Subsynchronous oscillation identification method based on environmental excitation data

Abstract

The application discloses a method for identifying subsynchronous oscillation based on environmental excitation data, and relates to the technical field of power system stability and control, and comprises the following steps: receiving environmental excitation data in a power system; inputting the environmental excitation data into a pre-established gated recurrent unit (GRU) model, and extracting subsynchronous oscillation features of the environmental excitation data; identifying the subsynchronous oscillation features by using an Attention mechanism, and then performing dimension transformation on the subsynchronous oscillation features by using a fully connected layer to obtain frequency and damping ratio parameters of the subsynchronous oscillation; the GRU model can realize preliminary extraction of subsynchronous oscillation mode features in the environmental excitation signal, the Attention mechanism can enable the model to selectively focus on specific time sequence features, thereby further improving the information utilization capability of the model, and finally the frequency and the damping ratio are output by performing dimension transformation on the fully connected layer.

Description

Technical Field

[0001] This invention relates to the field of power system stability and control technology, specifically a subsynchronous oscillation identification method based on environmental excitation data. Background Technology

[0002] Although some progress has been made in the research of subsynchronous oscillation related early warning methods, most algorithms start by analyzing the inducing factors of subsynchronous oscillations in power systems. However, the actual operation scenarios of power systems are complex and varied. Subsynchronous oscillation early warning methods based on the inducing factors of subsynchronous oscillations are often only applicable to specific power system operation scenarios, and their generalization faces difficulties. The process of mathematical analysis for specific scenarios also affects the real-time performance of the algorithm in actual use, and its performance in actual scenarios is often poor.

[0003] Traditional subsynchronous oscillation parameter estimation methods are mostly based on transient data at the time of subsynchronous oscillation occurrence. These traditional transient data-based methods are unsuitable for risk perception tasks related to subsynchronous oscillations. Current research has demonstrated that environmental excitation data under steady-state conditions in power systems possesses rich dynamic information. In the field of low-frequency power system oscillations, there are studies on oscillation parameter estimation based on environmental excitation data, successfully extracting low-frequency oscillation characteristic information from linearized state-space models of power systems under random load fluctuations. However, few studies have focused on subsynchronous oscillation parameter estimation based on environmental excitation data. While electromechanical (low-frequency) oscillations, subsynchronous torsional vibrations of mechanical shafts, and oscillations induced by electrical resonance and control in power grids differ significantly in their physical mechanisms and manifestations, they share a unified mathematical model and commonalities. Because the mathematical models of low-frequency oscillations and subsynchronous oscillations share a unified form, the characteristic information of subsynchronous oscillations may also exist in the power system signal model under environmental excitation. Summary of the Invention

[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide a subsynchronous oscillation identification method based on environmental excitation data.

[0005] The objective of this invention can be achieved through the following technical solution: a subsynchronous oscillation identification method based on environmental excitation data, the method comprising the following steps:

[0006] Receive environmental excitation data within the power system;

[0007] The environmental excitation data is input into a pre-established gated recurrent unit neural network (GRU) model to extract the subsynchronous oscillation characteristics of the environmental excitation data.

[0008] The subsynchronous oscillation features are identified using the Attention mechanism, and then the frequency and damping ratio parameters of the subsynchronous oscillation are obtained through dimensional transformation of the fully connected layer.

[0009] Preferably, the environmental stimulus types of the environmental stimulus data include white noise and colored noise.

[0010] Preferably, the GRU model includes a reset gate and an update gate to extract subsynchronous oscillation feature information contained in the environmental excitation data.

[0011] Preferably, the calculation formula for the reset door is as follows:

[0012] r t =σ(x t W xr +h t-1 W hr +b r )

[0013] In the formula, x t h represents the actual value of the input environmental stimulus data at time t. t-1 The hidden state output is the corresponding input of the environmental stimulus data at time t-1; W xr W hr b r These are the parameters that the model needs to train; the σ function ranges from 0 to 1, where 0 indicates discarding and 1 indicates retaining completely, r t To reset the output of the gate.

[0014] Preferably, the calculation formula for the update gate is as follows:

[0015] z t =σ(x t W xz +h t-1 W hz +b z )

[0016] In the formula, x t h represents the actual value of the input environmental stimulus data at time t. t-1 The hidden state output is the corresponding input of the environmental stimulus data at time t-1; W xz W hz b z These are the parameters that the model needs to train; the σ function has a range of 0 to 1, where 0 indicates discarding and 1 indicates retaining completely, z t To update the output of the gate.

[0017] Preferably, the GRU model calculates candidate hidden states and hidden states of subsynchronous oscillation features extracted from environmental excitation data.

[0018] Preferably, the candidate hidden state output The calculation is as follows:

[0019]

[0020] In the formula, x t h represents the actual value of the input environmental stimulus data at time t. t-1 The hidden state output is the corresponding input of the environmental stimulus data at time t-1; W xh W hh b h The parameters that the model needs to train; ⊙ represents element-wise multiplication; r t Reset the gate's output at the current time step; tanh is the activation function;

[0021] The hidden state output h t The calculation is as follows:

[0022]

[0023] In the formula, Let z be the candidate hidden state output corresponding to the environmental stimulus data input at time t. t Output the update gate for the current time step.

[0024] Preferably, the process of identifying subsynchronous oscillation features using the Attention mechanism includes the following steps:

[0025] Calculate the weighted scores of the feature information of the environmental stimulus data at each time step;

[0026] The calculated weight scores of the feature information are normalized to obtain the weights of the subsynchronous oscillation feature information at each time step.

[0027] The Attention value is calculated using the weights of the subsynchronous oscillation feature information obtained at each time step.

[0028] Preferably, the calculation process for the weighted scores of the feature information at each time step of the computational environment stimulus data is as follows:

[0029] e t =V T tanh(WS+Uh t )

[0030] In the formula, h t The hidden state output corresponding to the environmental stimulus data input at time t, e t The weight score for the subsynchronous oscillation feature information at time t in the final calculated environmental stimulus data; S is the output of the last time series of the GRU; V, W, and U are the parameters that the model needs to train; tanh is the activation function.

[0031] The Attention value is obtained by combining the weights of the synchronous oscillation feature information at each time step with h. t The weighted average is obtained, and the Attention value is passed through a fully connected layer to finally output the frequency and damping ratio of the subsynchronous oscillation.

[0032] The beneficial effects of this invention are:

[0033] The subsynchronous oscillation identification method based on environmental excitation data proposed in this invention can be used not only for risk warning before the occurrence of subsynchronous oscillations, but also for the analysis of subsynchronous oscillation phenomena in power systems under steady-state conditions. The proposed method possesses excellent generalizability and real-time performance. Compared with subsynchronous oscillation warning methods based on influencing factors, it is better compatible with the complex and ever-changing operating scenarios of actual power systems, and has excellent industrial practical value.

[0034] This invention is data-driven, requires no analysis of the signal model, and has good convenience and generalization, making it suitable for online use and model generalization;

[0035] This invention still achieves good estimation results in unfamiliar power system operation scenarios and has good generalization ability;

[0036] The time consumption of this invention is concentrated in the model training stage, which can be performed offline. The actual time spent on parameter estimation on the test set is very small. Therefore, it is suitable for online subsynchronous oscillation parameter estimation and risk perception tasks, and has good real-time performance. Attached Figure Description

[0037] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of the method of the present invention;

[0039] Figure 2 This is a model architecture diagram of the present invention;

[0040] Figure 3 This is a flowchart illustrating an implementation example of the present invention;

[0041] Figure 4 This is a structural diagram of the Guyuan wind power simulation system. Detailed Implementation

[0042] 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.

[0043] like Figure 1 As shown, the subsynchronous oscillation identification method based on environmental excitation data includes the following steps:

[0044] Receive environmental excitation data within the power system;

[0045] The environmental excitation data is input into a pre-established gated recurrent unit neural network (GRU) model to extract the subsynchronous oscillation characteristics of the environmental excitation data.

[0046] The subsynchronous oscillation features are identified using the Attention mechanism, and then the frequency and damping ratio parameters of the subsynchronous oscillation are obtained through dimensional transformation of the fully connected layer.

[0047] It should be further explained that, in the specific implementation process, the GRU model includes a reset gate and an update gate, which are used to extract the subsynchronous oscillation feature information contained in the environmental stimulus data, and to calculate the candidate hidden state and hidden state of the subsynchronous oscillation feature extracted from the environmental stimulus data.

[0048] The formula for calculating the door reset is as follows:

[0049] r t =σ(x t W xr +h t-1 W hr +b r )

[0050] In the formula, x t h represents the actual value of the input environmental stimulus data at time t. t-1 The hidden state output is the corresponding input of the environmental stimulus data at time t-1; W xr W hr b r These are the parameters that the model needs to train; the σ function ranges from 0 to 1, where 0 indicates discarding and 1 indicates retaining completely, r t To reset the output of the gate.

[0051] The formula for calculating the update gate is as follows:

[0052] z t =σ(x t W xz +h t-1W hz +b z )

[0053] In the formula, x t h represents the actual value of the input environmental stimulus data at time t. t-1 The hidden state output is the corresponding input of the environmental stimulus data at time t-1; W xz W hz b z These are the parameters that the model needs to train; the σ function has a range of 0 to 1, where 0 indicates discarding and 1 indicates retaining completely, z t To update the output of the gate.

[0054] Output of candidate hidden states corresponding to environmental stimulus data input at time t The calculation is as follows:

[0055]

[0056] In the formula, x t h represents the actual value of the input environmental stimulus data at time t. t-1 The hidden state output is the corresponding input of the environmental stimulus data at time t-1; W xh W hh b h The parameters that the model needs to train; ⊙ represents element-wise multiplication; r t Resets the gate's output at the current time step; tanh is the activation function.

[0057] The hidden state output h corresponding to the environmental stimulus data input at time t t The calculation is as follows:

[0058]

[0059] In the formula, Let z be the candidate hidden state output corresponding to the environmental stimulus data input at time t. t Output the update gate for the current time step.

[0060] It should be further explained that, in the specific implementation process, the process of identifying subsynchronous oscillation features using the Attention mechanism includes the following steps:

[0061] Calculate the weighted scores of the feature information of the environmental stimulus data at each time step;

[0062] The calculated weight scores of the feature information are normalized to obtain the weights of the subsynchronous oscillation feature information at each time step.

[0063] The Attention value is calculated using the weights of the subsynchronous oscillation feature information obtained at each time step.

[0064] The method for calculating the weighted scores of the feature information at each time step of the environmental stimulus data is as follows:

[0065] e t =V T tanh(WS+Uh t )

[0066] In the formula, h t The hidden state output corresponding to the environmental stimulus data input at time t, e t denoted as t, where t is the weight score of the subsynchronous oscillation feature information of the environmental stimulus data at time t; S is the output of the last time series of GRU; V, W, and U are the parameters that the model needs to train; and tanh is the activation function.

[0067] The process of normalizing the calculated feature information weights to obtain the weights of the subsynchronous oscillation feature information at each time step is as follows:

[0068]

[0069] In the formula, α t The weight value is the subsynchronous oscillation characteristic information of the environmental excitation data at time t, which is calculated.

[0070] The method for calculating the attention value is as follows:

[0071]

[0072] In the formula, C is the weight α of the subsynchronous oscillation characteristic information in the environmental excitation data calculated at each time point. t and hidden state h t The attention value obtained by weighted averaging is then used to output the frequency and damping ratio of the subsynchronous oscillation after passing through a fully connected layer.

[0073] In this embodiment, the implementation scheme is based on the Guyuan wind power simulation system. In order to simulate the complex and ever-changing uncertainties of the actual wind power system operation scenario and to test the generalization of the method described in this invention, this implementation scheme uses five scenario influencing factors as experimental scenario variables: environmental noise type, environmental noise fluctuation amplitude, environmental wind speed, system series compensation degree, and the proportional coefficient of the inner loop controller of the wind farm rotor side current. Various different wind power system operation scenarios are constructed.

[0074] The types of environmental stimulus data include white noise and colored noise.

[0075] White noise environmental excitation type refers to noise whose power spectral density is constant throughout the entire frequency domain. The generation method is as follows:

[0076]

[0077] In the formula, h(λ) (λ>0) is the impulse response function of G(s), and the power spectral densities of the input signal x(t) and the output signal y(t) have the following correspondence:

[0078] S y (ω)=|G(jω)| 2 S x (ω)

[0079] In the formula, G(jω) is the frequency response of the transfer function G(s).

[0080] Colored noise refers to noise with a non-flat power spectral density function. White noise can generate colored noise when passing through a channel due to the influence of the channel frequency. Therefore, a white noise signal x(t) can be converted into a colored noise signal y(t) through a linear transfer function G(s), as shown below:

[0081] S y (ω)=|G(jω)| 2 S x (ω)=|G(jω)| 2 ·1=G * (jω)G(jω)

[0082] In the formula, G * (jω) is the complex conjugate of G(jω). G(s) is the transfer function form of the shaping filter.

[0083] White noise can be passed through a shaping filter to obtain colored noise. The shaping filter can be modeled using a transfer function, which can be defined as follows:

[0084]

[0085] In the formula, P is the initial active load, nP is the fluctuation ratio of the active load, and the others are fixed parameters.

[0086] Apart from the types of environmental stimuli, other scenario variables are set as follows:

[0087] The environmental excitation fluctuation range is set between 7% and 13%; the wind speed range of the wind power simulation system is set between 6.5 m / s and 7.5 m / s, and the series compensation degree varies between 7% and 16%; the proportional coefficient K of the inner loop controller for the rotor side current of the wind farm is set between 7% and 16%. p The range of variation is set between 0.10 and 0.20.

[0088] The correct values ​​of the subsynchronous oscillation frequency and damping ratio for each scenario are calculated using the traditional subsynchronous oscillation identification method based on transient data in that scenario.

[0089] Example 1:

[0090] The present invention demonstrates the performance of parameter estimation under unfamiliar environmental excitation.

[0091] Set the proportional coefficient K of the inner loop controller for wind speed, series compensation degree, and wind farm rotor side current. p Within the aforementioned reasonable value range, using both white noise and colored noise as environmental excitation forms, data from power system operation scenarios with environmental excitation fluctuations of 7% to 11% were selected as the training set, and data from power system operation scenarios with environmental excitation fluctuations of 12% to 13% were selected as the test set. The constructed training set was used to train the model of the present invention to improve its performance. After training, the constructed test set was used to test the performance of the present invention.

[0092] The test set contains 100 sets of data. Testing showed that, under unfamiliar environmental conditions, this invention can be effectively used for subsynchronous oscillation frequency parameter estimation based on environmental excitation data. Calculations revealed that the minimum absolute error between the correct frequency values ​​in each set of data and the estimated values ​​from this invention was 0.0968 Hz, the maximum absolute error was 0.7421 Hz, and the overall average absolute error was 0.339365 Hz. Overall, this invention achieved good results in frequency parameter estimation under unfamiliar environmental conditions, with both the mean error and maximum error of the final identification results showing good performance.

[0093] In experimental environments stimulated by unfamiliar conditions, this invention also demonstrates good identification performance in the damping ratio parameter estimation task. Calculations show that the minimum absolute error between the correct damping ratio value and the estimated value from this invention in each test set is 0.0057, the maximum absolute error is 0.3013, and the overall average error is 0.113594. This invention performs excellently in the damping ratio parameter estimation task, with both the mean error and maximum error of the final identification results showing good performance.

[0094] Example 2:

[0095] The present invention demonstrates the effectiveness of parameter estimation in unfamiliar wind speed scenarios.

[0096] Set the environmental excitation, series compensation degree, and proportional coefficient K of the inner loop controller for the wind farm rotor side current. pWithin the aforementioned reasonable value range, data from power system operation scenarios with wind speeds of 6.5 m / s to 7.2 m / s were selected as the training set, and data from power system operation scenarios with wind speeds of 7.3 m / s to 7.5 m / s were selected as the test set. The constructed training set was used to train the model of the present invention to improve its performance. After training, the constructed test set was used to test the performance of the present invention.

[0097] The test set in this experimental environment also contained 100 sets of data. In a power system operation scenario with unfamiliar wind speeds, this invention achieved excellent results in the frequency parameter estimation task. After error calculation, the minimum absolute error of this invention in the frequency parameter estimation task was 0.000029Hz, the maximum absolute error was 0.1611Hz, and the overall mean error across the 100 sets of data was calculated to be 0.053583Hz. The final calculated frequency estimation error shows that this invention has excellent performance in the frequency parameter estimation task under unfamiliar wind speed scenarios, demonstrating good accuracy and outstanding performance.

[0098] In the task of estimating the damping ratio parameter under unfamiliar wind speed scenarios, the minimum absolute error of this invention was calculated to be 0.0047, the maximum absolute error was 0.2582, and the mean estimation error of 100 sets of data was calculated to be 0.108887. It can be seen that this invention performs excellently in the task of estimating the damping ratio parameter under unfamiliar wind speed scenarios, with both the maximum and mean errors within a relatively small range. No experimental results with excessively large errors were observed, demonstrating good parameter estimation accuracy in this task.

[0099] Example 3:

[0100] The present invention demonstrates the performance of parameter estimation in unfamiliar string complement scenarios.

[0101] Set the proportional coefficient K of the inner loop controller for environmental excitation, wind speed, and wind farm rotor side current. p Within the aforementioned reasonable range, data from power system operation scenarios with a series complement of 7% to 13% are selected as the training set, and data from power system operation scenarios with a series complement of 14% to 16% are selected as the test set. The constructed training set is used to train the model of the present invention to improve its performance. After training, the constructed test set is used to test the performance of the present invention.

[0102] The test set also contains 100 sets of test data. In unfamiliar string complement scenarios, this invention successfully estimated the frequency parameters based on environmental excitation data in each set of data. Calculations show that on the test set, the minimum absolute estimation error of this invention is 0.2376 Hz, the maximum absolute estimation error is 0.5903 Hz, and the overall mean estimation error is 0.429669 Hz. In summary, this invention can be effectively used for estimating subsynchronous oscillation frequency parameters based on environmental excitation data in unfamiliar string complement scenarios. The overall error is within an acceptable range, and no serious estimation errors were produced.

[0103] The present invention demonstrates good estimation performance in damping ratio estimation tasks under unfamiliar string complement scenarios. The minimum and maximum absolute estimation errors were calculated to be 0.000445 and 0.3323, respectively, and the overall mean estimation error for 100 sets of data was calculated to be 0.124958. Based on the calculated error results, the present invention exhibits good overall estimation performance and accuracy, and can be effectively used in damping ratio estimation tasks under unfamiliar string complement scenarios.

[0104] Example 4:

[0105] This invention is in unfamiliar K p Performance of parameter estimation in system scenarios.

[0106] Set the environmental excitation, wind speed, and series compensation degree within the aforementioned reasonable ranges, and select the proportional coefficient K of the inner loop controller for the rotor side current of the wind farm. p Data from power system operation scenarios within the range of 0.10 to 0.17 were used as the training set, K p Data from power system operation scenarios within the range of 0.18 to 0.20 were used as the test set. The constructed training set was used to train the model of the present invention to improve its performance. After training, the constructed test set was used to test the performance of the present invention.

[0107] Similarly, the test set contained 100 sets of data, and the final experimental results proved that the present invention is effective in unfamiliar K-type datasets. p Estimating the subsynchronous oscillation frequency parameters based on environmental excitation data is feasible in power system operation scenarios. The minimum absolute error on the test set is 0.0968Hz, and the maximum absolute error is 0.7421Hz. The overall mean error of this invention on the test set is calculated to be 0.339365Hz. Comprehensive error analysis shows that this invention is feasible in unfamiliar K-series power systems. p It demonstrates good estimation performance in frequency parameter estimation tasks under power system operation scenarios, showing excellent performance.

[0108] This invention is in unfamiliar K pThe damping ratio estimation performance is good in unfamiliar K-system scenarios. After detailed calculation, this invention performs well in these scenarios. p In the operating scenario of a power system, the minimum absolute error of the damping ratio parameter estimation is 0.0057, and the maximum absolute error is 0.3013. The damping ratio parameter estimation error of this invention on the test set remained between 0.0057 and 0.3013, demonstrating good performance stability. The overall mean error on the test set was calculated to be 0.113954, remaining at a low level. In summary, this invention performs well in unfamiliar K-series power systems. p In the context of power system operation, the damping ratio of subsynchronous oscillations can be estimated very effectively based on environmental excitation data. This invention demonstrates good accuracy and generalization in this task.

[0109] The test results of this invention in the above four different experimental environments show that the method of this invention is effective in the field of subsynchronous oscillation parameter estimation based on environmental excitation data and exhibits good performance. The final experimental results demonstrate that the method of this invention has good generalization ability in the task of subsynchronous oscillation parameter estimation based on environmental excitation data. In the four constructed experimental environments, it generalizes to unfamiliar environmental excitation, unfamiliar wind speed, unfamiliar series complement, and unfamiliar wind farm rotor-side current inner loop controller proportional coefficient K. p The method achieved good parameter estimation results in various power system operation scenarios. Furthermore, the method of this invention does not require analysis of specific signal models, eliminating the need for mathematical analysis of specific scenarios, which further facilitates the generalization and real-time application of the proposed method in practical scenarios. Additionally, the time consumption of the method described in this invention is mainly concentrated in the model training phase, which can be performed offline. In practical use, by calling the offline trained model proposed in this invention, the real-time requirements of online parameter estimation can be well met.

[0110] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0111] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A method for identifying subsynchronous oscillations based on environmental excitation data, characterized in that, The method includes the following steps: Receive environmental excitation data within the power system; The environmental excitation data is input into a pre-established gated recurrent unit neural network (GRU) model to extract the subsynchronous oscillation characteristics of the environmental excitation data. The subsynchronous oscillation features are identified using the Attention mechanism, and then the frequency and damping ratio parameters of the subsynchronous oscillation are obtained through dimensional transformation of the fully connected layer. The process of identifying subsynchronous oscillation features using the Attention mechanism includes the following steps: Calculate the weighted scores of the feature information of the environmental stimulus data at each time step; The calculated weight scores of the feature information are normalized to obtain the weights of the subsynchronous oscillation feature information at each time step. The Attention value is calculated using the weights of the subsynchronous oscillation feature information obtained at each time step; The calculation process for the weighted scores of the feature information at each time step of the computational environment stimulus data is as follows: In the formula, h t for t The hidden state output corresponding to the input environmental stimulus data at any given time, e t For the environmental stimulus data obtained in the final calculation t Weighted scores of time-series synchronous oscillation characteristic information; This is the final output value of the last timing sequence of the GRU; , , These are the parameters that the model needs to train; tanh is the activation function. The Attention value is obtained by combining the weights of the synchronous oscillation feature information at each time step with h. t The weighted average is obtained, and the Attention value is passed through a fully connected layer to finally output the frequency and damping ratio of the subsynchronous oscillation.

2. The subsynchronous oscillation identification method based on environmental excitation data according to claim 1, characterized in that, The environmental stimulus types in the environmental stimulus data include white noise and colored noise.

3. The subsynchronous oscillation identification method based on environmental excitation data according to claim 1, characterized in that, The GRU model includes a reset gate and an update gate, used to extract subsynchronous oscillation features contained in the environmental stimulus data.

4. The subsynchronous oscillation identification method based on environmental excitation data according to claim 3, characterized in that, The calculation formula for the reset door is as follows: In the formula, x t For the input environmental stimulus data in t The actual value at time h t-1 for t The hidden state output corresponding to the environmental stimulus data input at time -1; W xr W hr b r These are the parameters that the model needs to train; The function's range is 0 to 1; 0 indicates discarding, and 1 indicates retaining completely. t To reset the output of the gate.

5. The subsynchronous oscillation identification method based on environmental excitation data according to claim 3, characterized in that, The calculation formula for the update gate is as follows: In the formula, x t For the input environmental stimulus data in t The actual value at time h t-1 for t The hidden state output corresponding to the environmental stimulus data input at time -1; W xr W hr b r These are the parameters that the model needs to train; The function's range is from 0 to 1; 0 indicates discarding, and 1 indicates retaining completely. t To update the output of the gate.

6. The subsynchronous oscillation identification method based on environmental excitation data according to claim 1, characterized in that, The GRU model calculates the candidate hidden states and hidden states of the subsynchronous oscillation features extracted from the environmental excitation data.

7. The subsynchronous oscillation identification method based on environmental excitation data according to claim 6, characterized in that, The candidate hidden state output The calculation is as follows: In the formula, x t For the input environmental stimulus data in t The actual value at time h t-1 for t The hidden state output corresponding to the environmental stimulus data input at time -1; W xr W hr b r These are the parameters that the model needs to train; Represents element-wise multiplication; r t Reset the gate's output at the current time step; tanh is the activation function.

8. The subsynchronous oscillation identification method based on environmental excitation data according to claim 6, characterized in that, The hidden state output h t The calculation is as follows: In the formula, for t The output of the candidate hidden state corresponding to the input environmental stimulus data at each time step, z t Output the update gate for the current time step.

Images