A Multi-Objective Cooperative Optimization Scheduling Method and Device for Wind Farms Based on Digital Twin

Abstract

The present application provides a multi-objective collaborative optimization scheduling method and device for a wind farm based on digital twin. Among them, the dominant wind direction of the wind farm is divided according to the wind direction data of the wind farm, and several wind turbine groups are obtained by clustering the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction; several wind power scenarios of the wind farm are generated according to the wind speed data of the wind farm; the generated several wind power scenarios are reduced to obtain typical scenarios of the wind farm; for each typical scenario, a pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm meet the objective function is determined based on the reinforcement learning algorithm; the pre-scheduling instruction is used as the initial value of real-time optimization scheduling, and the distributed model predictive control algorithm is used to optimize the control parameters; under this typical scenario, each wind turbine is controlled to operate based on the optimized control parameters. By adopting the above method, the power generation efficiency and power generation power of the wind farm are improved.

Description

Technical Field

[0001] The present invention relates to the field of wind farm management and control, and in particular, to a multi-objective collaborative optimization scheduling method and device for a wind farm based on digital twin. Background Art

[0002] Wind power generation is widely used due to its advantages of cleanliness and renewability. The number and scale of global wind farms are increasing rapidly, making great contributions to energy utilization and environmental protection. Therefore, more and more wind farms are being put into use. In the prior art, when generating electricity through a wind farm, several wind turbines arranged in the wind farm are usually controlled to operate at a fixed yaw angle parameter. That is to say, no matter in which power generation environment and wind power scenario, the wind turbines work with the same yaw angle control strategy to output electricity for power generation.

[0003] However, it is found in the research that if each wind turbine works with the same yaw angle control strategy in any power generation environment and wind power scenario, it may occur that in some wind power scenarios, some wind turbines cannot operate normally at their specific yaw angles, or the maximum power generation output cannot be achieved when operating at their specific yaw angles, resulting in a reduction in the power generation efficiency of the wind farm. In addition, when a large number of wind turbines operate simultaneously, there is usually power loss caused by wake interference, which also leads to a reduction in the power generation power of the wind farm. Summary of the Invention

[0004] In view of this, an object of the present invention is to provide a multi-objective collaborative optimization scheduling method and device for a wind farm based on digital twin to improve the power generation efficiency and power generation power of the wind farm.

[0005] In a first aspect, an embodiment of the present application provides a multi-objective collaborative optimization scheduling method for a wind farm based on digital twin, and the method includes:

[0006] Dividing the dominant wind direction of the wind farm according to the wind direction data of the wind farm, and clustering the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction to obtain a plurality of wind turbine groups;

[0007] Generating a plurality of wind power scenarios of the wind farm according to the wind speed data of the wind farm; reducing the generated plurality of wind power scenarios to obtain the typical scenarios of the wind farm;

[0008] For each typical scenario, determining a pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm meet the objective function based on the reinforcement learning algorithm;

[0009] Taking the pre-scheduling instruction as the initial value of real-time optimal scheduling, and optimizing the control parameters by using the distributed predictive control algorithm;

[0010] Under this typical scenario, each wind turbine is controlled based on the optimized control parameters.

[0011] Optionally, the dominant wind direction of the wind farm is divided according to the wind direction data of the wind farm, and under the dominant wind direction, the wind turbines in the wind farm are grouped according to wake characteristics to obtain several wind turbine groups, including:

[0012] Adopt a method based on hierarchical clustering to divide the dominant wind direction from the wind direction data;

[0013] Under the dominant wind direction, combined with the specific layout of the wind farm, the wind turbines are regarded as the nodes of a directed graph, and the wake influence of the upstream turbines on the downstream turbines is regarded as the edges of the directed graph;

[0014] Based on the established directed graph above, use the community discovery algorithm to describe the community structure of the wind farm.

[0015] Optionally, generating several wind power scenarios for the wind farm according to the wind speed data of the wind farm includes:

[0016] Taking the randomly generated noise signal as the input random vector and the measured wind speed data of the wind farm as the sample data, use the generative adversarial neural network for scenario generation to obtain several wind power scenarios.

[0017] Optionally, reducing the generated several wind power scenarios to obtain the typical scenarios of the wind farm includes:

[0018] Use the unsupervised clustering method to reduce the generated several wind power scenarios to obtain the typical scenarios of the wind farm;

[0019] Optionally, for each typical scenario, based on the reinforcement learning algorithm, determine the pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm under this typical scenario meet the objective function, including:

[0020] Taking the output power and unit load as the objective function and the limit loads of the key components of the wind turbine and the limit angles of the yaw control as the constraint conditions, establish a reward mechanism;

[0021] According to the reward mechanism, obtain the pre-scheduling instructions of each wind turbine through self-learning to realize the pre-scheduling of the yaw control of the wind farm.

[0022] Optionally, the constraint conditions include the limit load limit and yaw angle limit of the wind turbine;

[0023] Among them, the extreme value distribution theory is used to simulate and obtain the limit load limit, and the yaw angle limit condition is that the load of the unit does not exceed the limit load within the limit angle range.

[0024] Optionally, taking the pre-scheduling instruction as the initial value of real-time optimal scheduling, and optimizing the control parameters by using the distributed model predictive control algorithm includes:

[0025] When optimizing the control parameters, the constraint conditions are used as constraints in the optimization process, the objective function is used as the optimization goal, and the overall optimization calculation of the wind farm is divided into parallel optimization calculations of multiple wind turbine subsystems, and the yaw scheduling instructions of each wind turbine are obtained in real time to realize the optimization of the control parameters.

[0026] Optionally, the method includes:

[0027] Soft-sensing the input wind speed through the unit state, verifying it in combination with the lidar wind measurement data, and constructing a digital twin model of the inflow wind speed of the wind farm;

[0028] Calculating the propagation law of the fluid in front of the computer, and constructing a digital twin model of the wind speed in front of the machine according to the input wind speed of the wind farm;

[0029] Adopting the finite element modeling method, combining the blade element theory and the momentum theory to obtain the aerodynamic characteristics of the wind turbine, and using the multi-body dynamics method to calculate the structural response of the wind turbine to construct a single-machine digital twin model;

[0030] Constructing a digital twin model of wake characteristics based on the dynamic meandering principle;

[0031] Combining the digital twin model of the inflow wind speed of the wind farm, the digital twin model of the wind speed in front of the machine, the single-machine digital twin model, the digital twin model of wake characteristics and the digital twin model of field-level energy management of the wind farm to construct the digital twin model of the wind farm.

[0032] Optionally, the method further includes:

[0033] Selecting the data with the highest correlation with the lidar wind measurement data among the unit state parameters as the input data, using the lidar wind measurement data as the output data, and constructing the digital twin model of the inflow wind speed of the wind farm by using the deep learning neural network method;

[0034] Using the large eddy simulation method to generate a neutral atmospheric boundary layer without wind turbines, placing the wind turbines in the neutral atmospheric boundary layer, and adopting the method of LES coupled with the actuator line model to study the interaction between the wind turbines and the atmospheric boundary layer, thereby constructing the digital twin model of the wind speed in front of the machine;

[0035] The aerodynamic performance of the impeller is analyzed using the blade element momentum theory, and the single-unit digital twin model is constructed based on the results of the aerodynamic performance analysis;

[0036] A passive tracer model of the wake is established based on the dynamic meandering wake model, and the passive tracer model is corrected for response based on the evolution process of wake deficit, wake meandering, the increase in turbulence intensity caused by the wake, the decrease in wind speed in the near-wake region, and the radial expansion of the wake to obtain the wake characteristic digital twin model;

[0037] The wind farm-level energy management twin model is used to dynamically adjust the power output of each wind turbine according to the real-time conditions of the wind farm and the grid demand, so as to optimize the energy distribution and operation efficiency of the entire wind farm.

[0038] In a second aspect, an embodiment of the present application provides a digital twin-based multi-objective collaborative optimization scheduling device for a wind farm, and the device includes:

[0039] A wind turbine clustering module, configured to divide the dominant wind direction of the wind farm according to the wind direction data of the wind farm; cluster the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction to obtain a number of wind turbine groups;

[0040] A typical scenario division module, configured to generate a number of wind power scenarios for the wind farm according to the wind speed data of the wind farm; reduce the generated number of wind power scenarios to obtain the typical scenarios of the wind farm;

[0041] A pre-scheduling instruction determination module, configured to, for each typical scenario, determine a pre-scheduling instruction that can make the output power and unit load of the wind farm digital twin model meet the objective function based on a reinforcement learning algorithm;

[0042] A control parameter optimization module, configured to use the pre-scheduling instruction as the initial value of real-time optimal scheduling, and optimize the control parameters using a distributed model predictive control algorithm;

[0043] A wind turbine control module, configured to control the operation of each wind turbine based on the optimized control parameters under this typical scenario.

[0044] Optionally, the step of dividing the dominant wind direction of the wind farm according to the wind direction data of the wind farm and clustering the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction to obtain a number of wind turbine groups includes:

[0045] Using a hierarchical clustering-based method to divide the dominant wind direction from the wind direction data;

[0046] Under the dominant wind direction, in combination with the specific layout of the wind farm, the wind turbines are regarded as the nodes of a directed graph, and the wake influence of the upstream turbines on the downstream turbines is regarded as the edges of the directed graph;

[0047] Based on the directed graph established above, a community discovery algorithm is used to describe the community structure of the wind farm.

[0048] Optionally, the generation of several wind power scenarios for the wind farm according to the wind speed data of the wind farm includes:

[0049] Using a randomly generated noise signal as the input random vector and the measured wind speed data of the wind farm as the sample data, a generative adversarial neural network is used for scenario generation to obtain several wind power scenarios.

[0050] Optionally, the reduction of the generated several wind power scenarios to obtain the typical scenarios of the wind farm includes:

[0051] Using an unsupervised clustering method to reduce the generated several wind power scenarios to obtain the typical scenarios of the wind farm;

[0052] Optionally, for each typical scenario, based on the reinforcement learning algorithm, a pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm meet the objective function is determined, including:

[0053] Taking the output power and unit load as the objective function and the key component loads of the wind turbines and the limit angles of yaw control as the constraint conditions, a reward mechanism is established;

[0054] According to the reward mechanism, through self-learning, the pre-scheduling instructions of each wind turbine are obtained to realize the pre-scheduling of the yaw control of the wind farm.

[0055] Optionally, the constraint conditions include the limit load limit and yaw angle limit of the wind turbines;

[0056] Among them, the extreme value distribution theory is used to simulate and obtain the limit load limit, and the yaw angle limit condition is that the unit load does not exceed the limit load within the limit angle range.

[0057] Optionally, taking the pre-scheduling instruction as the initial value of real-time optimal scheduling and using the distributed model predictive control algorithm to optimize the control parameters, including:

[0058] When optimizing the control parameters, the above-mentioned constraint conditions are used as the constraints in the optimization process, the objective function is used as the optimization goal, and the overall optimization calculation of the wind farm is divided into parallel optimization calculations of multiple wind turbine subsystems to obtain the yaw scheduling instructions of each wind turbine in real time, so as to realize the optimization of the control parameters.

[0059] Optionally, the device includes a model construction module, which is used for:

[0060] Soft-measure the input wind speed through the unit state, verify it in combination with the lidar wind measurement data, and construct a digital twin model of the wind farm inflow wind speed;

[0061] The propagation law of the fluid in front of the computer, and construct a digital twin model of the wind speed in front of the machine according to the input wind speed of the wind farm;

[0062] Adopt the finite element modeling method, combine the blade element theory and the momentum theory to obtain the aerodynamic characteristics of the wind turbine, and use the multi-body dynamics method to calculate the structural response of the wind turbine to construct a single-machine digital twin model;

[0063] Construct a digital twin model of wake characteristics based on the dynamic meandering principle;

[0064] Combine the digital twin model of the wind farm inflow wind speed, the digital twin model of the wind speed in front of the machine, the single-machine digital twin model, the digital twin model of the wake characteristics and the digital twin model of the wind farm field-level energy management to construct the digital twin model of the wind farm.

[0065] Optionally, the model construction module is specifically used for:

[0066] Select the data with the highest correlation with the lidar wind measurement data among the unit state parameters as the input data, use the lidar wind measurement data as the output data, and construct the digital twin model of the wind farm inflow wind speed by using the deep learning neural network method;

[0067] Use the large eddy simulation method to generate a neutral atmospheric boundary layer without wind turbines, place the wind turbines in the neutral atmospheric boundary layer, and adopt the method of LES coupled with the actuator line model to study the interaction between the wind turbines and the atmospheric boundary layer, so as to construct the digital twin model of the wind speed in front of the machine;

[0068] Adopt the blade element momentum theory to analyze the aerodynamic performance of the impeller, and construct the single-machine digital twin model based on the results of the aerodynamic performance analysis;

[0069] Establish a passive tracer model of the wake based on the dynamic meandering wake model, and perform response correction on the passive tracer model based on the evolution process of wake deficit, wake meandering, the increase in turbulence intensity caused by the wake, and the wind speed decrease and wake radial expansion in the near wake region to obtain the digital twin model of the wake characteristics;

[0070] The wind farm field-level energy management digital twin model is used to dynamically adjust the power output of each wind turbine according to the real-time conditions of the wind farm and the grid demand, so as to optimize the energy distribution and operation efficiency of the entire wind farm.

[0071] In a third aspect, an embodiment of the present application provides a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device runs, the processor communicates with the memory through the bus. When the machine-readable instructions are executed by the processor, the steps of the digital-twin-based multi-objective collaborative optimization scheduling method in any optional implementation manner of the first aspect are executed.

[0072] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the steps of the digital-twin-based multi-objective collaborative optimization scheduling method in any optional implementation manner of the first aspect are executed.

[0073] The technical solutions provided by the present application include but are not limited to the following beneficial effects:

[0074] First, the present application divides the dominant wind direction of the wind farm according to the wind direction data of the wind farm, and groups the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction to obtain several wind turbine groups; generates several wind power scenarios of the wind farm according to the wind speed data of the wind farm; reduces the generated several wind power scenarios to obtain the typical scenarios of the wind farm; in the above steps, by grouping the wind turbines in the wind farm to obtain several wind turbine groups, the unified optimization of the control strategies of the associated wind turbines can be realized, and the division of the typical wind power scenarios of the wind farm provides a basis for the subsequent optimization of the yaw angle under each wind power scenario.

[0075] Then, for each typical scenario, based on the reinforcement learning algorithm, a pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm under the typical scenario meet the objective function is determined; using the pre-scheduling instruction as the initial value of the real-time optimization scheduling, the distributed model predictive control algorithm is used to optimize the control parameters, which can determine the yaw angle control parameters adapted to the scenario for different wind power scenarios, and improve the adaptability and pertinence of the yaw angle control parameters.

[0076] Finally, under this typical scenario, controlling the operation of each wind turbine based on the optimized control parameters can avoid the power generation loss that may be caused by the diversity and complexity of the wind power scenarios and the wake interference between wind turbines.

[0077] Adopting the above solution, based on the wake interference situation of the wind farm, optimizing the yaw angle control parameters of the wind farm for different wind power scenarios, and then controlling the operation of each wind turbine based on the optimized control parameters, can reduce the power generation loss in the wind farm and improve the power generation efficiency and power generation power of the wind farm.

[0078] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following specifically presents preferred embodiments and, in conjunction with the accompanying drawings, provides a detailed description as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other relevant drawings can also be obtained based on these drawings.

[0080] Figure 1 Shows the flowchart of a multi-objective collaborative optimization scheduling method for a wind farm based on digital twin provided by Embodiment 1 of the present invention;

[0081] Figure 2 Shows the flowchart of a wind turbine grouping method provided by Embodiment 1 of the present invention;

[0082] Figure 3 Shows the flowchart of a pre-scheduling instruction generation method provided by Embodiment 1 of the present invention;

[0083] Figure 4 Shows the structural schematic diagram of a multi-objective collaborative optimization scheduling device for a wind farm based on digital twin provided by Embodiment 2 of the present invention;

[0084] Figure 5 Shows the structural schematic diagram of a computer device provided by Embodiment 3 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0085] In order to make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but only represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0086] Embodiment 1

[0087] For the convenience of understanding the present application, the following combines Figure 1The content described in the flowchart of a digital twin-based multi-objective collaborative optimization scheduling method for a wind farm provided by Embodiment 1 of the present invention is used to elaborate on Embodiment 1 of the present application in detail.

[0088] See Figure 1 as shown Figure 1 The flowchart of a digital twin-based multi-objective collaborative optimization scheduling method for a wind farm provided by Embodiment 1 of the present invention is shown, where the method includes steps S101 to S105:

[0089] S101: Divide the dominant wind directions of the wind farm according to the wind direction data of the wind farm, and group the wind turbines in the wind farm according to the wake characteristics under the dominant wind directions to obtain several groups of wind turbines.

[0090] Specifically, use the hierarchical clustering algorithm to divide the dominant wind directions, and use the community discovery algorithm (such as Louvain algorithm, Girvan-Newman algorithm, spectral clustering algorithm, etc.) to describe the community structure in the directed graph of the wind farm under the dominant wind directions, and group the wind turbines in the wind farm to obtain several groups of wind turbines.

[0091] S102: Generate several wind power scenarios for the wind farm according to the wind speed data of the wind farm; reduce the generated several wind power scenarios to obtain the typical scenarios of the wind farm.

[0092] Specifically, use a data-driven neural network model (such as variational autoencoder, long short-term memory neural network, generative adversarial neural network, etc.) for scenario generation, and use an unsupervised clustering method (including but not limited to self-organizing neural network SOM, K-means++ algorithm, Gaussian mixture model GMM, etc.) for scenario reduction to extract the typical scenarios of the wind farm output.

[0093] S103: For each typical scenario, determine the pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm meet the objective function based on the reinforcement learning algorithm.

[0094] Specifically, establish a reward mechanism with the output power and unit load as the objective function and the key component loads of the wind turbines and the limit angles of yaw control as the constraint conditions; according to the reward mechanism, obtain the pre-scheduling instructions of each wind turbine through self-learning to realize the pre-scheduling of the yaw control of the wind farm.

[0095] S104: Use the pre-scheduling instruction as the initial value of the real-time optimization scheduling, and optimize the control parameters by using the distributed model predictive control algorithm.

[0096] Specifically, when optimizing the control parameters, the constraint conditions are used as the constraints in the optimization process, and the objective function is used as the optimization goal. The overall optimization calculation of the wind farm is divided into parallel optimization calculations of multiple wind turbine subsystems, and the yaw scheduling instructions of each wind turbine are obtained in real time to achieve the optimization of the control parameters.

[0097] S105: Under this typical scenario, control the operation of each wind turbine based on the optimized control parameters.

[0098] Specifically, controlling each wind turbine to operate at the optimized control parameter refers to the yaw angle to achieve the minimum loss and maximum power generation under each typical scenario.

[0099] In an alternative embodiment, refer to Figure 2 as shown Figure 2 The flowchart of a wind turbine grouping method provided in Embodiment 1 of the present invention is shown. Among them, the dominant wind direction of the wind farm is divided according to the wind direction data of the wind farm, and the wind turbines in the wind farm are grouped according to the wake characteristics under the dominant wind direction to obtain several wind turbine groups, including steps S201 to S202:

[0100] S201: Use the method based on hierarchical clustering to divide the dominant wind direction from the wind direction data;

[0101] S202: Under the dominant wind direction, combined with the specific layout of the wind farm, regard the wind turbines as the nodes of a directed graph, and regard the wake influence of the upstream unit on the downstream unit as the edge of the directed graph;

[0102] S203: Based on the established directed graph above, use the community discovery algorithm to describe the community structure of the wind farm.

[0103] Specifically, use the hierarchical clustering algorithm to divide the dominant wind direction. First, preprocess all the wind direction data, limit the wind direction change between plus and minus 180 degrees. After standardization, regard all individual data points as one class, calculate the Euclidean distance and Fréchet distance between any two points, and construct a similarity metric matrix F(A,B):

[0104] F(A,B) = infmax{d(A(i),B(j))}

[0105] In the formula: A(i) and B(j) are two different classes, i and j are the serial numbers of a certain point in the class, that is, any two nodes. inf is used to find the lower bound, d is the Euclidean distance between two points. The physical meaning of the Fréchet distance is: the lower bound of the maximum value of the Euclidean distance between any two points in the two classes A(i) and B(j). Merge the two closest classes, update the similarity with the Fréchet distance, and repeat this step until all the data is clustered into one class to obtain the clustering graph.

[0106] Using a Louvain algorithm based on multi-level optimized modularity, combined with the specific layout of the wind farm and wake effect, the wind turbines in the farm are grouped into nodes. A single wind turbine can be regarded as a node of a directed graph, and the wake effect of the upstream turbine on the downstream turbine can be regarded as an edge of the directed graph. According to the wake effect, the set of directed edges in the farm can be obtained.

[0107] Define the modularity Q in the community structure of the directed graph as: the difference between the proportion of edges connecting the internal nodes of the community in the original directed network and the expected proportion of edges connecting the internal nodes of the community in the directed random network. The larger the modularity, the higher the quality of the clustering. The expression of Q is as follows:

[0108]

[0109] In the formula, i and j represent any two nodes in the directed graph, m represents the number of edges in the network, k i is the degree of node i, k j is the degree of node j, A ij is the element in the adjacency matrix, representing whether nodes i and j are connected; δ(i, j) is a judgment function. If i and j are in the same community, its value is 1, otherwise it is 0.

[0110] First, all nodes are divided into separate communities; second, a vertex is merged with its adjacent vertices. If the modularity at this time is greater than zero, this vertex is merged into the community where the adjacent vertices are located, and this step is repeated until the communities where all vertices are located no longer change; then, all communities are compressed into one vertex, the weights of all vertices in the original community are converted into the weight of the new vertex, and the weights of all edges between the original communities are converted into the weight of the new edge. Repeat the above three steps until the community no longer changes, and different groups of wind turbines in the entire wind farm are obtained.

[0111] In an alternative embodiment, the generating a plurality of wind power scenarios of the wind farm according to the wind speed data of the wind farm includes:

[0112] Using a randomly generated noise signal as the input random vector and the measured wind speed data of the wind farm as the sample data, a generative adversarial neural network is used for scenario generation to obtain a plurality of wind power scenarios.

[0113] Specifically, a generative adversarial neural network is used to generate scenarios for wind farm output, which includes a generator network and a discriminator network. The generator network inputs a random vector to generate simulated data similar to the sample data and passes the simulated data to the discriminator network. The input of the discriminator network is the sample data and the simulated data output by the generator network, and it outputs the probability that the simulated data belongs to the sample data. The random vector is a randomly generated noise signal, and the sample data is the measured data of the wind farm, including the time series wind speed information in the u, v, and w directions at different positions on the wind speed plane. By continuously updating the parameters of the generator and the discriminator, the generated data that conforms to the data samples is achieved.

[0114] The loss function Loss(G, D) of the generative adversarial neural network is as follows:

[0115]

[0116] Among them, n real is the number of real data, n fake is the number of generated data, x i is the real data, D(x i ) is the output result of the discriminator network based on the sample data, p is the noise data, G(p) represents the simulated data generated by the generator, and D(G(p)) represents the output result of the discriminator for the simulated data. The first term represents the loss of the samples from the real data on the discriminator, and the second term represents the loss of the samples from the generator. Fix the parameters of the generator network and update the parameters of the discriminator network, that is, calculate the JS divergence between the sample data and the simulated data; fix the parameters of the discriminator network and update the parameters of the generator network, that is, minimize the JS divergence to make the two data distribution laws closer. Alternately update the parameters of the two networks until the training ends.

[0117] In an optional implementation, the method of reducing the generated wind power scenarios to obtain the typical scenarios of the wind farm includes:

[0118] Using an unsupervised clustering method to reduce the generated wind power scenarios to obtain the typical scenarios of the wind farm.

[0119] Specifically, use the unsupervised K-means clustering method to reduce the above-generated wind power scenarios. Randomly select sample points in the sample set as the initial clustering centers, calculate the Euclidean distances from other sample points to the clustering centers, and select the remaining k3 - 1 clustering centers with probability P(x i ) The probability P(x i) is the ratio of the Euclidean distance from the i-th sample point to the initial clustering center to the sum of the Euclidean distances from all the remaining sample points to the initial clustering center. Set different numbers of clustering centers, perform clustering under different values, and select the value with the largest silhouette coefficient as the number of clustering centers k3. Calculate the mean of all data clusters and set it as the new clustering center, and repeat this step until the clustering center does not change. Select representative samples in the cluster as typical scenarios to achieve the reduction of complex scenarios.

[0120] In an alternative embodiment, refer to Figure 3 as shown in Figure 3 FIG. shows a flowchart of a pre-scheduling instruction generation method provided in the first embodiment of the present invention. Among them, for each typical scenario, based on the reinforcement learning algorithm, a pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm under this typical scenario meet the objective function is determined, including steps S301 to S302:

[0121] S301: Establish a reward mechanism with the output power and unit load as the objective function and the key component loads of the wind turbine and the limit angle of yaw control as the constraint conditions.

[0122] S302: According to the reward mechanism, obtain the pre-scheduling instructions for each wind turbine through self-learning to achieve the pre-scheduling of the yaw control of the wind farm.

[0123] Specifically, based on the reinforcement learning RL algorithm, for different said typical scenarios, with the output power and unit load of the digital twin model of the wind farm as the objective function and the key component loads of the wind turbine and the limit angle of yaw control as the constraint conditions, establish a reward mechanism, and through self-learning, obtain the yaw control angles of each unit in the field when the overall output power is the largest and the fan load is the smallest, so as to achieve the pre-scheduling of the yaw control of the wind farm; the key component loads include: the multi-modal load limits of the fan components (such as the out-of-plane bending moment at the blade root, the tower bottom bending moment, the intermediate shaft torque, etc.); obtain the reinforcement learning table, and set the initial yaw angle according to this table under different scenarios of the wind farm.

[0124] In an alternative embodiment, the constraint conditions include the limit load limit of the wind turbine and the yaw angle limit; among them, the extreme value distribution theory is used to simulate and obtain the limit load limit, and the yaw angle limit condition is that the unit load does not exceed the limit load within the limit angle range.

[0125] Specifically, the constraint conditions include the limit load limit and yaw angle limit of the wind turbine. The extreme value distribution theory is used to simulate the limit load. First, the limit load distribution is obtained by fitting the operating load data of the wind turbine (type I, type II, and type III distribution models can be used), the selected distribution model is evaluated using the normality test method and the model is verified. Finally, combined with the design service life of the wind turbine, the state probability of the limit load under different conditions of the unit is calculated by computer to obtain the limit load value. The yaw angle limit condition is that within the limit angle range, the load of the unit does not exceed the limit load.

[0126] In an alternative embodiment, taking the pre-scheduling instruction as the initial value of the real-time optimal scheduling, and optimizing the control parameters using the distributed predictive control algorithm includes:

[0127] When optimizing the control parameters, the aforementioned constraint conditions are used as constraints in the optimization process, the objective function is used as the optimization goal, and the overall optimization calculation of the wind farm is divided into parallel optimization calculations of multiple wind turbine subsystems to obtain the yaw scheduling instructions of each wind turbine in real time, thereby realizing the optimization of the control parameters.

[0128] Specifically, taking the yaw control parameters indicated by the pre-scheduling instruction as the initial value of the real-time optimal scheduling, and using the DMPC distributed predictive control algorithm to perform real-time optimization of the digital twin of the wind farm. Each wind turbine is regarded as a subsystem, and the controller of each wind turbine subsystem uses the wind farm model to optimize its corresponding input, and each sub-controller considers the information of the remaining subsystems in the calculation. In the optimization process, all wind turbine sub-controllers use the global objective function of the wind farm, use the aforementioned constraint conditions as constraints in the optimization process, divide the overall optimization calculation of a wind farm into parallel optimization calculations of multiple subsystems, and obtain the yaw scheduling instructions of each unit in the wind farm in real time, thereby realizing the optimization of the control parameters.

[0129] In an alternative embodiment, the method includes:

[0130] Soft-measure the input wind speed through the unit state, and verify it in combination with the lidar wind measurement data to construct a digital twin model of the inflow wind speed of the wind farm; calculate the propagation law of the fluid in front of the machine, and construct a digital twin model of the wind speed in front of the machine according to the input wind speed of the wind farm; use the finite element modeling method, combine the blade element theory and the momentum theory to obtain the aerodynamic characteristics of the wind turbine, and use the multi-body dynamics method to calculate the structural response of the wind turbine to construct a single-unit digital twin model; construct a digital twin model of the wake characteristics based on the dynamic meandering principle; combine the digital twin model of the inflow wind speed of the wind farm, the digital twin model of the wind speed in front of the machine, the single-unit digital twin model, the digital twin model of the wake characteristics and the digital twin model of the field-level energy management of the wind farm to construct the digital twin model of the wind farm.

[0131] Specifically, for the twin modeling of the wind speed in front of the wind turbines in a wind farm, soft measurement of the input wind speed is performed based on the unit status, and it is calibrated in combination with the wind measurement data of lidar to achieve precise measurement of the input wind speed of the wind farm. The propagation law of the fluid in front of the machine is calculated, and the inflow wind speed in front of the wind turbines is reconstructed according to the input wind speed of the wind farm; for the twin modeling of the dynamics of a single wind turbine, the finite element modeling method is adopted, and the aerodynamic characteristics of the wind turbine are obtained by combining the blade element theory and the momentum theory, and the structural response of the wind turbine is calculated using the multi-body dynamics method; a digital twin model of the wake characteristics of the wind farm is established based on the dynamic meandering principle; by combining the above models with the digital twin model of the wind farm-level energy management, the digital twin model of the wind farm is obtained.

[0132] Compile the above-mentioned wind farm twin model for the working conditions of parallel operation in the Linux RT environment, and perform local real-time simulation operation on a high-performance server with a multi-core CPU.

[0133] In an optional implementation, the method further includes:

[0134] Select the data with the highest correlation with the lidar wind measurement data among the unit status parameters as the input data, use the lidar wind measurement data as the output data, and construct the digital twin model of the inflow wind speed of the wind farm by using the deep learning neural network method.

[0135] Specifically, collect historical operation data as sample data, calibrate the lidar data among them, analyze the correlation between the unit status parameters (such as rotational speed, torque, pitch angle, power generation, blade root load, drive shaft load, etc.) in the historical operation data and the lidar wind measurement data, select the data with the highest correlation as the input data, use the lidar wind measurement data as the output data, and estimate the wind speed at the lidar measurement position (common value is 200 meters) as the output by using the deep learning neural network method.

[0136] Use the large eddy simulation method to generate a neutral atmospheric boundary layer without wind turbines, place the wind turbines in the neutral atmospheric boundary layer, and adopt the method of coupling the LES with the actuator line model to study the interaction between the wind turbines and the atmospheric boundary layer, thereby constructing the digital twin model of the wind speed in front of the machine.

[0137] Specifically, use the large eddy simulation method to generate a neutral atmospheric boundary layer without wind turbines, place the wind turbines in the neutral atmospheric boundary layer, and adopt the method of coupling the LES with the actuator line model to study the interaction between the wind turbines and the atmospheric boundary layer. The turbulence is divided into large-scale vortices and small-scale vortices by the filtering method. The large-scale vortices are directly solved by the N-S equation, and the small-scale vortices are simulated by the sub-grid model. The surface stress and temperature flux of the atmospheric boundary layer are simulated according to the surface roughness height. Reconstruct the inflow wind speed in front of the wind turbines through the above method, and analyze the evolution process of atmospheric turbulence in front of the wind turbines.

[0138] The blade element momentum theory is adopted to analyze the aerodynamic performance of the impeller, and the single-unit digital twin model is constructed based on the results of the aerodynamic performance analysis.

[0139] Specifically, the blade element momentum theory is adopted to analyze the aerodynamic performance of the impeller, realizing the refined modeling of the aerodynamic characteristics of the impeller. The blade element theory divides the entire blade into several aerodynamic characteristic regions, namely blade elements, along the radial direction according to the airfoil characteristic parameters such as the chord length and thickness of the blade; the momentum theory obtains the relationships between the forces, torques, momentum, and angular momentum acting on the impeller of the unit based on the law of conservation of momentum. By combining the blade element theory and the momentum theory, the overall aerodynamic characteristics of the impeller can be obtained by the superposition of the aerodynamic characteristics of several blade elements.

[0140] Based on the dynamic wake meandering (DWM) model, a passive tracer model of the wake is established, and the passive tracer model is corrected for response based on the evolution process of wake deficit, wake meandering, the increase in turbulence intensity caused by the wake, as well as the wind speed decrease and radial expansion of the wake in the near-wake region to obtain the digital twin model of the wake characteristics.

[0141] Specifically, a passive tracer model of the wake is established based on the DWM dynamic wake meandering model, considering the evolution process of wake deficit, wake meandering, and the increase in turbulence intensity caused by the wake; on the basis of DWM, the wind speed decrease and radial expansion of the wake in the near-wake region (pressure gradient region) are considered, and response correction is carried out to improve the accuracy of the far wake, calculate the overlapping regions of all wakes, and merge the wake deficits in the overlapping regions.

[0142] In the DWM model, small turbulent eddies with an eddy scale smaller than twice the wind turbine diameter are used to calculate the evolution of wake deficit, and large turbulent eddies with an eddy scale larger than twice the wind turbine diameter are used to calculate wake meandering (which increases the turbulence intensity and enhances the wind turbine load).

[0143] Based on the root sum square (RSS) method, wake deficits are superimposed axially. The transverse component (radial wake deficit) is superimposed by vector sum. The RSS method assumes that the local kinetic energy of the axial deficit in the merged wake is equal to the sum of the local energies of the axial deficits of each wake at a given wind data point. The RSS method is only applicable to scalar arrays. This method is applicable to axial deficits because the overlapping wakes may have similar axials. Since any given radial direction depends on the azimuth angle in the axisymmetric coordinate system, a vector sum is applied to the transverse component (radial wake deficit).

[0144] The wind farm-level energy management digital twin model is used to dynamically adjust the power output of each wind turbine according to the real-time conditions of the wind farm and the grid demand, so as to optimize the energy distribution and operation efficiency of the entire wind farm.

[0145] Specifically, according to the active load command given by the power grid, combined with the power output in the wind farm under the current yaw control, the multi-objective grey wolf algorithm is used to optimize the power output distribution of each unit. With the allowable power generation as the control target, using the digital twin model and real-time data of the wind turbine generator set, comprehensively considering the fatigue load of the wind turbine generator set under different power outputs, the output of the unit is dynamically adjusted by means of pitch control, increasing or decreasing the power of a single unit, so that the total output power of the wind farm meets the requirements of the power grid command.

[0146] The digital twin model of field-level energy management comprehensively manages and configures the energy of the wind farm, executes the active load command given by the power grid, and achieves the purpose of automatically controlling the grid-connected power of the wind farm. With the maximum allowable power generation as the control target, through distribution calculation, the grid-connected load of the wind farm is scheduled and controlled, so that the grid-connected load of the wind farm is automatically controlled within the rated capacity. The digital twin model of field-level energy management of the wind farm can dynamically adjust the power output of each wind turbine generator set according to the real-time conditions of the wind farm and the grid demand, so as to optimize the energy distribution and operation efficiency of the entire wind farm. According to the active load command given by the power grid, combined with the power output in the wind farm under the current yaw control, the multi-objective grey wolf algorithm is used to optimize the power output distribution of each unit. With the allowable power generation as the control target, using the digital twin model and real-time data of the wind turbine generator set, comprehensively considering the fatigue load of the wind turbine generator set under different power outputs, the output of the unit is dynamically adjusted by means of pitch control, increasing or decreasing the power of a single unit, so that the total output power of the wind farm meets the requirements of the power grid command. <>

[0147] In practical applications, the maximum information coefficient (MIC) method can be used to analyze the correlation between the unit state parameters (such as rotational speed, torque, pitch angle, power generation, blade root load, drive shaft load, etc.) and the radar wind measurement data in the historical operation data of the wind farm respectively. Taking the MIC correlation of rotational speed - radar wind measurement data as an example, the data points are distributed on a two-dimensional plane, and the two-dimensional plane is refined according to the grid of i rows and j columns, and the maximum mutual correlation information MIC[x,y] is calculated:

[0148]

[0149] Among them, I(x,y) is the joint probability between variables x and y, a and b are the number of grids divided in the horizontal and vertical directions respectively, B is a preset variable, usually set to the 0.6th power of the data volume; p(x,y) is the joint probability distribution function of variables x and y, p(x) is the marginal probability density function of variable x, and p(y) is the marginal probability density function of variable y. Normalize the maximum mutual information, and calculate the maximum mutual information values of the two types of data when i and j take different values as the MIC value. Replace the original data of the unit state parameters, and select the unit state parameter with the largest MIC value, which is the parameter with the strongest correlation with the radar wind measurement data.

[0150] Select the data with the highest correlation as the input data, use the radar wind measurement data as the output data, and use the LSTM time series dynamic deep learning neural network to estimate the wind speed at x3 meters in front of the machine (the measurement distance of the lidar, a common value is 200 meters). The LSTM contains a network layer composed of a series of gated units. Each gated unit can selectively control the flow of information and can learn whether this information needs to be remembered or forgotten. The cells in the LSTM network are the core units of the LSTM. It is a special neural network unit used to process long-term dependencies in sequence data. Each LSTM cell consists of a cell state and three gates, namely the input gate, the forget gate, and the output gate. These three gates determine the transmission degree of the input information, historical information, and output information at the current moment through certain operations and activation functions, so as to realize the memory of historical information and the processing of current information.

[0151] Set parameters such as the maximum number of training epochs, learning rate strategy, initial learning rate, learning rate decay factor, learning rate decay period, and gradient threshold, and train to obtain a neural network load model. During the operation of the model, use the operating state information of the unit as the input to predict the input wind speed 200 meters in front of the machine.

[0152] Use the large eddy simulation method to generate a neutral atmospheric boundary layer without wind turbines, place the wind turbines in the neutral atmospheric boundary layer, and adopt the method of coupling the LES and the actuator line model to study the interaction between the wind turbines and the atmospheric boundary layer. The turbulence is divided into large-scale vortices and small-scale vortices through the filtering method. The large-scale vortices are directly solved by the N-S equation, and the small-scale vortices are simulated by the sub-grid model. The surface stress and temperature flux of the atmospheric boundary layer are simulated according to the surface roughness height.

[0153] The turbulent momentum equation is:

[0154]

[0155] In the formula, t is time, x i1 and y j1 are the spatial coordinates in the i1 and j1 directions respectively, u i1 and u j1 are the velocity components in the i1 and j1 directions respectively. i1 represents the horizontal direction, j1 represents the vertical direction, p is the dynamic pressure, z is the surface roughness, f e is the volume force of the fan, Ω j is the rotation rate tensor, ρ0 is the air density, p0 is the initial value of the dynamic pressure; ε jk If it is an odd permutation, the value is -1, if it is an even permutation, the value is +1, and the others are 0; u k is the velocity component in the k direction, τ ijis the subgrid-scale stress, g is the acceleration due to gravity, and p b is the air buoyancy density.

[0156] Based on the blade element momentum theory (BEM), the entire blade is divided into several aerodynamic characteristic regions according to the airfoil characteristic parameters such as the blade chord length, thickness, and pitch angle, and finite element analysis and modeling are carried out accordingly. At a length r along the axial direction of the blade, a microelement with a length of dr is taken and called a blade element, whose chord length is l and pitch angle is β; the blade is divided into different microsegments along the wind turbine radius, and the aerodynamic lift and drag generated by each microsegment under the action of the wind speed are determined by its airfoil. It is assumed that the air flow between the microelements does not interfere with each other, and the forces and torques acting on all the microelements can be synthesized without loss into the resultant force and total torque borne by the blade. Then, the overall torque borne by the blade can be obtained by superposing the forces on all the blade elements. Suppose the blade element is acted upon by an oncoming wind with a relative velocity of v, generating an aerodynamic force dF perpendicular to the chord. Decompose this aerodynamic force along the direction of the oncoming wind and the direction perpendicular to it to obtain the lift dL and drag dD of the blade element under the oncoming wind:

[0157]

[0158] where ρ0 is the air density, C l , C d are the lift coefficient and drag coefficient of the blade element, respectively, obtained through aerodynamic performance analysis based on its airfoil information, r is the length of the blade along the axial direction, l is the chord length, and v is the relative velocity of the oncoming wind. Decompose the aerodynamic force dF along the directions parallel and perpendicular to the impeller rotation plane to obtain the local impeller thrust dF a and torque dF u , expressed in terms of the lift dL and drag dD:

[0159]

[0160] where is the angle between the oncoming wind and the impeller rotation plane. By calculating the forces on each microsegment through aerodynamic methods, the aerodynamic torque borne by the entire wind turbine blade can be obtained by superposing the forces on each microsegment, and its expression is:

[0161]

[0162] In the formula, F r is the total impeller thrust, Q r is the total impeller torque, R is the impeller radius, R hub is the hub radius, and V total is the wind speed at the blade. The structural dynamics response of a wind turbine with N degrees of freedom is described by the assumed mode method, and its dynamic equation is established using Kane's dynamic equation:

[0163] F(r)+FR F(r) = 0

[0164] where F(r) and F R F(r) is the generalized active force and generalized inertial force acting on the wind turbine system. The inertial force F1 of the tower and blades is obtained according to the following formula:

[0165]

[0166] where ρ(l) and a(l) are the linear density and linear acceleration vector along the center line of the tower and the axial direction of the blade respectively, and L is the total length of the tower or blade.

[0167] The inertial force F2 of the nacelle and hub is obtained according to the following formula:

[0168] F2 = v2(-m2a2) + w2(-H)

[0169] where v2, w2 and a2 are the partial linear velocity, partial angular velocity and linear acceleration of the nacelle and hub in the inertial system, m2 is the mass of the nacelle and hub, and H is the differential of the angular momentum of the nacelle and hub about the center of mass in the inertial system.

[0170] The generalized active force F(r) on the wind turbine system consists of the aerodynamic force F aero (r), the gravitational force F G (r), the elastic force F elastic (r), etc.:

[0171] F(r) = F aero (r) + F G (r) + F elastic (r)

[0172] where F aero (r) is calculated by the following formula:

[0173] F aero (r) = v2F3 + w2M

[0174] where F3 and M are the aerodynamic force and pitching moment acting on the blade in the aerodynamic module respectively.

[0175] The blade and the tower are simplified as a flexible cantilever beam with two degrees of freedom. The blade is connected to the rigid hub, and the tower is connected to the ground. A coordinate system is established based on orthogonal unit vectors to define the rigid structure reference system. The coordinate transformation between the coordinate systems of each component can be realized by the coordinate transformation matrix. According to the distributed characteristic parameters of the tower (such as diameter, density and stiffness, etc.), it is described by n1 generalized coordinates from bottom to top along the center of the tower. The deformation U(z3,t) of any point on the cantilever beam at any time can be written as:

[0176]

[0177] where z3 is the position along the cantilever direction, t is the time, and φ i (z3) is the vibration mode of the cantilever, is the generalized coordinate of the tower top or blade root; m ij , k ij represent the generalized mass and generalized stiffness respectively; c j (t) is the generalized coordinate with respect to the modal vibration mode φ i (z3), is the first derivative of c j (t), is the second derivative of c j (t). The force condition and displacement of any point on the wind turbine can be expressed by the above formula.

[0178] In the DWM model, small turbulent eddies with an eddy scale smaller than twice the rotor diameter are used to calculate the evolution of wake deficit, and large turbulent eddies with an eddy scale larger than twice the rotor diameter are used to calculate wake meandering (which increases the turbulence intensity and wind turbine load). The rotor wind speed is low-pass filtered, and the evolution of wake deficit under quasi-steady conditions is simulated through the thin shear layer approximation. The thin shear layer approximation of the Reynolds-averaged N-S equations in axisymmetric coordinates is used to simulate the evolution of wake deficit, and the eddy viscosity formula is used to capture the turbulence closure. The thin shear layer approximation discards the pressure term and assumes that the velocity gradient in the radial direction is much larger than the velocity gradient in the axial direction. The analytical expressions for momentum conservation and mass conservation are as follows:

[0179]

[0180] where v T is the eddy viscosity, V x and V r are the axial and radial velocities in the axisymmetric coordinate system, r0 is the rotor diameter, and r is the length of the blade along the axial direction:

[0181]

[0182] where and are the axial and radial wake deficits.

[0183]

[0184] where C nearwake is a user-defined parameter that determines the degree of wake radial expansion in the near-wake region; is the wind speed perpendicular to the rotor after low-pass filtering; C t (r) is the mean thrust coefficient, derived from the blade element momentum theory, and V x (x,r) is the axial velocity, is the axial wake deficit, V r (x, r) is the radial velocity, is the radial wake deficit.

[0185] Based on the root sum square (RSS) method, the wake deficits are superimposed axially. The lateral component (radial wake deficit) is superimposed by vector sum. The RSS method assumes that the local kinetic energy of the axial deficit in the combined wake is equal to the sum of the local energies of the axial deficits of each wake at a given wind data point. The RSS method is only applicable to scalar arrays. This method is applicable to the axial deficit because overlapping wakes may have similar axials. Since any given radial direction depends on the azimuth angle in an axisymmetric coordinate system, a vector sum is applied to the lateral component (radial wake deficit).

[0186] In this application, a wind farm digital twin model is constructed based on the digital twin models of the inflow wind speed of the wind farm, the wind speed in front of the machine, the single machine, the wake characteristics, and the wind farm-level energy management twin model. Then, an unsupervised deep learning neural network is used to generate typical scenarios. A two-step optimization scheduling method of pre-scheduling and real-time optimization scheduling is adopted to determine the optimal yaw control strategy for different typical scenarios respectively, and the yaw angles of the wind turbines in each scenario are controlled based on the optimal yaw control strategy to reduce the load of the units and reduce power loss. The established wind farm digital twin model avoids the defect of low accuracy in engineering wake modeling, reduces the optimization complexity through dominant wind direction division and typical scenario generation, effectively optimizes the output power and unit load of the wind farm through active yaw control, improves the energy utilization rate, and helps to reduce the operation cost of the wind farm.

[0187] Embodiment 2

[0188] See Figure 4 as shown in Figure 4 shows a schematic structural diagram of a wind farm multi-objective collaborative optimization scheduling device based on digital twin provided by Embodiment 2 of the present invention, wherein the device includes:

[0189] The wind turbine grouping module 401 is used to divide the dominant wind direction of the wind farm according to the wind direction data of the wind farm; group the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction to obtain several wind turbine groups;

[0190] The typical scenario division module 402 is used to generate several wind power scenarios of the wind farm according to the wind speed data of the wind farm; reduce the generated several wind power scenarios to obtain the typical scenarios of the wind farm;

[0191] The pre-scheduling instruction determination module 403 is configured to, for each typical scenario, determine a pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm in this typical scenario meet the objective function based on the reinforcement learning algorithm;

[0192] The control parameter optimization module 404 is configured to use the pre-scheduling instruction as the initial value of real-time optimal scheduling and optimize the control parameters by using the distributed model predictive control algorithm;

[0193] The wind turbine control module 405 is configured to control the operation of each wind turbine based on the optimized control parameters in this typical scenario.

[0194] In an optional implementation, the method for dividing the dominant wind direction of the wind farm according to the wind direction data of the wind farm and clustering the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction to obtain a number of wind turbine groups includes:

[0195] Adopt a method based on hierarchical clustering to divide the dominant wind direction from the wind direction data;

[0196] Under the dominant wind direction, combining the specific layout of the wind farm, regard the wind turbines as the nodes of a directed graph, and regard the wake influence of the upstream unit on the downstream unit as the edge of the directed graph;

[0197] Based on the established directed graph above, use the community detection algorithm to describe the community structure of the wind farm.

[0198] In an optional implementation, the method for generating a number of wind power scenarios of the wind farm according to the wind speed data of the wind farm includes:

[0199] Use the randomly generated noise signal as the input random vector and the measured wind speed data of the wind farm as the sample data, and adopt the generative adversarial neural network for scenario generation to obtain a number of wind power scenarios.

[0200] In an optional implementation, the method for reducing the generated number of wind power scenarios to obtain the typical scenarios of the wind farm includes:

[0201] Use the unsupervised clustering method to reduce the generated number of wind power scenarios to obtain the typical scenarios of the wind farm;

[0202] In an optional implementation, for each typical scenario, the method for determining a pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm in this typical scenario meet the objective function includes:

[0203] Taking the output power and the unit load as the objective functions, and taking the loads of the key components of the wind turbine and the limit angle of the yaw control as the constraints, a reward mechanism is established;

[0204] According to the reward mechanism, through self-learning, the pre-scheduling instructions for each wind turbine are obtained, and the pre-scheduling of the yaw control of the wind farm is realized.

[0205] In an alternative embodiment, the constraints include the limit load limit and the yaw angle limit of the wind turbine;

[0206] Among them, the extreme value distribution theory is used to simulate and obtain the limit load limit, and the yaw angle limit condition is that the unit load does not exceed the limit load within the limit angle range.

[0207] In an alternative embodiment, taking the pre-scheduling instruction as the initial value of the real-time optimal scheduling, and using the distributed predictive control algorithm to optimize the control parameters, including:

[0208] When optimizing the control parameters, the above constraints are used as the constraints in the optimization process, the objective function is used as the optimization goal, and the overall optimization calculation of the wind farm is divided into parallel optimization calculations of multiple wind turbine subsystems, and the yaw scheduling instructions for each wind turbine are obtained in real time to realize the optimization of the control parameters.

[0209] In an alternative embodiment, the device includes a model construction module, and the model construction module is used for:

[0210] Soft-sensing the input wind speed through the unit state, and combining with the lidar wind measurement data for verification to construct a digital twin model of the inflow wind speed of the wind farm;

[0211] Calculating the propagation law of the fluid in front of the computer, and constructing a digital twin model of the wind speed in front of the machine according to the input wind speed of the wind farm;

[0212] Adopting the finite element modeling method, combining the blade element theory and the momentum theory to obtain the aerodynamic characteristics of the wind turbine, and using the multi-body dynamics method to calculate the structural response of the wind turbine to construct a single-machine digital twin model;

[0213] Constructing a digital twin model of the wake characteristics based on the dynamic meandering principle;

[0214] Combining the digital twin model of the inflow wind speed of the wind farm, the digital twin model of the wind speed in front of the machine, the single-machine digital twin model, the digital twin model of the wake characteristics and the digital twin model of the field-level energy management of the wind farm to construct the digital twin model of the wind farm.

[0215] In an alternative embodiment, the model construction module is specifically used for:

[0216] Select the data with the highest correlation with the radar wind measurement data among the unit status parameters as the input data, use the radar wind measurement data as the output data, and construct the digital twin model of the inflow wind speed of the wind farm by using the deep learning neural network method;

[0217] Use the large eddy simulation method to generate a neutral atmospheric boundary layer without wind turbines, place the wind turbines in the neutral atmospheric boundary layer, and adopt the method of coupling the LES with the actuator line model to study the interaction between the wind turbines and the atmospheric boundary layer, thereby constructing the digital twin model of the wind speed in front of the machine;

[0218] Adopt the blade element momentum theory to conduct aerodynamic performance analysis on the impeller, and construct the digital twin model of the single machine based on the results of the aerodynamic performance analysis;

[0219] Establish a passive tracer model of the wake based on the dynamic meandering wake model, and perform response correction on the passive tracer model based on the evolution process of the wake deficit, wake meandering, the increase in the turbulence intensity caused by the wake, as well as the wind speed decrease and radial expansion of the wake in the near-wake region to obtain the digital twin model of the wake characteristics;

[0220] The digital twin model of the wind farm field-level energy management is used to dynamically adjust the power output of each wind turbine according to the real-time conditions of the wind farm and the grid demand, so as to optimize the energy distribution and operation efficiency of the entire wind farm.

[0221] Embodiment III

[0222] Based on the same application concept, see Figure 5 as shown, Figure 5 shows a schematic structural diagram of a computer device provided in Embodiment III of the present invention. Among them, as Figure 5 shown, a computer device 500 provided in Embodiment III of the present application includes:

[0223] A processor 501, a memory 502, and a bus 503. The memory 502 stores machine-readable instructions executable by the processor 501. When the computer device 500 runs, communication is carried out between the processor 501 and the memory 502 through the bus 503. When the machine-readable instructions are run by the processor 501, the steps of the multi-objective collaborative optimization scheduling method for a wind farm based on digital twin shown in the above Embodiment I are executed.

[0224] Embodiment IV

[0225] Based on the same application concept, the present application embodiment also provides a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program is run by a processor, the steps of the multi-objective collaborative optimization scheduling method for a wind farm based on digital twin described in any one of the above embodiments are executed.

[0226] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the systems and devices described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0227] The computer program product for performing multi-objective collaborative optimization scheduling of a wind farm based on digital twins provided by an embodiment of the present invention includes a computer-readable storage medium storing program codes, and the instructions included in the program codes can be used to execute the methods described in the foregoing method embodiments. For specific implementation, reference can be made to the method embodiments, and will not be elaborated herein.

[0228] The multi-objective collaborative optimization scheduling device of a wind farm based on digital twins provided by an embodiment of the present invention can be specific hardware on a device or software or firmware installed on the device, etc. For the device provided by an embodiment of the present invention, the implementation principle and the technical effects produced are the same as those in the foregoing method embodiments. For a brief description, for the parts not mentioned in the device embodiment, reference can be made to the corresponding content in the foregoing method embodiments. Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the systems, devices, and units described above can all refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0229] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For another example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some communication interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0230] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place, or can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0231] In addition, each functional unit in the embodiments provided by the present invention can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0232] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0233] It should be noted that: similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. In addition, the terms "first", "second", "third", etc. are only used for descriptive distinction and cannot be understood as indicating or implying relative importance.

[0234] Finally, it should be noted that: the above-mentioned embodiments are only specific implementation manners of the present invention, used to illustrate the technical solutions of the present invention, rather than limiting it. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: any person skilled in the art within the technical scope disclosed by the present invention can still modify the technical solutions described in the foregoing embodiments, or can easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes, or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A multi-objective collaborative optimization scheduling method for a wind farm based on digital twin, characterized in that, The method includes: Dividing the dominant wind direction of the wind farm according to the wind direction data of the wind farm, and clustering the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction to obtain several wind turbine groups, so as to realize the unified optimization of the control strategies of associated wind turbines; Generating several wind power scenarios of the wind farm according to the wind speed data of the wind farm; reducing the generated several wind power scenarios to obtain the typical scenarios of the wind farm; For each typical scenario, determining a pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm meet the objective function based on the reinforcement learning algorithm, wherein the pre-scheduling instruction is a yaw scheduling instruction; Taking the pre-scheduling instruction as the initial value of real-time optimal scheduling, and optimizing the control parameters by using the distributed model predictive control algorithm; Under this typical scenario, controlling the operation of each wind turbine based on the optimized control parameters; The dividing the dominant wind direction of the wind farm according to the wind direction data of the wind farm, and clustering the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction to obtain several wind turbine groups includes: Adopting a method based on hierarchical clustering to divide the dominant wind direction from the wind direction data; Under the dominant wind direction, combining the specific layout of the wind farm, regarding the wind turbines as the nodes of a directed graph, and regarding the wake influence of the upstream unit on the downstream unit as the edge of the directed graph; Based on the established directed graph, using the community discovery algorithm to describe the community structure of the wind farm.

2. The method according to claim 1, characterized in that, The generating several wind power scenarios of the wind farm according to the wind speed data of the wind farm includes: Taking the randomly generated noise signal as the input random vector and the measured wind speed data of the wind farm as the sample data, and using the generative adversarial neural network for scenario generation to obtain several wind power scenarios.

3. The method according to claim 1, characterized in that The reducing the generated several wind power scenarios to obtain the typical scenarios of the wind farm includes: Using the unsupervised clustering method to reduce the generated several wind power scenarios to obtain the typical scenarios of the wind farm.

4. The method according to claim 1, wherein The determining a pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm meet the objective function based on the reinforcement learning algorithm for each typical scenario includes: Taking the output power and unit load as the objective function, and taking the key component loads of the wind turbines and the limit angles of yaw control as the constraint conditions, and establishing a reward mechanism; According to the reward mechanism, obtaining the pre-scheduling instructions of each wind turbine through self-learning to realize the pre-scheduling of the yaw control of the wind farm.

5. The method according to claim 4, characterized in that, The constraint conditions include the limit load limit and yaw angle limit of the wind turbines; Among them, the limit load limit is simulated by using the extreme value distribution theory, and the yaw angle limit condition is that the unit load does not exceed the limit load within the limit angle range.

6. The method according to claim 4, characterized in that, The taking the pre-scheduling instruction as the initial value of real-time optimal scheduling, and optimizing the control parameters by using the distributed model predictive control algorithm includes: When optimizing the control parameters, the above-mentioned constraint conditions are used as constraints in the optimization process, the above-mentioned objective function is used as the optimization goal, and the overall optimization calculation of the wind farm is divided into parallel optimization calculations of multiple wind turbine subsystems, and the yaw scheduling instructions of each wind turbine are obtained in real time to realize the optimization of the control parameters.

7. The method according to claim 1, characterized in that, The method includes: Soft-sensing the input wind speed through the unit state, and verifying it in combination with the lidar wind measurement data to construct a digital twin model of the inflow wind speed of the wind farm; Calculating the propagation law of the pre-turbine fluid, and constructing a digital twin model of the pre-turbine wind speed according to the input wind speed of the wind farm; Adopting the method of finite element modeling, combining the blade element theory and the momentum theory to obtain the aerodynamic characteristics of the wind turbine, and using the multi-body dynamics method to calculate the structural response of the wind turbine to construct a single-unit digital twin model; Constructing a digital twin model of wake characteristics based on the dynamic meandering principle; Combining the digital twin model of the inflow wind speed of the wind farm, the digital twin model of the pre-turbine wind speed, the single-unit digital twin model, the digital twin model of wake characteristics and the digital twin model of the wind farm-level energy management of the wind farm to construct the digital twin model of the wind farm.

8. The method according to claim 7, characterized in that, The method also includes: Selecting the data with the highest correlation with the lidar wind measurement data among the unit state parameters as the input data, using the lidar wind measurement data as the output data, and constructing the digital twin model of the inflow wind speed of the wind farm by using the deep learning neural network method; Using the large eddy simulation method to generate a neutral atmospheric boundary layer without wind turbines, placing the wind turbines in the neutral atmospheric boundary layer, and adopting the method of LES coupled with the actuator line model to study the interaction between the wind turbines and the atmospheric boundary layer, thereby constructing the digital twin model of the pre-turbine wind speed; Adopting the blade element momentum theory to analyze the aerodynamic performance of the impeller, and constructing the single-unit digital twin model based on the results of the aerodynamic performance analysis; Establishing a passive tracer model of the wake based on the dynamic meandering wake model, and correcting the response of the passive tracer model based on the evolution process of wake deficit, wake meandering, the increase in turbulence intensity caused by the wake, the wind speed decrease in the near wake region and the radial expansion of the wake to obtain the digital twin model of wake characteristics; The wind farm-level energy management digital twin model is used to dynamically adjust the power output of each wind turbine according to the real-time conditions of the wind farm and the grid demand, so as to optimize the energy distribution and operation efficiency of the entire wind farm.

9. A multi-objective collaborative optimization scheduling device for a wind farm based on digital twin, characterized in that, The device includes: A wind turbine grouping module, which is used to divide the dominant wind direction of the wind farm according to the wind direction data of the wind farm; group the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction to obtain several wind turbine groups, so as to realize the unified optimization of the control strategies of the associated wind turbines; A typical scenario division module, which is used to generate several wind power scenarios of the wind farm according to the wind speed data of the wind farm; reduce the generated several wind power scenarios to obtain the typical scenarios of the wind farm; A pre-scheduling instruction determination module, which is used to determine, for each typical scenario, a pre-scheduling instruction that can make the output power and unit load of the digital twin model of the wind farm meet the objective function based on the reinforcement learning algorithm, where the pre-scheduling instruction is a yaw scheduling instruction; A control parameter optimization module, which is used to use the pre-scheduling instruction as the initial value of real-time optimal scheduling and optimize the control parameters by using the distributed model predictive control algorithm; A wind turbine control module, which is used to control the operation of each wind turbine based on the optimized control parameters under this typical scenario; Dividing the dominant wind direction of the wind farm according to the wind direction data of the wind farm, and clustering the wind turbines in the wind farm according to the wake characteristics under the dominant wind direction to obtain a number of wind turbine groups, including: Using a hierarchical clustering-based method to divide the dominant wind direction from the wind direction data; Under the dominant wind direction, combining the specific layout of the wind farm, regarding the wind turbines as the nodes of a directed graph, and regarding the wake influence of the upstream unit on the downstream unit as the edge of the directed graph; Based on the established directed graph, using a community detection algorithm to describe the community structure of the wind farm.

Images