Large-scale power flow sample solving method based on physically-aware neural network

Abstract

The application discloses a large-scale power flow sample solving method based on a physically informed neural network. The application simulates random operation conditions of a high-proportion new energy power system by a Monte Carlo method to generate a power flow calculation sample training set which is not solved, then constructs a physically informed loss function based on a power flow equation, and performs self-supervised training on the neural network on the power flow calculation sample training set by using the physically informed loss function to obtain a trained neural network, and finally solves large-scale power flow samples to be solved by using the obtained neural network. The application considers random fluctuations of power system operation conditions after high-proportion new energy access, and trains the neural network through self-supervised learning to complete rapid solving of large-scale power flow samples.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power systems, and relates to an efficient solution method for large-scale power flow samples, in particular, a large-scale power flow sample solution method based on a physically informed neural network. BACKGROUND

[0002] The power flow calculation based on the steady state is the most basic and important link in the analysis of power systems. With the continuous increase of the penetration rate of renewable energy, both the supply and demand sides of power energy present high uncertainty. The operation complexity and randomness of high-proportion new energy power systems are increasingly deepening. The massive operation conditions under the double fluctuations of sources and loads put forward greater requirements for the calculation frequency and efficiency of large-scale systems under the N-1 criterion. The traditional power flow calculation method cannot balance the calculation accuracy, speed and cost, and is difficult to meet the challenges of fast calculation of power flow under high-proportion renewable energy power systems, so it is urgent to propose a new efficient solution method for large-scale power flow samples. SUMMARY

[0003] The technical problem to be solved by the application is to overcome the deficiencies of the prior art, and to provide a large-scale power flow sample solution method based on a physically informed neural network, which can quickly solve the power flow samples of high-proportion new energy power systems with strong randomness.

[0004] To this end, the application adopts the following technical solution: a large-scale power flow sample solution method based on a physically informed neural network, comprising the following steps:

[0005] S1, simulating the random operation conditions of a high-proportion new energy power system by a Monte Carlo method to generate an unsolved power flow calculation sample training set;

[0006] S2, constructing a physically informed loss function based on a power flow equation, and using the physically informed loss function to perform self-supervised training on the neural network on the power flow calculation sample training set to obtain a well-trained neural network;

[0007] S3, using the neural network obtained in S2 to solve the large-scale power flow sample to be solved.

[0008] Further, S1 comprises:

[0009] S11 simulates the output of a wind turbine using a Weibull distribution:

[0010]

[0011]

[0012] In the formula, c and k are the scale parameter and shape parameter of the Weibull distribution respectively, V is the wind speed, P is the active power output of the wind turbine at time t, and Pt Pnom is the rated power of the unit; V t V is the wind speed at the generator terminal; V ci V is the cut-in wind speed; V τ V is the rated wind speed; V co V is the cut-out wind speed; A, B, C are the parameters of the wind turbine;

[0013] S12 uses the uniform distribution random simulation load fluctuation in [0.8, 1.2] based on the original system power flow data;

[0014] S13 uses the Monte Carlo method to randomly sample the node load and wind speed to generate the unsolved power flow sample training set.

[0015] Further, the S2 comprises:

[0016] S21 constructs the corresponding feature matrix H and adjacency matrix A for each power flow sample data in the training set containing N nodes:

[0017]

[0018] A=Y

[0019] wherein, is the injection power complex vector of the power flow sample, and the dimension is N x 1; is the node voltage per unit complex vector of the power flow sample, and the dimension is N x 1; respectively, are 0-1 vectors representing node types, and 1 indicates that the corresponding node belongs to the type, and the dimensions are both N x 1.

[0020] S22 initializes the multi-layer perceptron neural network The neural network with θ as the weight matrix takes the feature matrix H and the adjacency matrix A as input, and outputs the predicted value of the node voltage per unit complex vector Meanwhile, the cascaded neural network is constructed to construct the physical informed loss function:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] wherein, is the prediction value of the node voltage per unit complex vector output by the neural network; w i respectively, are the corresponding weights of the sub-loss components loss i . represents the prediction value of the power injection complex vector of the power flow sample calculated according to the prediction value of the node voltage per unit complex vector and the power flow equation; represents the j-th element of the vector Real(), Imag() respectively represent taking the real part and the imaginary part of the vector; reflects the degree of violation of the power flow equation by the prediction value of the neural network; reflects the degree of violation of the given value by the voltage amplitude of the PV and the balance node predicted by the neural network; reflects the degree of violation of the given value by the voltage phase angle of the balance node predicted by the neural network.

[0028] S23 uses the mature open source optimizer Adam to implement the optimization problem, that is, to train the weight value θ of the neural network:

[0029] min L(θ)

[0030] The training is completed to obtain the corresponding neural network

[0031] Further, S3 uses the trained neural network to solve the large-scale power flow sample to be solved one by one to determine the power system power flow state under different operating conditions.

[0032] The present application has the following beneficial effects:

[0033] The present application first simulates the random operating conditions of the high-proportion new energy power system by the Monte Carlo method to generate an unsolved power flow calculation sample training set, then constructs a physical loss function based on the power flow equation, and uses the physical loss function to perform self-supervised training on the neural network on the power flow calculation sample training set to obtain a trained neural network, and finally uses the obtained neural network to solve the large-scale power flow sample to be solved. Considering the random fluctuations of the operating conditions of the power system after the high-proportion new energy is connected, the neural network is trained to quickly solve the large-scale power flow sample, which better balances the calculation accuracy, speed and cost, and copes with the fast calculation of the power flow under the high-proportion renewable energy power system, and serves as the basis for the safety check of the power system. It has important significance for the safe and stable operation of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is the flowchart of the large-scale power flow sample solving method based on the physical informed neural network of the present application;

[0035] Figure 2 Structure diagram of IEEE 39-node system used for example;

[0036] Figure 3 Structure diagram of neural network with three-layer perceptron;

[0037] Figure 4 Comparison diagram of solver solution result and solution result of the method for a certain power flow sample. DETAILED DESCRIPTION

[0038] The application will be further described below in conjunction with the accompanying drawings of the specification.

[0039] As Figure 1 shown, the large-scale power flow sample solving method based on physically-aware neural network, characterized in that, comprising the following steps:

[0040] S1, simulating random operation conditions of high-proportion new energy power system by Monte Carlo method to generate a training set of unsolved power flow calculation samples;

[0041] S2, constructing a physically-aware loss function based on power flow equation, and using the physically-aware loss function to perform self-supervised training on the neural network based on the training set of power flow calculation samples to obtain a well-trained neural network;

[0042] S3, using the neural network obtained in S2 to solve the large-scale power flow sample to be solved.

[0043] The following is a specific embodiment:

[0044] The embodiment provides a large-scale power flow sample solving method based on physically-aware neural network.

[0045] The example system adopts a standard IEEE 39-node test system, and the topological structure is as shown in Figures 2 to 4 .

[0046] S11, first, the Weibull distribution is used to simulate the output of the wind turbine at each node:

[0047]

[0048]

[0049] In the formula: wherein, c and k are the scale parameter and shape parameter of the Weibull distribution respectively, V is the wind speed; P is the active power output of the wind turbine at time t; P t is the rated power of the unit; V t is the wind speed at the machine end; V ci is the starting wind speed; V τ is the rated wind speed; V co is the cut-off wind speed; A, B, and C are wind turbine parameters.

[0050] S12, then on the basis of the original system power flow data, using the uniform distribution of the random simulation load fluctuation in the range of [0.8, 1.2].

[0051] S13, finally using Monte Carlo method random sampling of each node load and wind speed, generate 60000 groups of unsolved power flow samples, in which 42000 groups of power flow samples are randomly selected as the training set of neural network, and the remaining 18000 samples are as the test set to be solved.

[0052] S21, first, for each training set containing N nodes of power flow sample data preprocessing, respectively construct its corresponding feature matrix H and adjacency matrix A:

[0053]

[0054] A=Y

[0055] wherein, is the injection power complex vector of the power flow sample, the dimension is N x 1; is the node voltage per unit complex vector of the power flow sample, the dimension is N x 1; respectively, 0-1 vector representing node type, 1 indicates that the corresponding node belongs to the type, the dimension is N x 1.

[0056] S22, initialize multi-layer perceptron (MLP) neural network The weight matrix of the neural network is initialized with standard normal distribution. The neural network with θ as the weight matrix outputs the predicted value of the node voltage per unit complex vector by inputting the feature matrix H and the adjacency matrix A

[0057] Then, the cascaded neural network is constructed to construct the physical informed loss function:

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] wherein, is the predicted value of the node voltage per unit complex vector output by the neural network; w irespectively, are the corresponding weights of the sub-loss composition item loss i . represents the predicted value of the power injection complex vector of the power flow sample calculated according to the node voltage norm complex vector prediction value and the power flow equation; represents the jth element of the vector ; Real(), Imag() respectively represent taking the real part and the imaginary part of the vector; reflects the degree of violation of the power flow equation by the neural network prediction value; reflects the degree of violation of the given value by the voltage amplitude of the PV, the balance node predicted by the neural network; reflects the degree of violation of the given value by the voltage phase angle of the balance node predicted by the neural network.

[0065] S23, using the mature open source optimizer Adam to realize the optimization problem, that is, training the weight value θ of the neural network:

[0066] min L(θ).

[0067] After training, the corresponding neural network

[0068] S3, using the neural network trained in step S23, the large-scale power flow sample to be solved is solved one by one. The solution result of a certain power flow sample and the solution result of the professional solver are as shown in Figure 3 . The solver takes about 0.75s to solve the sample, and the method takes only 0.09s to solve. The comparison result shows that the method can significantly improve the solving efficiency while accurately solving the power flow sample.

[0069] The above only describes specific embodiments of the present application, but the technical features of the present application are not limited to this. Any changes or modifications made by those skilled in the art within the scope of the present application are covered by the patent scope of the present application.

Claims

1. A method for solving large-scale power flow samples based on a physically-informed neural network, characterized in that, Includes the following steps: S1, simulates the stochastic operating conditions of a high-proportion renewable energy power system using the Monte Carlo method to generate an unsolved power flow calculation sample training set; S2. Construct a physical informed loss function based on the power flow equation, and use the physical informed loss function to perform self-supervised training on the power flow calculation sample training set to obtain a well-trained neural network. S21 For each containing N Using the power flow sample data from the training set of each node, construct its corresponding feature matrix H and adjacency matrix A: ； ； in, Let the complex vector of the injected power of this power flow sample have dimension . ; Let be the complex vector of the per-unit nodal voltages of this power flow sample, with dimension . ; These are 0-1 vectors representing node types, where 1 indicates that the corresponding node belongs to that type, and their dimensions are both 0-1. ; S22 Initialize the multilayer perceptron neural network Multilayer perceptron MLP by The neural network with weight matrix H takes the feature matrix H and admittance matrix Y as input and outputs the predicted value of the complex vector of per-unit voltage of the nodes. Simultaneously, a cascaded neural network is constructed to generate a physical knowledge loss function: ； ； ； ； ； ； in, This is the predicted value of the complex vector of per-unit node voltages output by the neural network; Each is a component of the sub-loss The corresponding weights; This represents the predicted value of the complex vector of injected power in the power flow sample, calculated based on the complex vector prediction of the node voltage per unit value and the power flow equation. Representing vectors The j Each element; Real() and Imag() represent taking the real and imaginary parts of the vector, respectively; This reflects the degree to which the neural network predictions violate the power flow equations; This reflects the degree of deviation between the PV and the voltage amplitude of the slack node predicted by the neural network and the given values; This reflects the degree to which the voltage phase angle of the equilibrium node predicted by the neural network deviates from the given value; S23 uses the mature open-source optimizer Adam to solve the optimization problem, specifically the optimization of the weights of a neural network. Conduct training: ； The corresponding neural network is obtained after training. ; S3. Using the neural network obtained in step S2, solve the large-scale power flow sample to be solved.

2. The method for solving large-scale power flow samples based on a physically aware neural network according to claim 1, characterized in that, The S1 includes: S11 uses a Weibull distribution to simulate wind turbine output: ， ； In the formula: where, c and k These are the scale parameter and shape parameter of the Weibull distribution, respectively. V Wind speed; P For the wind turbine at all times t Those who have made meritorious contributions; P t This refers to the rated power of the unit. V t Wind speed at the generator end; V ci To activate the wind speed; V τ Rated wind speed; V co To cut off wind speed; A , B , C These are the fan parameters; S12 uses a uniformly distributed random simulation of load fluctuations in the range [0.8, 1.2] based on the original system power flow data. S13 uses the Monte Carlo method to randomly sample the load and wind speed at each node, generating an unsolved power flow sample training set.

3. The method for solving large-scale power flow samples based on a physically aware neural network according to claim 1, characterized in that, The S3 also includes: Using the neural network trained in S2 The large-scale power flow samples to be solved are solved one by one to determine the power flow state of the power system under different operating conditions.

Images