A method for correcting generator parameters of a collapse-proof multi-timescale generative adversarial network

Abstract

The present invention proposes a method for correcting generator parameters using a multi-time-scale generative adversarial network to prevent collapse. The method samples the random noise signal and the real signal of the generator at different time scales, and inputs the sampled generator signal data into the generator; then the sampled random noise signal of the generator, the real signal of the generator, and the generated signal generated by the generator are input into the discriminator; an anti-collapse threshold is set to judge whether the probability that the judgment signal obtained in the discriminator is a real signal is greater than the anti-collapse threshold, and the data greater than the anti-collapse threshold is stored in the model database, and the data less than the anti-collapse threshold is re-learned through a conditionally constructed generative adversarial network. The proposed method can effectively solve the problem of generator parameter correction, and sampling at multiple time scales can continuously optimize the obtained results, realize the generation and discrimination functions of the conditionally constructed generative adversarial network, and improve the anti-collapse capability in the generator parameter correction process.

Description

Technical Field

[0001] The invention belongs to the field of power system and artificial intelligence application, relates to an artificial intelligence method, and is suitable for correcting parameters of power system generators. Background Art

[0002] The dispatching, operation, safety assessment and stability analysis of power systems all rely on dynamic simulation results. The accuracy of the simulation results will directly affect the safe operation and economic dispatch of the power system. Therefore, it is particularly important to ensure that the simulation system can correctly reflect the dynamic characteristics of the actual system. A large number of calculation and analysis results show that there is a large difference between the simulation results and the measured data, and the inaccuracy of the generator model parameters is considered to be the main cause of the error. In summary, the correction of generator model parameters is an indispensable part of ensuring the safe and stable operation of the power system.

[0003] Most of the existing generator models in various power plants use typical mathematical models. Whether the parameters are based on manufacturer data, typical data or field measured data, the effectiveness of their dynamic characteristics needs to be further evaluated. In order to verify and obtain a more accurate generator model, it is common to use equipment field testing methods to obtain its parameters. However, regular field testing of generator model parameters has greatly increased the workload and cost of power plants, and the normal operation of the generator has also been affected to a certain extent. At the same time, the parameters measured on site will also be affected by various errors.

[0004] Therefore, the present invention proposes a method for correcting the parameters of a generator using a multi-time scale generative adversarial network to prevent collapse, which can solve the problems of workload and cost burden of power plants. At the same time, the collapse prevention idea, multi-time scale sampling method and generative adversarial network are combined, different time scales are selected, and the retention degree of the sampled data is different, so that the obtained results are more detailed, accurate and true; the continuous learning of the generator parameters by the conditional construction generative adversarial network is continuously improved; the collapse prevention threshold L is set, and the collapse prevention idea is applied to the generator parameter correction. If the probability that the judgment signal obtained by the discriminator D is a true signal is greater than the collapse prevention threshold L, the generator parameter matrix corresponding to the generated signal is saved in the generator parameter model database, otherwise the real signal data of the generator is re-input into the generator operation model as the random noise signal data of the generator, and is re-learned through the conditional construction generative adversarial network instead of being directly output as a result, pretending that the generator parameters have been learned in place, and the collapse prevention process can effectively improve the accuracy of the generator parameter correction, making the generator parameter correction result more reliable.

[0005] In summary, generator parameter correction requires an anti-collapse multi-time scale generative adversarial network method to coordinate the control of the entire system, improve the learning ability of the generative adversarial network, and achieve a more realistic correction effect on the generator parameters. Summary of the invention

[0006] The present invention proposes a method for correcting generator parameters using a collapse-proof multi-time-scale generative adversarial network. The method combines the collapse-proof idea, the multi-time-scale sampling method and the generative adversarial network for correcting the generator parameters. The steps of the proposed method during use are as follows:

[0007] Step (1): The proposed anti-collapse idea is used for generator parameter correction, and the anti-collapse threshold L is set. If the probability that the judgment signal obtained in the discriminator D is a true signal is greater than the anti-collapse threshold L, the generator parameter matrix corresponding to the generated signal output by the discriminator D is stored in the generator parameter model database; if the probability that the judgment signal obtained in the discriminator D is a true signal is less than the anti-collapse threshold L, the generator parameter data is re-input into the condition to construct a generative adversarial network for learning, instead of being directly output as a result;

[0008] The conditional construction generative adversarial network consists of a generator G and a discriminator D. Both the generator G and the discriminator D are deep neural network systems. The energy of the deep neural network system in the conditional construction generative adversarial network with anti-collapse idea is:

[0009] E(v,h)=-a T vb T hv T Wh (1)

[0010] Where W is the weight of the connection between the visible unit and the hidden unit, v is the visible unit, a is the bias of the visible unit, h is the hidden unit, b is the bias of the hidden unit, E(v,h) is the energy of the deep neural network system, and T is the transpose operation of the matrix;

[0011] The joint probability distribution of (v,h) is:

[0012]

[0013] Where O(v,h) is the joint probability distribution of (v,h), e is the logarithm with the irrational number e as the base, and Z is the partition function, whose normalization constant is used to ensure that the sum of the probability distribution is 1. The calculation formula of the partition function is:

[0014]

[0015] in, is the sum of all possible visible units and hidden units;

[0016] The edge probability of the visible layer is summed by the hidden units and is expressed as:

[0017]

[0018] in, is the sum of all hidden units;

[0019] The activation probabilities of each node in the hidden layer and visible layer are described as:

[0020]

[0021]

[0022] Among them, w i,j is the weight matrix composed of the connections between hidden units and visible units, h j is the jth hidden unit, v i is the i-th visible unit, b j is the bias of the jth hidden unit, a i is the bias of the ith visible unit; P(h j =1|v) is the activation probability of the hidden unit when the state of the visible unit is given; P(v i =1|h) is the activation probability of the visible unit when the state of the hidden unit is given; σ represents the logical function, which is described as:

[0023]

[0024] The action selection probability is:

[0025]

[0026] Where r is the probability of action selection, ln( ) is the logarithm of the natural logarithm with the irrational number e as the base, c q is the action probability factor, q is the probability of action r after update, and q is described as:

[0027]

[0028] Among them, k s is the probability state conversion factor, s is the conversion state with probability q;

[0029] The specific process of conditional construction of generative adversarial network is as follows: a set of adversarial deep neural networks is set up, including generator G and discriminator D; during the training process, the random noise active power signal, random noise reactive power signal, real active power signal and real reactive power signal of the generator are input into generator G, and generator G attempts to generate generated signal data that can be identified as real signal data by discriminator D, and then the random noise active power signal, random noise reactive power signal, real active power signal and real reactive power signal of the generator and the generated active power signal and generated reactive power signal generated by generator G are provided to discriminator D, and discriminator D distinguishes the generated signal data and real signal data generated by generator G, and generator G and discriminator D constitute a dynamic "game process";

[0030] The conditional construction of the generative adversarial network defines the loss function of the discriminator D as the cross entropy of a standard binary classification problem. The label corresponding to the real data is 1, and the label corresponding to the generated data is 0. When optimizing the discriminator D, the total expectation should be maximized, that is:

[0031]

[0032] Wherein, D is the discriminant function, G is the generating function, V(D,G) is the discriminant function of the discriminator D, x is the random noise signal, the real signal and the generated signal data of the generator input to the discriminator D, D(x(is the probability that x comes from the real signal data of the generator instead of the random noise signal and the generated signal data, logD(x) is the loss function of the discriminator D, z is the random noise signal and the real signal data of the generator input to the generator G, G(z) is the generated signal data generated by the generator G, D(G(z)) is the probability that the judgment signal obtained in the discriminator D is the real signal, log(1-D(G(z))) is the loss function of the generator G, P data (x) is the real signal data distribution of the generator, P z (z) is the random noise signal data distribution of the generator, The expected value of the real signal data of the generator is input to the discriminator D. The expected value of the generated signal data generated by the generator G is input to the discriminator D. is the maximum total expectation for optimizing the discriminator D;

[0033] When optimizing the generator G, the function value should be minimized, that is:

[0034]

[0035] in, The minimum total value for optimizing the generator G;

[0036] The value function of the conditional construction of the generative adversarial network by optimizing the discriminator D is:

[0037]

[0038] in, is the value function for optimizing the discriminator D;

[0039] The learning process of conditionally constructed generative adversarial networks is a min-max optimization process, and the optimal generator G model is:

[0040]

[0041] Among them, G* represents the optimal generator G model, arg represents a function that finds a set of parameters for a function, When the total value optimized by the discriminator D is the largest and the total value optimized by the generator G is the smallest, G and D are selected;

[0042] The optimal discriminator D model is:

[0043] D*=P data (x)logD(x)+P G (x)log(1-D(x)) (14)

[0044] Among them, D* represents the optimal discriminator D model, P G (x) is the data distribution of the generated signal generated by the generator G;

[0045] Set the collapse prevention threshold L, and provide the random noise active power signal, random noise reactive power signal, real active power signal, real reactive power signal of the generator and the generated active power signal and generated reactive power signal generated by the generator G to the discriminator D. The probability that the judgment signal outputted from the discriminator D is a real signal is that when the probability that the judgment signal obtained from the discriminator D is a real signal is greater than the collapse prevention threshold L, the generator parameter matrix corresponding to the generated signal is stored in the generator parameter model database; when the probability that the judgment signal obtained from the discriminator D is a real signal is less than the collapse prevention threshold L, the real active power signal and real reactive power signal of the generator are used as the random noise active power signal and random noise reactive power signal of the generator and input into the generator operation model, and the generative adversarial network is re-constructed through conditions for learning, which is specifically described as follows:

[0046]

[0047] Among them, y zl is the generator parameter output instruction, y outis an instruction to store the generator parameter matrix corresponding to the generated signal into the generator parameter model database, y stu is the instruction to input the real signal of the generator into the generator operation model as a random noise signal, L is the collapse prevention threshold;

[0048] Step (2): Apply the proposed method to calibrate the generator parameters;

[0049] The random noise active power signal, random noise reactive power signal, real active power signal and real reactive power signal of the generator are input into the generator G, and the random noise active power signal, random noise reactive power signal, real active power signal and real reactive power signal of the generator and the generated active power signal and generated reactive power signal generated by the generator G are provided to the discriminator D. Considering the active power accuracy and reactive power accuracy of the signal comprehensively, the total accuracy of the real signal data and the generated signal data is:

[0050]

[0051] Where X is the sequence of the real signal data of the generator, Y is the sequence of the generated signal data of the generator, S(X,Y) is the total accuracy of the real signal data and the generated signal data of the generator, S P (X,Y) is the active power accuracy of the real signal data and the generated signal data of the generator, S Q (X, Y) is the reactive power accuracy of the real signal data and the generated signal data of the generator;

[0052] Considering the amplitude and waveform of active power and reactive power, the accuracy of generator model parameters during generator parameter correction is:

[0053]

[0054] Among them, S P is the accuracy of active power amplitude and waveform, S Q is the accuracy of reactive power amplitude and waveform, Val is the accuracy of generator model parameters;

[0055] The active power and reactive power error values ​​before and after the generator parameter correction are evaluated by the mean square percentage error and the maximum error. The formula is:

[0056]

[0057] Δy max =max|y km -y ke | (19) Among them, M SPE is the mean square percentage error of the active power and reactive power error values ​​before and after correction, Δymax is the maximum error of active power and reactive power before and after correction, y km is the real signal data of the kth sampling step, y ke is the random noise signal data of the kth sampling step, n step is the number of sampling steps;

[0058] Step (3): The proposed method is used for data sampling of generator parameters. When using the conditionally constructed generative adversarial network to correct the generator parameters, different sampling steps are selected, that is, different time scales are selected. Compared with sampling at a single time scale, the application of multiple time scales enables the signal data to be learned more fully. Through the continuous learning of the system by the conditionally constructed generative adversarial network at different time scales, the cognitive ability of the conditionally constructed generative adversarial network is improved, and it can output more accurate data results.

[0059] Compared with the prior art, the present invention has the following advantages and effects:

[0060] (1) The present invention uses a generative adversarial network in a conditional generative adversarial network method to correct the generator parameters. The specific process is as follows: a group of adversarial deep neural networks are set, including a generator G and a discriminator D; during the training process, the random noise active power signal, random noise reactive power signal, real active power signal, and real reactive power signal of the generator are input into the generator G, and the generator G attempts to generate generated signal data that can be identified as real signal data by the discriminator D, and then the random noise active power signal, random noise reactive power signal, real active power signal, and real reactive power signal of the generator and the generated active power signal and generated reactive power signal generated by the generator G are provided to the discriminator D, and the discriminator D distinguishes the generated signal data generated by the generator G from the real signal data. The generator G and the discriminator D constitute a dynamic "game process". In the process of continuous "game", the correction of the generator parameters is achieved, and the correction result obtained is more accurate.

[0061] (2) The present invention proposes an anti-collapse idea for the conditional construction of a generative adversarial network to correct the parameters of the generator, so that the learning process of the conditional construction of the generative adversarial network is more perfect, and the cognitive ability of the conditional construction of the generative adversarial network is continuously improved. The anti-collapse threshold L is set, and the random noise active power signal, random noise reactive power signal, real active power signal, real reactive power signal of the generator and the generated active power signal and generated reactive power signal generated by the generator G are provided to the discriminator D. The probability that the judgment signal outputted from the discriminator D is a real signal, when the probability that the judgment signal obtained from the discriminator D is a real signal is greater than the anti-collapse threshold L, the generator parameter matrix corresponding to the generated signal is stored in the generator parameter model database; when the probability that the judgment signal obtained from the discriminator D is a real signal is less than the anti-collapse threshold L, the real signal of the generator is input into the generator operation model as the random noise signal of the generator, and the conditional construction of the generative adversarial network is re-learned. The anti-collapse process can effectively improve the correction accuracy of the generator parameters, make the conditional construction of the generative adversarial network learn the generator parameters more fully, and make the generator parameter correction result more reliable.

[0062] (3) The multi-time scale sampling method proposed in the present invention is used for data sampling of generator parameters. When the conditionally constructed generative adversarial network is used to correct the generator parameters, different sampling steps are selected, that is, different time scales are selected. Compared with sampling at a single time scale, the application of multiple time scales enables the signal data to be learned more fully. Through the continuous learning of the system by the conditionally constructed generative adversarial network at different time scales, the cognitive ability of the conditionally constructed generative adversarial network is improved, and it can output more accurate data results. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flow chart of generator parameter correction of the method of the present invention.

[0064] Figure 2 It is a schematic diagram of the collapse-proof multi-time-scale generative adversarial network control of the method of the present invention. DETAILED DESCRIPTION

[0065] The present invention proposes a method for correcting parameters of a generator of a multi-time-scale generative adversarial network for collapse prevention, which is described in detail in conjunction with the accompanying drawings as follows:

[0066] Figure 1 The following is a flow chart of the generator parameter correction method of the present invention. First, the generator operation parameter matrix is ​​input into the generator operation model. The signal data is sampled at a random time scale k, and then accumulated from k value until k+n stepValues ​​are taken as time scales in turn to sample the random noise signal and the real signal data. Secondly, the random noise signal and the real signal data of the generator are input into the generator operation model, and the random noise active power signal, the random noise reactive power signal, the real active power signal, and the real reactive power signal are output. The random noise active power signal, the random noise reactive power signal, the real active power signal, and the real reactive power signal data of the generator are input into the generator G, and the generated active power signal and the generated reactive power signal data generated by the generator G are output. Then the random noise active power signal, the random noise reactive power signal, the real active power signal, the real reactive power signal of the generator and the generated active power signal and the generated reactive power signal data generated by the generator G are input into the discriminator D, and the discriminator D obtains the probability that the judgment signal is a real signal as D(G(z)), and compares D(G(z)) with the collapse prevention threshold L. Finally, if the probability D(G(z)) that the judgment signal obtained by the discriminator D is a true signal is greater than the anti-collapse threshold L, the generator parameter matrix corresponding to the generated signal is saved in the generator parameter model database; otherwise, the true signal data of the generator is re-input into the generator operation model as the random noise signal data of the generator, and the generative adversarial network is re-constructed through conditions for learning.

[0067] Figure 2 This is a schematic diagram of the anti-collapse multi-time scale generative adversarial network control of the method of the present invention. First, the noise signal and the real signal data are sampled at a specified time scale. Then, the sampled data is input into the generator G for learning. Finally, the noise signal, the generated signal, and the real signal data are all input into the discriminator D for judgment. If the probability that the judgment signal obtained by the discriminator D is a real signal is greater than the anti-collapse threshold L, the generator parameter matrix corresponding to the generated signal is saved in the generator parameter model database, and the data less than the anti-collapse threshold is re-conditioned to construct a generative adversarial network for learning.

Claims

1. A method for correcting parameters of generators in a multi-timescale generative adversarial network for collapse prevention, characterized in that: This method combines the anti-collapse idea, multi-time scale sampling method and generative adversarial network for the correction of generator parameters. The steps of the proposed method in use are: Step (1): The proposed anti-collapse idea is used for generator parameter correction, and the anti-collapse threshold L is set. If the probability that the judgment signal obtained in the discriminator D is a true signal is greater than the anti-collapse threshold L, the generator parameter matrix corresponding to the generated signal output by the discriminator D is stored in the generator parameter model database; if the probability that the judgment signal obtained in the discriminator D is a true signal is less than the anti-collapse threshold L, the generator parameter data is re-input into the condition to construct a generative adversarial network for learning, instead of being directly output as a result; The conditional construction generative adversarial network consists of a generator G and a discriminator D. Both the generator G and the discriminator D are deep neural network systems. The energy of the deep neural network system in the conditional construction generative adversarial network with anti-collapse idea is: E(v,h)=-a T v-b T h-v T Wh (1) Where W is the weight of the connection between the visible unit and the hidden unit, v is the visible unit, a is the bias of the visible unit, h is the hidden unit, b is the bias of the hidden unit, E(v,h) is the energy of the deep neural network system, and T is the transpose operation of the matrix; The joint probability distribution of (v,h) is: Where O(v,h) is the joint probability distribution of (v,h), e is the logarithm with the irrational number e as the base, and Z is the partition function, whose normalization constant is used to ensure that the sum of the probability distribution is 1. The calculation formula of the partition function is: in, is the sum of all possible visible units and hidden units; The edge probability of the visible layer is summed by the hidden units and is expressed as: in, is the sum of all hidden units; The activation probabilities of each node in the hidden layer and visible layer are described as: Among them, w i,j is the weight matrix composed of the connections between hidden units and visible units, h j is the jth hidden unit, v i is the i-th visible unit, b j is the bias of the jth hidden unit, a i is the bias of the ith visible unit; P(h j =1|v) is the activation probability of the hidden unit when the state of the visible unit is given; P(v i =1|h) is the activation probability of the visible unit when the state of the hidden unit is given; σ represents the logical function, which is described as: The action selection probability is: Where r is the probability of action selection, ln() is the logarithm of the natural logarithm with the irrational number e as the base, c q is the action probability factor, q is the probability of action r after update, and q is described as: Among them, k s is the probability state conversion factor, s is the conversion state with probability q; The specific process of conditional construction of generative adversarial network is as follows: a set of adversarial deep neural networks is set up, including generator G and discriminator D; during the training process, the random noise active power signal, random noise reactive power signal, real active power signal and real reactive power signal of the generator are input into generator G, and generator G attempts to generate generated signal data that can be identified as real signal data by discriminator D, and then the random noise active power signal, random noise reactive power signal, real active power signal and real reactive power signal of the generator and the generated active power signal and generated reactive power signal generated by generator G are provided to discriminator D, and discriminator D distinguishes the generated signal data and real signal data generated by generator G, and generator G and discriminator D constitute a dynamic "game process"; The conditional construction of the generative adversarial network defines the loss function of the discriminator D as the cross entropy of a standard binary classification problem. The label corresponding to the real data is 1, and the label corresponding to the generated data is 0. When optimizing the discriminator D, the total expectation should be maximized, that is: Wherein, D is the discriminant function, G is the generating function, V(D,G) is the discriminant function of the discriminator D, x is the random noise signal, the real signal and the generated signal data of the generator input to the discriminator D, D(x) is the probability that x comes from the real signal data of the generator instead of the random noise signal and the generated signal data, logD(x) is the loss function of the discriminator D, z is the random noise signal and the real signal data of the generator input to the generator G, G(z) is the generated signal data generated by the generator G, D(G(z)) is the probability that the judgment signal obtained in the discriminator D is the real signal, log(1-D(G(z))) is the loss function of the generator G, P data (x) is the real signal data distribution of the generator, P z (z) is the random noise signal data distribution of the generator, The expected value of the real signal data of the generator is input to the discriminator D. The expected value of the generated signal data generated by the generator G is input to the discriminator D. is the maximum total expectation for optimizing the discriminator D; When optimizing the generator G, the function value should be minimized, that is: in, The minimum total value for optimizing the generator G; The value function of the conditional construction of the generative adversarial network by optimizing the discriminator D is: in, is the value function for optimizing the discriminator D; The learning process of conditionally constructed generative adversarial networks is a min-max optimization process, and the optimal generator G model is: Among them, G* represents the optimal generator G model, arg represents a function that finds a set of parameters for a function, When the total value optimized by the discriminator D is the largest and the total value optimized by the generator G is the smallest, G and D are selected; The optimal discriminator D model is: D*=P data (x)logD(x)+P G (x)log(1-D(x)) (14) Among them, D* represents the optimal discriminator D model, P G (x) is the data distribution of the generated signal generated by the generator G; Set the collapse prevention threshold L, and provide the random noise active power signal, random noise reactive power signal, real active power signal, real reactive power signal of the generator and the generated active power signal and generated reactive power signal generated by the generator G to the discriminator D. The probability that the judgment signal outputted from the discriminator D is a real signal is that when the probability that the judgment signal obtained from the discriminator D is a real signal is greater than the collapse prevention threshold L, the generator parameter matrix corresponding to the generated signal is stored in the generator parameter model database; when the probability that the judgment signal obtained from the discriminator D is a real signal is less than the collapse prevention threshold L, the real active power signal and real reactive power signal of the generator are used as the random noise active power signal and random noise reactive power signal of the generator and input into the generator operation model, and the generative adversarial network is re-constructed through conditions for learning, which is specifically described as follows: Among them, y zl Output command for generator parameters, y out is an instruction to store the generator parameter matrix corresponding to the generated signal into the generator parameter model database, y stu is the instruction to input the real signal of the generator into the generator operation model as a random noise signal, L is the collapse prevention threshold; Step (2): Apply the proposed method to calibrate the generator parameters; The random noise active power signal, random noise reactive power signal, real active power signal and real reactive power signal of the generator are input into the generator G, and the random noise active power signal, random noise reactive power signal, real active power signal and real reactive power signal of the generator and the generated active power signal and generated reactive power signal generated by the generator G are provided to the discriminator D. Considering the active power accuracy and reactive power accuracy of the signal comprehensively, the total accuracy of the real signal data and the generated signal data is: Where X is the sequence of the real signal data of the generator, Y is the sequence of the generated signal data of the generator, S(X,Y) is the total accuracy of the real signal data and the generated signal data of the generator, S P (X,Y) is the active power accuracy of the real signal data and the generated signal data of the generator, S Q (X, Y) is the reactive power accuracy of the real signal data and the generated signal data of the generator; Considering the amplitude and waveform of active power and reactive power, the accuracy of generator model parameters during generator parameter correction is: Among them, S P is the accuracy of active power amplitude and waveform, S Q is the accuracy of reactive power amplitude and waveform, Val is the accuracy of generator model parameters; The active power and reactive power error values ​​before and after the generator parameter correction are evaluated by the mean square percentage error and the maximum error. The formula is: Δy max =max|y km -y ke | (19) Among them, M SPE is the mean square percentage error of the active power and reactive power error values ​​before and after correction, Δy max is the maximum error of active power and reactive power before and after correction, y km is the real signal data of the kth sampling step, y ke is the random noise signal data of the kth sampling step, n step is the number of sampling steps; Step (3): The proposed method is used for data sampling of generator parameters. When using the conditionally constructed generative adversarial network to correct the generator parameters, different sampling steps are selected, that is, different time scales are selected. Compared with sampling at a single time scale, the application of multiple time scales enables the signal data to be learned more fully. Through the continuous learning of the system by the conditionally constructed generative adversarial network at different time scales, the cognitive ability of the conditionally constructed generative adversarial network is improved, and it can output more accurate data results.

Images