Fan energy management method and system based on multi-dimensional monitoring

Abstract

The invention discloses a fan energy management method and system based on multi-dimensional monitoring, and relates to the technical field of fan data monitoring and control, and the method comprises the steps: obtaining multi-dimensional time sequence data in historical monitoring data to construct a sample set; constructing a generating capacity prediction model, collecting to-be-predicted data in real time, transmitting the to-be-predicted data into the generating capacity prediction model, and obtaining the overall predicted generating capacity; acquiring a wind power plant operation area and a constraint wind speed, screening node fans in the wind power plant operation area by taking the constraint wind speed as a screening condition to obtain a node fan set corresponding to the wind power plant operation area, and calculating a load balance value by utilizing the multi-dimensional attribute of each node fan; and calculating a load distribution coefficient value of each fan based on the load balance value, and carrying out load adjustment on the node fans according to the load distribution coefficient values and the overall predicted generating capacity. According to the invention, the prediction of the overall generating capacity and the load adjustment of each fan can be effectively realized, so that the power generation stability is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine data monitoring and control, and more specifically to a wind turbine energy management method and system based on multi-dimensional monitoring. Background Art

[0002] In the context of global energy transformation, wind power generation, as a clean and sustainable energy utilization method, occupies an increasingly important position in the power supply system. The efficient operation and optimized management of wind farms are of key significance to improving the competitiveness of wind power and promoting the large-scale development of renewable energy.

[0003] As the attention to environmental protection and sustainable development continues to increase, countries have committed to promoting the transformation of energy structure to renewable energy. Wind power generation has become one of the important ways to achieve this goal due to its advantages such as abundant resources and no pollution. The scale of wind farms continues to expand, and the installed capacity continues to grow. The operation and management level of wind farms directly affects the status and role of wind power in the energy market. For example, in some European countries, wind power has become one of the main sources of electricity, and the operating efficiency and reliability of its wind farms are crucial to ensuring national energy security.

[0004] However, the intermittent and volatile nature of wind power poses a challenge to the stable operation of the power system. The change in wind speed is not controlled by humans, which may cause large fluctuations in the power generation of wind turbines at any time. This uncertainty makes it difficult for wind farms to coordinate with the power grid, such as difficulty in meeting the real-time power balance requirements of the power grid and possible fluctuations in the voltage and frequency of the power grid. Therefore, how to effectively manage wind farms and improve the stability and predictability of their power generation is an urgent problem that technicians in this field need to solve. Summary of the invention

[0005] In view of this, the present invention provides a wind turbine energy management method and system based on multi-dimensional monitoring, which can effectively realize the prediction of overall power generation and load regulation of each wind turbine, thereby ensuring the stability of power generation.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A fan energy management method based on multi-dimensional monitoring, comprising:

[0008] Collect various historical monitoring data of wind farms, filter out multi-dimensional time series data, and construct sample sets based on multi-dimensional time series data and power generation at each sampling moment;

[0009] Construct a neural network model, use the sample set to train the neural network model, and obtain a full-precision model; perform binary quantization on the neural network model parameters to obtain a quantized model, train the quantized model based on the full-precision model, add the model-to-model error to the loss function of the quantized model, and obtain a power generation prediction model; collect the data to be predicted in real time, pass it into the power generation prediction model, and obtain the overall predicted power generation;

[0010] Obtaining a wind farm operation area and a constrained wind speed, wherein the wind farm operation area includes at least one node wind turbine, and each of the node wind turbines has a multi-dimensional attribute for characterizing a power generation state of a corresponding unit area in the current wind farm operation area;

[0011] The node wind turbines in the wind farm operation area are screened using the constrained wind speed as a screening condition to obtain a set of node wind turbines corresponding to the wind farm operation area, and a load balancing value of each node wind turbine is calculated using the multi-dimensional attributes of each node wind turbine in the set of node wind turbines;

[0012] Based on the load balancing value of each node wind turbine, a load distribution coefficient value of each node wind turbine is calculated, and according to the load distribution coefficient value and the overall predicted power generation, the load of the node wind turbine is adjusted.

[0013] Preferably, constructing a sample set based on multidimensional time series data and power generation at each sampling moment specifically includes: combining multidimensional data at the same sampling moment into one sampling data to obtain a sampling data sequence; dividing the sampling data sequence according to a preset time window to obtain multiple time window samples, and putting them into the sample set; the power generation corresponding to each time window sample at the last sampling moment is used as a label for the time window sample.

[0014] Preferably, in the constructed neural network model, the neural network model includes a compound attention module and a temporal convolution module; the compound attention module includes a trend segment preprocessing layer, a pooling layer and a compound perception attention layer; wherein the trend segment preprocessing layer is used to divide each input sample from the time series dimension into multiple trend samples and then pass them into the pooling layer; the pooling layer is used to perform maximum pooling and average pooling on each segment of trend samples respectively, and then pass the pooling features obtained by adding the maximum pooling features and the average pooling features into the compound perception attention layer; the compound perception attention layer is used to learn the pooling features and obtain the compound attention vector of the last layer; the temporal convolution module includes a temporal convolution layer and a multi-layer perception layer; the temporal convolution layer is used to receive the output results of the compound attention module, use one-dimensional convolution with different expansion values ​​in each layer to realize temporal information transmission, and pass the temporal information into the multi-layer perception layer; the multi-layer perception layer obtains the prediction results through multi-layer perception learning.

[0015] Preferably, the step of performing binary quantization on the neural network model parameters to obtain the quantized model is to perform binary quantization on the weights and biases of the neural network model to obtain the binarized weights and biases as initial parameters of the quantized model; and the step of training the quantized model based on the full-precision model includes:

[0016] Get the mean weight of each layer in the full-precision model as the balancing factor for each layer;

[0017] In the forward propagation, the last layer of the compound attention module in the quantized model uses the binary Sigmoid activation function to calculate the compound attention vector. The binarized weights of other layers are multiplied by the activation vector of the previous layer and added with the bias, and then the Relu activation function is used to obtain the initial activation vector of each layer. The initial activation vector of each layer is binary quantized and then multiplied by the equalization factor of the corresponding layer to obtain the final activation vector of each layer for the next layer.

[0018] In back propagation, the gradient of the binarized weight is first calculated according to the loss function, and then the gradient is clipped as the gradient value of the floating-point gradient to update the weight.

[0019] Preferably, the use of the multi-dimensional attributes of each node fan in the node fan set to calculate the load balancing value of each node fan specifically includes: using the multi-dimensional attributes of each node fan in the node fan set to calculate the state deviation of each attribute of each node fan; calculating the load balancing value of each node fan based on the state deviation and multi-dimensional attributes of each node fan.

[0020] Preferably, the calculation of the load distribution coefficient value of each node wind turbine specifically includes: calculating the load balancing value of each node wind turbine in the set of node wind turbines corresponding to each wind farm operating area; averaging the load balancing values ​​in the corresponding wind farm operating area to obtain the load mean of the wind farm operating area; and using the load mean of each wind farm operating area to calculate the load distribution coefficient value of each node wind turbine.

[0021] Preferably, the method of calculating the load distribution coefficient value of each wind turbine node by using the load mean value of each wind farm operation area specifically includes:

[0022]

[0023] Among them, g is the load distribution coefficient value, b is ji is the i-th node wind turbine in the j-th wind farm operation area, b j ′ is the load mean value of the j-th wind farm operation area, m is the maximum number of wind turbines in the j-th wind farm operation area, and k is the maximum value of the wind farm operation area.

[0024] Preferably, the multidimensional attributes include power coefficient, fan vibration frequency and component temperature;

[0025] Taking the constrained wind speed as a screening condition, the node wind turbines in the wind farm operation area are screened to obtain a node wind turbine set corresponding to the wind farm operation area, including: obtaining the wind turbine vibration frequency of each node wind turbine in the wind farm operation area; screening out the node wind turbines whose wind turbine vibration frequency meets the screening condition to form the node wind turbine set.

[0026] A fan energy management system based on multi-dimensional monitoring, comprising:

[0027] The sample set construction module collects various historical monitoring data of the wind farm, filters out multi-dimensional time series data, and constructs a sample set based on the multi-dimensional time series data and the power generation at each sampling moment;

[0028] The power generation prediction module builds a neural network model, trains the neural network model using a sample set, and obtains a full-precision model; performs binary quantization on the neural network model parameters to obtain a quantized model, trains the quantized model based on the full-precision model, adds the model-to-model error to the loss function of the quantized model, and obtains a power generation prediction model; collects the data to be predicted in real time, and transmits it to the power generation prediction model to obtain the overall predicted power generation;

[0029] A constraint acquisition module is used to acquire a wind farm operation area and a constrained wind speed, wherein the wind farm operation area includes at least one node wind turbine, and each node wind turbine has a multi-dimensional attribute for characterizing a power generation state of a corresponding unit area in the current wind farm operation area;

[0030] A load balancing value calculation module, using the constrained wind speed as a screening condition, screens the node wind turbines in the wind farm operation area to obtain a set of node wind turbines corresponding to the wind farm operation area, and calculates the load balancing value of each node wind turbine by using the multi-dimensional attributes of each node wind turbine in the set of node wind turbines;

[0031] The load adjustment module calculates the load distribution coefficient value of each node wind turbine based on the load balancing value of each node wind turbine, and performs load adjustment on the node wind turbine according to the load distribution coefficient value and the overall predicted power generation.

[0032] Through the above technical solutions, it can be known that compared with the prior art, the present invention discloses a wind turbine energy management method and system based on multi-dimensional monitoring. By collecting various historical monitoring data of wind farms and screening out multi-dimensional time series data to construct a sample set, it can comprehensively integrate data information of different dimensions and different time points, so that subsequent model training can be based on rich and relevant data, making the best use of existing data resources and laying the foundation for accurate energy management; the constructed neural network model includes a composite attention module and a time series convolution module. The composite attention module can process input samples from multiple angles through the collaborative work of trend segment preprocessing layer, pooling layer and composite perception attention layer, so as to better capture trend characteristics and key information in the data; the time series convolution module uses different expansion The one-dimensional convolution of the value is used to realize the transmission of time series information, and then combined with the multi-layer perception layer learning to obtain the prediction result. The overall architecture enables the model to have strong learning and prediction capabilities for time series related data such as wind turbine power generation, which can improve the accuracy of power generation prediction; the node wind turbine set is screened out based on the constrained wind speed, and then the load balancing value of each node wind turbine is calculated, and the load distribution coefficient value is calculated on this basis. Finally, the node wind turbine is load adjusted according to the load distribution coefficient value and the overall predicted power generation. This method of implementing load adjustment based on multi-dimensional state evaluation can make the load of each wind turbine more reasonable and balanced while ensuring the power generation, avoid overload or underload of some wind turbines, and improve the power generation efficiency of the entire wind farm. At the same time, it also helps to extend the service life of the wind turbine and reduce equipment maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0034] Figure 1 A diagram of the method steps provided by the present invention;

[0035] Figure 2 A schematic diagram of the structure provided by the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.

[0037] The embodiment of the present invention discloses a fan energy management method based on multi-dimensional monitoring, such as Figure 1 As shown, including:

[0038] Collect various historical monitoring data of wind farms, filter out multi-dimensional time series data, and construct sample sets based on multi-dimensional time series data and power generation at each sampling moment;

[0039] Construct a neural network model, use the sample set to train the neural network model, and obtain a full-precision model; perform binary quantization on the neural network model parameters to obtain a quantized model, train the quantized model based on the full-precision model, add the model-to-model error to the loss function of the quantized model, and obtain a power generation prediction model; collect the data to be predicted in real time, pass it into the power generation prediction model, and obtain the overall predicted power generation;

[0040] Obtain the wind farm operation area and constrained wind speed. The wind farm operation area contains at least one node wind turbine. Each node wind turbine has multi-dimensional attributes, which are used to characterize the power generation state of the corresponding unit area in the current wind farm operation area.

[0041] Taking the constrained wind speed as the screening condition, the node wind turbines in the wind farm operation area are screened to obtain the node wind turbine set corresponding to the wind farm operation area. The load balancing value of each node wind turbine is calculated by using the multi-dimensional attributes of each node wind turbine in the node wind turbine set.

[0042] Based on the load balancing value of each node wind turbine, the load distribution coefficient value of each node wind turbine is calculated, and the load of the node wind turbine is adjusted according to the load distribution coefficient value and the overall predicted power generation.

[0043] Among them, historical monitoring data includes different types of long-term continuous data, most of which have different magnitudes, units, forms, etc. In addition, there are redundancy and bad features in the data features collected by multiple sensors, which do not contribute to the prediction accuracy of the model. Therefore, it is necessary to preprocess the historical monitoring data as follows to filter out multi-dimensional time series data to reduce the interference of environmental factors when the sensor collects data and improve the prediction accuracy:

[0044] Removing noise features: Using the Savitzky-Golay smoothing filter method, remove abnormal data in each type of monitoring data, and fill in the data according to the fitted polynomial to obtain the denoised data;

[0045] Standardized data features: After denoising, each type of monitoring data is standardized according to the mean and standard deviation;

[0046] Screening data features: Perform curve fitting on each type of standardized monitoring data, calculate the regression evaluation index value, obtain the monitoring data type whose regression evaluation index value is greater than the screening threshold, and obtain the data features for predicting the overall power generation.

[0047] In a specific embodiment, constructing a sample set based on multidimensional time series data and power generation at each sampling moment specifically includes: combining multidimensional data at the same sampling moment into one sampling data to obtain a sampling data sequence; dividing the sampling data sequence according to a preset time window to obtain multiple time window samples, and putting them into the sample set; the power generation corresponding to each time window sample at the last sampling moment is used as a label for the time window sample.

[0048] Specifically, each sample data is represented as {d1, d2, ..., d M} M×C It means that there are M sampling moments in total, and monitoring data of C characteristic dimensions are screened out at each sampling moment. For example, wind speed, wind direction, temperature, air pressure and wind turbine maintenance record data are screened out from various historical monitoring data, then C=5.

[0049] The time window length is an empirical value, which is adjusted according to the training results during the training process. The time window length is set to 30, that is, the monitoring array at every 30 sampling moments constitutes a time window sample.

[0050] In a specific embodiment, in constructing a neural network model, the neural network model includes a composite attention module and a temporal convolution module; the composite attention module includes a trend segment preprocessing layer, a pooling layer and a composite perception attention layer; wherein the trend segment preprocessing layer is used to divide each input sample from the time series dimension into multiple trend samples and then pass them into the pooling layer; the pooling layer is used to perform maximum pooling and average pooling on each segment of trend samples respectively, and then pass the pooling features obtained by adding the maximum pooling features and the average pooling features into the composite perception attention layer; the composite perception attention layer is used to learn the pooling features and obtain the composite attention vector of the last layer; the temporal convolution module includes a temporal convolution layer and a multi-layer perception layer; the temporal convolution layer is used to receive the output results of the composite attention module, use one-dimensional convolution with different expansion values ​​in each layer to realize temporal information transmission, and pass the temporal information into the multi-layer perception layer; the multi-layer perception layer obtains the prediction results through multi-layer perception learning.

[0051] In a specific embodiment, the length of the composite attention vector of the last layer is consistent with the feature dimension of the input sample; the input sample is weighted according to the composite attention vector to obtain the output result of the composite attention module.

[0052] Since each time window sample is time series data, in order to accurately obtain the trend changes caused by the change of data over time, this step divides the input samples into time periods again in the trend segment preprocessing layer to obtain N trend samples, which are expressed as follows:

[0053]

[0054] Among them, Split(·) represents the time period division operation, Hconcat(·) represents the splicing operation of the time series dimension, and D H×C Represents the input sample, that is, each time window sample in the sample set; each trend sample There are samples.

[0055] Specifically, in the pooling layer, each trend sample is input into the maximum pooling layer and the average pooling layer respectively, and the maximum pooling feature and the average pooling feature are obtained, thereby generating two different spatial context descriptions. Then, the pooling feature obtained by adding the maximum pooling feature and the average pooling feature is obtained according to the following formula:

[0056]

[0057] Among them, pool(·) represents the pooling operation, Avg(·) represents the average pooling operation, and Max(·) represents the maximum pooling operation. Represents the obtained pooled features.

[0058] In a specific embodiment, binary quantization is performed on the neural network model parameters to obtain a quantized model, and binary quantization is performed on the weights and biases of the neural network model to obtain the binarized weights and biases as initial parameters of the quantized model; training the quantized model based on the full-precision model includes:

[0059] Get the mean weight of each layer in the full-precision model as the balancing factor for each layer;

[0060] In the forward propagation, the last layer of the compound attention module in the quantized model uses the binary Sigmoid activation function to calculate the compound attention vector. The binarized weights of other layers are multiplied by the activation vector of the previous layer and added with the bias, and then the Relu activation function is used to obtain the initial activation vector of each layer. The initial activation vector of each layer is binary quantized and then multiplied by the equalization factor of the corresponding layer to obtain the final activation vector of each layer for the next layer.

[0061] In back propagation, the gradient of the binarized weight is first calculated according to the loss function, and then the gradient is clipped as the gradient value of the floating-point gradient to update the weight.

[0062] The pooled features are passed to the composite perception attention layer to calculate the composite attention vector. The composite perception attention layer is built based on the Multi-Layer Perception (MLP), which can be composed of an input layer, multiple hidden layers, and a fully connected layer as an output, or it can include only one fully connected layer; in the last layer, the weights are multiplied by the pooled features of the previous layer, and the composite attention vector is obtained after activation. It should be noted that if there is a bias in the last layer, the weights are multiplied by the pooled features of the previous layer and then the bias is added, and the composite attention vector is obtained after activation.

[0063] The length of the composite attention vector is consistent with the feature dimension C of the input sample, indicating the weight distribution of the multi-dimensional features. Finally, the input sample is weighted according to the composite attention vector to obtain the output result of the composite attention module, that is, the monitoring data of each feature dimension at each sampling moment in the input sample is multiplied by the composite attention vector of the same feature dimension to obtain the weighted sample as the output result of the composite attention module and passed into the temporal convolution module.

[0064] The temporal convolution module includes a temporal convolution layer and a multi-layer perception layer; the temporal convolution layer is constructed according to the temporal convolutional neural network (TCN), which is used to receive the output results of the composite attention module, and use one-dimensional convolution with different expansion values ​​in each layer to realize the transmission of temporal information, and pass the temporal information to the multi-layer perception layer; the multi-layer perception layer is constructed according to the multi-layer perceptron MLP. Through multi-layer perception learning, the single neuron in the last layer captures the entire time series information and outputs the prediction result.

[0065] Among them, the binary Sigmoid activation function includes: when the activation value obtained by using the Sigmoid activation function is greater than or equal to the preset threshold, the activation is 1; otherwise, it is suppressed to 0; binary quantization of weights, biases and initial activation vectors is to convert the values ​​into two numerical values ​​of 1 and -1.

[0066] Adding the inter-model error to the loss function of the quantized model is to increase the information entropy of the weights of each layer of the full-precision model and the quantized model and the mean square error of the output of each layer on the basis of the mean square error between the prediction result of the quantized model and the label of the power generation.

[0067] Considering that in the process of binary quantization (referred to as binarization) of the neural network model parameters, the model parameters are converted from high-bit expression form (such as 32-bit floating point type) to binary low-bit quantization form, although the memory usage of the model is significantly reduced, it will bring obvious precision loss, and it is difficult to maintain the model prediction performance. Therefore, this embodiment uses the sample set to perform two training processes on the constructed neural network model: the first time, no quantization is performed on the model parameters, and the model parameters are all floating point types. The mean error between the overall predicted power generation and the label of the power generation is used as the loss function for training. After reaching the maximum number of iterations or prediction accuracy, the training is stopped to obtain the trained model as the full-precision model; the second time, the parameters of the neural network model with the same structure are binary quantized to obtain a quantized model, and the model parameters are optimized during the training process based on the same sample set. In order to improve the prediction accuracy and prediction performance, an equalization factor is constructed based on the full-precision model trained for the first time, and the information entropy and output error of each layer between the full-precision model and the quantized model trained for the second time are added to the loss function, the quantized model parameters are optimized, and the quantized model trained for the second time is used as the final power generation prediction model.

[0068] Specifically, the neural network model parameters are binary quantized to obtain a quantized model, which is to binary quantize the weights and biases of the neural network model, and obtain the binarized weights and biases as the initial parameters of the quantized model. Then, the quantized model is trained based on the full-precision model, including:

[0069] ① Obtain the mean weight of each layer in the full-precision model as the balancing factor for each layer.

[0070] It should be noted that the model parameters in the full-precision model are all floating-point data. The mean weight of each layer in the full-precision model is obtained, including: for the convolution layer, the mean weight of all convolution kernels in the current layer is obtained; for the fully connected layer, the mean weight of all neurons in the current layer is obtained.

[0071] ② In the forward propagation, the last layer of the compound attention module in the quantized model uses the binary Sigmoid activation function to calculate the compound attention vector. The binarized weights of other layers are multiplied by the activation vector of the previous layer and added with the bias, and then the Relu activation function is used to obtain the initial activation vector of each layer. The initial activation vector of each layer is binary quantized and then multiplied by the equalization factor of the corresponding layer to obtain the final activation vector of each layer for the next layer.

[0072] It should be noted that the weights in the neural network model are used to represent the connection strength between neurons in each layer; the bias is used to correctly classify samples and ensure that the output value cannot be activated arbitrarily; the activation function plays the role of nonlinear mapping and can limit the output amplitude of neurons to a certain range.

[0073] Binary quantization of weights, biases, and initial activation vectors is performed by converting the values ​​to 1 and -1 using the following sign function:

[0074]

[0075] Among them, x1 represents the weight value, bias value or the value in the initial activation vector. The last layer of the composite attention module uses the Sigmoid activation function, and the other layers use the Relu activation function.

[0076] The binary Sigmoid activation function is based on the activation value obtained by the Sigmoid activation function for binary quantization. If the activation value is greater than or equal to the preset threshold, the activation is 1; otherwise, it is suppressed to 0. The formula is as follows:

[0077]

[0078] Among them, x2 represents the value to be activated, that is, the value after the binarized weight of the last layer is multiplied by the pooling feature of the previous layer.

[0079] Among them, the loss function Loss of the quantized model includes three parts: the mean square error between the prediction result and the remaining life label, the information entropy of the weights of each layer of the full-precision model and the quantized model, and the mean square error of the output of each layer.

[0080] In a specific embodiment, using the multidimensional attributes of each node fan in the node fan set, calculating the load balancing value of each node fan specifically includes: using the multidimensional attributes of each node fan in the node fan set to calculate the state deviation of each attribute of each node fan; calculating the load balancing value of each node fan based on the state deviation and multidimensional attributes of each node fan.

[0081] Among them, the state deviation is used to measure the degree of deviation between the current actual operating state of the wind turbine and the ideal operating state. The ideal operating state refers to the state in which the wind turbine can achieve the highest efficiency, longest service life and most stable operation under various constraints. For example, for a wind turbine, its ideal operating state may be that at the rated wind speed, the output power is the rated power, the health indicators such as vibration and temperature are within the normal range, and the power factor meets the requirements of the power grid. The state deviation quantifies the gap between the current state of the wind turbine and this ideal state by comprehensively considering the multi-dimensional attributes of the wind turbine (such as power coefficient, wind turbine vibration frequency and component temperature).

[0082] b ji =w ji *X ji ;

[0083] Among them, b jiis the state deviation of the i-th node wind turbine in the j-th wind farm operation area, w ji is the coefficient of deviation, X ji It is the attribute value corresponding to the i-th node wind turbine among all the node wind turbines corresponding to the j-th wind farm operation area.

[0084] The deviation coefficient step requires first calculating the partial score, partial score ratio, partial score mean and finite value. The specific calculation is as follows:

[0085]

[0086] Among them, y ji is the partial score, representing the absolute difference between the true attribute value and the average attribute value, p avg is the number of wind turbines a corresponding to the j-th wind farm operation area. j The average value of the power coefficient, fan vibration frequency and component temperature, p ji is the attribute value corresponding to the i-th node wind turbine among all the node wind turbines corresponding to the j-th wind farm operation area, max(p)-min(p) is the difference between the maximum and minimum values ​​of the same attribute value, and the node wind turbine set a corresponding to the j-th wind farm operation area is calculated. j All partial values ​​{y j1 ,y j2 ,...,y ji}(i=1,...,m, m represents the number of wind turbines in the node corresponding to the j-th wind farm operation area). After calculating the partial score, the partial score ratio is calculated:

[0087]

[0088] Among them, p ji is a j The attribute value of each node fan in (q ji 、v ji and c ji )’s partial score percentage.

[0089] After obtaining the partial value ratio of the attribute value of each node wind turbine, the attribute (q j 、v j and c j )’s partial mean:

[0090]

[0091] Among them, e ji is the partial mean.

[0092] Calculate the properties (q j 、vj and c j ) has a finite value q ji :

[0093]

[0094] The finite value q is calculated ji After that, the deviation coefficients of each attribute of the wind turbine at the i-th node in the j-th wind farm operation area are calculated respectively.

[0095]

[0096] Where m represents the number of node wind turbines corresponding to the j-th wind farm operation area. It is the sum of the finite values ​​of wind turbines at each node in the operating area of ​​the wind farm.

[0097] In a specific embodiment, calculating the load distribution coefficient value of each node wind turbine specifically includes: calculating the load balancing value of each node wind turbine in the set of node wind turbines corresponding to each wind farm operating area; averaging the load balancing values ​​in the corresponding wind farm operating area to obtain the load mean of the wind farm operating area; using the load mean of each wind farm operating area to calculate the load distribution coefficient value of each node wind turbine.

[0098] In a specific embodiment, using the load average of each wind farm operation area, the load distribution coefficient value of each node wind turbine is calculated specifically including:

[0099]

[0100] Among them, g is the load distribution coefficient value, b is ji is the i-th node wind turbine in the j-th wind farm operation area, b j ′ is the load mean value of the j-th wind farm operation area, m is the maximum number of wind turbines in the j-th wind farm operation area, and k is the maximum value of the wind farm operation area.

[0101] In a specific embodiment, the multi-dimensional attributes include power coefficient, fan vibration frequency, and component temperature;

[0102] Taking the constrained wind speed as a screening condition, the node wind turbines in the wind farm operation area are screened to obtain a node wind turbine set corresponding to the wind farm operation area, including: obtaining the wind turbine vibration frequency of each node wind turbine in the wind farm operation area; screening out the node wind turbines whose wind turbine vibration frequencies meet the screening conditions to form a node wind turbine set.

[0103] A fan energy management system based on multi-dimensional monitoring, such as Figure 2 As shown, including:

[0104] The sample set construction module collects various historical monitoring data of the wind farm, filters out multi-dimensional time series data, and constructs a sample set based on the multi-dimensional time series data and the power generation at each sampling moment;

[0105] The power generation prediction module builds a neural network model, trains the neural network model using a sample set, and obtains a full-precision model; performs binary quantization on the neural network model parameters to obtain a quantized model, trains the quantized model based on the full-precision model, adds the model-to-model error to the loss function of the quantized model, and obtains a power generation prediction model; collects the data to be predicted in real time, and transmits it to the power generation prediction model to obtain the overall predicted power generation;

[0106] A constraint acquisition module is used to acquire the wind farm operation area and the constrained wind speed. The wind farm operation area contains at least one node wind turbine. Each node wind turbine has multi-dimensional attributes, which are used to characterize the power generation state of the corresponding unit area in the current wind farm operation area.

[0107] The load balancing value calculation module uses the constrained wind speed as a screening condition to screen the node wind turbines in the wind farm operation area, obtains the node wind turbine set corresponding to the wind farm operation area, and uses the multi-dimensional attributes of each node wind turbine in the node wind turbine set to calculate the load balancing value of each node wind turbine;

[0108] The load adjustment module calculates the load distribution coefficient value of each node wind turbine based on the load balancing value of each node wind turbine, and adjusts the load of the node wind turbine according to the load distribution coefficient value and the overall predicted power generation.

[0109] In a specific embodiment, the load of the node wind turbine is adjusted according to the load distribution coefficient value and the overall predicted power generation. Specifically, the load value is distributed to the wind turbines that do not meet the screening conditions according to the maximum wind speed they can withstand, and the relevant components inside the wind turbine that affect the wind speed are adjusted (adjusting the blade pitch angle or rotation speed, etc.). For the node wind turbines in the screened node wind turbine set, the corresponding load is adjusted according to the load distribution coefficient value and the overall predicted power generation.

[0110] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0111] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fan energy management method based on multi-dimensional monitoring, characterized in that: include: Collect various historical monitoring data of wind farms, filter out multi-dimensional time series data, and construct sample sets based on multi-dimensional time series data and power generation at each sampling moment; Construct a neural network model, use the sample set to train the neural network model, and obtain a full-precision model; perform binary quantization on the neural network model parameters to obtain a quantized model, train the quantized model based on the full-precision model, add the model-to-model error to the loss function of the quantized model, and obtain a power generation prediction model; collect the data to be predicted in real time, pass it into the power generation prediction model, and obtain the overall predicted power generation; Obtaining a wind farm operation area and a constrained wind speed, wherein the wind farm operation area includes at least one node wind turbine, and each of the node wind turbines has a multi-dimensional attribute for characterizing a power generation state of a corresponding unit area in the current wind farm operation area; The node wind turbines in the wind farm operation area are screened using the constrained wind speed as a screening condition to obtain a set of node wind turbines corresponding to the wind farm operation area, and a load balancing value of each node wind turbine is calculated using the multi-dimensional attributes of each node wind turbine in the set of node wind turbines; Based on the load balancing value of each node wind turbine, a load distribution coefficient value of each node wind turbine is calculated, and according to the load distribution coefficient value and the overall predicted power generation, the load of the node wind turbine is adjusted.

2. A wind turbine energy management method based on multi-dimensional monitoring according to claim 1, characterized in that: The method of constructing a sample set based on multidimensional time series data and power generation at each sampling moment specifically includes: combining multidimensional data at the same sampling moment into one sampling data to obtain a sampling data sequence; dividing the sampling data sequence according to a preset time window to obtain multiple time window samples, and putting them into the sample set; and using the power generation corresponding to each time window sample at the last sampling moment as a label for the time window sample.

3. The wind turbine energy management method based on multi-dimensional monitoring according to claim 1 is characterized in that: In the constructed neural network model, the neural network model includes a composite attention module and a temporal convolution module; the composite attention module includes a trend segment preprocessing layer, a pooling layer and a composite perception attention layer; The trend segment preprocessing layer is used to divide each input sample into multiple trend samples from the time series dimension and then pass them into the pooling layer; the pooling layer is used to perform maximum pooling and average pooling on each trend sample segment, and then pass the pooling feature obtained by adding the maximum pooling feature and the average pooling feature into the composite perception attention layer; the composite perception attention layer is used to learn the pooling feature and obtain the composite attention vector of the last layer; the temporal convolution module includes a temporal convolution layer and a multi-layer perception layer; the temporal convolution layer is used to receive the output result of the composite attention module, use one-dimensional convolution with different expansion values ​​in each layer to realize temporal information transmission, and pass the temporal information into the multi-layer perception layer; the multi-layer perception layer obtains the prediction result through multi-layer perception learning.

4. The wind turbine energy management method based on multi-dimensional monitoring according to claim 1 is characterized in that: The binary quantization of the neural network model parameters to obtain the quantized model is to binary quantize the weights and biases of the neural network model to obtain the binarized weights and biases as the initial parameters of the quantized model; The full-precision model-based quantization model training includes: Get the mean weight of each layer in the full-precision model as the balancing factor for each layer; In the forward propagation, the last layer of the compound attention module in the quantized model uses the binary Sigmoid activation function to calculate the compound attention vector. The binarized weights of other layers are multiplied by the activation vector of the previous layer and added with the bias, and then the Relu activation function is used to obtain the initial activation vector of each layer. The initial activation vector of each layer is binary quantized and then multiplied by the equalization factor of the corresponding layer to obtain the final activation vector of each layer for the next layer. In back propagation, the gradient of the binarized weight is first calculated according to the loss function, and then the gradient is clipped as the gradient value of the floating-point gradient to update the weight.

5. The wind turbine energy management method based on multi-dimensional monitoring according to claim 1 is characterized in that: The method of utilizing the multidimensional attributes of each node fan in the node fan set to calculate the load balancing value of each node fan specifically includes: utilizing the multidimensional attributes of each node fan in the node fan set to calculate the state deviation of each attribute of each node fan; and calculating the load balancing value of each node fan based on the state deviation and multidimensional attributes of each node fan.

6. The wind turbine energy management method based on multi-dimensional monitoring according to claim 1 is characterized in that: The calculation of the load distribution coefficient value of each node wind turbine specifically includes: calculating the load balancing value of each node wind turbine in the set of node wind turbines corresponding to each wind farm operating area; averaging the load balancing values ​​in the corresponding wind farm operating area to obtain the load mean of the wind farm operating area; and using the load mean of each wind farm operating area to calculate the load distribution coefficient value of each node wind turbine.

7. A wind turbine energy management method based on multi-dimensional monitoring according to claim 6, characterized in that: The method of calculating the load distribution coefficient value of each wind turbine node by using the load mean value of each wind farm operation area specifically includes: Among them, g is the load distribution coefficient value, b is ji is the i-th node wind turbine in the j-th wind farm operation area, b j ′ is the load mean value of the j-th wind farm operation area, m is the maximum number of wind turbines in the j-th wind farm operation area, and k is the maximum value of the wind farm operation area.

8. The wind turbine energy management method based on multi-dimensional monitoring according to claim 6, characterized in that: The multidimensional attributes include power coefficient, fan vibration frequency and component temperature; The node wind turbines in the wind farm operation area are screened using the constrained wind speed as a screening condition to obtain a set of node wind turbines corresponding to the wind farm operation area, including: obtaining a wind turbine vibration frequency of each node wind turbine in the wind farm operation area; Node fans whose fan vibration frequencies meet the screening condition are screened out to form the node fan set.

9. A wind turbine energy management system based on multi-dimensional monitoring, applying a wind turbine energy management method based on multi-dimensional monitoring as described in any one of claims 1 to 8, characterized in that: include: The sample set construction module collects various historical monitoring data of the wind farm, filters out multi-dimensional time series data, and constructs a sample set based on the multi-dimensional time series data and the power generation at each sampling moment; The power generation prediction module builds a neural network model, trains the neural network model using a sample set, and obtains a full-precision model; performs binary quantization on the neural network model parameters to obtain a quantized model, trains the quantized model based on the full-precision model, adds the model-to-model error to the loss function of the quantized model, and obtains a power generation prediction model; collects the data to be predicted in real time, and transmits it to the power generation prediction model to obtain the overall predicted power generation; A constraint acquisition module is used to acquire a wind farm operation area and a constrained wind speed, wherein the wind farm operation area includes at least one node wind turbine, and each node wind turbine has a multi-dimensional attribute for characterizing a power generation state of a corresponding unit area in the current wind farm operation area; A load balancing value calculation module, using the constrained wind speed as a screening condition, screens the node wind turbines in the wind farm operation area to obtain a set of node wind turbines corresponding to the wind farm operation area, and calculates the load balancing value of each node wind turbine by using the multi-dimensional attributes of each node wind turbine in the set of node wind turbines; The load adjustment module calculates the load distribution coefficient value of each node wind turbine based on the load balancing value of each node wind turbine, and performs load adjustment on the node wind turbine according to the load distribution coefficient value and the overall predicted power generation.

Images