Island state detection method and system for large-scale wind power

Abstract

The present invention provides an island state detection method and system for large-scale wind power, which relates to the technical field of wind power generation, including dividing a wind farm into a core area and a boundary area, combining the power flow direction of the boundary area and the frequency response curve of the core area, constructing a three-dimensional correlation feature matrix including network impedance characteristics, power flow characteristics and frequency response characteristics; generating an expanded feature matrix based on weather condition information in historical operation data of the wind farm; constructing a training data set using the expanded feature matrix, inputting the training data set into a hybrid deep learning network, and obtaining an island state detection model based on the extracted spatial features and time series features and the corresponding island state labels; judging the island state based on the island state detection model, outputting an island state warning signal and triggering a safety protection mechanism of the wind farm.

Description

Technical Field

[0001] The present invention relates to wind power generation technology, and in particular to an island state detection method and system for large-scale wind power. Background Art

[0002] At present, wind farm island detection methods mainly include passive and active methods. Passive detection methods rely on monitoring changes in grid parameters, such as frequency, voltage, power, etc., to determine whether an island occurs. However, the changes in these parameters in certain circumstances (such as when the load fluctuates greatly) may be similar to the changes in the island state, leading to misjudgment. In addition, the passive method has a low sensitivity for island detection of small-capacity wind farms and is difficult to meet actual needs.

[0003] Active detection methods determine the islanding state by injecting specific disturbance signals into the power grid and analyzing the system's response to the disturbance. Most existing active methods focus on a single feature, such as impedance, power direction, or frequency response. However, a single feature is easily affected by noise and system operating conditions, resulting in reduced reliability of the detection results. Moreover, these methods usually lack effective use of historical data and fail to fully tap the potential information contained in the data, thus limiting detection accuracy and adaptability. In addition, existing technologies also lack intelligent adjustment strategies in the selection of disturbance signals and parameter settings, making it difficult to adapt to the complex and changing power grid environment. Summary of the invention

[0004] The embodiments of the present invention provide a method and system for detecting an island state of large-scale wind power, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] Provides an island status detection method for large-scale wind power, including:

[0007] The wind farm is divided into a core area and a boundary area, an active disturbance injection unit is arranged in the boundary area, the active disturbance injection unit is controlled to inject a small disturbance signal into the power grid, the voltage and current data of the core area and the boundary area are collected in real time, the network impedance characteristics are obtained by detecting the response characteristics of the core area to the injected disturbance, and a three-dimensional correlation characteristic matrix including network impedance characteristics, power flow characteristics and frequency response characteristics is constructed by combining the power flow direction of the boundary area and the frequency response curve of the core area;

[0008] Based on the weather condition information in the historical operation data of the wind farm, the three-dimensional correlation feature matrix is ​​feature expanded to generate an expanded feature matrix; a training data set is constructed using the expanded feature matrix, and the training data set is input into a hybrid deep learning network including a convolutional neural network and a long short-term memory network, the spatial features of the expanded feature matrix are extracted through the convolutional neural network, the temporal features of the expanded feature matrix are extracted through the long short-term memory network, the spatial features and the temporal features are fused to obtain a fused feature vector, and the hybrid deep learning network is trained based on the fused feature vector and the corresponding island state label to obtain an island state detection model;

[0009] The real-time monitoring data of the wind farm is input into the island state detection model, and the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector is calculated based on the island state detection model. When the matching degree is greater than a preset threshold, it is determined that the wind farm has an island state, an island state warning signal is output, and a safety protection mechanism of the wind farm is triggered.

[0010] In an optional embodiment,

[0011] The wind farm is divided into a core area and a boundary area, an active disturbance injection unit is arranged in the boundary area, the active disturbance injection unit is controlled to inject a small disturbance signal into the power grid, voltage and current data of the core area and the boundary area are collected in real time, network impedance characteristics are obtained by detecting the response characteristics of the core area to the injected disturbance, and a three-dimensional correlation feature matrix including network impedance characteristics, power flow characteristics and frequency response characteristics is constructed in combination with the power flow direction of the boundary area and the frequency response curve of the core area. The steps include:

[0012] The wind farm is divided into a core area and a boundary area based on the electrical distance. The normalized electrical distance is obtained by calculating the ratio of the equivalent impedance from each wind turbine to the booster station to the base impedance. The wind turbine group whose normalized electrical distance is less than a preset threshold is divided into the core area, and the wind turbine group whose normalized electrical distance is greater than the preset threshold is divided into the boundary area.

[0013] A plurality of active disturbance injection units are arranged in the boundary area, wherein the active disturbance injection units include a power converter and an energy storage device, and the active disturbance injection units are controlled to inject an attenuated sinusoidal disturbance signal having a preset frequency and a preset amplitude into the power grid, and the attenuated sinusoidal disturbance signal is used as an input excitation signal;

[0014] Collecting a response voltage and a response current of the core area after receiving an input excitation signal, and calculating a network impedance characteristic of the core area according to a ratio of the response voltage to the response current, wherein the network impedance characteristic includes an impedance amplitude characteristic and a phase angle characteristic;

[0015] collecting active power and reactive power in the boundary area, and constructing a power flow vector in the boundary area, wherein the power flow vector is used to characterize the energy exchange state between the wind farm and the external power grid;

[0016] Performing wavelet transform on the frequency response curve of the core area to extract the frequency response characteristics of the core area, wherein the frequency response characteristics include scale coefficients and translation coefficients;

[0017] The network impedance characteristics, power flow vector and frequency response characteristics are combined to construct a three-dimensional correlation feature matrix.

[0018] In an optional embodiment,

[0019] Based on the weather condition information in the historical operation data of the wind farm, the three-dimensional correlation feature matrix is ​​feature expanded to generate an expanded feature matrix; a training data set is constructed using the expanded feature matrix, and the training data set is input into a hybrid deep learning network including a convolutional neural network and a long short-term memory network, and the spatial features of the expanded feature matrix are extracted by the convolutional neural network, and the temporal features of the expanded feature matrix are extracted by the long short-term memory network, and the spatial features and the temporal features are fused to obtain a fused feature vector, and the steps of training the hybrid deep learning network based on the fused feature vector and the corresponding island state label to obtain an island state detection model include:

[0020] Collecting weather condition information from historical operation data of the wind farm, the weather condition information including wind speed, temperature, humidity and air pressure; normalizing the weather condition information, aligning the normalized weather condition information with the three-dimensional correlation feature matrix in the time dimension, and generating an expanded feature matrix;

[0021] Constructing a hybrid deep learning network, wherein the hybrid deep learning network includes a convolutional neural network and a long short-term memory network, wherein the convolutional neural network includes a plurality of convolutional layers, a normalization layer, and an activation function layer, and the long short-term memory network includes a plurality of memory layers;

[0022] The expanded feature matrix is ​​input into a convolutional neural network, and the spatial feature vector of the expanded feature matrix is ​​extracted through the multiple convolutional layers. The output of each convolutional layer is subjected to feature enhancement through a normalization layer and an activation function layer to obtain a feature sequence after spatial dimension compression;

[0023] Inputting the feature sequence compressed by the spatial dimension into a long short-term memory network, extracting the temporal feature vector of the feature sequence through the multiple memory layers, wherein the memory layer includes a forget gate, an input gate, and an output gate;

[0024] The spatial feature vector and the temporal feature vector are concatenated to obtain a fused feature vector, and the fused feature vector is subjected to feature dimension reduction and anti-overfitting processing through a fully connected layer and a random dropout layer; based on the fused feature vector and a preset island state label, the hybrid deep learning network is trained using a multiple loss function to obtain an island state detection model.

[0025] In an optional embodiment,

[0026] The steps of constructing a hybrid deep learning network, wherein the hybrid deep learning network includes a convolutional neural network and a long short-term memory network, wherein the convolutional neural network includes a plurality of convolutional layers, a normalization layer, and an activation function layer, and wherein the long short-term memory network includes a plurality of memory layers include:

[0027] In a convolutional neural network, the importance weights of different convolutional layers are adaptively calculated by a global context encoder, and the global context encoder generates a channel attention map based on the statistical information of the input features, and the channel attention map is used to adjust the fusion weights of the output features of different convolutional layers;

[0028] A feature interaction module is constructed between convolutional layers. The feature interaction module includes a feature resampling unit and an adaptive fusion unit. Dynamic information exchange of output features of different convolutional layers is realized through a transformation matrix. The parameters of the transformation matrix are updated through a back propagation algorithm. The update step is proportional to the feature importance. The adaptive fusion unit automatically adjusts the information exchange intensity according to the feature correlation.

[0029] A bidirectional memory path is set in the long short-term memory network, the memory layers of the forward path and the reverse path share parameters, and the bidirectional timing information is dynamically integrated through a state fusion gate, and the state fusion gate calculates the fusion weight based on the current input and the historical state;

[0030] A time-aware mechanism is embedded in the memory layer to encode the time intervals between adjacent moments into a continuous vector, and the continuous vector is integrated into the calculation of the forget gate and the input gate. The time interval encoding is generated by a learnable time embedding layer.

[0031] In an optional embodiment,

[0032] Based on the fused feature vector and the preset island state label, the steps of training the hybrid deep learning network using multiple loss functions to obtain the island state detection model include:

[0033] Based on the fused feature vector and the preset island state label, a feature importance evaluation index is calculated through feature mutual information calculation, feature correlation analysis, and feature prediction gain evaluation; a dynamic feature weight is generated through a softmax function with adjustable smoothness, and the dynamic feature weight is used to weight the feature importance evaluation index;

[0034] Performing time series enhancement processing on the input data according to the feature importance evaluation index to obtain an enhanced feature sequence, selecting time series-related samples from the enhanced feature sequence to construct positive sample pairs, and using a contrastive learning loss function to calculate the cosine similarity between the positive sample pairs, wherein the contrastive learning loss function includes a positive sample pair similarity calculation item and a negative sample pair similarity calculation item;

[0035] Constructing a multi-layer attention network to extract the correlation between features to obtain a feature correlation score, combining the feature correlation score with the feature importance evaluation index to calculate a feature fusion weight, and using the feature fusion weight to adjust the classification cross entropy loss function to obtain a combined loss function, wherein the combined loss function includes a cross entropy loss term and a knowledge distillation constraint term;

[0036] The gradient scaling coefficient of each layer is calculated by combining the network layer depth information of the hybrid deep learning network and the weighted feature importance evaluation index; in the process of network back propagation, the partial derivative of the loss function of each layer with respect to the parameter is obtained as the original gradient, and the original gradient is corrected by using the gradient scaling coefficient to obtain the corrected gradient;

[0037] Calculate the sample sampling probability by combining the prediction confidence of the hybrid deep learning network with the feature importance evaluation index, and construct a training batch based on the sample sampling probability;

[0038] A multiple loss function is constructed based on the contrastive learning loss function and the combined loss function, and the constructed training batches are iteratively trained using the multiple loss function to update the parameters of the hybrid deep learning network until a preset training termination condition is met.

[0039] In an optional embodiment,

[0040] The prediction confidence of the hybrid deep learning network is combined with the feature importance evaluation index to calculate the sample sampling probability, and the step of constructing a training batch based on the sample sampling probability includes:

[0041] Obtain the prediction confidence of the hybrid deep learning network for the training samples, and combine the feature importance evaluation index to calculate the initial sampling probability of each training sample in a nonlinear combination manner. The nonlinear combination uses a sigmoid activation function to map the probability value to a preset interval; dynamically adjust the initial sampling probability based on an adaptive weight coefficient, the adaptive weight coefficient changes exponentially with the training rounds, and introduces a normalization coefficient to ensure the stability of the number of batch samples;

[0042] According to the prediction confidence, the training samples are divided into easy-to-classify samples, medium-level samples and difficult samples, and the difficult samples are subjected to sample enhancement processing, wherein the sample enhancement processing adopts a weighted combination of the gradient of the loss function and the importance of the features;

[0043] An initial training batch is constructed based on the dynamically adjusted initial sampling probability and the enhanced difficult samples, and a balance index of the initial training batch is determined. When the balance index does not meet the preset conditions, the training batch is reconstructed until a training batch that meets the balance requirements is constructed.

[0044] In an optional embodiment,

[0045] The steps of inputting the real-time monitoring data of the wind farm into the island state detection model, calculating the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector based on the island state detection model, determining that the wind farm has an island state when the matching degree is greater than a preset threshold, outputting an island state warning signal and triggering a safety protection mechanism of the wind farm include:

[0046] Collecting real-time monitoring data of the wind farm, the real-time monitoring data including voltage, current, active power, reactive power and phase angle parameters; performing feature extraction on the real-time monitoring data to construct a detection feature vector including voltage feature components, current feature components, active power feature components, reactive power feature components and phase angle feature components;

[0047] The detection feature vector is input into an island state detection model, and the island state detection model includes a feature matching degree calculation unit and a weight optimization unit; the feature matching degree calculation unit calculates the feature matching degree between the voltage feature component, the current feature component, the active power feature component, the reactive power feature component and the phase angle feature component and the corresponding component of the preset island feature vector respectively; the weight optimization unit adaptively adjusts the weight coefficient of each feature component based on the historical island event data, and performs weighted calculation on the feature matching degree of each feature component and the corresponding weight coefficient to obtain a comprehensive feature matching degree;

[0048] The comprehensive feature matching degree is compared with the preset island state warning threshold to determine the island state warning level of the wind farm, and corresponding protective measures are triggered according to the island state warning level, including: when the comprehensive feature matching degree is greater than the first warning level threshold, a first-level protective measure of reducing the active output of the wind turbine and increasing reactive support is triggered; when the comprehensive feature matching degree is greater than the second warning level threshold, a second-level protective measure of switching the island operation mode and emergency adjustment of voltage and frequency is triggered; when the comprehensive feature matching degree is greater than the third warning level threshold, a third-level protective measure of executing emergency shutdown and disconnecting the grid-connected switch is triggered;

[0049] The execution results of the protective measures and the island state detection results are added to the historical island event data to optimize the weight coefficient of the weight optimization unit and realize the adaptive optimization of the island state detection model.

[0050] According to a second aspect of the embodiments of the present invention,

[0051] Provides island status detection system for large-scale wind power, including:

[0052] The first unit is used to divide the wind farm into a core area and a boundary area, set an active disturbance injection unit in the boundary area, control the active disturbance injection unit to inject a small disturbance signal into the power grid, collect voltage and current data of the core area and the boundary area in real time, obtain network impedance characteristics by detecting the response characteristics of the core area to the injected disturbance, and construct a three-dimensional correlation feature matrix including network impedance characteristics, power flow characteristics and frequency response characteristics in combination with the power flow direction of the boundary area and the frequency response curve of the core area;

[0053] The second unit is used to perform feature expansion on the three-dimensional correlation feature matrix based on weather condition information in historical operation data of the wind farm to generate an expanded feature matrix; construct a training data set using the expanded feature matrix, input the training data set into a hybrid deep learning network including a convolutional neural network and a long short-term memory network, extract spatial features of the expanded feature matrix through the convolutional neural network, extract temporal features of the expanded feature matrix through the long short-term memory network, fuse the spatial features and the temporal features to obtain a fused feature vector, and train the hybrid deep learning network based on the fused feature vector and the corresponding island state label to obtain an island state detection model;

[0054] The third unit is used to input the real-time monitoring data of the wind farm into the island state detection model, calculate the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector based on the island state detection model, and determine that the wind farm has an island state when the matching degree is greater than a preset threshold, output an island state warning signal and trigger the safety protection mechanism of the wind farm.

[0055] According to a third aspect of the embodiments of the present invention,

[0056] An electronic device is provided, comprising:

[0057] processor;

[0058] a memory for storing processor-executable instructions;

[0059] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0060] A fourth aspect of the embodiments of the present invention is:

[0061] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0062] The present invention constructs a three-dimensional correlation feature matrix by combining network impedance characteristics, power flow characteristics and frequency response characteristics, and uses weather condition information to expand features, so that the features of the input deep learning model are more comprehensive and rich, and can more comprehensively reflect the characteristics of the island state of the wind farm, thereby improving the accuracy of island state detection and effectively reducing the probability of misjudgment and missed judgment.

[0063] The present invention adopts a hybrid deep learning network of convolutional neural network and long short-term memory network, which can effectively extract spatial features and temporal features and fuse them, so that the model has stronger adaptability and robustness to different working conditions and disturbances, and can maintain high detection performance even in complex power grid environments.

[0064] The present invention can realize real-time detection and early warning of island status by collecting data in real time and using a trained deep learning model for rapid calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of a method for detecting an island state of large-scale wind power according to an embodiment of the present invention;

[0066] Figure 2 It is a structural schematic diagram of an island state detection system for large-scale wind power according to an embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0069] Figure 1 FIG. 1 is a flow chart of a method for detecting an island state of large-scale wind power according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0070] S1. Divide the wind farm into a core area and a boundary area, set an active disturbance injection unit in the boundary area, control the active disturbance injection unit to inject a small disturbance signal into the power grid, collect voltage and current data of the core area and the boundary area in real time, obtain network impedance characteristics by detecting the response characteristics of the core area to the injected disturbance, and construct a three-dimensional correlation feature matrix including network impedance characteristics, power flow characteristics and frequency response characteristics by combining the power flow direction of the boundary area and the frequency response curve of the core area;

[0071] S2. Based on the weather condition information in the historical operation data of the wind farm, the three-dimensional correlation feature matrix is ​​feature expanded to generate an expanded feature matrix; a training data set is constructed using the expanded feature matrix, and the training data set is input into a hybrid deep learning network including a convolutional neural network and a long short-term memory network, the spatial features of the expanded feature matrix are extracted through the convolutional neural network, the temporal features of the expanded feature matrix are extracted through the long short-term memory network, the spatial features and the temporal features are fused to obtain a fused feature vector, and the hybrid deep learning network is trained based on the fused feature vector and the corresponding island state label to obtain an island state detection model;

[0072] S3. Input the real-time monitoring data of the wind farm into the island state detection model, and calculate the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector based on the island state detection model. When the matching degree is greater than the preset threshold, it is determined that the wind farm has an island state, and an island state warning signal is output and the safety protection mechanism of the wind farm is triggered.

[0073] Exemplarily, first, the wind farm is divided into a core area and a boundary area. The core area usually selects an area containing more wind turbines, and the boundary area is located outside the core area and connected to the external power grid. An active disturbance injection unit is configured in the boundary area, which can be a controllable power electronic device, such as an inverter. The active disturbance injection unit is controlled to inject a small disturbance signal into the power grid. The disturbance signal can be a change in frequency or amplitude. The disturbance amplitude is set to a small range of fluctuations that does not affect the stable operation of the power grid, such as a voltage amplitude change within 0.5% of the rated voltage, and a frequency change within 0.1Hz of the rated frequency. At the same time, sensors are deployed in the core area and the boundary area to collect voltage and current data in real time. Assume that a wind farm is divided into 5 areas, of which area 1 and area 2 are core areas, and area 3, area 4 and area 5 are boundary areas. A sinusoidal disturbance signal with an amplitude of 0.2% of the rated voltage is injected into boundary area 3.

[0074] Next, the response characteristics of the core area to the injected disturbance are analyzed. By comparing the changes in the voltage and current data in the core area before and after the disturbance is injected, the network impedance characteristics can be obtained. For example, the phase difference change of the voltage and current in the core area before and after the disturbance injection can be calculated to characterize the change in network impedance. At the same time, the power flow direction in the boundary area is used to determine whether the power flows to the external power grid or to the inside of the wind farm. In addition, by analyzing the curve of the change of the core area frequency over time, the frequency response characteristics of the core area can be obtained. For example, the amplitude and rate of frequency change can be analyzed. Assume that after the disturbance is injected, the voltage phase of the core area 1 lags behind the current phase by 0.1 degrees, the power flow direction of the boundary area 3 changes from flowing to the external power grid to flowing to the inside of the wind farm, and the frequency of the core area 2 drops by 0.05Hz in 0.1 seconds.

[0075] Then, the network impedance characteristics, power flow characteristics and frequency response characteristics are combined to construct a three-dimensional correlation characteristic matrix. For example, the network impedance change, power flow direction and frequency change rate can be used as the three dimensions of the three-dimensional matrix. Assuming that the network impedance change is quantified as 1, the power flow direction changes from flowing to the external power grid to flowing to the inside of the wind farm is quantified as -1, and the frequency change rate is quantified as -2, a three-dimensional matrix containing these characteristic values ​​can be constructed.

[0076] In order to improve the detection accuracy, the weather condition information in the historical operation data of the wind farm is used to expand the three-dimensional correlation feature matrix. For example, weather data such as wind speed, wind direction, and temperature can be added to the feature matrix to generate an expanded feature matrix. Assuming that the wind speed at that time was 10m / s, the wind direction was due north, and the temperature was 25℃, these data can be added to the feature matrix.

[0077] The training data set is constructed using the extended feature matrix, and the training data set is input into the hybrid deep learning network for training. The network contains a convolutional neural network (CNN) and a long short-term memory network (LSTM). CNN is used to extract the spatial features of the extended feature matrix, and LSTM is used to extract the temporal features of the extended feature matrix. The spatial features and temporal features are fused to obtain a fused feature vector. The hybrid deep learning network is trained based on the fused feature vector and the corresponding island state label to obtain an island state detection model.

[0078] In practical applications, the real-time monitoring data of the wind farm is input into the trained island state detection model, and the model will calculate the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector. The preset island feature vector can be obtained by analyzing the known island state data. When the matching degree is greater than the preset threshold, it is determined that the wind farm has an island state, and an island state warning signal is output and the safety protection mechanism of the wind farm is triggered. Assuming that the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector is 0.95, and the preset threshold is 0.9, it is determined that the wind farm has an island state.

[0079] The present invention constructs a three-dimensional correlation feature matrix including network impedance characteristics, power flow characteristics and frequency response characteristics, and expands the characteristics in combination with weather condition information, so as to more comprehensively reflect the characteristics of the islanding state of the wind farm, thereby improving the detection accuracy; the hybrid deep learning network of CNN and LSTM can effectively extract the spatial and temporal characteristics of the feature matrix, and enhance the robustness of the model to different working conditions and noise.

[0080] In an optional embodiment,

[0081] The wind farm is divided into a core area and a boundary area, an active disturbance injection unit is arranged in the boundary area, the active disturbance injection unit is controlled to inject a small disturbance signal into the power grid, voltage and current data of the core area and the boundary area are collected in real time, network impedance characteristics are obtained by detecting the response characteristics of the core area to the injected disturbance, and a three-dimensional correlation feature matrix including network impedance characteristics, power flow characteristics and frequency response characteristics is constructed in combination with the power flow direction of the boundary area and the frequency response curve of the core area. The steps include:

[0082] The wind farm is divided into a core area and a boundary area based on the electrical distance. The normalized electrical distance is obtained by calculating the ratio of the equivalent impedance from each wind turbine to the booster station to the base impedance. The wind turbine group whose normalized electrical distance is less than a preset threshold is divided into the core area, and the wind turbine group whose normalized electrical distance is greater than the preset threshold is divided into the boundary area.

[0083] A plurality of active disturbance injection units are arranged in the boundary area, wherein the active disturbance injection units include a power converter and an energy storage device, and the active disturbance injection units are controlled to inject an attenuated sinusoidal disturbance signal having a preset frequency and a preset amplitude into the power grid, and the attenuated sinusoidal disturbance signal is used as an input excitation signal;

[0084] Collecting a response voltage and a response current of the core area after receiving an input excitation signal, and calculating a network impedance characteristic of the core area according to a ratio of the response voltage to the response current, wherein the network impedance characteristic includes an impedance amplitude characteristic and a phase angle characteristic;

[0085] collecting active power and reactive power in the boundary area, and constructing a power flow vector in the boundary area, wherein the power flow vector is used to characterize the energy exchange state between the wind farm and the external power grid;

[0086] Performing wavelet transform on the frequency response curve of the core area to extract the frequency response characteristics of the core area, wherein the frequency response characteristics include scale coefficients and translation coefficients;

[0087] The network impedance characteristics, power flow vector and frequency response characteristics are combined to construct a three-dimensional correlation feature matrix.

[0088] For example, in order to more accurately obtain the operating status of a wind farm, a wind farm state perception method based on a three-dimensional correlation feature matrix is ​​proposed. This method injects a small disturbance signal at the boundary of the wind farm, combines the impedance response of the core area of ​​the wind farm, the power flow in the boundary area, and the frequency response of the core area, and constructs a three-dimensional correlation feature matrix, thereby achieving comprehensive perception of the operating status of the wind farm.

[0089] First, the wind farm is divided into a core area and a boundary area. The equivalent impedance from each wind turbine to the booster station is calculated, and then the equivalent impedance of each wind turbine is compared with the preset reference impedance to obtain the normalized electrical distance. A threshold is set, such as 0.8. The wind turbine group with a normalized electrical distance less than 0.8 is defined as the core area, and the wind turbine group with a normalized electrical distance greater than 0.8 is defined as the boundary area. For example, there are 100 wind turbines in a wind farm. After calculation, the normalized electrical distance of 70 wind turbines is less than 0.8, which is divided into the core area; the normalized electrical distance of 30 wind turbines is greater than 0.8, which is divided into the boundary area.

[0090] Next, an active disturbance injection unit is set in the boundary area. Each active disturbance injection unit consists of a power converter and an energy storage device. The active disturbance injection unit is controlled to inject a decaying sinusoidal disturbance signal with a preset frequency and amplitude into the power grid. For example, the frequency of the injection signal is set to 5 Hz, the amplitude is 0.5% of the rated voltage, and the decay time constant is 0.1 second. This disturbance signal is used as the input excitation signal.

[0091] Then, the response voltage and response current of the core area after receiving the input excitation signal are collected. The network impedance characteristics of the core area are calculated by the ratio of the response voltage and the response current, including the impedance amplitude characteristics and the phase angle characteristics. For example, 100 data points are collected in one cycle, and the impedance amplitude and phase angle sequence are calculated.

[0092] At the same time, the active power and reactive power data of the boundary area are collected. The collected active power and reactive power data are combined into a vector to construct a power flow vector of the boundary area, which is used to characterize the energy exchange state between the wind farm and the external power grid. For example, 100 active power and reactive power data points are collected in one cycle to form a 2×100 power flow vector.

[0093] In addition, the frequency response curve of the core area is subjected to wavelet transformation to extract the frequency response characteristics of the core area, including the scale coefficient and the translation coefficient. For example, the frequency response curve is decomposed into 5 layers using the Daubechies wavelet basis function to obtain 5 groups of scale coefficients and translation coefficients.

[0094] Finally, the network impedance features (impedance amplitude and phase angle), power flow vector (active power and reactive power), and frequency response features (scaling coefficient and translation coefficient) obtained in the previous steps are combined to construct a three-dimensional correlation feature matrix. The dimension of this matrix is ​​determined by the dimension of each feature. For example, assuming that the impedance feature dimension is 2×100, the power flow vector dimension is 2×100, and the frequency response feature dimension is 10×100, the dimension of the final three-dimensional correlation feature matrix is ​​22×100.

[0095] The present invention improves the accuracy of wind farm state perception: by combining multiple feature information to construct a three-dimensional correlation feature matrix, it can more comprehensively reflect the operating state of the wind farm and avoid the one-sidedness caused by single feature information, thereby improving the accuracy of state perception; enhances the wind farm fault diagnosis capability: the three-dimensional correlation feature matrix contains rich wind farm operation information, which can be used to identify and diagnose various types of wind farm faults, such as line faults, wind turbine faults, etc., thereby improving the reliability and stability of the wind farm; improves the wind farm operation efficiency: through accurate perception of the wind farm operation state, the wind farm control strategy can be optimized, such as adjusting the wind turbine output, thereby improving the wind farm operation efficiency and reducing the operation cost.

[0096] In an optional embodiment,

[0097] Based on the weather condition information in the historical operation data of the wind farm, the three-dimensional correlation feature matrix is ​​feature expanded to generate an expanded feature matrix; a training data set is constructed using the expanded feature matrix, and the training data set is input into a hybrid deep learning network including a convolutional neural network and a long short-term memory network, and the spatial features of the expanded feature matrix are extracted by the convolutional neural network, and the temporal features of the expanded feature matrix are extracted by the long short-term memory network, and the spatial features and the temporal features are fused to obtain a fused feature vector, and the steps of training the hybrid deep learning network based on the fused feature vector and the corresponding island state label to obtain an island state detection model include:

[0098] Collecting weather condition information from historical operation data of the wind farm, the weather condition information including wind speed, temperature, humidity and air pressure; normalizing the weather condition information, aligning the normalized weather condition information with the three-dimensional correlation feature matrix in the time dimension, and generating an expanded feature matrix;

[0099] Constructing a hybrid deep learning network, wherein the hybrid deep learning network includes a convolutional neural network and a long short-term memory network, wherein the convolutional neural network includes a plurality of convolutional layers, a normalization layer, and an activation function layer, and the long short-term memory network includes a plurality of memory layers;

[0100] The expanded feature matrix is ​​input into a convolutional neural network, and the spatial feature vector of the expanded feature matrix is ​​extracted through the multiple convolutional layers. The output of each convolutional layer is subjected to feature enhancement through a normalization layer and an activation function layer to obtain a feature sequence after spatial dimension compression;

[0101] Inputting the feature sequence compressed by the spatial dimension into a long short-term memory network, extracting the temporal feature vector of the feature sequence through the multiple memory layers, wherein the memory layer includes a forget gate, an input gate, and an output gate;

[0102] The spatial feature vector and the temporal feature vector are concatenated to obtain a fused feature vector, and the fused feature vector is subjected to feature dimension reduction and anti-overfitting processing through a fully connected layer and a random dropout layer; based on the fused feature vector and a preset island state label, the hybrid deep learning network is trained using a multiple loss function to obtain an island state detection model.

[0103] For example, historical operation data of a wind farm is collected to obtain weather information, including wind speed, temperature, humidity, and air pressure. For example, data can be collected every 10 minutes for a month to obtain a data set containing four types of weather information, including wind speed, temperature, humidity, and air pressure, as well as a timestamp. Assume that the data collected at a certain time point is: wind speed 15m / s, temperature 25℃, humidity 70%, and air pressure 101kPa.

[0104] Normalize the collected weather information. The purpose of normalization is to eliminate the influence of different feature dimensions. Common normalization methods include minimum-maximum normalization and standardization. For example, the minimum-maximum normalization method is used to normalize the wind speed to between 0 and 1, the temperature to between 0 and 1, and the humidity and air pressure are also processed in the same way. Assume that after normalization, the above data becomes: wind speed 0.6, temperature 0.8, humidity 0.7, air pressure 0.95.

[0105] Construct a three-dimensional correlation feature matrix. This matrix contains the correlation between the wind turbines in the wind farm and their changes over time. For example, the power correlation coefficients between the wind turbines can be calculated based on their geographical locations and historical operating data, and these coefficients can be arranged in chronological order to form a three-dimensional matrix. Assume that the matrix dimensions are (time step, number of generators, number of generators), where the time step is 1 month and the number of generators is 10.

[0106] The normalized weather condition information is aligned with the three-dimensional correlation feature matrix in the time dimension to generate an expanded feature matrix. Specifically, the weather condition information of each time step is copied multiple times so that the number is equal to the square of the number of generator sets, and then these weather condition information are added as new dimensions to each layer of the three-dimensional correlation feature matrix. In this way, the dimension of the expanded feature matrix becomes (time step, number of generator sets, number of generator sets, number of weather condition information). For example, the above normalized weather condition information (0.6, 0.8, 0.7, 0.95) is copied 100 times and added to each time step of the three-dimensional correlation feature matrix to generate an expanded feature matrix.

[0107] Construct a hybrid deep learning network that includes a convolutional neural network (CNN) and a long short-term memory network (LSTM). CNN is used to extract spatial features of the expanded feature matrix, and LSTM is used to extract temporal features. The CNN part includes multiple convolutional layers, normalization layers, and activation function layers. For example, three convolutional layers can be used, each followed by a batch normalization layer and a ReLU activation function layer. The LSTM part includes multiple memory layers, each of which includes a forget gate, an input gate, and an output gate. For example, two LSTM layers can be used.

[0108] The expanded feature matrix is ​​input into CNN, and the spatial feature vector is extracted through multiple convolutional layers. The output of each convolutional layer is enhanced through the normalization layer and activation function layer to obtain a feature sequence with compressed spatial dimension.

[0109] The feature sequence compressed in the spatial dimension is input into LSTM, and the time series feature vector is extracted through multiple memory layers.

[0110] The spatial feature vector extracted by CNN and the temporal feature vector extracted by LSTM are concatenated to obtain a fused feature vector.

[0111] The fused feature vector is processed for feature dimension reduction and anti-overfitting through a fully connected layer and a random dropout layer. For example, a fully connected layer and a random dropout layer can be used, and the random dropout rate is set to 0.5.

[0112] Based on the fused feature vector and the preset island state label, a hybrid deep learning network is trained using multiple loss functions to obtain an island state detection model. For example, a weighted sum of a cross entropy loss function and a mean square error loss function can be used as the loss function.

[0113] By integrating weather condition information and the correlation between wind turbines, the present invention can more comprehensively characterize the operating status of wind farms, thereby improving the accuracy of island status detection; the deep learning model can automatically learn features and reduce dependence on artificial feature engineering, thereby enhancing the robustness of the model; the hybrid deep learning network can process large amounts of data in parallel, thereby improving the efficiency of island status detection.

[0114] In an optional embodiment,

[0115] The steps of constructing a hybrid deep learning network, wherein the hybrid deep learning network includes a convolutional neural network and a long short-term memory network, wherein the convolutional neural network includes a plurality of convolutional layers, a normalization layer, and an activation function layer, and wherein the long short-term memory network includes a plurality of memory layers include:

[0116] In a convolutional neural network, the importance weights of different convolutional layers are adaptively calculated by a global context encoder, and the global context encoder generates a channel attention map based on the statistical information of the input features, and the channel attention map is used to adjust the fusion weights of the output features of different convolutional layers;

[0117] A feature interaction module is constructed between convolutional layers. The feature interaction module includes a feature resampling unit and an adaptive fusion unit. Dynamic information exchange of output features of different convolutional layers is realized through a transformation matrix. The parameters of the transformation matrix are updated through a back propagation algorithm. The update step is proportional to the feature importance. The adaptive fusion unit automatically adjusts the information exchange intensity according to the feature correlation.

[0118] A bidirectional memory path is set in the long short-term memory network, the memory layers of the forward path and the reverse path share parameters, and the bidirectional timing information is dynamically integrated through a state fusion gate, and the state fusion gate calculates the fusion weight based on the current input and the historical state;

[0119] A time-aware mechanism is embedded in the memory layer to encode the time intervals between adjacent moments into a continuous vector, and the continuous vector is integrated into the calculation of the forget gate and the input gate. The time interval encoding is generated by a learnable time embedding layer.

[0120] Exemplarily, first, construct the convolutional neural network part. This part contains multiple convolutional layers, normalization layers, and activation function layers. Each convolutional layer uses a convolution kernel to extract local features of the input data, the normalization layer standardizes the features, and the activation function layer introduces nonlinearity to enhance the expressiveness of the model. Taking image data as an example, input a 224x224x3 color image. The first convolutional layer uses 64 convolution kernels of size 3x3, a step size of 1, and a padding of 1 to extract 64 feature maps. Subsequent convolutional layers can adjust parameters such as the number, size, step size, and padding of convolution kernels as needed.

[0121] In order to adaptively fuse the output features of different convolutional layers, a global context encoder is introduced. The encoder generates a channel attention map based on the statistical information of the input features (such as the mean and standard deviation). The channel attention map reflects the importance of different channels and is used to adjust the fusion weights of the output features of different convolutional layers. For example, if the output feature of a convolutional layer is given a higher weight in the global context encoder, then this feature will play a greater role in the subsequent fusion process.

[0122] A feature interaction module is built between convolutional layers to facilitate information exchange between features at different levels. The feature interaction module includes a feature resampling unit and an adaptive fusion unit. The feature resampling unit implements dynamic information exchange between output features of different convolutional layers through a transformation matrix, such as adjusting feature maps to the same size using upsampling or downsampling operations. The parameters of the transformation matrix are updated through a back-propagation algorithm, and the update step is proportional to the importance of the feature, that is, the more important the feature, the larger the parameter update amplitude. The adaptive fusion unit automatically adjusts the information exchange intensity according to the feature correlation. For example, the higher the correlation between two feature maps, the greater the information exchange intensity between them.

[0123] Next, the long short-term memory network part is constructed. This part contains multiple memory layers for processing time series information. In order to better capture the bidirectional dependencies of time series information, a bidirectional memory path is set, namely the forward path and the reverse path. The forward path processes information from the past to the present, and the reverse path processes information from the future to the present. The memory layers of the forward path and the reverse path share parameters to reduce the number of model parameters. The bidirectional time series information is dynamically integrated through the state fusion gate, which calculates the fusion weight based on the current input and the historical state. For example, if the current input is more relevant to the historical state of the forward path, the state fusion gate assigns a higher weight to the forward path.

[0124] A time-aware mechanism is embedded in the memory layer to better handle time series data with uneven time intervals. The time intervals between adjacent moments are encoded as a continuous vector and the continuous vector is integrated into the calculation of the forget gate and the input gate. The time interval encoding is generated by a learnable time embedding layer. For example, a vector with a time interval of 1 hour is represented as [0.1, 0.2, 0.3], and a vector with a time interval of 2 hours is represented as [0.2, 0.4, 0.6].

[0125] Finally, the output of the convolutional neural network is used as the input of the long short-term memory network to obtain the final hybrid deep learning network.

[0126] Through the global context encoder and feature interaction module, the present invention can adaptively fuse the output features of different convolutional layers and promote information exchange of features at different levels, thereby improving the expressiveness of features; through the bidirectional memory path and time perception mechanism, the bidirectional dependency of timing information and the influence of time intervals can be better captured, thereby enhancing the timing modeling capability; by combining the advantages of convolutional neural networks and long short-term memory networks, data with spatiotemporal characteristics can be effectively processed, and better performance can be achieved in various tasks.

[0127] In an optional embodiment,

[0128] Based on the fused feature vector and the preset island state label, the steps of training the hybrid deep learning network using multiple loss functions to obtain the island state detection model include:

[0129] Based on the fused feature vector and the preset island state label, a feature importance evaluation index is calculated through feature mutual information calculation, feature correlation analysis, and feature prediction gain evaluation; a dynamic feature weight is generated through a softmax function with adjustable smoothness, and the dynamic feature weight is used to weight the feature importance evaluation index;

[0130] Performing time series enhancement processing on the input data according to the feature importance evaluation index to obtain an enhanced feature sequence, selecting time series-related samples from the enhanced feature sequence to construct positive sample pairs, and using a contrastive learning loss function to calculate the cosine similarity between the positive sample pairs, wherein the contrastive learning loss function includes a positive sample pair similarity calculation item and a negative sample pair similarity calculation item;

[0131] Constructing a multi-layer attention network to extract the correlation between features to obtain a feature correlation score, combining the feature correlation score with the feature importance evaluation index to calculate a feature fusion weight, and using the feature fusion weight to adjust the classification cross entropy loss function to obtain a combined loss function, wherein the combined loss function includes a cross entropy loss term and a knowledge distillation constraint term;

[0132] The gradient scaling coefficient of each layer is calculated by combining the network layer depth information of the hybrid deep learning network and the weighted feature importance evaluation index; in the process of network back propagation, the partial derivative of the loss function of each layer with respect to the parameter is obtained as the original gradient, and the original gradient is corrected by using the gradient scaling coefficient to obtain the corrected gradient;

[0133] Calculate the sample sampling probability by combining the prediction confidence of the hybrid deep learning network with the feature importance evaluation index, and construct a training batch based on the sample sampling probability;

[0134] A multiple loss function is constructed based on the contrastive learning loss function and the combined loss function, and the constructed training batches are iteratively trained using the multiple loss function to update the parameters of the hybrid deep learning network until a preset training termination condition is met.

[0135] For example, in order to build a high-precision island state detection model, an island detection method based on a hybrid deep learning network and multiple loss functions is proposed. This method integrates multiple features and optimizes the model training process through strategies such as dynamic feature weighting, temporal enhancement, contrastive learning, multi-layer attention mechanism, gradient correction and sample sampling, thereby improving the generalization ability and detection accuracy of the model.

[0136] First, various operating data in the power system are collected, such as voltage, current, frequency, active power, reactive power, etc., and these data are preprocessed, such as data cleaning and normalization, to obtain the initial feature set. Then, according to the actual operating status of the power system, the island status label at each moment is manually marked, such as "island" or "normal".

[0137] Next, the feature importance evaluation index between the fused feature vector and the preset island state label is calculated. Specifically, three methods, namely feature mutual information calculation, feature correlation analysis and feature prediction gain evaluation, are used to calculate the three feature importance evaluation indexes. For example, the higher the mutual information value between a feature and the island state label, the higher the importance of the feature. Then, a softmax function with adjustable smoothness is used to generate dynamic feature weights based on these three feature importance evaluation indicators. For example, assuming that the three indicators are 0.8, 0.5, and 0.2 respectively, after the softmax function, the corresponding weights may be 0.6, 0.3, and 0.1.

[0138] Then, the time series enhancement processing is performed according to the feature importance evaluation index. For example, if the importance of a certain feature is high, the change trend of the feature in the time series will be considered more when constructing the training sample. Assuming that the value of an important feature at time t is x(t), x(t), x(t+1) can be used as an enhanced feature sequence. Select time-related samples from the enhanced feature sequence to construct positive sample pairs. For example, (x(t-1), x(t)) and (x(t), x(t+1)) are used as a positive sample pair. Then, the contrastive learning loss function is used to calculate the cosine similarity between the positive sample pairs. The loss function contains the similarity calculation items of the positive sample pairs and the negative sample pairs, and aims to maximize the similarity between the positive sample pairs and minimize the similarity between the negative sample pairs. For example, assuming that there are two positive sample pairs with cosine similarities of 0.9 and 0.7 respectively, and a negative sample pair with a cosine similarity of 0.2, the value of the contrastive learning loss function will be relatively small.

[0139] At the same time, a multi-layer attention network is constructed to extract the correlation between features and obtain the feature correlation score. For example, if a feature has a high correlation with multiple other features, the correlation score of the feature will be relatively high. The feature correlation score is combined with the feature importance evaluation index to calculate the feature fusion weight. For example, if a feature is both important and highly correlated with other features, its fusion weight will be relatively high. The classification cross entropy loss function is adjusted using the feature fusion weight to obtain a combined loss function. The combined loss function includes a cross entropy loss term and a knowledge distillation constraint term. For example, assuming that the true label of a sample is "island" and the model predicts a probability of 0.8, the cross entropy loss will be relatively small.

[0140] In addition, the gradient scaling coefficient of each layer is calculated by combining the network layer depth information of the hybrid deep learning network and the weighted feature importance evaluation index. During the network back propagation process, the partial derivative of the loss function of each layer with respect to the parameter is obtained as the original gradient, and the original gradient is corrected using the gradient scaling coefficient to obtain the corrected gradient. For example, if the gradient scaling coefficient of a layer is large, the original gradient of the layer will be amplified.

[0141] Finally, the prediction confidence of the hybrid deep learning network is combined with the feature importance evaluation index to calculate the sample sampling probability. For example, if the prediction confidence of a sample is high and the feature importance is high, the probability of the sample being sampled will be relatively high. Construct a training batch based on the sample sampling probability. Construct a multiple loss function based on the contrastive learning loss function and the combined loss function, and use the multiple loss functions to iteratively train the constructed training batches and update the parameters of the hybrid deep learning network until the preset training termination conditions are met, such as reaching the maximum number of iterations or the loss function converges.

[0142] For example, suppose there are three features: voltage, current, and frequency, and their importance evaluation indexes are 0.8, 0.6, and 0.4, respectively. After the softmax function, the corresponding dynamic feature weights are 0.5, 0.3, and 0.2, respectively. The voltage value at time t is 1.0, 0.9 at time t-1, and 1.1 at time t+1. Construct positive sample pairs (0.9, 1.0) and (1.0, 1.1). Suppose the feature association scores extracted by the multi-layer attention network are 0.7, 0.5, and 0.3, respectively. Combining the importance evaluation index and the association score, the feature fusion weights are calculated to be 0.6, 0.4, and 0.2, respectively. Suppose the true label of a sample is "island" and the model predicts a probability of 0.9, then the cross entropy loss will be small. Suppose the gradient scale factor of a layer is 1.2, then the original gradient of the layer will be amplified by 1.2 times. Suppose the prediction confidence of a sample is 0.8 and the feature importance evaluation index is 0.7, then the probability of the sample being sampled will be high.

[0143] By integrating multiple features, dynamic feature weighting and multi-layer attention mechanism, the present invention can more comprehensively capture the operating status information of the power system, thereby improving the accuracy of island detection; through strategies such as timing enhancement, contrastive learning and gradient correction, the robustness of the model to noise and abnormal data can be enhanced, and the generalization ability of the model can be improved; through sample sampling strategy, the training time can be reduced and the training efficiency can be improved.

[0144] In an optional embodiment,

[0145] The prediction confidence of the hybrid deep learning network is combined with the feature importance evaluation index to calculate the sample sampling probability, and the step of constructing a training batch based on the sample sampling probability includes:

[0146] Obtain the prediction confidence of the hybrid deep learning network for the training samples, and combine the feature importance evaluation index to calculate the initial sampling probability of each training sample in a nonlinear combination manner. The nonlinear combination uses a sigmoid activation function to map the probability value to a preset interval; dynamically adjust the initial sampling probability based on an adaptive weight coefficient, the adaptive weight coefficient changes exponentially with the training rounds, and introduces a normalization coefficient to ensure the stability of the number of batch samples;

[0147] According to the prediction confidence, the training samples are divided into easy-to-classify samples, medium-level samples and difficult samples, and the difficult samples are subjected to sample enhancement processing, wherein the sample enhancement processing adopts a weighted combination of the gradient of the loss function and the importance of the features;

[0148] An initial training batch is constructed based on the dynamically adjusted initial sampling probability and the enhanced difficult samples, and a balance index of the initial training batch is determined. When the balance index does not meet the preset conditions, the training batch is reconstructed until a training batch that meets the balance requirements is constructed.

[0149] In order to clearly explain a method for constructing training batches based on prediction confidence and feature importance, the specific implementation method and technical details are described in detail below, and data examples are provided.

[0150] First, obtain the prediction confidence of the hybrid deep learning network for the training samples. Suppose we have an image classification task, and use a hybrid deep learning network to classify 1,000 image samples to obtain the probability that each sample belongs to each category. For example, for the first sample, the network predicts that the probability of it belonging to category A is 0.9, the probability of it belonging to category B is 0.05, and the probability of it belonging to category C is 0.05. This 0.9 is the prediction confidence of the sample.

[0151] Next, we combine the feature importance evaluation index. Assume that we obtain the importance score of each feature through some method (such as gradient back propagation). For example, for image classification tasks, features can be pixels in the image or higher-level features such as edges, textures, etc. Assume that we obtain the importance score of each pixel, ranging from 0 to 1.

[0152] Then, the initial sampling probability of each training sample is calculated by nonlinear combination. The prediction confidence and feature importance are nonlinearly combined, for example, the confidence is multiplied by the average value of the feature importance. Then, the combined value is mapped to a preset interval, such as between 0.1 and 0.9, using the sigmoid activation function. Assuming that the confidence of the first sample is 0.9 and the average value of the feature importance is 0.8, the combined value is 0.72, and after sigmoid mapping, the initial sampling probability is 0.67.

[0153] The initial sampling probability is dynamically adjusted based on the adaptive weight coefficient. The adaptive weight coefficient is set to change exponentially with the training round. For example, the initial weight coefficient is 0.5, and it is multiplied by a decay factor, such as 0.95, after each round of training. A normalization coefficient is introduced to ensure that the number of batch samples is stable. For example, the sampling probability of all samples is divided by their sum so that the sum is 1. Assuming the training batch size is 64, the sampling probability of each sample is multiplied by 64.

[0154] The training samples are divided into easy-to-classify samples, medium-level samples, and difficult samples according to the prediction confidence. For example, samples with a confidence greater than 0.8 are easy-to-classify samples, samples with a confidence between 0.5 and 0.8 are medium-level samples, and samples with a confidence less than 0.5 are difficult samples.

[0155] Perform sample enhancement on difficult samples. Sample enhancement is performed by using a weighted combination of the loss function gradient and feature importance. For example, random rotation, translation, scaling, etc. are performed on difficult samples, and the magnitude of the operation is proportional to the loss function gradient and feature importance. Assuming that the loss function gradient of a difficult sample is large and some features are of high importance, the sample is rotated and translated to a greater extent.

[0156] The initial training batch is constructed based on the dynamically adjusted initial sampling probability and the enhanced difficult samples. For example, the roulette algorithm is used to randomly sample according to the sampling probability of each sample to form the initial training batch.

[0157] Determine the balance index of the initial training batch. The balance index can be the ratio of the number of samples in each category. For example, if the goal is to have roughly the same number of samples in each category, the variance of the number of samples in each category can be calculated. When the balance index does not meet the preset conditions, such as the variance is greater than a certain threshold, the training batch is rebuilt until a training batch that meets the balance requirements is built.

[0158] Data case: Assume there are 10 samples, the prediction confidences are [0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1, 0.05], the feature importance is 0.8, the initial weight coefficient is 0.5, the decay factor is 0.95, and the training batch size is 4. After the above steps, the sampling probability of each sample can be obtained, and the training batch can be constructed accordingly.

[0159] The present invention can reduce training time and improve training efficiency by giving priority to samples with large amounts of information for training; by enhancing difficult samples, the model's recognition ability for difficult samples can be improved, thereby improving the model's generalization ability; by dynamically adjusting the sampling probability and the control of the balance index, the balance of the training batch can be ensured, avoiding overfitting of the model to certain categories.

[0160] In an optional embodiment,

[0161] The steps of inputting the real-time monitoring data of the wind farm into the island state detection model, calculating the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector based on the island state detection model, determining that the wind farm has an island state when the matching degree is greater than a preset threshold, outputting an island state warning signal and triggering a safety protection mechanism of the wind farm include:

[0162] Collecting real-time monitoring data of the wind farm, the real-time monitoring data including voltage, current, active power, reactive power and phase angle parameters; performing feature extraction on the real-time monitoring data to construct a detection feature vector including voltage feature components, current feature components, active power feature components, reactive power feature components and phase angle feature components;

[0163] The detection feature vector is input into an island state detection model, and the island state detection model includes a feature matching degree calculation unit and a weight optimization unit; the feature matching degree calculation unit calculates the feature matching degree between the voltage feature component, the current feature component, the active power feature component, the reactive power feature component and the phase angle feature component and the corresponding component of the preset island feature vector respectively; the weight optimization unit adaptively adjusts the weight coefficient of each feature component based on the historical island event data, and performs weighted calculation on the feature matching degree of each feature component and the corresponding weight coefficient to obtain a comprehensive feature matching degree;

[0164] The comprehensive feature matching degree is compared with the preset island state warning threshold to determine the island state warning level of the wind farm, and corresponding protective measures are triggered according to the island state warning level, including: when the comprehensive feature matching degree is greater than the first warning level threshold, a first-level protective measure of reducing the active output of the wind turbine and increasing reactive support is triggered; when the comprehensive feature matching degree is greater than the second warning level threshold, a second-level protective measure of switching the island operation mode and emergency adjustment of voltage and frequency is triggered; when the comprehensive feature matching degree is greater than the third warning level threshold, a third-level protective measure of executing emergency shutdown and disconnecting the grid-connected switch is triggered;

[0165] The execution results of the protective measures and the island state detection results are added to the historical island event data to optimize the weight coefficient of the weight optimization unit and realize the adaptive optimization of the island state detection model.

[0166] For example, real-time monitoring data of a wind farm is collected. Through sensors and data acquisition systems, parameters such as voltage, current, active power, reactive power, and phase angle of the wind farm are collected in real time. For example, the data collected at a certain moment is: voltage 10kV, current 100A, active power 1MW, reactive power 0.5MVar, phase angle 30°. These data will be stored and used for subsequent feature extraction and analysis.

[0167] The collected real-time monitoring data is feature extracted to construct a detection feature vector. Parameters such as voltage, current, active power, reactive power and phase angle are analyzed to extract characteristic components that can reflect the island state. For example, the fluctuation range of voltage amplitude, current harmonic content, active power drop rate, reactive power change trend and phase angle offset degree can be extracted as characteristic components. Assume that the extracted characteristic vector is: [0.1, 0.2, 0.3, 0.4, 0.5], which represent the characteristic components of voltage, current, active power, reactive power and phase angle respectively.

[0168] The extracted detection feature vector is input into the island state detection model. The model includes a feature matching degree calculation unit and a weight optimization unit. The feature matching degree calculation unit calculates the matching degree between each feature component in the detection feature vector and the corresponding component of the preset island feature vector. The preset island feature vector is pre-set based on historical island event data, for example: [0.15, 0.25, 0.35, 0.45, 0.55]. The matching degree can be calculated in a variety of ways, such as calculating the absolute value or relative value of the difference between the two components.

[0169] The weight optimization unit adaptively adjusts the weight coefficient of each characteristic component based on the historical island event data. For example, if the historical data shows that the voltage characteristic component is more important for the judgment of the island state, it is given a higher weight coefficient. Assume that the initial weight coefficient is: [0.2, 0.2, 0.2, 0.2, 0.2]. The weight optimization unit adjusts the weight coefficient according to the historical data, for example, the adjusted weight coefficient is: [0.3, 0.1, 0.2, 0.2, 0.2]. Then, the feature matching degree of each characteristic component is weightedly calculated with the corresponding weight coefficient to obtain the comprehensive feature matching degree. Assume that the calculated comprehensive feature matching degree is 0.8.

[0170] The calculated comprehensive feature matching degree is compared with the preset island state warning threshold to determine the island state warning level of the wind farm. Assume that the preset three-level warning thresholds are 0.6, 0.8 and 0.9 respectively. Since the comprehensive feature matching degree of 0.8 is greater than the first warning level threshold of 0.6 and less than the third warning level threshold of 0.9, and equal to the second warning level threshold of 0.8, the secondary protection measures are triggered, i.e. switching the island operation mode and emergency voltage and frequency adjustment.

[0171] The execution results of the protection measures and the islanding state detection results are added to the historical islanding event data to optimize the weight coefficient of the weight optimization unit and realize the adaptive optimization of the islanding state detection model. For example, the execution of the secondary protection measures and whether the wind farm successfully avoided the islanding accident are recorded. These data will be used to update the weight coefficient, for example, increase the weight coefficient of the voltage characteristic component to 0.4, to improve the accuracy of future islanding state detection.

[0172] The present invention can more accurately identify the islanding state of a wind farm and reduce the occurrence of misjudgment and missed judgment through feature extraction and matching degree calculation; based on the comparison of the comprehensive feature matching degree with the preset threshold, a graded early warning of the islanding state can be achieved, and corresponding protective measures can be triggered according to the warning level to improve the safety and stability of the wind farm; by adding the execution results of the protective measures and the islanding state detection results to the historical data, the weight coefficient can be continuously optimized to improve the adaptability and robustness of the model.

[0173] Figure 2 FIG. 1 is a schematic diagram of a structure of an island state detection system for large-scale wind power according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0174] The first unit is used to divide the wind farm into a core area and a boundary area, set an active disturbance injection unit in the boundary area, control the active disturbance injection unit to inject a small disturbance signal into the power grid, collect voltage and current data of the core area and the boundary area in real time, obtain network impedance characteristics by detecting the response characteristics of the core area to the injected disturbance, and construct a three-dimensional correlation feature matrix including network impedance characteristics, power flow characteristics and frequency response characteristics in combination with the power flow direction of the boundary area and the frequency response curve of the core area;

[0175] The second unit is used to perform feature expansion on the three-dimensional correlation feature matrix based on weather condition information in historical operation data of the wind farm to generate an expanded feature matrix; construct a training data set using the expanded feature matrix, input the training data set into a hybrid deep learning network including a convolutional neural network and a long short-term memory network, extract spatial features of the expanded feature matrix through the convolutional neural network, extract temporal features of the expanded feature matrix through the long short-term memory network, fuse the spatial features and the temporal features to obtain a fused feature vector, and train the hybrid deep learning network based on the fused feature vector and the corresponding island state label to obtain an island state detection model;

[0176] The third unit is used to input the real-time monitoring data of the wind farm into the island state detection model, calculate the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector based on the island state detection model, and determine that the wind farm has an island state when the matching degree is greater than a preset threshold, output an island state warning signal and trigger the safety protection mechanism of the wind farm.

[0177] According to a third aspect of the embodiments of the present invention,

[0178] An electronic device is provided, comprising:

[0179] processor;

[0180] a memory for storing processor-executable instructions;

[0181] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0182] A fourth aspect of the embodiments of the present invention is:

[0183] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0184] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting islanding status of large-scale wind power, characterized in that: include: The wind farm is divided into a core area and a boundary area, an active disturbance injection unit is arranged in the boundary area, the active disturbance injection unit is controlled to inject a small disturbance signal into the power grid, the voltage and current data of the core area and the boundary area are collected in real time, the network impedance characteristics are obtained by detecting the response characteristics of the core area to the injected disturbance, and a three-dimensional correlation characteristic matrix including network impedance characteristics, power flow characteristics and frequency response characteristics is constructed by combining the power flow direction of the boundary area and the frequency response curve of the core area; Based on the weather condition information in the historical operation data of the wind farm, feature expansion is performed on the three-dimensional correlation feature matrix to generate an expanded feature matrix; A training data set is constructed using an expanded feature matrix, and the training data set is input into a hybrid deep learning network including a convolutional neural network and a long short-term memory network, and spatial features of the expanded feature matrix are extracted through the convolutional neural network, and temporal features of the expanded feature matrix are extracted through the long short-term memory network, and the spatial features and the temporal features are fused to obtain a fused feature vector, and the hybrid deep learning network is trained based on the fused feature vector and the corresponding island state label to obtain an island state detection model; The real-time monitoring data of the wind farm is input into the island state detection model, and the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector is calculated based on the island state detection model. When the matching degree is greater than a preset threshold, it is determined that the wind farm has an island state, an island state warning signal is output, and a safety protection mechanism of the wind farm is triggered.

2. The method according to claim 1, characterized in that: The wind farm is divided into a core area and a boundary area, an active disturbance injection unit is arranged in the boundary area, the active disturbance injection unit is controlled to inject a small disturbance signal into the power grid, voltage and current data of the core area and the boundary area are collected in real time, network impedance characteristics are obtained by detecting the response characteristics of the core area to the injected disturbance, and a three-dimensional correlation feature matrix including network impedance characteristics, power flow characteristics and frequency response characteristics is constructed in combination with the power flow direction of the boundary area and the frequency response curve of the core area. The steps include: The wind farm is divided into a core area and a boundary area based on the electrical distance. The normalized electrical distance is obtained by calculating the ratio of the equivalent impedance from each wind turbine to the booster station to the base impedance. The wind turbine group whose normalized electrical distance is less than a preset threshold is divided into the core area, and the wind turbine group whose normalized electrical distance is greater than the preset threshold is divided into the boundary area. A plurality of active disturbance injection units are arranged in the boundary area, wherein the active disturbance injection units include a power converter and an energy storage device, and the active disturbance injection units are controlled to inject an attenuated sinusoidal disturbance signal having a preset frequency and a preset amplitude into the power grid, and the attenuated sinusoidal disturbance signal is used as an input excitation signal; Collecting a response voltage and a response current of the core area after receiving an input excitation signal, and calculating a network impedance characteristic of the core area according to a ratio of the response voltage to the response current, wherein the network impedance characteristic includes an impedance amplitude characteristic and a phase angle characteristic; collecting active power and reactive power in the boundary area, and constructing a power flow vector in the boundary area, wherein the power flow vector is used to characterize the energy exchange state between the wind farm and the external power grid; Performing wavelet transform on the frequency response curve of the core area to extract the frequency response characteristics of the core area, wherein the frequency response characteristics include scale coefficients and translation coefficients; The network impedance characteristics, power flow vector and frequency response characteristics are combined to construct a three-dimensional correlation feature matrix.

3. The method according to claim 1, characterized in that Based on the weather condition information in the historical operation data of the wind farm, feature expansion is performed on the three-dimensional correlation feature matrix to generate an expanded feature matrix; The steps of constructing a training data set by using an extended feature matrix, inputting the training data set into a hybrid deep learning network including a convolutional neural network and a long short-term memory network, extracting spatial features of the extended feature matrix by using the convolutional neural network, extracting temporal features of the extended feature matrix by using the long short-term memory network, fusing the spatial features and the temporal features to obtain a fused feature vector, and training the hybrid deep learning network based on the fused feature vector and the corresponding island state label to obtain an island state detection model include: Collecting weather condition information from historical operation data of the wind farm, the weather condition information including wind speed, temperature, humidity and air pressure; normalizing the weather condition information, aligning the normalized weather condition information with the three-dimensional correlation feature matrix in the time dimension, and generating an expanded feature matrix; Constructing a hybrid deep learning network, wherein the hybrid deep learning network includes a convolutional neural network and a long short-term memory network, wherein the convolutional neural network includes a plurality of convolutional layers, a normalization layer, and an activation function layer, and the long short-term memory network includes a plurality of memory layers; The expanded feature matrix is ​​input into a convolutional neural network, and the spatial feature vector of the expanded feature matrix is ​​extracted through the multiple convolutional layers. The output of each convolutional layer is subjected to feature enhancement through a normalization layer and an activation function layer to obtain a feature sequence after spatial dimension compression; Inputting the feature sequence compressed by the spatial dimension into a long short-term memory network, extracting the temporal feature vector of the feature sequence through the multiple memory layers, wherein the memory layer includes a forget gate, an input gate, and an output gate; The spatial feature vector and the temporal feature vector are concatenated to obtain a fused feature vector, and the fused feature vector is subjected to feature dimension reduction and anti-overfitting processing through a fully connected layer and a random dropout layer; based on the fused feature vector and a preset island state label, the hybrid deep learning network is trained using a multiple loss function to obtain an island state detection model.

4. The method according to claim 3, characterized in that: The steps of constructing a hybrid deep learning network, wherein the hybrid deep learning network includes a convolutional neural network and a long short-term memory network, wherein the convolutional neural network includes a plurality of convolutional layers, a normalization layer, and an activation function layer, and wherein the long short-term memory network includes a plurality of memory layers include: In a convolutional neural network, the importance weights of different convolutional layers are adaptively calculated by a global context encoder, and the global context encoder generates a channel attention map based on the statistical information of the input features, and the channel attention map is used to adjust the fusion weights of the output features of different convolutional layers; A feature interaction module is constructed between convolutional layers. The feature interaction module includes a feature resampling unit and an adaptive fusion unit. Dynamic information exchange of output features of different convolutional layers is realized through a transformation matrix. The parameters of the transformation matrix are updated through a back propagation algorithm. The update step is proportional to the feature importance. The adaptive fusion unit automatically adjusts the information exchange intensity according to the feature correlation. A bidirectional memory path is set in the long short-term memory network, the memory layers of the forward path and the reverse path share parameters, and the bidirectional timing information is dynamically integrated through a state fusion gate, and the state fusion gate calculates the fusion weight based on the current input and the historical state; A time-aware mechanism is embedded in the memory layer to encode the time intervals between adjacent moments into a continuous vector, and the continuous vector is integrated into the calculation of the forget gate and the input gate. The time interval encoding is generated by a learnable time embedding layer.

5. The method according to claim 3, characterized in that: Based on the fused feature vector and the preset island state label, the steps of training the hybrid deep learning network using multiple loss functions to obtain the island state detection model include: Based on the fused feature vector and the preset island state label, a feature importance evaluation index is calculated through feature mutual information calculation, feature correlation analysis, and feature prediction gain evaluation; a dynamic feature weight is generated through a softmax function with adjustable smoothness, and the dynamic feature weight is used to weight the feature importance evaluation index; Performing time series enhancement processing on the input data according to the feature importance evaluation index to obtain an enhanced feature sequence, selecting time series-related samples from the enhanced feature sequence to construct positive sample pairs, and using a contrastive learning loss function to calculate the cosine similarity between the positive sample pairs, wherein the contrastive learning loss function includes a positive sample pair similarity calculation item and a negative sample pair similarity calculation item; Constructing a multi-layer attention network to extract the correlation between features to obtain a feature correlation score, combining the feature correlation score with the feature importance evaluation index to calculate a feature fusion weight, and using the feature fusion weight to adjust the classification cross entropy loss function to obtain a combined loss function, wherein the combined loss function includes a cross entropy loss term and a knowledge distillation constraint term; The gradient scaling coefficient of each layer is calculated by combining the network layer depth information of the hybrid deep learning network and the weighted feature importance evaluation index; in the process of network back propagation, the partial derivative of the loss function of each layer with respect to the parameter is obtained as the original gradient, and the original gradient is corrected by using the gradient scaling coefficient to obtain the corrected gradient; Calculate the sample sampling probability by combining the prediction confidence of the hybrid deep learning network with the feature importance evaluation index, and construct a training batch based on the sample sampling probability; A multiple loss function is constructed based on the contrastive learning loss function and the combined loss function, and the constructed training batches are iteratively trained using the multiple loss function to update the parameters of the hybrid deep learning network until a preset training termination condition is met.

6. The method according to claim 5, characterized in that The prediction confidence of the hybrid deep learning network is combined with the feature importance evaluation index to calculate the sample sampling probability, and the step of constructing a training batch based on the sample sampling probability includes: Obtain the prediction confidence of the hybrid deep learning network for the training samples, and combine the feature importance evaluation index to calculate the initial sampling probability of each training sample in a nonlinear combination manner. The nonlinear combination uses a sigmoid activation function to map the probability value to a preset interval; dynamically adjust the initial sampling probability based on an adaptive weight coefficient, the adaptive weight coefficient changes exponentially with the training rounds, and introduces a normalization coefficient to ensure the stability of the number of batch samples; According to the prediction confidence, the training samples are divided into easy-to-classify samples, medium-level samples and difficult samples, and the difficult samples are subjected to sample enhancement processing, wherein the sample enhancement processing adopts a weighted combination of the gradient of the loss function and the importance of the features; An initial training batch is constructed based on the dynamically adjusted initial sampling probability and the enhanced difficult samples, and a balance index of the initial training batch is determined. When the balance index does not meet the preset conditions, the training batch is reconstructed until a training batch that meets the balance requirements is constructed.

7. The method according to claim 1, characterized in that The steps of inputting the real-time monitoring data of the wind farm into the island state detection model, calculating the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector based on the island state detection model, determining that the wind farm has an island state when the matching degree is greater than a preset threshold, outputting an island state warning signal and triggering a safety protection mechanism of the wind farm include: Collecting real-time monitoring data of the wind farm, the real-time monitoring data including voltage, current, active power, reactive power and phase angle parameters; performing feature extraction on the real-time monitoring data to construct a detection feature vector including voltage feature components, current feature components, active power feature components, reactive power feature components and phase angle feature components; The detection feature vector is input into an island state detection model, and the island state detection model includes a feature matching degree calculation unit and a weight optimization unit; the feature matching degree calculation unit calculates the feature matching degree between the voltage feature component, the current feature component, the active power feature component, the reactive power feature component and the phase angle feature component and the corresponding component of the preset island feature vector respectively; the weight optimization unit adaptively adjusts the weight coefficient of each feature component based on the historical island event data, and performs weighted calculation on the feature matching degree of each feature component and the corresponding weight coefficient to obtain a comprehensive feature matching degree; The comprehensive feature matching degree is compared with the preset island state warning threshold to determine the island state warning level of the wind farm, and corresponding protective measures are triggered according to the island state warning level, including: when the comprehensive feature matching degree is greater than the first warning level threshold, a first-level protective measure of reducing the active output of the wind turbine and increasing reactive support is triggered; when the comprehensive feature matching degree is greater than the second warning level threshold, a second-level protective measure of switching the island operation mode and emergency adjustment of voltage and frequency is triggered; when the comprehensive feature matching degree is greater than the third warning level threshold, a third-level protective measure of executing emergency shutdown and disconnecting the grid-connected switch is triggered; The execution results of the protective measures and the island state detection results are added to the historical island event data to optimize the weight coefficient of the weight optimization unit and realize the adaptive optimization of the island state detection model.

8. An island state detection system for large-scale wind power, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to divide the wind farm into a core area and a boundary area, set an active disturbance injection unit in the boundary area, control the active disturbance injection unit to inject a small disturbance signal into the power grid, collect voltage and current data of the core area and the boundary area in real time, obtain network impedance characteristics by detecting the response characteristics of the core area to the injected disturbance, and construct a three-dimensional correlation feature matrix including network impedance characteristics, power flow characteristics and frequency response characteristics in combination with the power flow direction of the boundary area and the frequency response curve of the core area; The second unit is used to perform feature expansion on the three-dimensional correlation feature matrix based on weather condition information in the historical operation data of the wind farm to generate an expanded feature matrix; A training data set is constructed using an expanded feature matrix, and the training data set is input into a hybrid deep learning network including a convolutional neural network and a long short-term memory network, and spatial features of the expanded feature matrix are extracted through the convolutional neural network, and temporal features of the expanded feature matrix are extracted through the long short-term memory network, and the spatial features and the temporal features are fused to obtain a fused feature vector, and the hybrid deep learning network is trained based on the fused feature vector and the corresponding island state label to obtain an island state detection model; The third unit is used to input the real-time monitoring data of the wind farm into the island state detection model, calculate the matching degree between the detection feature vector of the real-time monitoring data and the preset island feature vector based on the island state detection model, and determine that the wind farm has an island state when the matching degree is greater than a preset threshold, output an island state warning signal and trigger the safety protection mechanism of the wind farm.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images