A microgrid impedance identification method based on a deep neural network

Abstract

The application discloses a micro-grid impedance identification method based on a deep neural network, and comprises the following steps: first, an equivalent impedance model of a micro-grid is established, then impedance data and voltage and current data of the micro-grid under different conditions are obtained through the equivalent impedance model of the micro-grid, so as to form a training data set; then, an improved deep neural network model is trained through swing exploration with the aid of the training data set; finally, bus voltage and output current of the micro-grid are collected in real time, and are input into the improved deep neural network model after training, so that corresponding impedance values are output.

Description

Technical Field

[0001] This invention belongs to the field of power system technology, and more specifically, relates to a microgrid impedance identification method based on deep neural networks. Background Technology

[0002] With the development of systems science, the interaction and stable operation of systems, taking microgrids and power systems as examples, have received widespread attention. Due to the maturity of distributed energy technologies, more and more microgrid systems, with photovoltaic power as their core, are being connected to the power system, injecting a large amount of energy into it. However, because microgrid systems are susceptible to impedance instability and fluctuations caused by frequent changes in the natural environment, large-scale connection of microgrids can easily lead to a series of phenomena such as asynchronous oscillations, and even cause the power system to eventually collapse, posing a very serious threat to the stable operation of the power system.

[0003] Rapidly obtaining the current power grid system impedance and using this as the primary basis for optimizing power grid control strategies has become paramount in the grid connection and stable operation of microgrids. Therefore, an accurate, stable, and rapid impedance measurement method is a crucial component of ensuring stable system operation and an important direction for future systems science research.

[0004] Traditional methods for measuring power grid impedance are categorized into active, passive, and quasi-passive methods. Active measurement involves applying various short-duration disturbance signals to the power grid to obtain the system's disturbance response and analyze it, then using algorithms to calculate impedance information. This method is affected by the type and duration of the disturbance signal, resulting in varying accuracy, and can impact grid stability and power quality. The biggest difference between passive and active impedance measurement is that passive impedance measurement does not require injecting any disturbance signals into the system. Instead, it uses the inherent information at each measurement point as the data foundation and employs various algorithms to estimate the current system impedance.

[0005] The above analysis shows that existing impedance identification techniques cannot simultaneously achieve ideal results in terms of identification speed, identification accuracy, and maintaining system stability. Therefore, it is necessary to conduct in-depth research on impedance identification techniques. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a microgrid impedance identification method based on deep neural networks. This method utilizes neural networks for microgrid impedance identification, offering advantages such as fast identification speed, high accuracy, and minimal need for parameter adjustments and manual interaction.

[0007] To achieve the above-mentioned objectives, the present invention provides a microgrid impedance identification method based on a deep neural network, characterized by comprising the following steps:

[0008] (1) Establish the equivalent impedance model of the microgrid.

[0009]

[0010]

[0011]

[0012] Among them, Z v_busj Z represents the impedance of the j-th voltage-controlled converter module in the microgrid, m represents the number of voltage-controlled converter modules, and Z... v_bus Z represents the impedance of all voltage-controlled converter modules in the microgrid; i_busk Z represents the impedance of the k-th current-controlled converter module in the microgrid, and n represents the number of current-controlled converter modules. i_bus The impedance represents the impedance of all current-controlled converter modules in the microgrid; D is the steady-state value of the duty cycle of the switching transistor in the microgrid; L is the inductance value in the microgrid; C is the capacitance value in the microgrid; R is the resistance value of the inductor coil in the microgrid; U is the bus voltage of the microgrid; and I is the output current of the microgrid.

[0013] (2) Obtain the training dataset

[0014] Based on the equivalent impedance model of the microgrid, impedance data, voltage and current data of the microgrid under different conditions are collected to form a training dataset.

[0015] (3) Constructing an improved deep neural network model

[0016] FCNNNet was chosen as the main network, with N fully connected layers. The voltage and current data in the training dataset were used as inputs, and the output was the impedance prediction value.

[0017] Replace the original ReLU activation function in the FCNNNet network with the GULX activation function;

[0018] Replace the original loss function in the FCNNNet network with a multi-connection loss function;

[0019]

[0020] in, Let be the predicted impedance value of the output corresponding to the i-th set of input data. Let α be the actual impedance value corresponding to the i-th set of input data. i Let be the decay factor for the i-th input data group, u be the total number of training iterations for each data group, and N be the number of fully connected layers.

[0021] (4) Training an improved deep neural network model through a swing exploration method

[0022] (4.1) Introduce a fully random weight change mechanism to initialize the improved deep neural network model: use the OCDW probability distribution function f(x) to change the network weights;

[0023]

[0024] Where ω is the network weight, n is the number of neurons in each fully connected layer, and η, ι, κ, and λ are constants;

[0025] (4.2) Input the voltage, current and corresponding impedance of the first group into the improved deep neural network model to obtain the corresponding impedance prediction value; then reinitialize the network weights according to the method described in step (4.1), and re-input the voltage and current data of the first group to obtain the corresponding impedance prediction value. Then, follow this process. When the voltage and current data of the first group have been iterated a total of u times, proceed to step (4.3).

[0026] (4.3) Substitute the u predicted impedance values ​​of the same set of outputs and the corresponding impedance values ​​of the inputs into the multi-connection loss function to calculate the loss function value of the first set of voltage and current data after training, denoted as loss1.

[0027] (4.4) Reinitialize the network weights according to the method described in step (4.1), then input the voltage and current data of the second group, and process the voltage and current data of the second group according to the method described in steps (4.2)-(4.3) to obtain the loss function value loss2. Then, continue in this manner until the training dataset is traversed.

[0028] (4.5) Substitute the loss function value corresponding to each group of data in the training dataset into the following formula:

[0029]

[0030] Where ω is the network weight, b is the network offset, ρ is the differential coefficient, L is the mean error, and ΔL is the change in loss.

[0031] (4.6) Determine whether the change in loss ΔL is within the preset interval (0,1). If it is, the training ends and proceeds to step (5); otherwise, return to step (4.1) and use the training dataset for the next round of training.

[0032] (5) Impedance identification of microgrids

[0033] The bus voltage and output current of the microgrid are collected in real time and input into the improved deep neural network model after training, so as to output the corresponding impedance value.

[0034] The objective of this invention is achieved as follows:

[0035] This invention relates to a microgrid impedance identification method based on deep neural networks. First, an equivalent impedance model of the microgrid is established. Then, impedance data, voltage, and current data of the microgrid under different conditions are obtained through the equivalent impedance model to form a training dataset. Next, an improved deep neural network model is trained using the training dataset through a swing-exploration method. Finally, the bus voltage and output current of the microgrid are collected in real time and input into the improved deep neural network model after training, thereby outputting the corresponding impedance value.

[0036] Meanwhile, the microgrid impedance identification method based on deep neural networks of this invention also has the following beneficial effects:

[0037] (1) By using the microgrid impedance identification method of deep neural network, the system impedance was identified, and the identification accuracy was greatly improved.

[0038] (2) The microgrid impedance identification method using deep neural networks in the identification process has higher accuracy compared with the active injection identification widely used in the Americas, and does not affect the power quality of the power system. At the same time, due to the fast forward propagation of neural networks, its identification speed is also greatly improved;

[0039] (3) The impedance equivalence achieved by using a novel microgrid impedance model, compared with the traditional equivalent model proposed by Thevenin et al. in France, introduces the dynamic characteristics of the system, realizes a more realistic real-time representation of the system impedance, greatly improves the application rate of the algorithm, and saves costs. Attached Figure Description

[0040] Figure 1 This is a flowchart of the microgrid impedance identification method based on deep neural networks of the present invention; Detailed Implementation

[0041] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0042] Example

[0043] For ease of description, the relevant technical terms appearing in the specific implementation method will be explained first:

[0044] FCNNNet (Full Convolution Nature Neural Network): A fully convolutional natural neural network Figure 1 This is a flowchart of the microgrid impedance identification method based on deep neural networks of the present invention.

[0045] In this embodiment, as Figure 1 As shown, the present invention provides a microgrid impedance identification method based on a deep neural network, comprising the following steps:

[0046] S1. Establish the equivalent impedance model of the microgrid.

[0047]

[0048]

[0049]

[0050] Among them, Z v_busj Z represents the impedance of the j-th voltage-controlled converter module in the microgrid, m represents the number of voltage-controlled converter modules, and Z... v_bus Z represents the impedance of all voltage-controlled converter modules in the microgrid; i_busk Z represents the impedance of the k-th current-controlled converter module in the microgrid, and n represents the number of current-controlled converter modules. i_bus This represents the impedance of all current-controlled converter modules in the microgrid; D is the steady-state value of the duty cycle of the switching transistor in the microgrid; L is the inductance value in the microgrid; C is the capacitance value in the microgrid; R is the resistance value of the inductor coil in the microgrid; U is the bus voltage of the microgrid; and I is the output current of the microgrid. This method can introduce nonlinear components such as inductors and capacitors to better represent the nonlinear characteristics between voltage, current, and impedance.

[0051] S2. Obtain the training dataset

[0052] Based on the equivalent impedance model of the microgrid, impedance data, voltage and current data of the microgrid under different conditions are collected to form a training dataset.

[0053] Table 1 shows a partial training data table:

[0054]

[0055]

[0056]

[0057]

[0058] S3. Construct an improved deep neural network model

[0059] FCNNNet was chosen as the main network, with N=10 fully connected layers. Voltage and current data from the training dataset were used as inputs, and the output was the impedance prediction value.

[0060] The original ReLU activation function in the FCNNNet network is replaced with the GULX activation function, as shown in the formula:

[0061]

[0062] Where e is the natural logarithm, β is the decay factor, and x is the input quantity;

[0063] Replace the original loss function in the FCNNNet network with a multi-connection loss function;

[0064]

[0065] in, Let z be the predicted impedance value of the output corresponding to the i-th set of input data. i Let α be the actual impedance value corresponding to the i-th set of input data. i α is the decay factor for the i-th input data group, u is the total number of training iterations for each data group, and N is the number of fully connected layers; i Typically, 0.95 is used, and u is typically 256;

[0066] S4. Train the improved deep neural network model using a swing-and-explore approach.

[0067] S4.1. Introduce a fully random weight change mechanism to initialize the improved deep neural network model: Use the OCDW probability distribution function f(x) to change the network weights;

[0068]

[0069] Where P(ω) represents the probability of generating ω, ω is the network weight, n is the number of neurons in each fully connected layer, usually within a preset interval (0,16), and η, ι, κ, λ are constants, typically taken as 4, 8, 2.3, ...

[0070] S4.2 Input the voltage, current and corresponding impedance of the first group into the improved deep neural network model to obtain the corresponding impedance prediction value; then reinitialize the network weights according to the method described in step S4.1, and re-input the voltage and current data of the first group to obtain the corresponding impedance prediction value. Then, in this way, when the voltage and current data of the first group have been iterated a total of u times, proceed to step S4.3.

[0071] S4.3 Substitute the u predicted impedance values ​​of the same set of outputs and the corresponding input impedance values ​​into the multi-connection loss function to calculate the loss function value of the first set of voltage and current data after training, denoted as loss1;

[0072] S4.4 Reinitialize the network weights according to the method described in step S4.1, then input the voltage and current data of the second group, and process the voltage and current data of the second group according to the method described in steps S4.2-S4.3 to obtain the loss function value loss2. Then, repeat this process until the training dataset is completely traversed.

[0073] S4.5 Substitute the loss function value corresponding to each group of data in the training dataset into the following formula:

[0074]

[0075] Where ω is the network weight, b is the network offset, ρ is the differential coefficient, L is the mean error, and ΔL is the change in loss; where ω is within the preset interval (0,1); b is within the preset interval (0,1); and ρ is within the preset interval (0.3,0.8).

[0076] S4.6 Determine whether the change in loss ΔL is within the preset interval (0,1). If yes, training ends and proceed to step S5; otherwise, return to step S4.1 and use the training dataset for the next round of training.

[0077] S5. Impedance Identification of Microgrids

[0078] The bus voltage and output current of the microgrid are collected in real time and input into the improved deep neural network model after training, so as to output the corresponding impedance value.

[0079] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A microgrid impedance identification method based on deep neural networks, characterized in that, Includes the following steps: (1) Establish the equivalent impedance model of the microgrid; ； ； ； in, Indicating the first in a microgrid The impedance of a voltage-controlled converter module Indicates the number of voltage-controlled converter modules. This represents the impedance of all voltage-controlled converter modules in the microgrid; Indicating the first in a microgrid The impedance of a current-controlled converter module Indicates the number of current-controlled converter modules This represents the impedance of all current-controlled converter modules in the microgrid; It is the steady-state value of the duty cycle of the switching transistor in the microgrid. It is the inductance value in a microgrid. It is the capacitance value in the microgrid. It is the resistance value of the inductor coil in the microgrid. This is the bus voltage of the microgrid. This refers to the output current of the microgrid. (2) Obtain the training dataset; Based on the equivalent impedance model of the microgrid, impedance data, voltage and current data of the microgrid under different conditions are collected to form a training dataset. (3) Construct an improved deep neural network model; FCNNNet was chosen as the main network component. The fully connected layer takes voltage and current data from the training dataset as input and outputs impedance prediction values. Replace the original ReLU activation function in the FCNNNet network with the GULX activation function; The GULX activation function is: ； in, It is the natural logarithm. It is the attenuation factor. It is the input quantity; Replace the original loss function in the FCNNNet network with a multi-connection loss function; ； in, For the first The input data corresponds to the predicted output impedance value. For the first The input data corresponds to the actual impedance value. For the first Attenuation factor under the input data set The total number of times to train for each set of data. It is the number of fully connected layers; (4) Train the improved deep neural network model through a swing exploration method; (4.1) Introducing a fully random weight change mechanism to initialize the improved deep neural network model: using the OCDW probability distribution function. Change the network weights; ； in, It is network weight. It refers to the number of network layers. It is the number of neurons in each fully connected layer. It is a constant; (4.2) Input the voltage, current and corresponding impedance of the first group into the improved deep neural network model to obtain the corresponding impedance prediction value; then reinitialize the network weights according to the method described in step (4.1), and re-input the voltage and current data of the first group to obtain the corresponding impedance prediction value. Then, repeat this process until the voltage and current data of the first group have been iterated for a total of 100 iterations. After that, proceed to step (4.3). (4.3) Output the same group Substituting the predicted impedance value and the corresponding input impedance value into the multi-connected loss function, the loss function value after training the first set of voltage and current data is calculated, denoted as . ; (4.4) Reinitialize the network weights according to the method described in step (4.1), then input the voltage and current data of the second group, and process the voltage and current data of the second group according to the methods described in steps (4.2)-(4.3) to obtain the loss function value. And so on, until the entire training dataset has been traversed; (4.5) Substitute the loss function value corresponding to each group of data in the training dataset into the following formula: ； in, It is network weight. It is the network offset. These are differential coefficients. It is the mean error. It is the change in loss; (4.6) Determine the change in loss Is it within the preset range? If yes, then training ends and proceed to step (5); otherwise, return to step (4.1) and use the training dataset for the next round of training. (5) Impedance identification of microgrids; The bus voltage and output current of the microgrid are collected in real time and input into the improved deep neural network model after training, so as to output the corresponding impedance value.

Images