When the traditional loss function is directly used, the ACC in all cases can reach more than 95%, but the SEN values ​​are 97.91%, 94.05%, 98.68% and 93.65% respectively. The influence of each faulty sample on the reliability index cannot be ignored. In this module, the larger the value of SEN, the fewer faulty samples are misjudged. Obviously, when using the traditional loss function, especially for cases 2 and 4, the effectiveness of successful / faulty sample identification is not satisfactory, and the values ​​are all below 95%. When the loss function proposed in the present invention is adopted, it can be seen from Table 4 that the SEN values ​​of Examples 1 to 4 can be increased to 99.42%, 97.99%, 99.02%, and 95.92%, respectively. The SEN values ​​in each example have been improved, and all have reached more than 95%. Since the imbalance ratio of success / failure samples is different in different examples, the setting of the weight β is also different. The threshold selected in this chapter for classification is 0.4. From the above results, it can be seen that the load shedding sensitive loss function of the present invention can effectively improve the identification accuracy of fault samples to more than 95%, meeting the accuracy requirements of industrial applications.

[0130] Table 4 Classification results of “success / failure” samples under different examples

[0131]

[0132] In order to verify the advancement of the present invention, it is compared with the most advanced data-driven reliability evaluation method, and the performance comparison results are shown in Table 5. It can be seen from Table 5 that the relative errors of LOLP and EENS calculated by MLKNN are much larger than those calculated by the proposed method. The minimum relative error of MLKNN is much larger than the maximum relative error of the proposed method. The main reason is that the proposed method has divided the problem into subtasks to significantly reduce the complexity. In addition, a specific learning strategy is further proposed to ensure the accuracy of the calculation. In addition, due to the direct mapping function of the deep neural network, the proposed method shows faster computing performance. MLKNN must compare a large number of samples with each other, which is very time-consuming. In short, compared with the most advanced data-driven methods, the proposed method shows significant advantages.

[0133] Compared with the benchmark values, from the perspective of computational accuracy, the reliability indicators estimated by the proposed method are very close to the exact values ​​through the high-precision identification of fault samples by deep neural networks. The relative errors of LOLP and EENS are no more than 3.3% and 0.45%, respectively. From the perspective of computational time, the proposed method greatly reduces the computational time by avoiding unnecessary power flow optimization. The speedup ratio of the proposed method ranges from 6.2 to 23.4. Therefore, when considering both transmission line faults and complex AC power flows, the data-driven framework proposed in this chapter can quickly and accurately calculate the reliability indicators of large power grids.

[0134] Table 5 Performance comparison between the present invention and the classical method for large power grid reliability assessment

[0135]

[0136] In summary, the present invention proposes a power grid reliability assessment method based on fine-grained data-driven. Compared with the existing data-driven operation reliability assessment method, the buffer introduced in the present invention can effectively identify the problem sample, that is, the system state where the flow exceeds the limit, and reduce the learning complexity of identifying fault samples. By pre-identifying a part of the successful samples through the buffer, the sample imbalance problem can also be effectively alleviated. In addition, the reliability assessment framework proposed in the present invention can increase the calculation speed by 4-23 times while meeting the calculation accuracy requirements. Compared with previous data-driven methods, the proposed method can comprehensively consider AC flow, branch random failures, etc., and is more guaranteed in calculation accuracy.

[0137] In the description of the present invention, it is necessary to understand that the terms "coaxial", "bottom", "one end", "top", "middle", "the other end", "upper", "one side", "top", "inside", "outside", "front", "center", "both ends" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0138] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0139] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A power grid reliability assessment method based on fine-grained data drive, characterized in that: The method comprises: Obtaining power grid system status data, and randomly extracting a portion of the power grid system status data according to the variance coefficient using the Monte Carlo method; The extracted power grid system status data is preprocessed to calculate the power difference of the system DC power flow before and after the power grid system is disconnected; the power of the DC power flow of each branch is calculated based on the admittance matrix and the voltage phase angle, expressed as P k =B k δ k ; The power difference of the DC power flow before and after the disconnection of different branches is calculated by the transfer distribution factor, which is expressed as The power difference vector of the DC power flow of all branches before and after the branch is disconnected is obtained, expressed as ΔP DC =P a -P b ; Among them, P k represents the power of the DC power flow on branch k, B k represents the admittance matrix on branch k, δ k represents the voltage phase angle on branch k, represents the power change of the DC power flow on branch k caused by the interruption of branch l, DF k-l represents the change in DC power flow on branch k caused by the interruption of branch l, P l represents the DC power flow on branch l, l≠k, k, l∈[1,2,...,L], L represents the total number of branches in the power grid system, ΔP DC The power difference vector of the DC power flow of all branches before and after the branch is disconnected, P a The power vector representing the DC power flow of all branches before the branch is disconnected, P b The power vector representing the DC power flow of all branches after the branch is disconnected; The power difference, active power vector and reactive power vector of the load, the maximum active power and reactive power of the unit, and the minimum active output and reactive output of the unit are obtained, and are input into the data-driven optimal power flow module as the input feature vector of the sample to be tested; the data-driven optimal power flow module is trained using a load shedding sensitive loss function, and the load shedding sensitive loss function is expressed as: <h2 style=";text-align:left;direction:ltr">L<h2 style=";text-align:left;direction:ltr"> OPF <h2 style=";text-align:left;direction:ltr"> =βL1+L2 Among them, L OPF represents the load shedding sensitivity loss function, β represents the weight factor, L1 represents the loss function of the successful sample, N1 represents the number of successful samples, Ω s represents the successful sample set, y i represents the actual value of the i-th problem sample, y i represents the predicted value of the i-th problem sample; L2 represents the loss function of the fault sample, N2 represents the number of fault samples, Ω f represents a set of fault samples; The power flow regression module calculates the regression result of the sample to be tested, and the power flow classification module predicts the classification result of the sample to be tested, and outputs the regression result and the classification result as the output feature vector of the sample to be tested; When the regression result exceeds the corresponding preset threshold, or / and the mark of the classification result is a problem mark, the sample to be tested is regarded as a problem sample; Input the input feature vector of the problem sample into the data-driven optimal power flow module, predict the classification result of the problem sample, and output the classification result as the output feature vector of the problem sample to determine the faulty sample; The minimum load shedding analysis is performed on the fault sample to calculate the active power of the inflow branch and the outflow branch, as well as the voltage amplitude and the voltage phase angle, which are expressed as: st P i (V,δ)-PD i +C i =0,i∈ND Q i (V,δ)-QD i =0,i∈ND 0≤C i ≤PD i ,i∈ND Among them, C i represents the amount of load loss at node i; ND represents the number of load nodes in the power grid system, P i (V,δ) represents the injected active power at node i, Q i (V,δ) represents the injected reactive power at node i; V i is the voltage amplitude at node i, δ i represents the voltage phase angle at node i, δ j represents the voltage phase angle at node j, δ ij represents the voltage phase difference from node i to node j, δ ij =δ i -δ j ;PD i with QD i Respectively represent the active load and reactive load at node i; PG i With QG i are the active power and reactive power of node i respectively; G ij is the real part of the node admittance matrix from node i to node j, B ij is the imaginary part of the node admittance matrix from node i to node j; is the minimum active output of the unit at node i; is the maximum active output of the unit at node i; is the minimum reactive power output of the unit at node i, is the maximum reactive power output of the unit at node i, NG represents the number of all generator nodes in the system, T k (V,δ) represents the transmission power on branch k, represents the maximum transmission power on branch k, L represents the number of all branches in the power grid system, V i min is the minimum voltage amplitude at node i, V i max is the maximum voltage amplitude at node i, and N represents the total number of nodes in the system; According to the calculation results of the minimum load shedding analysis of the fault sample, the uncertainty of the fault sample is calculated in combination with the variance coefficient, and the reliability evaluation result of the fault sample is output; The variance coefficient is expressed as: Among them, α represents the variance coefficient; represents the estimated value of the reliability index; Indicates estimated value The variance of .

Images