Network construction type converter frequency control method based on physical model and data fusion

Abstract

The invention discloses a network construction type converter frequency control method based on a physical model and data fusion, and belongs to the field of power systems. According to the method, the problem that an existing network-forming converter is insufficient in adaptability in control parameter design is solved, historical data and real-time operation data of the frequency are fused, a small-signal physical model is combined, and the physical model and prediction data jointly drive the converter to output the frequency. Specifically, the method comprises the following steps: firstly, preprocessing historical data, segmenting the preprocessed historical data into a vector set and endowing the vector set with a time axis; secondly, constructing an LSTM network, and realizing time sequence prediction through a forgetting gate, an input gate, memory unit updating and an output gate; establishing a small-signal physical model according to the large-signal physical model of the network-forming converter; and finally, fusing LSTM prediction data with the physical model, converting LSTM prediction output data into a small signal form, and giving weight compensation to a feedback channel to realize frequency control.

Description

Technical Field

[0001] This invention relates to the field of power system technology, specifically to a frequency control method for grid-type converters based on physical models and data fusion. Background Technology

[0002] With the rapid development of renewable energy, distributed energy has gradually become an important component of modern power systems. As a core device for grid connection of new energy sources, grid-connected converters have been widely used in microgrids and distributed generation systems due to their technical advantages in maintaining grid stability. However, as the performance of source-side equipment gradually degrades and the dynamic characteristics of the power grid become increasingly complex, the design of control parameters for grid-connected converters faces the problem of insufficient adaptability. This problem may lead to a decline in system dynamic performance, or even cause grid stability failure in some cases, thereby affecting the normal operation of the entire distribution network. Therefore, this invention proposes a frequency control method for grid-connected converters based on physical models and data fusion. Summary of the Invention

[0003] The purpose of this invention is to provide a frequency control method for grid-type converters based on physical models and data fusion. Based on historical frequency data and real-time operating data, combined with a small-signal control model at the physical level, the method improves the stability of frequency control, avoids the problem of mismatched control parameters, and solves the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a frequency control method for a grid-type converter based on physical model and data fusion. The method is based on historical frequency data and real-time operating data, combined with the small-signal physical model of the grid-type converter, to propose a frequency control method for the grid-type converter. The frequency control method for the grid-type converter includes the following steps: Step 1: Historical Data Preprocessing The historical data of the system operation is processed, including calculating the mean and standard deviation, marking outliers, handling outliers, and assigning the preprocessed vector set to the time axis; Step 2: Establish a Long Short-Term Memory Network This includes establishing a forget gate, establishing an input gate, defining the principles for updating memory units, and establishing an output gate; Step 3: Establishing the small-signal physical model of the grid-type converter This includes establishing the large-signal physical model of the grid converter and deriving the small-signal physical model of the grid converter; Step 4: Fuse LSTM prediction data with physical model This includes acquiring LSTM prediction data and fusing LSTM prediction data with grid-type converters.

[0005] Furthermore, in step 1, the historical data is divided into segments of length [missing information]. N Given a vector set containing both vector and matrix data, assuming the data in the vector set is in column vector format, the method for detecting null values, abnormally large values, or abnormally small values ​​for each vector set includes the following steps: Step 1-1, Calculate the mean and standard deviation: For all values ​​in the preprocessed data, calculate the mean and standard deviation; Define the valid data of the vector set as The data is , where n represents the total number of data; The mean of this data and standard deviation The calculation is as follows:

[0006]

[0007] In the formula, x i Is The Middle i One element; Step 1-2, mark outliers: For each data point, calculate its standardized value. Typically, if the standardized value of a data point is greater than a certain threshold, it is considered an outlier or outlier, expressed as follows:

[0008] If the above conditions are met, the data will be marked as an outlier; If the data contains null values, the program will mark them as "NaN", which is also considered an outlier. Steps 1-3, handling outliers: For outlier handling, if the outlier is a marginal value, it is assigned a mean; if the outlier is not marginal, linear interpolation of nearby data points is used, i.e.:

[0009] It is worth noting that the edge refers to a vector. The serial number is 1 or n Element; Steps 1-4: Assign the preprocessed vector set to the time axis: Based on the sampling step size of historical data, the first data point of the vector set is defined as the zero point of time; Then, time is allocated to each data point according to the sampling step size, forming a timeline of the data.

[0010] Furthermore, step 2 includes the following steps: Step 2-1, Create the forget gate: The forget gate determines whether the state information from the previous time step is forgotten, and it is expressed as follows:

[0011] in, f t The output of the forget gate; and b f These are the weights and biases of the forget gate, respectively; h t-1 and v t These are the memory information from the previous moment and the input data from the current moment, respectively. S Use the Sigmoid activation function; Step 2-2, establish the input gate: The input gate determines how the current input vector set will update the memory cell, expressed as:

[0012]

[0013] in, i t The output of the input gate; Candidate memory states; , , b i and b C For the corresponding weights and biases; Steps 2-3: Define the principles for updating memory units: The current memory state of the neural network is updated based on the data from the input gate and the data retained after passing through the forget gate, which is expressed as:

[0014] In the formula, C t represent t The state of memory at any given moment; similarly C t-1 represent t The memory state at time -1; Steps 2-4: Create the output gate: The output gate does not represent the output of the LSTM, but rather the output of the memory cell, which is represented as:

[0015]

[0016] in, o t This is the output of the output gate; and b o These are the weights and biases of the forget gate, respectively; h t This refers to the memory information at the current moment.

[0017] Furthermore, in step 3, the method for establishing the large-signal physical model of the grid-type converter is as follows: The grid-connected model of a grid-connected converter can be equivalent to an AC voltage source; Define the voltage amplitude of the grid converter as U v The real-time phase angle is The impedance of the connection line connecting the converter to the common grid connection point is expressed as a complex number. ,in R e Represents the line resistance. X e It is the line reactance. j It is an imaginary unit; therefore, the active power delivered by a grid-connected converter to the common grid connection point is... P and reactive power Q Expressed as:

[0018]

[0019] In the formula, U p The voltage amplitude at the common grid connection point; The real-time phase angle of the public grid connection point; For complex numbers The modulus.

[0020] At the same time, the control branches of the grid-type converter, including the active power-phase angle and reactive power-voltage control branches, can be uniformly represented as follows:

[0021]

[0022] In the formula, g p and g q These represent the static gains of the active power-phase angle and reactive power-voltage control branches, respectively. J and D These represent the inertia coefficient and the droop coefficient, respectively. sRepresents the Laplace operator; , , and These represent reference values ​​for active power, reactive power, frequency, and voltage, respectively.

[0023] Furthermore, in step 3, the method for establishing and deriving the small-signal physical model of the grid converter is as follows: Based on the large-signal physical model of the grid converter, a small-signal physical model of the grid converter is established, with the control branch as the forward path and the power delivered by the converter as the feedback path. For the feedback channel, the power acquisition process passes through a low-pass filter before entering the main control branch. Therefore, the small-signal representations of the active and reactive power feedback channels are as follows:

[0024]

[0025] In the formula, Small signal form representing variables; This represents the cutoff frequency of the low-pass filter; The small signal representing the power angle.

[0026] Furthermore, the physical model for small signals is as follows:

[0027] .

[0028] Further, in step 4, obtaining the LSTM prediction data specifically involves: In the frequency prediction task of grid-type converters, the historical frequency data of grid-type converters are used as the input of LSTM. Define input vector V =[ V 1, V 2,…, V N ] T A frequency operation dataset for a grid-type converter; Arrange the dataset according to n Each sampling period is divided into a vector set, i.e., an arbitrary vector set. V k ( k The frequency data of (=1,2,…,n) are represented as follows: V k =[ 1, 2,…, n ] T; Predicted output vector W k =[ w 1, w 2,…, w n ] T For a grid-type converter in n Prediction frequency for each sampling period; The output of LSTM prediction data is to transform the predicted output vector into a small signal form, that is... Y k =[ y 1, y 2,…, y n ] T The specific calculation method is as follows: ; In the formula, Y k This represents the small-signal form of the k-th predicted output vector; A reference value indicating frequency; w 1, w 2,…, w n This represents the future predicted by the LSTM model. n Angular frequency value for each sampling period; n Indicates the number of sampling periods.

[0029] Furthermore, the small-signal form of the predicted output vector transformation will inherit the vector set. V k The timeline indicates that after the current prediction data output time series ends, the output vector corresponding to the next set of vectors will continue to output its prediction data.

[0030] Furthermore, in the LSTM prediction output step, historical data in the database is updated along with real-time sampling data from the grid-type converter, thereby achieving rapid dynamic prediction with a minimum data latency of [missing information]. n One sampling period.

[0031] Furthermore, in step 4, the driving method for fusing the small-signal physical model of the grid-type converter with the LSTM prediction data is as follows: Based on the elements output by the Long Short-Term Memory Network, which are the predicted values ​​of frequency, the values ​​are integrated to transform them into the predicted values ​​of the small-signal power angle. These values ​​are then weighted and compensated to the feedback channel of the small-signal power angle to achieve frequency control.

[0032] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes a Long Short-Term Memory (LSTM) network to perform multi-step prediction of frequency change trends, enabling earlier detection of changes on both the grid and source sides, thus achieving faster frequency regulation. Simultaneously, establishing large-signal and small-signal physical models for the grid-connected converter helps predict its performance under different operating conditions. Furthermore, fitting and predicting historical frequency data using an LSTM network, combined with closed-loop compensation using the physical model, effectively reduces frequency errors and enhances the robustness and stability of the power system under dynamic conditions. By fusing the LSTM prediction data with the physical model, the invention overcomes the problems of traditional physical models relying on fixed parameters and having poor adaptability, achieving adaptive frequency control over a wider operating range. Attached Figure Description

[0033] Figure 1 This is a flowchart of the frequency control method for a grid-type converter based on physical model and data fusion according to the present invention. Figure 2 This is a schematic diagram of the prediction output of the LSTM of the present invention; Figure 3 This is a schematic diagram of the response of the grid-type converter of the present invention to the LSTM output. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Insufficient adaptability of control parameter design in grid-connected converters can lead to a decline in system dynamic performance and, in some cases, even cause grid stability failures, thus affecting the normal operation of the entire distribution system. To address this issue, please refer to [link to relevant documentation / reference]. Figures 1-3 This embodiment provides the following technical solution: A frequency control method for grid-type converters based on physical models and data fusion, wherein the method is a frequency control method for grid-type converters based on historical frequency data and real-time operating data, combined with a small-signal physical model; The frequency control method for the grid-type converter includes the following steps: Step 1: Historical Data Preprocessing The system's historical data is processed, including calculating the mean and standard deviation, marking outliers, handling outliers, and assigning the preprocessed vector set to the time axis. Step 2: Construct a Long Short-Term Memory (LSTM) network This includes establishing a forget gate, establishing an input gate, defining the principles for updating memory units, and establishing an output gate; Step 3: Establish the small-signal physical model of the grid-type converter, including: Step 3-1: Establish the large-signal physical model of the grid-type converter; Step 3-2: Derive the small-signal physical model of the grid converter; Step 4: Integrate LSTM prediction data with the physical model, including: Step 4-1: Obtain the prediction data from the LSTM; Step 4-2: Integrate the physical model of the network-type converter small signal and LSTM data.

[0036] The technical effects of the above-mentioned solution are as follows: First, by calculating the mean and standard deviation, and labeling and processing outliers, a clean and accurate data foundation can be provided for subsequent model training, thereby improving the predictive performance of the model. Second, by constructing a Long Short-Term Memory (LSTM) network, the time-series characteristics of frequency data can be captured, enabling multi-step prediction of frequency change trends. This method can detect changes on the grid side and source side earlier, thus achieving faster frequency regulation. Third, the small-signal physical model of the grid-type converter includes both a large-signal physical model and a small-signal physical model. The large-signal physical model describes the electrical characteristics of the converter under a wide range of changes, while the small-signal physical model focuses on subtle dynamic changes near a certain operating point. This combination allows the control strategy to respond more accurately to various operating conditions, enhancing the robustness of the power system. Finally, by performing closed-loop compensation between the LSTM prediction output and the physical model, frequency errors can be effectively reduced. By combining data-driven and model-driven methods, not only is the prediction accuracy improved, but the stability of the power system under dynamic operating conditions is also enhanced.

[0037] In summary, by preprocessing data, constructing an LSTM, establishing a small-signal physical model, and fusing the two, high-precision, adaptive frequency control was achieved. This precise control not only improves the accuracy of frequency control but also optimizes the control strategy, reduces unnecessary control actions, and improves the overall efficiency of the power system, thereby overcoming the problems of traditional physical models relying on fixed parameters and having poor adaptability.

[0038] In step 1, the historical data is divided into segments of length [length missing]. N Given a vector set containing both vector and matrix data, assuming the data in the vector set is in column vector format, the method for detecting null values, abnormally large values, or abnormally small values ​​for each vector set includes the following steps: Step 1-1, Calculate the mean and standard deviation: For all values ​​in the preprocessed data, calculate the mean and standard deviation; Define the valid data of the vector set as The data is ,in n Represents the total number of data points; The mean of this data and standard deviation The calculation is as follows:

[0039]

[0040] In the formula, x i Is The Middle i One element; Step 1-2, mark outliers: For each data point, calculate its standardized value. Typically, if the standardized value of a data point is greater than a certain threshold, it is considered an outlier or outlier, expressed as follows:

[0041] If the above conditions are met, the data will be marked as an outlier; If the data contains null values, the program will mark them as "NaN", which is also considered an outlier. Steps 1-3, handling outliers: For handling outliers, if the outlier is a marginal value, a mean is assigned between it and the nearest data point; if the outlier is not marginal, linear interpolation of the nearest data point is used, i.e.:

[0042] It is worth noting that the edge refers to a vector. The serial number is 1 or n Element; Steps 1-4: Assign the preprocessed vector set to the time axis: Based on the sampling step size of historical data, the first data point of the vector set is defined as the zero point of time; Then, time is allocated to each data point according to the sampling compensation, forming a timeline of the data.

[0043] The technical effects of the above solution are as follows: By dividing historical data into vector sets and calculating the mean and standard deviation of the data in each vector set, the statistical characteristics of the data in each vector set can be accurately understood, providing a foundation for subsequent data analysis and processing. Null values, outliers, and outliers are detected and marked, effectively identifying outliers and missing values ​​in the data, thereby improving the quality and reliability of the data. The preprocessed vector sets are assigned to a time axis, and time is allocated to each data point according to the sampling step size of the historical data to form complete time series data, which helps to better explore the temporal correlation and patterns in the data. By preprocessing historical data, the quality of the data can be effectively improved, avoiding the impact on subsequent modeling and decision-making.

[0044] Step 2 includes the following steps: Step 2-1, Create the forget gate: The forget gate determines whether the state information from the previous time step is forgotten, and it is expressed as follows:

[0045] in, f t The output of the forget gate; and b f These are the weights and biases of the forget gate, respectively; h t-1 and v t These are the memory information from the previous moment and the input data from the current moment, respectively. S Use the Sigmoid activation function; Step 2-2, establish the input gate: The input gate determines how the current input vector set will update the memory cell, expressed as:

[0046]

[0047] in, i t The output of the input gate; Candidate memory states; , , b i and b C For the corresponding weights and biases; Steps 2-3: Define the principles for updating memory units: The current memory state of the neural network is updated based on the data from the input gate and the data retained after passing through the forget gate, which is expressed as:

[0048] In the formula, C t represent t The state of memory at any given moment; similarly C t-1 represent t The memory state at time -1; Steps 2-4: Create the output gate: The output gate does not represent the output of the LSTM, but rather the output of the memory cell, which is represented as:

[0049]

[0050] in, o t This is the output of the output gate; and b o These are the weights and biases of the forget gate, respectively; h t This refers to the memory information at the current moment.

[0051] The technical effects of the above solution are as follows: The establishment of the forget gate enables LSTM to determine whether the state information of the previous time step is forgotten. Through the selective forgetting mechanism, useful information can be retained and useless information can be discarded, improving the efficiency and accuracy of information processing. The establishment of the input gate determines how the current input vector set updates the memory unit, enabling the model to precisely control the input of information, avoiding unnecessary information interference, and improving the accuracy of memory unit updates. The memory unit update principle is based on the input gate and the data retained by the forget gate, dynamically updating the current memory state of the neural network, so that the memory state can better reflect the current data characteristics and enhance the model's adaptability to data changes. The establishment of the output gate controls the output of the memory unit, enabling LSTM to stably output memory information, avoiding over-exposure or under-exposure of information, and improving the stability and reliability of the output. In summary, through the synergistic effect of the forget gate, input gate, memory unit update and output gate, LSTM can effectively process complex sequence information, capture long-term dependencies in the data, and improve the processing capability of time series data.

[0052] In step 3, the method for establishing the large-signal physical model of the grid-type converter is as follows: The grid-connected model of a grid-connected converter can be equivalent to an AC voltage source; Define the voltage amplitude of the grid converter as U v The real-time phase angle is The impedance of the connection line connecting the converter to the common grid connection point is expressed as a complex number. ,in R e Represents the line resistance. X e It is the line reactance. j It is an imaginary unit; therefore, the active power delivered by a grid-connected converter to the common grid connection point is... P and reactive power Q Expressed as:

[0053]

[0054] In the formula, U p The voltage amplitude at the common grid connection point; The real-time phase angle of the public grid connection point; For complex numbers The modulus.

[0055] At the same time, the control branches of the grid-type converter, including the active power-phase angle and reactive power-voltage control branches, can be uniformly represented as follows:

[0056]

[0057] In the formula, g p and g q These represent the static gains of the active power-phase angle and reactive power-voltage control branches, respectively. J and D These represent the inertia coefficient and the droop coefficient, respectively. s Represents the Laplace operator; , , and These represent reference values ​​for active power, reactive power, frequency, and voltage, respectively.

[0058] The technical effects of the above-mentioned technical solution are as follows: by accurately calculating the active and reactive power delivered by the grid-connected converter to the common grid connection point, it is helpful to better understand and control the output characteristics of the converter. The established control branch can dynamically adjust the operating state of the converter. Through the adjustment of the control branch, the converter can quickly respond to changes in the power grid and maintain the stable operation of the power grid.

[0059] In step 3, the method for deriving the small-signal physical model of the grid converter is as follows: Based on the large-signal physical model of the grid converter, a small-signal physical model of the grid converter is established, with the control branch as the forward path and the power delivered by the converter as the feedback path. For the feedback channel, the power acquisition process passes through a low-pass filter before entering the main control branch. Therefore, the small-signal representations of the active and reactive power feedback channels are as follows:

[0060]

[0061] In the formula, Small signal form representing variables; This represents the cutoff frequency of the low-pass filter; The small signal representing the power angle; The physical model for small signals is as follows:

[0062] .

[0063] The technical effects of the above solution are as follows: Based on the large-signal physical model of the grid-type converter, a small-signal physical model is established, which can accurately analyze the dynamic characteristics of the converter under small-signal disturbances, and help to better understand and control the operating state of the converter. Through the synergistic effect of the control branch and the feedback channel, the small-signal physical model can quickly respond to small-signal disturbances, improve the dynamic response speed of the converter, and ensure the stable operation of the system under small disturbances. The small-signal physical model of the control branch can accurately describe the relationship between the power angle and the voltage and power changes, thereby realizing precise control of the converter output.

[0064] In step 4, the prediction data of the LSTM is obtained, specifically as follows: In the frequency prediction task of grid-type converters, the historical frequency data of grid-type converters are used as the input of LSTM. Define input vector V =[ V 1, V 2,…, V N ] T A frequency operation dataset for a grid-type converter; Arrange the dataset according to n Each sampling period is divided into a vector set, i.e., an arbitrary vector set. V k ( k The frequency data of (=1,2,…,n) are represented as follows: V k =[ 1, 2,…, n ] T ; Predicted output vector W k =[ w 1, w 2,…, w n ] T For a grid-type converter in n Prediction frequency for each sampling period; The output of LSTM prediction data is to transform the predicted output vector into a small signal form, that is... Y k =[ y 1, y 2,…, y n ] T The specific calculation method is as follows:

[0065] In the formula, Y k This represents the small-signal form of the k-th predicted output vector; A reference value indicating frequency; w 1, w 2,…, w n This represents the future predicted by the LSTM model. n Angular frequency value for each sampling period; n Indicates the number of sampling periods; The small-signal form of the predicted output vector transformation will inherit the vector set. V k The timeline shows that after the current prediction data output time series ends, the output vector corresponding to the next set of vectors will continue to output its prediction data. A schematic diagram of LSTM prediction output is shown below. Figure 2 As shown; The historical data in the database will be updated along with the real-time sampling data from the grid-type converter, thereby enabling rapid dynamic prediction with a minimum data latency of [missing information]. n The sampling period ensures that the prediction model can make predictions based on the latest data, enhancing the model's real-time performance and adaptability.

[0066] The technical effects of the above solution are as follows: by using the historical frequency data of the grid-type converter as input to the LSTM, it is possible to accurately predict the future frequency changes of the converter, thereby improving the accuracy of frequency prediction. The LSTM model can quickly respond to frequency changes and capture the dynamic characteristics of the frequency in a timely manner, ensuring the stable operation of the converter under frequency disturbances. The predicted output vector is converted into a small signal form, which is convenient for integration with the physical model to achieve more precise control and analysis. The output vector inherits the time axis of the vector set, ensuring the temporal continuity of the predicted data, which is helpful for time series analysis and subsequent control decisions.

[0067] In step 4, the driving method for fusing the small-signal physical model of the network converter and LSTM data is as follows: The elements output by the Long Short-Term Memory (LSTM) network, which are the predicted frequency values, are integrated to transform them into predicted values ​​for the small-signal power angle. These predicted values ​​are then weighted and fed back to the small-signal power angle feedback channel to achieve frequency control. Specifically, as follows... Figure 3 As shown, and These represent the small-signal feedback weights of the power angle in the physical model and the small-signal compensation weights of the power angle driven by data, respectively.

[0068] By integrating physical models and data-driven approaches, frequency control of the grid-type converter was achieved.

[0069] The technical effects of the above solution are as follows: by integrating the frequency prediction value output by LSTM, it is converted into the prediction value of the small power angle signal, enabling the grid-type converter to accurately predict frequency changes and respond accordingly, thereby improving the accuracy of frequency control. The prediction value is assigned to the time axis and inherited, realizing the continuity and real-time nature of the prediction data in time, ensuring that the converter can grasp the frequency changes in real time and adjust accordingly.

[0070] Working principle: First, the historical frequency data of the grid-type converter is preprocessed to ensure data quality and reliability. Then, a Long Short-Term Memory (LSTM) network is constructed to predict frequency change trends in multiple steps. Through historical data preprocessing and LSTM multi-step prediction, changes on the grid side and source side can be detected earlier, thus achieving faster frequency regulation. Second, a large-signal physical model and a small-signal physical model of the grid-type converter are established. The large-signal physical model describes the electrical characteristics of the converter under large-scale variations, while the small-signal physical model focuses on subtle dynamic changes near a certain operating point, enabling more accurate analysis of the converter's behavior under small-scale disturbances. Finally, the prediction output of the LSTM is fused with the physical model. By using the historical frequency data of the grid-type converter as the input of the LSTM, the frequency prediction value output by the LSTM is integrated to transform it into the predicted value of the small-signal power angle. Weights are then assigned to compensate the feedback channel of the small-signal power angle, enabling frequency control based on the predicted value, thereby achieving rapid dynamic regulation.

[0071] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A frequency control method for a grid-type converter based on physical models and data fusion, characterized in that, The method is based on historical frequency data and real-time operating data, combined with the small-signal physical model of the grid converter, to propose a frequency control method for the grid converter. The frequency control method for the grid-type converter includes the following steps: Step 1: Historical Data Preprocessing The historical data of the system operation is processed, including calculating the mean and standard deviation, marking outliers, handling outliers, and assigning the preprocessed vector set to the time axis; Step 2: Establish a Long Short-Term Memory Network This includes establishing a forget gate, establishing an input gate, defining the principles for updating memory units, and establishing an output gate; Step 3: Establishing the small-signal physical model of the grid-type converter This includes establishing the large-signal physical model of the grid converter and deriving the small-signal physical model of the grid converter; Step 4: Fuse LSTM prediction data with physical model This includes acquiring LSTM prediction data and fusing LSTM prediction data with grid-type converters.

2. The frequency control method for a grid-type converter based on physical model and data fusion according to claim 1, characterized in that, In step 1, the historical data is divided into segments of length [missing information]. N Given a vector set containing both vector and matrix data, assuming the data in the vector set is in column vector format, the method for detecting null values, abnormally large values, or abnormally small values ​​for each vector set includes the following steps: Step 1-1, Calculate the mean and standard deviation: For all values ​​in the preprocessed data, calculate the mean and standard deviation; Define the valid data of the vector set as The data is ,in n Represents the total number of data points; The mean of this data and standard deviation The calculation is as follows: ； ； In the formula, x i Is The Middle i One element; Step 1-2, mark outliers: For each data point, calculate its standardized value. Typically, if the standardized value of a data point is greater than a certain threshold, it is considered an outlier or outlier, expressed as follows: ； If the above conditions are met, the data will be marked as an outlier; If the data contains null values, the program will mark them as "NaN", which is also considered an outlier. Steps 1-3, handling outliers: For outlier handling, if the outlier is a marginal value, it is directly assigned the mean; if the outlier is not marginal, linear interpolation of nearby data points is used, i.e.: ； It is worth noting that the edge refers to a vector. The serial number is 1 or n Element; Steps 1-4: Assign the preprocessed vector set to the time axis: Based on the sampling step size of historical data, the first data point of the vector set is defined as the zero point of time; Then, time is allocated to each data point according to the sampling step size, forming a timeline of the data.

3. The frequency control method for a grid-type converter based on physical model and data fusion according to claim 1, characterized in that, Step 2 includes the following steps: Step 2-1, Create the forget gate: The forget gate determines whether the state information from the previous time step is forgotten, and it is expressed as follows: ； in, f t The output of the forget gate; and b f These are the weights and biases of the forget gate, respectively; h t-1 and v t These are the memory information from the previous moment and the input data from the current moment, respectively. S Use the Sigmoid activation function; Step 2-2, establish the input gate: The input gate determines how the current input vector set will update the memory cell, expressed as: ； ； in, i t The output of the input gate; Candidate memory states; , , b i and b C For the corresponding weights and biases; Steps 2-3: Define the principles for updating memory units: The current memory state of the neural network is updated based on the data from the input gate and the data retained after passing through the forget gate, which is expressed as: ； In the formula, C t represent t The state of memory at any given moment; similarly C t-1 represent t The memory state at time -1; Steps 2-4: Create the output gate: The output gate does not represent the output of the LSTM, but rather the output of the memory cell, which is represented as: ； ； in, o t This is the output of the output gate; and b o These are the weights and biases of the forget gate, respectively; h t This refers to the memory information at the current moment.

4. The frequency control method for a grid-type converter based on physical model and data fusion according to claim 1, characterized in that, In step 3, the method for establishing the large-signal physical model of the grid-type converter is as follows: The grid-connected model of a grid-connected converter can be equivalent to an AC voltage source; Define the voltage amplitude of the grid converter as U v The real-time phase angle is The impedance of the converter connected to the common grid connection point is expressed as a complex number. ,in R e Represents the line resistance. X e It is the line reactance. j It is the imaginary unit; Therefore, the active power delivered by the grid-connected converter to the common grid connection point P and reactive power Q Expressed as: ； ； In the formula, U p The voltage amplitude at the common grid connection point; The real-time phase angle of the public grid connection point; For complex numbers The modulus; At the same time, the control branches of the grid-type converter, including the active power-phase angle and reactive power-voltage control branches, can be uniformly represented as follows: ； ； In the formula, g p and g q These represent the static gains of the active power-phase angle and reactive power-voltage control branches, respectively. J and D These represent the inertia coefficient and the droop coefficient, respectively. s Represents the Laplace operator; , , and These represent reference values ​​for active power, reactive power, frequency, and voltage, respectively.

5. The frequency control method for a grid-type converter based on physical model and data fusion according to claim 1, characterized in that, In step 3, the method for deriving the small-signal physical model of the grid converter is as follows: Based on the large-signal physical model of the grid converter, a small-signal physical model of the grid converter is established with the control branch as the forward path and the converter output power as the feedback path. For the feedback channel, the power acquisition process passes through a low-pass filter before entering the main control branch. Therefore, the small-signal representations of the active and reactive power feedback channels are as follows: ； ； In the formula, Small signal form representing variables; This represents the cutoff frequency of the low-pass filter; The small signal representing the power angle.

6. The frequency control method for a grid-type converter based on physical model and data fusion according to claim 5, characterized in that, For the small-signal physical model, specifically: ； 。 7. The frequency control method for a grid-type converter based on physical model and data fusion according to claim 1, characterized in that, In step 4, the prediction data of LSTM is obtained, specifically as follows: In the frequency prediction model of the grid converter, the historical frequency data of the grid converter is used as the input of LSTM; Define input vector V =[ V 1, V 2,…, V N ] T A frequency operation dataset for a grid-type converter; Arrange the dataset according to n Each sampling period is divided into a vector set, i.e., an arbitrary vector set. V k ( k The frequency data of (=1,2,…,n) are represented as follows: V k =[ 1, 2,…, n ] T ; Predicted output vector W k =[ w 1, w 2,…, w n ] T For a grid-type converter in n Prediction frequency for each sampling period; The output of LSTM prediction data is to transform the predicted output vector into a small signal form, that is... Y k =[ y 1, y 2,…, y n ] T The specific calculation method is as follows: ； In the formula, Y k This represents the small-signal form of the k-th predicted output vector; A reference value indicating frequency; w 1, w 2,…, w n This represents the future predicted by the LSTM model. n Angular frequency value for each sampling period; n Indicates the number of sampling periods.

8. The frequency control method for a grid-type converter based on physical model and data fusion according to claim 7, characterized in that, The small-signal form of the predicted output vector transformation will inherit the vector set. V k The timeline indicates that after the current prediction data output time series ends, the output vector corresponding to the next set of vectors will continue to output its prediction data.

9. The frequency control method for a grid-type converter based on physical model and data fusion according to claim 6, characterized in that, In the prediction output step of LSTM, historical data in the database is updated with real-time sampling data from the grid-type converter, thereby achieving fast dynamic prediction with a minimum data latency of [missing information]. n One sampling period.

10. The frequency control method for a grid-type converter based on physical model and data fusion according to claim 1, characterized in that, In step 4, the driving method for fusing the small-signal physical model of the grid-type converter with LSTM prediction data is as follows: The frequency prediction value is obtained by taking the discrete integral of the vector set with time axis output by the long short-term memory network. This vector set is then weighted and compensated to the feedback channel of the power angle small signal to achieve hybrid drive frequency control.

Images