New energy unit wide frequency impedance fast measurement method and system based on transfer learning

Abstract

The application discloses a new energy unit wide-frequency impedance fast measurement method and system based on transfer learning, and comprises an offline training stage: based on the electrical characteristics and topological structure of an actual new energy unit, a standard simulation model is constructed; a single-sine signal injection method is used to perform wide-frequency impedance scanning on the standard simulation model in all working conditions, so that offline impedance data sets are obtained; and a pre-training model is trained. In an online measurement stage: a model to be trained is constructed; the single-sine signal injection method is used to measure the impedance data of the actual new energy unit at key frequency points online, so that online impedance data sets are obtained; and the transfer learning technology is used to fine-tune the model to be trained, so that the wide-frequency impedance characteristics are quickly reconstructed. The application combines the transfer learning, reduces the online measurement frequency points by 80% to 90%, and meanwhile, maintains the high signal-to-noise ratio of the single-sine signal. The application solves the contradiction between the precision and the efficiency in the wide-frequency impedance measurement, and is particularly suitable for the application of frequent fluctuations of the grid-connected working conditions of the new energy unit.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy power generation, in particular to a new energy unit wide-frequency impedance fast measurement method and system based on transfer learning. BACKGROUND

[0002] Under the background of energy structure transformation and sustainable development, new energy power generation technology represented by wind energy and solar energy has received unprecedented attention and development. The power system is developing towards high proportion of renewable energy access and high proportion of power electronic equipment application, which has triggered a series of stability problems, seriously threatening the safe and stable operation of the power system and the efficient consumption of new energy. Therefore, it is an urgent need for current engineering application to realize the stability analysis and online evaluation of new energy grid-connected system.

[0003] Impedance analysis method is one of the mainstream methods for stability analysis of new energy grid-connected system, the core of which is to accurately obtain the wide-frequency impedance characteristics of new energy power generation equipment and power grid, and then use impedance stability criterion to analyze the stability of grid-connected system.

[0004] Impedance measurement is a method of obtaining impedance characteristics by injecting small signal disturbance into the measured system, and measurement accuracy and efficiency are two important indicators for evaluating impedance measurement performance. At present, impedance measurement methods are mainly divided into passive and active measurement. Passive measurement method uses the inherent harmonic or noise signal of the measured system port to estimate the impedance characteristics, without external disturbance source, which is low in cost but limited in measurement accuracy. In contrast, active measurement method injects self-designed disturbance signal into the measured system through external equipment to obtain more accurate impedance characteristics, which is higher in measurement accuracy and wider in application range, so it has been widely studied and applied.

[0005] According to the type of disturbance signal, active measurement method is mainly divided into single-sine signal injection method and wideband signal injection method.

[0006] Single-sine signal injection method injects single frequency disturbance signal point by point, which has the advantages of high signal-to-noise ratio and high measurement accuracy, but the measurement efficiency is low in full frequency range, which is difficult to meet the rapid measurement demand under dynamic working conditions. Wideband signal injection method injects disturbance signal containing multiple frequency components at one time, which can significantly improve the measurement efficiency, but the measurement accuracy decreases due to high signal complexity and limited disturbance amplitude. Therefore, the existing methods are difficult to achieve effective balance between measurement accuracy and efficiency, and it is difficult to meet the wide-frequency impedance fast measurement demand of new energy unit under dynamic working conditions. SUMMARY

[0007] In view of the defects in the prior art, the purpose of the present application is to provide a new energy unit wide-frequency impedance fast measurement method and system based on transfer learning.

[0008] According to one aspect of the present application, a new energy unit wide frequency impedance fast measurement method based on transfer learning is provided, comprising:

[0009] Offline training stage:

[0010] Based on the electrical characteristics and topological structure of the actual new energy unit, a standard simulation model of the new energy unit is constructed;

[0011] The single sinusoidal signal injection method is used to scan the wide frequency impedance of the standard simulation model under all operating conditions, and an offline impedance dataset is obtained;

[0012] Based on the offline impedance dataset, a pre-trained model is trained;

[0013] Online measurement stage:

[0014] Based on the pre-trained model, a to-be-trained model is constructed;

[0015] The single sinusoidal signal injection method is used to measure the impedance data of the actual new energy unit at key frequency points online, and an online impedance dataset is obtained;

[0016] Based on the online impedance dataset, the transfer learning technology is used to fine-tune the parameters of the to-be-trained model;

[0017] The fine-tuned to-be-trained model is used to measure the wide frequency impedance.

[0018] Preferably, the standard simulation model of the new energy unit is constructed based on the electrical characteristics and topological structure of the actual new energy unit, comprising:

[0019] The standard simulation model is constructed based on the Simulink or PSCAD simulation platform;

[0020] The main circuit topology, electrical parameters and control strategy are set to be consistent with the actual new energy unit;

[0021] The control parameters are set to ensure that the standard simulation model can operate stably in the full operating condition range.

[0022] Preferably, the single sinusoidal signal injection method is used to scan the wide frequency impedance of the standard simulation model under all operating conditions, and an offline impedance dataset is obtained, comprising:

[0023] The operating condition range is determined, which includes the active power output and the reactive power output range, the active power output range is 0-1 p.u., and the reactive power output range is the maximum capacitive reactive power to the maximum inductive reactive power required by the power grid unit output;

[0024] The frequency sweep range is determined, which covers at least half of the frequency band below the switching frequency;

[0025] The active power resolution, the reactive power resolution and the sweep frequency resolution are set to ensure that impedance data can be collected;

[0026] Under the working condition range, the sweep frequency range and the resolution, impedance data of the standard simulation model is collected by using a single sinusoidal signal injection method, to obtain an offline impedance data set.

[0027] Preferably, the pre-trained model is a BP neural network model.

[0028] Preferably, the pre-trained model is a BP neural network model.

[0029] The structure of the to-be-trained model is consistent with that of the pre-trained model, including an input layer, a hidden layer and an output layer.

[0030] The hidden layer parameters of the pre-trained model are inherited.

[0031] The output layer parameters are randomly initialized.

[0032] Preferably, the selection process of the key frequency points includes:

[0033] The boundary frequency points include the lower limit frequency point and the upper limit frequency point of the offline sweep frequency range.

[0034] The basic frequency points are selected uniformly at logarithmic intervals within the sweep frequency range composed of the boundary frequency points.

[0035] According to the stationarity of impedance characteristic changes, the density of the selected basic frequency points is increased or decreased.

[0036] Preferably, the pre-trained model is a BP neural network model.

[0037] According to the data amount and the complexity of impedance characteristics of the online impedance data set, the learning rate and the training rounds are set, the learning rate ranges from 1e-5 to 1e-3, and the training rounds range from 50 to 200.

[0038] The to-be-trained model is optimized by using the learning rate and the training rounds.

[0039] Preferably, the effectiveness of the pre-trained model and the to-be-trained model after fine-tuning is verified by a determination coefficient.

[0040] Preferably, the new energy unit includes a direct-drive wind turbine, a doubly-fed wind turbine, a photovoltaic power generation unit and an energy storage unit, and a new energy power generation system composed of the same; the wideband impedance includes traditional positive and negative sequence impedance, improved positive and negative sequence impedance and dq domain impedance.

[0041] According to a second aspect of the present application, a new energy unit wide frequency impedance fast measurement system based on transfer learning is provided, comprising an offline training module and an online measurement module.

[0042] The offline training module comprises:

[0043] The simulation unit: based on the electrical characteristics and topological structure of the actual new energy unit, a standard simulation model of the new energy unit is constructed.

[0044] The offline data unit: using single sine signal injection method, the standard simulation model is scanned for full working condition wide frequency impedance, and an offline impedance data set is obtained.

[0045] The pre-training unit: based on the offline impedance data set, a pre-training model is trained.

[0046] The online measurement module comprises:

[0047] The to-be-trained unit: based on the pre-training model, a to-be-trained model is constructed.

[0048] The online data unit: using single sine signal injection method, the impedance data of the actual new energy unit at the key frequency points are measured online, and an online impedance data set is obtained.

[0049] The fine-tuning unit: based on the online impedance data set, the parameters of the to-be-trained model are fine-tuned using transfer learning technology.

[0050] The measurement unit: using the fine-tuned to-be-trained model, the wide frequency impedance is measured.

[0051] Compared with the prior art, the present application has at least one of the following beneficial effects:

[0052] The new energy unit wide frequency impedance fast measurement method based on transfer learning of the present application, by stage-by-stage fusion of single sine signal injection and transfer learning technology, while retaining the high signal-to-noise ratio advantage of single sine signal injection, significantly reduces the online measurement frequency points, effectively solves the contradiction between precision and efficiency in wide frequency impedance measurement, realizes the rapid and accurate reconstruction of wide frequency impedance characteristics, provides reliable technical support for stability analysis and online evaluation of new energy grid-connected system, and effectively guarantees the safe and stable operation of new energy grid-connected system. Especially suitable for wide frequency impedance fast measurement in photovoltaic power generation unit, wind turbine and other new energy grid-connected working condition frequent fluctuation scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0053] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments, with reference to the accompanying drawings:

[0054] Figure 1A basic flowchart of a new energy unit wide-frequency impedance fast measurement method based on transfer learning in an embodiment;

[0055] Figure 2 An electrical structure and a control structure of a direct-drive permanent magnet synchronous wind turbine (DDPMSG-WTG) in an embodiment;

[0056] Figure 3 A wide-frequency admittance characteristic measurement result of an actual direct-drive permanent magnet synchronous wind turbine (DDPMSG-WTG) under a rated operating condition in an embodiment;

[0057] Figure 4 A structural schematic diagram of a new energy unit wide-frequency impedance fast measurement system based on transfer learning in an embodiment. DETAILED DESCRIPTION

[0058] The application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that, for those skilled in the art, without departing from the concept of the application, a number of modifications and improvements can be made. These all belong to the protection scope of the application.

[0059] The traditional single-sine signal injection method has the advantages of high signal-to-noise ratio and high measurement accuracy, but the full-frequency measurement efficiency is low, and it is difficult to meet the rapid measurement demand under dynamic operating conditions. The wideband signal injection method can significantly improve the measurement efficiency by injecting a disturbance signal containing multiple frequency components at one time, but the measurement accuracy is reduced due to high signal complexity and limited disturbance amplitude. In summary, the existing methods are difficult to achieve effective balance between measurement accuracy and efficiency. The embodiment of the application proposes a new energy unit wide-frequency impedance fast measurement method based on transfer learning, which greatly improves the measurement efficiency while ensuring the measurement accuracy, and effectively solves the contradiction between measurement accuracy and efficiency.

[0060] Referring to Figure 1 FIG. 1 shows a basic flowchart of a new energy unit wide-frequency impedance fast measurement method based on transfer learning in an embodiment, which measurement method adopts the following steps:

[0061] First stage: offline training stage:

[0062] Offline training stage:

[0063] S1. Based on the electrical characteristics and topological structure of the actual new energy unit, a standard simulation model of the new energy unit is constructed;

[0064] S2. The single-sine signal injection method is used to perform full-condition wide-frequency impedance scanning on the standard simulation model to obtain an offline impedance data set;

[0065] S3. training the pre-trained model based on the offline impedance dataset;

[0066] the online measurement phase:

[0067] S4. constructing a to-be-trained model based on the pre-trained model;

[0068] S5. measuring impedance data of the actual new energy unit at key frequency points in an online manner by using a single-sine signal injection method to obtain an online impedance dataset;

[0069] S6. fine-tuning parameters of the to-be-trained model based on the online impedance dataset by using a transfer learning technique;

[0070] S7. measuring wide-frequency impedance by using the fine-tuned to-be-trained model.

[0071] Specifically, the input and output data format of the fine-tuned to-be-trained model is {input: U d , U q , I d , I q , f; output: Z RS}. Wherein, U d , U q , I d , I q , f are the port dq-axis voltage, current and disturbance frequency of the actual new energy unit under real-time operating conditions; and Z RS is the port impedance of the actual new energy unit under real-time operating conditions. According to actual application requirements, the impedance type can be selected as positive and negative sequence impedance or dq-domain impedance, and the impedance form can be selected as real part and imaginary part, or amplitude and phase angle.

[0072] It should be noted that the fine-tuned to-be-trained model is not directly measured by the single-sine signal injection method at each frequency point. Its essence is the wide-frequency impedance characteristics of the actual new energy unit under real-time operating conditions. By inputting U d , U q , I d , I q and the frequency f, the full-band impedance characteristics are predicted, and the time-consuming process of measuring each frequency point is saved.

[0073] The above embodiment fuses the single-sine signal injection and the transfer learning technique in stages, while retaining the high signal-to-noise ratio advantage of the single-sine signal injection, significantly reduces the online measurement frequency points, and effectively solves the contradiction between precision and efficiency in wide-frequency impedance measurement.

[0074] Reference Figure 2As shown, the actual new energy unit of an embodiment of the present application is the electrical structure and control structure diagram of a direct-drive permanent magnet synchronous wind turbine (DDPMSG-WTG). Based on this structure, the embodiment implements step S1 described above to construct a standard simulation model of the direct-drive permanent magnet synchronous wind turbine.

[0075] Due to the restriction of commercial secrets and other factors, the control system of the direct-drive permanent magnet synchronous wind turbine usually presents a "black / gray box" characteristic, and the internal detailed control structure and parameters cannot be obtained. The embodiment can obtain the main circuit topology, electrical parameters and the basic control structure (such as V dc Q control, PQ control, etc.) of the system from the system operator. Based on the above structure and parameter information, a standard simulation model is constructed on the Simulink and PSCAD simulation platforms. Specifically, the standard simulation model is consistent with the actual new energy unit in terms of the main circuit topology, electrical parameters and the basic control structure adopted; for the control parameters that cannot be obtained, common control parameter design methods can be used for design to ensure that the standard simulation model can stably operate in the full operating condition range.

[0076] As an example, the control parameter design method can refer to the following literature, as follows (Reference: [1] Dai Jinshui, Lv Jing. Design of double closed-loop PI regulator for grid-connected inverter of wind power generation [J]. Electrical and Electrical, 2011, (09): 5-9.)

[0077] An embodiment of the present application implements S2 to perform full operating condition wide frequency impedance scanning on the standard simulation model by using a single sinusoidal signal injection method to obtain an offline impedance data set. Specifically, the following steps can be used:

[0078] S21, determine the operating condition range, including the active power output and the reactive power output range.

[0079] To ensure that the pre-trained model established subsequently is applicable to the entire steady state operating interval and meets the reactive power requirements of the power grid, the active power output range is set to 0-1 p.u., the reactive power output range is set to the maximum capacitive reactive power to the maximum inductive reactive power required by the power grid for unit output, and has sufficient resolution.

[0080] It should be noted that the numerical range referred to by "sufficient resolution" needs to be determined according to the specific scene requirements, and the core principle is to avoid missing the key characteristics of the impedance characteristics due to insufficient frequency scanning resolution.

[0081] The resolution of different numerical ranges can adapt to the impedance characteristic differences of different new energy units (such as wind power and photovoltaic), and retain the universality and flexibility of the technical solution.

[0082] Exemplary: in the frequency band where the impedance characteristic changes significantly, the sweep frequency resolution needs to be high enough (e.g. ≤5Hz, more finely, e.g. ≤1Hz); in the frequency band where the impedance characteristic changes gently, the sweep frequency resolution can be appropriately reduced (e.g. ≥20Hz).

[0083] Of course, the acquisition, labeling cost of impedance data and the training cost of the model in actual application also need to be considered, and the sweep frequency resolution cannot be too fine.

[0084] Exemplary: the active power output range is set to 0-1 p.u., the reactive power output range is set to -0.3-0.3 p.u., and the resolution is set to 0.1 p.u., a total of 70 operating conditions are obtained.

[0085] S22, determine the sweep frequency range, i.e. the frequency range of the disturbance signal injected into the new energy unit port.

[0086] The sweep frequency range here covers at least the lower half of the switching frequency band. Similarly, it also needs to have sufficient resolution.

[0087] Exemplary: the sweep frequency range is set to 1-1001Hz, the resolution of the low frequency band (1-200Hz) is set to 1Hz, the resolution of the high frequency band (201-1001Hz) is set to 25Hz, and a total of 233 frequency points are obtained.

[0088] S23, measure the impedance data of the standard simulation model within the setting of the operating condition range, sweep frequency range and resolution, and obtain the offline impedance data set.

[0089] It should be noted that: considering the low efficiency requirement of offline measurement, the single-sine signal injection method with the highest measurement accuracy can be preferred for frequency point-by-frequency point measurement. The format of the offline impedance data set obtained by measurement is {input: U d , U q , I d , I q , f; output: Z SM}. Wherein, U d , U q , I d , I q , f are the dq-axis voltage, current and disturbance frequency of the standard simulation model port, a total of five inputs; Z SM is the conventional positive and negative sequence impedance of the standard simulation model port, expressed as real and imaginary parts, a total of four outputs.

[0090] The S21-S23 in the above embodiment constructs a high-quality and high-coverage offline impedance dataset by fine design of working conditions and sweep parameters, in combination with a high-precision data acquisition method. This provides a reliable pre-training basis for subsequent steps, so that only a small number of key frequency points are needed in the online measurement stage to quickly reconstruct the wideband impedance characteristics, thereby improving the online measurement efficiency by more than 80% while ensuring the accuracy, and effectively solving the contradiction between the impedance measurement accuracy and efficiency under the dynamic working conditions of new energy units.

[0091] In a preferred embodiment of the present application, the pre-trained model used is a BP neural network model. The BP neural network model has several significant advantages.

[0092] In terms of the balance between efficiency and cost, the BP neural network can achieve a determination coefficient (which reflects the prediction accuracy of the model) of more than 99.99% at a relatively low training cost in the wideband impedance model identification task of new energy units, due to its lightweight structure and efficient training characteristics. In contrast, if other network models (such as CNN, RNN, etc.) are used, not only is the improvement in model prediction accuracy very limited, but also several times the training cost needs to be invested, resulting in a marginal benefit much lower than the increase in cost, thereby significantly reducing the engineering cost-effectiveness.

[0093] In terms of task characteristics and model adaptability, the inputs (including port dq voltage, current, and disturbance frequency) and outputs (port impedance) of the wideband impedance identification task of new energy units are independent and have no spatial and temporal correlation, which essentially belongs to a multiple-input multiple-output static regression problem. The BP neural network has high adaptability to this task characteristic and is sufficient to meet the demand for high accuracy. Other networks (such as CNN, RNN) have advantages in designing data with spatial and temporal continuity or graphical dependence, but they cannot take advantage of these advantages in this task, and instead, the risk of overfitting and deployment complexity increase due to the exponential increase in the number of parameters.

[0094] In terms of the friendliness of transfer learning, the shallow structure of the BP neural network makes it easier to inherit and fine-tune parameters in transfer learning, making the transfer process more stable and efficient.

[0095] Further, in a preferred embodiment, step S3 is implemented. Based on the offline impedance dataset, a pre-trained model of the BP neural network is trained. Specifically, the following steps can be used:

[0096] S31, the offline impedance dataset is normalized and divided into a training set and a test set.

[0097] For example, the Z-score normalization method is used for normalization, and the impedance data of 70 working conditions are evenly divided into a training set and a test set in order.

[0098] S32, a BP neural network is constructed, the number of neurons of the input and output layers is consistent with the dimension of the input and output features respectively, and the number of layers of the hidden layer, the number of neurons of each layer of the hidden layer, and the training algorithm and hyperparameters need to be adjusted and optimized according to the performance of the neural network. After completing the network construction and parameter setting, the BP neural network constructed is trained based on the training set to obtain a pre-trained model.

[0099] For example, the number of neurons of the input layer and the output layer of the BP neural network is 5 and 4 respectively; the number of layers of the hidden layer is 5, and the number of neurons of each hidden layer is 8, 10, 12, 14 and 12 respectively; the training algorithm is set to Levenberg-Marquardt algorithm, the learning rate is set to 0.003, and the number of training rounds is set to 800.

[0100] S33, the effectiveness of the pre-trained model is verified and evaluated by using the coefficient of determination to guide the structure optimization and parameter adjustment of the pre-trained model, and to ensure the reliability and applicability of the pre-trained model. The calculation formula of the coefficient of determination is as follows:

[0101]

[0102] Wherein, n is the sample capacity; y i is the impedance measurement value obtained by sweeping, is the impedance prediction value output by the model, is the average value of the sample impedance measurement value. The range of the coefficient of determination is 0-1, the closer the coefficient of determination is to 1, the better the effectiveness of the pre-trained model. The above effectiveness includes accuracy and generalization, which are represented by the coefficients of determination calculated based on the training set and the test set respectively. In this embodiment, the accuracy and generalization of the pre-trained model are as high as 99.99996%, which indicates that the model has high fitting precision and excellent generalization ability, and provides a highly reliable pre-training basis for the subsequent use of transfer learning technology.

[0103] After the pre-trained model is trained, the online measurement phase is entered. In a preferred embodiment, step S4 is implemented to construct a to-be-trained model with pre-trained initialization of the structure and the parameters of the hidden layer and random initialization of the parameters of the output layer. Specifically:

[0104] A new BP neural network is constructed, which has the same architecture as the pre-trained model and covers the input layer, the hidden layer and the output layer. The input layer serves as a channel for information transmission, and the number of nodes thereof matches the dimension of the input features, and does not need to be initialized. The initialization process only involves the weights and biases of the hidden layer and the output layer. Specifically, the parameters of the hidden layer are inherited from the pre-trained model, and the parameters of the output layer are randomly initialized to form the to-be-trained model.

[0105] In an embodiment of the present application, step S5 is implemented to measure the impedance data of the actual new energy unit at the key frequency points by using the single-sine signal injection method, so as to obtain the online impedance data set. Specifically, the following steps can be used:

[0106] S51, select the key frequency points.

[0107] First, determine the boundary frequency points: the starting frequency point and the ending frequency point covering the above-mentioned sweep frequency range;

[0108] Then, uniformly sample in logarithmic intervals: within the determined sweep frequency range boundary, uniformly select the basic frequency points in logarithmic intervals;

[0109] Finally, implement dynamic adjustment of frequency band density: for the selected basic frequency points, reduce the density of frequency points in the low frequency band and the high frequency band where the impedance characteristic changes relatively gently; and increase the density of frequency points in the working frequency band where the impedance characteristic changes significantly.

[0110] For example, 24 key frequency points are selected, which are 1, 6, 11, 21, 31, 41, 52, 62, 72, 82, 92, 102, 122, 142, 162, 182, 201, 326, 451, 576, 701, 826, 951, and 1001 Hz.

[0111] S52, inject single-sine disturbance signals of each key frequency point into the port of the actual new energy unit in turn, and measure the online impedance data set. It should be ensured that the data format of the online impedance data set is completely consistent with that of the offline impedance data set.

[0112] Among them, the characteristics of the single-sine disturbance signal are determined by three parameters: amplitude, frequency, and initial phase.

[0113] Regarding the amplitude, the amplitude of the disturbance must be controlled within a sufficiently small range to ensure that it does not affect the stable operation of the unit. At the same time, in order to meet the measurement requirements, it is necessary to ensure that there is a large enough signal-to-noise ratio. Generally, the disturbance amplitude is set to between 1% and 5% of the voltage or current amplitude of the injected point. In terms of frequency, in step S52, the frequency is set to the selected key frequency points in turn, for example, the 24 key frequency points selected in the above embodiment. As for the initial phase, it is generally set to 0 degrees.

[0114] It should be noted that the generation method of the single-sine disturbance signal in step S23 is the same as above, and the difference is that S23 is full-band sweep (such as 233 points in the embodiment).

[0115] Of course, in some other embodiments, the measurement accuracy requirement is low, and S52 can use the multi-sine signal injection method to improve the measurement efficiency when implemented.

[0116] Compared with the traditional single / multi-sine perturbation injection method, the number of frequency points measured online is reduced by about 89.70% by injecting single-sine perturbation signals of key frequency points into the port of the actual new energy unit, and the measurement efficiency is improved.

[0117] In an embodiment of the present application, step S6 is implemented, and the fine-tuning of the to-be-trained model parameters is performed based on the online impedance data set using the transfer learning technology. Specifically, the following steps can be used:

[0118] S61, set the learning rate and training rounds.

[0119] Since the actual new energy unit and the standard simulation model only differ in control parameters and structural details, as shown in Table 1, they belong to weakly heterogeneous or even homogeneous systems, so a small learning rate and training rounds should be used to optimize the to-be-trained model during parameter fine-tuning to avoid overfitting and reduce model performance.

[0120] For example, the initial values of the learning rate and the training rounds can be preset according to the amount of online data, impedance characteristics complexity, etc. The learning rate is 1e-5~1e-3, and the training rounds are 50~200.

[0121] It should be noted that the main purpose of the "lower learning rate and training rounds" set here is to prevent overfitting and maintain the knowledge in the pre-training stage. Therefore, the above numerical range is not fixed, but should be dynamically adjusted according to the evaluation of the model training effect. For example:

[0122] If the model shows high accuracy but poor generalization ability, it may mean that overfitting has occurred, and the learning rate should be reduced at this time.

[0123] By observing the trend of the loss function with the training period, if the loss function has stabilized before reaching the preset maximum training period, the training process can be terminated in advance.

[0124] S62, fine-tune the parameters of the to-be-trained model based on the online impedance data to obtain the wideband impedance characteristic model of the actual new energy unit under the current operating condition.

[0125] Further, in a specific implementation, the learning rate is set to 0.00001, and the training rounds are set to 80. Compared with the traditional data-driven wideband impedance identification method, the training rounds are reduced by 90%, significantly improving the efficiency of model training, and further improving the overall efficiency of impedance measurement of the present application.

[0126] Table 1 Parameters of standard simulation model and actual new energy unit

[0127]

[0128]

[0129] Similarly, in an embodiment, the coefficient of determination is used to verify and evaluate the effectiveness of the obtained wideband impedance characteristic model to guide the adjustment of learning rate and training rounds in the fine-tuning process, and to ensure that the measurement method proposed in the application has high enough measurement accuracy.

[0130] The offline training stage in the above embodiment is based on a standard simulation model of a certain type of new energy unit, and the common law of its wideband impedance characteristics is obtained through pre-training to construct a general knowledge base of its impedance characteristics. The online stage is aimed at the characteristic deviation of the actual new energy unit caused by the difference in control structure, parameters, etc. Through fine-tuning, the reuse of pre-training knowledge is realized, and the engineering problem of "one unit one model" is solved.

[0131] In practical applications, new energy units cannot operate safely and stably under some extreme conditions. Not only can the wideband impedance characteristics of the unit under extreme conditions be measured, but also the safety of equipment and personnel may be threatened. The phased strategy of the above embodiment completes the impedance scanning of the full working condition and full frequency band in the simulation environment through the offline measurement stage, effectively avoiding the risk of extreme condition measurement of the actual new energy unit, while ensuring the comprehensiveness and reliability of the measurement.

[0132] In addition, in a specific embodiment, the model trained by fine-tuning is used for measurement, and the obtained wideband admittance characteristic measurement results of the actual DDPMSG-WTG under rated conditions are as shown in Figure 3 The effectiveness of the obtained wideband admittance characteristic model is as high as 99.91%, and the entire measurement process takes no more than 5 seconds, indicating that the impedance measurement method proposed in the embodiment of the application effectively solves the contradiction between the measurement accuracy and efficiency of the wideband impedance of the new energy unit.

[0133] In some specific embodiments, the representation form of impedance can adopt the form of imaginary part and real part or amplitude and phase angle to meet different application requirements.

[0134] In some specific embodiments, the single-sine signal injection method is used to measure impedance data point by point; when the measurement accuracy requirement is low, the multi-sine signal injection method can be used to improve the measurement efficiency.

[0135] The application scope of the application includes but is not limited to the wideband impedance fast measurement of direct-drive wind turbine generators, doubly-fed wind turbine generators, photovoltaic power generation units, energy storage units and new energy power generation systems composed of them. The wideband impedance includes traditional positive and negative sequence impedance, improved positive and negative sequence impedance and dq domain impedance, and admittance is also applicable.

[0136] Based on the same inventive concept, in other embodiments of the present application, a new energy unit wide frequency impedance fast measurement system based on transfer learning is provided, as shown in the accompanying drawings, comprising an offline training module and an online measurement module. Figure 4

[0137] The offline training module comprises:

[0138] The simulation unit: based on the electrical characteristics and topological structure of the actual new energy unit, a standard simulation model of the new energy unit is constructed;

[0139] The offline data unit: using single sine signal injection method, the standard simulation model is scanned for full working condition wide frequency impedance, and the offline impedance data set is obtained;

[0140] The pre-training unit: based on the offline impedance data set, the pre-training model is trained;

[0141] The online measurement module comprises:

[0142] The to-be-trained unit: based on the pre-training model, the to-be-trained model is constructed;

[0143] The online data unit: using single sine signal injection method, the impedance data of the actual new energy unit at the key frequency points is measured online, and the online impedance data set is obtained;

[0144] The fine-tuning unit: based on the online impedance data set, the transfer learning technology is used to fine-tune the parameters of the to-be-trained model;

[0145] The measurement unit: using the fine-tuned to-be-trained model, the wide frequency impedance is measured.

[0146] In the above examples of the present application, each module / unit can refer to the implementation technology of the corresponding steps of the above-mentioned embodiments of the new energy unit wide frequency impedance fast measurement method based on transfer learning, which will not be repeated here.

[0147] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various modifications or modifications within the scope of the claims, which does not affect the essential content of the present application. The above preferred features can be combined for use in the case of not conflicting with each other.​

Claims

1. A method for fast measurement of wideband impedance of new energy units based on transfer learning, characterized in that, The method comprises the following steps: An offline training phase: Based on the electrical characteristics and topological structure of the actual new energy unit, a standard simulation model of the new energy unit is constructed; Single-sine signal injection method is used to scan the wide frequency impedance of the standard simulation model in all operating conditions, and an offline impedance data set is obtained; Based on the offline impedance data set, a pre-trained model is trained; An online measurement phase: Based on the pre-trained model, a to-be-trained model is constructed; Single-sine signal injection method is used to measure the impedance data of the actual new energy unit at key frequency points, and an online impedance data set is obtained; Based on the online impedance data set, the parameters of the to-be-trained model are fine-tuned using transfer learning technology; The fine-tuned to-be-trained model is used to measure the wide frequency impedance; The single-sine signal injection method is used to scan the wide frequency impedance of the standard simulation model in all operating conditions, and an offline impedance data set is obtained, which comprises the following steps: Determine the operating condition range, which includes the active power output and reactive power output range, the active power output range is 0-1 p.u., and the reactive power output range is the maximum capacitive reactive power to the maximum inductive reactive power required by the power grid; Determine the sweep range, which covers at least half of the lower frequency band of the switching frequency; Set the active power resolution, reactive power resolution and sweep resolution to ensure that impedance data can be collected; Under the conditions of the operating condition range, sweep range and resolution, impedance data of the standard simulation model is collected point by point using single-sine signal injection method, and an offline impedance data set is obtained; Based on the pre-trained model, a to-be-trained model is constructed, which comprises the following steps: The structure of the to-be-trained model is consistent with that of the pre-trained model, including input layer, hidden layer and output layer; Inherit the hidden layer parameters of the pre-trained model; Randomly initialize the output layer parameters; The selection process of the key frequency points comprises the following steps: Select the boundary frequency points, including the lower limit frequency point and the upper limit frequency point of the offline sweep range; Uniformly select the basic frequency points in the sweep range composed of the boundary frequency points according to logarithmic interval; According to the smoothness of impedance characteristics, increase or decrease the density of the selected basic frequency points.

2. The method of claim 1, wherein the method is based on transfer learning. Based on the electrical characteristics and topological structure of the actual new energy unit, a standard simulation model of the new energy unit is constructed, which comprises the following steps: Based on the Simulink or PSCAD simulation platform, a standard simulation model is constructed; Set the main circuit topology, electrical parameters and control strategy to be consistent with the actual new energy unit; Set the control parameters to ensure that the standard simulation model can operate stably in the full operating condition range.

3. The method of claim 1, wherein the method is based on transfer learning. The pre-trained model is a BP neural network model.

4. The method of claim 1, wherein the method is based on transfer learning. Based on the online impedance data set, the parameters of the to-be-trained model are fine-tuned using transfer learning technology, which comprises the following steps: According to the data volume and impedance characteristic complexity of the online impedance data set, set the learning rate and training rounds, the learning rate range is 1e-5~1e-3, and the training rounds range is 50~200 times; Optimize the to-be-trained model using the learning rate and training rounds.

5. The method of claim 1, wherein the method is based on transfer learning. The effectiveness of the pre-trained model and the to-be-trained model after fine-tuning is verified by a coefficient of determination.

6. The method of claim 1, wherein the method is based on transfer learning. The new energy unit includes a direct-drive wind turbine, a double-fed wind turbine, a photovoltaic power generation unit, an energy storage unit and a new energy power generation system composed of the same; the wideband impedance includes traditional positive and negative sequence impedance, improved positive and negative sequence impedance and Dq domain impedance.

7. A new energy unit wide frequency impedance fast measurement system based on transfer learning, used to realize the new energy unit wide frequency impedance fast measurement method based on transfer learning in claim 1. The method comprises an offline training module and an online measurement module. The offline training module comprises: An emulation unit: based on the electrical characteristics and topology structure of an actual new energy unit, a standard emulation model of the new energy unit is constructed; An offline data unit: a single-sine signal injection method is used to perform full-condition wide-frequency impedance scanning on the standard emulation model to obtain an offline impedance dataset; A pre-training unit: based on the offline impedance dataset, a pre-training model is trained; The online measurement module comprises: A to-be-trained unit: based on the pre-training model, a to-be-trained model is constructed; An online data unit: a single-sine signal injection method is used to measure the impedance data of an actual new energy unit at key frequency points online to obtain an online impedance dataset; A fine-tuning unit: based on the online impedance dataset, a transfer learning technique is used to fine-tune the parameters of the to-be-trained model; A measurement unit: the to-be-trained model after fine-tuning is used to measure the wide-frequency impedance.

Images