A power system disturbance lowest frequency prediction method based on a physical information embedded layer

Abstract

The application discloses a power system disturbance minimum frequency prediction method based on a physical information embedded layer, divides a system frequency response physical model minimum frequency expression into different parts, constructs physical information embedded layer features, including first feature information, first weight information and first bias information, and further acquires a physical information embedded layer output function; the output function of the physical information embedded layer and system operation feature data are normalized and then combined, serving as an input layer embedded long short-term memory network, and a novel power system minimum frequency prediction model based on the physical information embedded layer is constructed. The method deeply fuses the frequency prediction physical model and the deep learning algorithm, enhances the ability of a data-driven model to mine physical knowledge, and can quickly and accurately predict the system minimum frequency after disturbance.

Description

A method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer Technical Field

[0001] This invention relates to the field of new energy power systems, and in particular to a method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer. Background Technology

[0002] With the large-scale grid connection of new energy sources, a large number of synchronous generators with rotational inertia are being replaced by new energy power sources, resulting in a decrease in the inertia of the power system and a weakening of its primary frequency regulation capability. Against this backdrop, the risk of frequency instability after power system disturbances will increase.

[0003] Currently, frequency analysis methods for power system disturbances are mainly divided into physical model methods, data-driven methods, and physical-data fusion methods. New power systems, due to the diversity of frequency regulation units and their complex structures, struggle to balance computational accuracy and efficiency with physical model methods. In recent years, the rise of artificial intelligence has provided a feasible data-driven method for predicting power system frequency response. However, because data-driven methods rely solely on large amounts of training data for modeling, the lack of physical knowledge hinders further improvement in prediction accuracy. Physical-data fusion methods combine the prediction results of physical models and data-driven models, achieving complementary advantages. Some deep fusion methods can also fully explore the physical knowledge hidden within the physical model, further enhancing the prediction accuracy of the data-driven model. Applying physical-data fusion methods to power system frequency prediction can achieve more accurate power system frequency forecasts, better meeting frequency prediction requirements. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a method for predicting the lowest frequency of power system disturbances based on a physical information embedding layer. By constructing a physical information embedding layer, the frequency prediction physical model is deeply integrated with a deep learning algorithm, which enhances the ability of the data-driven model to mine physical knowledge and can quickly and accurately predict the lowest frequency of the system after disturbance.

[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows:

[0006] A method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer, the method comprising:

[0007] Multiple disturbance information is acquired. Each disturbance information is obtained by simulating a certain fault condition in the power system through a simulation model. The disturbance information includes system operation information and actual frequency response values. The system operation information includes system operation data before and after the disturbance caused by the fault condition.

[0008] A physical model of the system frequency response is established based on multiple disturbance information, and the lowest frequency function is obtained based on the physical model of the system frequency response.

[0009] The lowest frequency function is transformed into physical information embedded layer features, which include first feature information, first weight information and first bias information.

[0010] Calculate the output function of the physical information embedding layer based on its characteristics;

[0011] The output function of the physical information embedded layer and the system operation information are combined and input into the long short-term memory network model to obtain the lowest frequency prediction model based on the long short-term memory network model.

[0012] Furthermore, a frequency response database is constructed based on the disturbance information;

[0013] The frequency response database is used as sample data for training the lowest frequency prediction model, resulting in a trained lowest frequency prediction model.

[0014] In some embodiments, establishing a physical model of the system frequency response based on multiple disturbance information, and obtaining the lowest frequency function based on the physical model of the system frequency response further includes:

[0015] The system frequency response physical model is identified based on multiple disturbance information. The parameter information includes the system frequency response function coefficients, angular frequency, oscillation decay time constant, damping decay time constant, and initial phase angle.

[0016] Based on the parameter information, a physical model of the system frequency response is constructed. The lowest frequency is then substituted into the physical model to obtain the lowest frequency function, which is expressed by formula (1). Formula (1) is as follows:

[0017]

[0018] In formula (1), f minSFR The lowest frequency calculated from the physical model of the system's frequency response, t minSFR f is the extreme time calculated from the physical model of the system's frequency response. N For the rated frequency, P d The active disturbance is represented by T1, the damping decay time constant is T2, and the oscillation decay time constant is ω2. ω2 is the angular frequency, θ is the initial phase angle, and C0, C1, and C2 are the coefficients of the system frequency response function.

[0019] In some embodiments, the parameter information for identifying the physical model of the system frequency response based on multiple disturbance information includes:

[0020] Given a sampling time, the frequency value corresponding to the sampling time is calculated based on the time domain function of the system frequency response, and is expressed by formula (2), which is as follows:

[0021]

[0022] In formula (2), t is the given sampling time, and f t The frequency value at time t;

[0023] The actual value of the frequency response in formula (2) is fitted using the least squares method. The least squares method uses the sum of squared errors as the objective function, which is expressed by formula (3), as follows:

[0024]

[0025] In formula (3), For parameter information, This represents the first system frequency dynamic response data, which is calculated from the system frequency response physical model. r，k This represents the dynamic frequency response data of the second system, which is obtained through simulation using a simulation model. N1 represents the number of sampling points.

[0026] In some embodiments, the lowest frequency function is converted into physical information embedded layer features, wherein the physical information embedded layer features include first feature information, first weight information, and first bias information, including:

[0027] In formula (1) and P d The first feature information converted into the embedded layer features of physical information is represented by formula (4), which is as follows:

[0028]

[0029] In formula (4), X PI This is the first feature information of the embedded layer of physical information;

[0030] The C0, C1, and C2 in formula (1) are converted into the first weight information of the physical information embedded layer features, which is expressed by formula (5):

[0031] W PI =[w0, w1, w2] =[C0f N C1f N C2f N ],

[0032] In formula (5), W PI w0, w1, and w2 are the first weight information of the physical information embedded layer features, and w0, w1, and w2 are the three components in the first weight information of the physical information embedded layer features.

[0033] And, in formula (1) f N The first bias information, converted into physical information embedded layer features, is expressed by formula (6):

[0034] b PI =[f N ],

[0035] In formula (6), b PI This is the first bias information of the embedded features of physical information.

[0036] In some embodiments, the output function of the physical information embedding layer is calculated based on the physical information embedding layer features, and is expressed by formula (7), which is as follows:

[0037] f PIL =W PI X PI +b PI ,

[0038] In formula (7), f PIL This is the output function of the physical information embedding layer.

[0039] In some embodiments, the output function of the physical information embedding layer and the system operation information are combined and input into the long short-term memory network model to obtain the lowest frequency prediction model based on the long short-term memory network model, including:

[0040] The output function of the physical information embedding layer and the system operation information are normalized to obtain the first physical information embedding layer data and the first system operation information.

[0041] The first physical information embedded layer data and the first system operation information are merged to obtain the first dataset;

[0042] Using the first dataset as the input layer of the Long Short-Term Memory network model, the lowest frequency prediction model is obtained.

[0043] In some embodiments, the frequency response database is used as sample data for training the lowest frequency prediction model, resulting in a trained lowest frequency prediction model including:

[0044] The perturbation information in the frequency response database is divided into a training set and a test set according to a certain ratio;

[0045] The training set and the test set are normalized to obtain the second dataset. The second dataset includes the processed test set and the processed training set. The processed training set includes the first lowest frequency actual value, and the processed test set includes the second lowest frequency actual value.

[0046] The first lowest frequency actual value is used as the training label, and the second weight information and second bias information of the long short-term memory network model are updated using the Adam algorithm.

[0047] The long short-term memory network model is trained in a supervised manner using the processed training set to obtain the first prediction result, which is the first lowest frequency prediction value of the processed training set.

[0048] Adjust the hyperparameters of the Long Short-Term Memory network model based on the first prediction results;

[0049] Furthermore, the trained Long Short-Term Memory Network model is validated using the processed test set to obtain a second prediction result, which is the second lowest frequency prediction value of the processed test set.

[0050] The second prediction result is denormalized to obtain the third lowest frequency prediction value.

[0051] The minimum frequency error of the Long Short-Term Memory network model is calculated based on the actual value of the second minimum frequency and the predicted value of the third minimum frequency.

[0052] Determine whether the lowest frequency error falls within the preset threshold range;

[0053] If so, it means that the Long Short-Term Memory network model has been trained.

[0054] If not, update the hyperparameters of the Long Short-Term Memory network model and simultaneously update the third lowest frequency prediction value, recalculating the lowest frequency error until the lowest frequency error falls within the preset threshold range.

[0055] In some embodiments, system operating data includes the voltage and phase angle of each bus in the power system, the frequency, mechanical power and electromagnetic power of each generator, the active power of the load and the system power deficit.

[0056] In some embodiments, the simulation model is PSD-BPA.

[0057] In some embodiments, the power system is an extended IEEE-39 node system.

[0058] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0059] This invention proposes a power system disturbance minimum frequency prediction method based on a physical information embedding layer. It fully utilizes physical information by transforming terms containing power disturbances and extreme time in the minimum frequency function of the system frequency response physical model into input features of a neural network, and transforming terms containing parameters to be identified into weights of the neural network. This constructs a physical information embedding layer. By establishing a system frequency response physical model, the minimum frequency function is transformed into physical information embedding layer features, which are then combined with system operation information and input into the prediction model. The proposed physical information embedding layer features can correct the physical model parameters through network training, uncovering more accurate physical knowledge and obtaining better preliminary prediction results than the original physical model. Compared with purely data-driven methods, the corrected physical model can more accurately capture the dynamic changes in system frequency, and the preliminary prediction results significantly improve the minimum frequency prediction performance of the fusion model. Furthermore, by using time-domain simulation software to perform batch simulations of various power system fault conditions, a rich frequency response database covering multiple fault scenarios is constructed, providing strong support for subsequent model training and optimization and enhancing the model's generalization ability. Furthermore, this embodiment fuses the output of the physical information embedding layer with system operation information as input to the Long Short-Term Memory (LSTM) network. This preserves the physical mechanism information while leveraging the LTM network's ability to model time-series data, enabling a better characterization of the complex dynamic frequency response of power systems and resulting in superior prediction performance. On the other hand, validation based on the extended IEEE-39 node system demonstrates its applicability to large-scale, complex new power systems, providing reliable predictive support for power system frequency stability analysis and control. In summary, the method proposed in this embodiment integrates physical mechanism and data-driven modeling techniques, exhibiting significant advantages in improving prediction accuracy and enhancing generalization ability, and possessing important application value for power system dynamic analysis and control. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 is a schematic diagram of steps S101 to S106 of the power system disturbance minimum frequency prediction method based on physical information embedded layer according to a specific implementation.

[0062] Figure 2 is a schematic diagram of steps S201 to S210 of the power system disturbance minimum frequency prediction method based on physical information embedded layer according to a specific implementation.

[0063] Figure 3 is a schematic diagram of the lowest frequency prediction model structure based on the physical information embedded layer described in the specific implementation;

[0064] Figure 4 is a flowchart illustrating the power system disturbance minimum frequency prediction method based on physical information embedded layer according to a specific implementation. Detailed Implementation

[0065] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] This invention provides a method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer. By constructing a physical information embedded layer, the frequency prediction physical model is deeply integrated with a deep learning algorithm, which enhances the ability of the data-driven model to mine physical knowledge and can quickly and accurately predict the lowest frequency of the system after a disturbance.

[0067] Please refer to Figures 1 and 4. This embodiment provides a method for predicting the lowest frequency of power system disturbances based on a physical information embedding layer. The method includes:

[0068] S101. Obtain multiple disturbance information. Each disturbance information is obtained by simulating a certain fault condition in the power system through a simulation model. The disturbance information includes system operation information and actual frequency response values. The system operation information includes system operation data before and after the disturbance caused by the fault condition.

[0069] S102. Establish a physical model of the system frequency response based on multiple disturbance information, and obtain the lowest frequency function based on the physical model of the system frequency response.

[0070] S103. Convert the lowest frequency function into physical information embedded layer features, the physical information embedded layer features including first feature information, first weight information and first bias information;

[0071] S104. Calculate the output function of the physical information embedding layer based on the characteristics of the physical information embedding layer;

[0072] S105. The output function of the physical information embedded layer and the system operation information are merged and input into the long short-term memory network model to obtain the lowest frequency prediction model based on the long short-term memory network model.

[0073] Furthermore, a frequency response database is constructed based on the disturbance information;

[0074] S106. Use the frequency response database as sample data for training the lowest frequency prediction model to obtain the trained lowest frequency prediction model.

[0075] In this embodiment, the specific steps of the power system disturbance minimum frequency prediction method based on the physical information embedded layer are as follows:

[0076] Step 1: Use time-domain simulation software to perform batch simulations of various fault conditions in the new power system to generate disturbance information of the power system. Construct a frequency response database by selecting system operating data before and after the disturbance and actual values ​​of the frequency response.

[0077] Step 2: Using the actual frequency response values ​​and system operating data obtained in Step 1, identify the physical model parameters of the system frequency response, establish the physical model of the system frequency response, and obtain the lowest frequency function after system disturbance. Divide the lowest frequency function into different parts to construct the physical information embedded layer features, including the first feature information, the first weight information, and the first bias information, and further calculate the physical information embedded layer output function;

[0078] Step 3: Reconstruct the output function of the physical information embedding layer and the system operation information obtained in Step 1. Normalize and merge the two, then use them as the input layer to embed a Long Short-Term Memory (LSTM) network, constructing a power system minimum frequency prediction model based on the physical information embedding layer. With the support of a frequency response database, train the proposed power system minimum frequency prediction model based on the physical information embedding layer.

[0079] Step 4: Input the online power system operation data into the frequency prediction model trained in Step 3 to achieve the prediction of the lowest frequency of the new power system.

[0080] This embodiment proposes a power system disturbance minimum frequency prediction method based on a physical information embedding layer. It fully utilizes physical information by transforming terms containing power disturbances and extreme time in the minimum frequency function of the system frequency response physical model into input features of a neural network, and transforming terms containing parameters to be identified into weights of the neural network. This constructs a physical information embedding layer. By establishing a system frequency response physical model, the minimum frequency function is transformed into physical information embedding layer features, which are then combined with system operation information and input into the prediction model. The proposed physical information embedding layer features can correct the physical model parameters through network training, uncovering more accurate physical knowledge and obtaining better preliminary prediction results than the original physical model. Compared with purely data-driven methods, the corrected physical model can more accurately capture the dynamic changes in system frequency, and the preliminary prediction results significantly improve the minimum frequency prediction performance of the fusion model. By using time-domain simulation software to perform batch simulations of various power system fault conditions, a rich frequency response database covering multiple fault conditions is constructed, providing strong support for subsequent model training and optimization and enhancing the model's generalization ability. Furthermore, this embodiment fuses the output of the physical information embedding layer with system operation information as input to the Long Short-Term Memory (LSTM) network. This preserves the physical mechanism information while leveraging the LTM network's ability to model time-series data, enabling a better characterization of the complex dynamic frequency response of power systems and resulting in superior prediction performance. On the other hand, validation based on the extended IEEE-39 node system demonstrates its applicability to large-scale, complex new power systems, providing reliable predictive support for power system frequency stability analysis and control. In summary, the method proposed in this embodiment integrates physical mechanism and data-driven modeling techniques, exhibiting significant advantages in improving prediction accuracy and enhancing generalization ability, and possessing important application value for power system dynamic analysis and control.

[0081] In some embodiments, establishing a physical model of the system frequency response based on multiple disturbance information, and obtaining the lowest frequency function based on the physical model of the system frequency response further includes:

[0082] The system frequency response physical model is identified based on multiple disturbance information. The parameter information includes the system frequency response function coefficients, angular frequency, oscillation decay time constant, damping decay time constant, and initial phase angle.

[0083] Based on the parameter information, a physical model of the system frequency response is constructed. The lowest frequency is then substituted into the physical model to obtain the lowest frequency function, which is expressed by formula (1). Formula (1) is as follows:

[0084]

[0085] In formula (1), f minSFRThe lowest frequency calculated from the physical model of the system's frequency response, t minSFR f is the extreme time calculated from the physical model of the system's frequency response. N For the rated frequency, P d The active disturbance is represented by T1, the damping decay time constant is T2, and the oscillation decay time constant is ω2. ω2 is the angular frequency, θ is the initial phase angle, and C0, C1, and C2 are the coefficients of the system frequency response function.

[0086] In this embodiment, the actual frequency response value and system operating data obtained by the aforementioned method are used to identify the physical model parameters of the system frequency response, establish the physical model of the system frequency response, and obtain the lowest frequency expression after system disturbance. The lowest frequency expression is divided into different parts to construct the physical information embedded layer features, including first feature information, first weight information, and first bias information, and the physical information embedded layer output function is further calculated.

[0087] The lowest frequency time-domain expression of the physical model of the system frequency response is shown in equation (1):

[0088]

[0089] In equation (1), f minSFR The lowest frequency calculated from the physical model of the system's frequency response, t minSFR f is the extreme time calculated from the physical model of the system's frequency response. N For the rated frequency, P d The active disturbance is represented by T1, the damping decay time constant is T2, and the oscillation decay time constant is ω2. ω2 is the angular frequency, θ is the initial phase angle, and C0, C1, and C2 are the coefficients of the system frequency response function.

[0090] In some embodiments, the parameter information for identifying the physical model of the system frequency response based on multiple disturbance information includes:

[0091] Given a sampling time, the frequency value corresponding to the sampling time is calculated based on the time domain function of the system frequency response, and is expressed by formula (2), which is as follows:

[0092]

[0093] In formula (2), t is the given sampling time, and f t The frequency value at time t;

[0094] The actual value of the frequency response in formula (2) is fitted using the least squares method. The least squares method uses the sum of squared errors as the objective function, which is expressed by formula (3), as follows:

[0095]

[0096] In formula (3), For parameter information, This represents the first system frequency dynamic response data, which is calculated from the system frequency response physical model. r，k This represents the dynamic frequency response data of the second system, which is obtained through simulation using a simulation model. N1 represents the number of sampling points.

[0097] In this embodiment, the values ​​of C0, C1, C2, T1, T2, ω2, and θ are obtained by least squares identification. The specific identification process is as follows:

[0098] Given a sampling time t, calculate the frequency value f corresponding to the sampling time t using the time-domain expression of the system frequency response shown in equation (2). t .

[0099]

[0100] The system's actual frequency response data is fitted using the least squares method to obtain the system's frequency response model parameters. The objective function of the least squares method is the sum of squared errors, as shown in equation (3):

[0101]

[0102] In equation (3), This represents the system frequency dynamic response data calculated from the physical model (i.e., the first system frequency dynamic response data); f r，k N1 represents the system frequency dynamic response data obtained from the full-time domain simulation (i.e., the second system frequency dynamic response data); N1 represents the number of sampling points.

[0103] In some embodiments, the lowest frequency function is converted into physical information embedded layer features, wherein the physical information embedded layer features include first feature information, first weight information, and first bias information, including:

[0104] In formula (1) and P d The first feature information converted into the embedded layer features of physical information is represented by formula (4), which is as follows:

[0105]

[0106] In formula (4), XPI This is the first feature information of the embedded layer of physical information;

[0107] The C0, C1, and C2 in formula (1) are converted into the first weight information of the physical information embedded layer features, which is expressed by formula (5):

[0108] W PI =[w0, w1, w2] =[C0f N C1f N C2f N ],

[0109] In formula (5), W PI w0, w1, and w2 are the first weight information of the physical information embedded layer features, and w0, w1, and w2 are the three components in the first weight information of the physical information embedded layer features.

[0110] And, in formula (1) f N The first bias information, converted into physical information embedded layer features, is expressed by formula (6):

[0111] b PI =[f N ],

[0112] In formula (6), b PI This is the first bias information of the embedded features of physical information.

[0113] In this embodiment, equation (1) is divided into different parts to construct a physical information embedding layer. The specific steps are as follows:

[0114] First, extract equation (1) from... and P d The system consists of three parts, each containing active power perturbation and extreme value time information. The values ​​of these three parts vary significantly across different samples, making them unsuitable for conversion into parameters of a neural network layer. Therefore, they are transformed into input features X for a physical information embedding layer. PI (That is, the first feature information of the embedded layer of physical information), as shown in Equation (4).

[0115]

[0116] The remaining part of equation (1) consists of some physical parameters C0, C1, and C2 whose exact values ​​are unknown. These parts are then converted into the first weight information W of the physical information embedding layer. PI and the first bias information b PI The purpose is to use neural network training to correct unknown physical parameters and further explore the physical knowledge hidden in the parameters. The specific expressions of the constructed physical information embedded layer features and parameters are shown in (5)-(6).

[0117] W PI =[w0, w1, w2] =[C0f N C1f N C2f N ],

[0118] b PI =[f N ],

[0119] In equation (5), w0, w1, and w2 are three components of the first weight information of the physical information embedded layer features.

[0120] In some embodiments, the output function of the physical information embedding layer is calculated based on the physical information embedding layer features, and is expressed by formula (7), which is as follows:

[0121] f PIL =W PI X PI +b PI ,

[0122] In formula (7), f PIL This is the output function of the physical information embedding layer.

[0123] In some embodiments, the output function of the physical information embedding layer and the system operation information are combined and input into the long short-term memory network model to obtain the lowest frequency prediction model based on the long short-term memory network model, including:

[0124] The output function of the physical information embedding layer and the system operation information are normalized to obtain the first physical information embedding layer data and the first system operation information.

[0125] The first physical information embedded layer data and the first system operation information are merged to obtain the first dataset;

[0126] Using the first dataset as the input layer of the Long Short-Term Memory network model, the lowest frequency prediction model is obtained.

[0127] In this embodiment, the output of the physical information embedding layer and the system operation characteristic data obtained by the aforementioned method are reconstructed, normalized, and then merged. These are then used as the input layer to embed a Long Short-Term Memory (LSTM) network, constructing a novel power system minimum frequency prediction model based on the physical information embedding layer. With the support of a frequency response database, the proposed novel power system minimum frequency prediction model based on the physical information embedding layer is trained.

[0128] Referring to Figure 2, in some embodiments, the frequency response database is used as sample data for training the lowest frequency prediction model, resulting in a trained lowest frequency prediction model including:

[0129] S201. Divide the disturbance information in the frequency response database into a training set and a test set according to a certain ratio;

[0130] S202. Normalize the training set and the test set to obtain the second dataset. The second dataset includes the processed test set and the processed training set. The processed training set includes the first lowest frequency actual value, and the processed test set includes the second lowest frequency actual value.

[0131] S203. Use the first lowest frequency actual value as the training label, and use the Adam algorithm to update the second weight information and the second bias information of the long short-term memory network model.

[0132] S204. Supervise the long short-term memory network model using the processed training set to obtain the first prediction result, which is the first lowest frequency prediction value of the processed training set.

[0133] S205. Adjust the hyperparameters of the long short-term memory network model based on the first prediction result;

[0134] Furthermore, the trained Long Short-Term Memory Network model is validated using the processed test set to obtain a second prediction result, which is the second lowest frequency prediction value of the processed test set.

[0135] S206. Perform inverse normalization on the second prediction result to obtain the third lowest frequency prediction value.

[0136] S207. Calculate the minimum frequency error of the long short-term memory network model based on the actual value of the second minimum frequency and the predicted value of the third minimum frequency.

[0137] S208. Determine whether the lowest frequency error is within the range of a preset threshold.

[0138] S209. If yes, it means that the Long Short-Term Memory network model has been trained.

[0139] S210. If not, update the hyperparameters of the Long Short-Term Memory network model and update the third lowest frequency prediction value simultaneously. Recalculate the lowest frequency error until the lowest frequency error is within the range of the preset threshold.

[0140] In this embodiment, the specific training process can be understood in conjunction with Figure 3:

[0141] The frequency response dataset from the aforementioned method is divided into training and test sets according to a certain ratio, and the data is reconstructed and normalized. The processed physical information embedded layer output and system operation feature data are used as input features of the Long Short-Term Memory (LSTM) network, and the processed actual lowest frequency value (i.e., the first actual lowest frequency value) is used as the training label. The Adam algorithm is used to update the second weight information and second bias information of the LTM network for supervised training. Simultaneously, the hyperparameters of the model need to be adjusted based on the model's prediction results on the training set (i.e., the first prediction result) to ensure that the trained model has good predictive performance.

[0142] Input features from the test set are fed into the trained model to obtain the lowest frequency predicted value (i.e., the second lowest frequency predicted value), and the prediction result (i.e., the second prediction result) is denormalized. The lowest frequency prediction error of the model is calculated using the denormalized lowest frequency predicted value (i.e., the third lowest frequency predicted value) and the actual lowest frequency value (i.e., the second lowest frequency actual value), which is used to evaluate the model's prediction accuracy. If the model's prediction accuracy meets the preset requirements, the model is saved. If it does not meet the requirements, hyperparameters are reselected and training is performed.

[0143] In some embodiments, system operating data includes the voltage and phase angle of each bus in the power system, the frequency, mechanical power and electromagnetic power of each generator, the active power of the load and the system power deficit.

[0144] In some embodiments, the simulation model is PSD-BPA.

[0145] In some embodiments, the power system is an extended IEEE-39 node system.

[0146] Furthermore, the following example can be derived from the above scheme:

[0147] The simulation software used was PSD-BPA, and the test system adopted an extended IEEE-39-bus system. The initial total active load of this system was 6160.1 MW, the rated frequency was 50 Hz, the base capacity was 100 MVA, and the base voltage was 345 kV. Wind farms with rated capacities of 571.2 MW and 550.8 MW were connected at nodes 36 and 37, respectively. A photovoltaic power station with a rated capacity of 70.35 MW was connected at node 35, while the output power of the turbine generator at node 35 was reduced to 585 MW.

[0148] A PSD-BPA software program was written in Python to generate samples for model training in batches. To simulate different system operating scenarios, the initial system load was set to vary between 80% and 105%, with a variation increment of 5%. Power disturbances were introduced by adding load to a single bus and changing generator output. The disturbance range for a single bus load was 1.5%–6.5% of the initial system load, with each disturbance increment of 0.5%. After removing the two generators under load on buses 31 and 39, the wind farm connected to buses 36 and 37, and the photovoltaic power station and generator connected to bus 35, one of the remaining five generators was randomly selected to change its output each time. The generator output variation range was 50%–100% of its original output, with each output variation increment of 10%. Based on this, 1434 samples were generated, and the sample set was divided into training and test sets in an 8:2 ratio. The hyperparameters of the Long Short-Term Memory (LSTM) model were selected using grid search. The selected LSTM model had 3 layers, with 100, 100, and 50 neurons per layer, respectively.

[0149] To verify the superiority of the proposed lowest frequency prediction model based on the embedded physical information layer (named the PILSTM model), four fusion models were designed as comparison models, referencing different literature. By connecting the lowest frequency calculation results of the physical model and the LSTM model in different ways, three fusion models were obtained: the LSTM-serial fusion model, the LSTM-parallel fusion model, and the LSTM-guided fusion model. In addition to the above three shallow fusion models, this paper also designed the LSTM-variable weight loss function model. This model adds weights reflecting the magnitude of the error between the LSTM model's predicted value and the physical model's calculated value, based on the traditional MSE loss function. To evaluate the prediction accuracy of the proposed model, mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) were introduced as evaluation metrics. Furthermore, to facilitate comparison between the prediction accuracy of the pure data-driven model and the fusion model, the prediction error of the LSTM model is also presented. The prediction errors of different models are the average values ​​of five random experiments, as shown in Table 1.

[0150] Table 1 Comparison of lowest frequency prediction errors of different models

[0151]

[0152] Table 1 shows that the errors of the physics-data fusion method are all lower than those of the pure data-driven LSTM model. This indicates that incorporating physical knowledge improves the prediction accuracy of the data-driven model. Meanwhile, there are significant differences in prediction accuracy among the different fusion methods in Table 1. Among all fusion models, the PILSTM model has the best prediction performance, indicating that the PILSTM model indirectly optimizes the parameters to be identified in the physics model through network training, introducing more accurate physical knowledge into the LSTM model and obtaining a better lowest frequency prediction result.

[0153] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0154] This invention proposes a power system disturbance minimum frequency prediction method based on a physical information embedding layer. It fully utilizes physical information by transforming terms containing power disturbances and extreme time in the minimum frequency function of the system frequency response physical model into input features of a neural network, and transforming terms containing parameters to be identified into weights of the neural network. This constructs a physical information embedding layer. By establishing a system frequency response physical model, the minimum frequency function is transformed into physical information embedding layer features, which are then combined with system operation information and input into the prediction model. The proposed physical information embedding layer features can correct the physical model parameters through network training, uncovering more accurate physical knowledge and obtaining better preliminary prediction results than the original physical model. Compared with purely data-driven methods, the corrected physical model can more accurately capture the dynamic changes in system frequency, and the preliminary prediction results significantly improve the minimum frequency prediction performance of the fusion model. Furthermore, by using time-domain simulation software to perform batch simulations of various power system fault conditions, a rich frequency response database covering multiple fault scenarios is constructed, providing strong support for subsequent model training and optimization and enhancing the model's generalization ability. Furthermore, this embodiment fuses the output of the physical information embedding layer with system operation information as input to the Long Short-Term Memory (LSTM) network. This preserves the physical mechanism information while leveraging the LTM network's ability to model time-series data, enabling a better characterization of the complex dynamic frequency response of power systems and resulting in superior prediction performance. On the other hand, validation based on the extended IEEE-39 node system demonstrates its applicability to large-scale, complex new power systems, providing reliable predictive support for power system frequency stability analysis and control. In summary, the method proposed in this embodiment integrates physical mechanism and data-driven modeling techniques, exhibiting significant advantages in improving prediction accuracy and enhancing generalization ability, and possessing important application value for power system dynamic analysis and control.

[0155] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0156] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0157] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer, characterized in that, The method includes: acquiring multiple disturbance information, each disturbance information being obtained by simulating a fault condition in a power system using a simulation model; the disturbance information including system operation information and actual frequency response values, the system operation information including system operation data before and after the disturbance caused by the fault condition; establishing a physical model of the system frequency response based on the multiple disturbance information; obtaining a minimum frequency function based on the physical model of the system frequency response; converting the minimum frequency function into physical information embedding layer features, the physical information embedding layer features including first feature information, first weight information, and first bias information; calculating the output function of the physical information embedding layer based on the physical information embedding layer features; merging the output function of the physical information embedding layer and the system operation information, and inputting it into a long short-term memory network model to obtain a result based on the long short-term memory network. The memory network model has a minimum frequency prediction model; and a frequency response database is constructed based on the disturbance information; the frequency response database is used as sample data for the minimum frequency prediction model and trained to obtain a trained minimum frequency prediction model; a system frequency response physical model is established based on multiple disturbance information, and the minimum frequency function is obtained based on the system frequency response physical model. The minimum frequency function is further obtained by: identifying the parameter information of the system frequency response physical model based on multiple disturbance information, wherein the parameter information includes the system frequency response function coefficients, angular frequency, oscillation decay time constant, damping decay time constant, and initial phase angle; a system frequency response physical model is constructed based on the parameter information, and the minimum frequency is substituted into the system frequency response physical model to obtain the minimum frequency function, wherein the minimum frequency function is expressed by formula (1), and formula (1) is as follows: In formula (1), The lowest frequency calculated from the physical model of the system's frequency response. The extreme time is calculated from the physical model of the system's frequency response. For the rated frequency, For active disturbance, The damping decay time constant is The oscillation decay time constant is Angular frequency, The initial phase angle, The system frequency response function coefficients are used; the lowest frequency function is converted into physical information embedded layer features, the physical information embedded layer features include first feature information, first weight information and first bias information including: the coefficients of formula (1) are used to convert the lowest frequency function ... 、 and The first feature information converted into the physical information embedded layer feature is represented by formula (4), which is as follows: In formula (4), The first feature information of the embedded layer features of the physical information; the first feature information of the physical information is the first feature information of the embedded layer features in formula (1). The first weight information, converted into physical information embedded layer features, is expressed by formula (5): In formula (5), This is the first weight information of the embedded layer features of the physical information. The three components are the first weight information of the embedded layer features of the physical information; and the ... The first bias information, converted into physical information embedded layer features, is expressed by formula (6): In formula (6), This is the first bias information of the embedded layer features of the physical information.

2. The method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer according to claim 1, characterized in that, The parameter information of the physical model of the system frequency response identified based on the multiple disturbance information includes: given a sampling time, calculating the frequency value corresponding to the sampling time based on the time domain function of the system frequency response, expressed by formula (2), which is as follows: In formula (2), For a given sampling time, In order to be in The frequency value at time; the actual value of the frequency response in the formula (2) is fitted using the least squares method, which uses the sum of squared errors as the objective function, and is expressed by formula (3), which is as follows: In formula (3), For parameter information, ， This represents the first system frequency dynamic response data, which is calculated from the system frequency response physical model. This represents the frequency dynamic response data of the second system, which was obtained through simulation using a simulation model. Indicates the number of sampling points.

3. The method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer according to claim 1, characterized in that, The output function of the physical information embedding layer is calculated based on the physical information embedding layer features, and is expressed by formula (7), which is as follows: In formula (7), This is the output function of the embedded layer of physical information.

4. The method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer according to claim 3, characterized in that, The process of merging the output function of the physical information embedding layer and the system operation information, and inputting them into a long short-term memory network model to obtain a minimum frequency prediction model based on the long short-term memory network model includes: normalizing the output function of the physical information embedding layer and the system operation information to obtain first physical information embedding layer data and first system operation information; merging the first physical information embedding layer data and the first system operation information to obtain a first dataset; and using the first dataset as the input layer of the long short-term memory network model to obtain the minimum frequency prediction model.

5. The method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer according to claim 4, characterized in that, Using the frequency response database as sample data for training the lowest frequency prediction model to obtain a trained lowest frequency prediction model includes: dividing the perturbation information in the frequency response database into a training set and a test set according to a certain ratio; normalizing the training set and the test set to obtain a second dataset, the second dataset including a processed test set and a processed training set, the processed training set including a first actual lowest frequency value, and the processed test set including a second actual lowest frequency value; using the first actual lowest frequency value as a training label, and using the Adam algorithm to update the second weight information and the second bias information of the long short-term memory network model; using the processed training set to perform supervised training on the long short-term memory network model to obtain a first prediction result, the first prediction result being the first lowest frequency prediction value of the processed training set. The hyperparameters of the Long Short-Term Memory (LSTM) network model are adjusted according to the first prediction result; and the trained LSM network model is validated using the processing test set to obtain a second prediction result, which is the second lowest frequency prediction value of the processing test set; the second prediction result is denormalized to obtain a third lowest frequency prediction value; the lowest frequency error of the LSM network model is calculated based on the actual value of the second lowest frequency and the third lowest frequency prediction value; it is determined whether the lowest frequency error is within a preset threshold range; if yes, the LSM network model is considered to have been trained; if no, the hyperparameters of the LSM network model are updated, and the third lowest frequency prediction value is updated synchronously, and the lowest frequency error is recalculated until the lowest frequency error is within the preset threshold range.

6. The method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer according to claim 1, characterized in that, The system operation data includes the voltage and phase angle of each bus in the power system, the frequency, mechanical power and electromagnetic power of each generator, the active power of the load and the power deficit of the system.

7. The method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer according to claim 1, characterized in that, The simulation model is PSD-BPA.

8. The method for predicting the lowest frequency of power system disturbances based on a physical information embedded layer according to claim 1, characterized in that, The power system is an extended IEEE-39 node system.

Images