Space-time three-dimensional neural network calculation method based on flow field distribution characteristics of wind power plant

Abstract

The invention discloses a space-time three-dimensional neural network calculation method based on flow field distribution characteristics of a wind power plant, and belongs to the technical field of wind power. The method comprises the following steps: acquiring position data of an anemometer tower, a wind turbine generator and a wind speed measuring instrument and wind speed and wind direction data of corresponding positions by using a communication system at a set time resolution; constructing a space-time three-dimensional neural network calculation model of the flow field of the wind power plant according to the space position relation among the anemometer tower, the wind turbine generator and the wind speed measuring instrument and the space-time correlation among the wind speed and the wind direction; and based on the space-time three-dimensional neural network calculation model of the wind power plant flow field, providing a training strategy of the space-time three-dimensional neural network calculation model. The problem of unsafe and unstable power grid operation caused by unreasonable switching-out of the distributed photovoltaic power generation system in the prior art is solved. According to the space-time three-dimensional neural network model structure, a mathematical calculation formula considering the space-time distribution characteristics of neurons is provided, and a new calculation method is provided for accurate calculation of the space-time distribution characteristics of the flow field of the wind power plant.

Description

Technical Field

[0001] The invention belongs to the technical field of wind power, and in particular relates to a spatiotemporal three-dimensional neural network calculation method based on the flow field distribution characteristics of a wind farm. Background Art

[0002] With the national "carbon peak" and "carbon neutrality" goals, wind power generation has been developed and utilized on a large scale in recent years. Due to the natural randomness and uncertainty of wind power generation, accurate and real-time calculation of the spatiotemporal distribution characteristics of the wind farm flow field is of great help to the power prediction and optimized operation of the wind farm. However, for wind farms in complex terrain, the flow field distribution characteristics are affected by factors such as the incoming wind speed and direction, terrain, and the positional relationship between wind turbines, making it extremely difficult to accurately calculate the flow field distribution characteristics of wind farms in complex terrain.

[0003] At present, the calculation of flow field distribution characteristics of wind farms mainly adopts three methods: field measurement, numerical simulation and wind tunnel experiment. The field measurement method requires a lot of measurement equipment and measurement time. It is usually used in scientific research and is rarely used in actual engineering. The numerical simulation method combines the terrain and climate information of the wind farm and uses numerical simulation software to calculate the flow field distribution characteristics of the wind farm. The numerical simulation method has the advantages of low calculation cost and the ability to calculate the flow field distribution characteristics of the wind farm under all working conditions. However, the numerical simulation method has the disadvantages of low calculation accuracy and long calculation time. The wind tunnel experiment method compresses the field prototype according to a certain ratio, and then measures the flow field distribution characteristics of the wind farm through wind tunnel experiments. The measurement cost of the wind tunnel experiment is relatively low, but there are certain differences between the wind tunnel experiment environment and the real environment, which makes the wind tunnel measurement results differ greatly from the real flow field distribution.

[0004] The above three wind farm flow field distribution calculation methods can calculate the distribution characteristics of the wind farm flow field to a certain extent, but these three methods cannot describe the spatiotemporal variation characteristics of the wind farm flow field and cannot accurately calculate the spatiotemporal distribution characteristics of the wind farm flow field. Summary of the invention

[0005] The purpose of the present invention is to solve the problem of unsafe and unstable power grid operation caused by unreasonable cut-out of distributed photovoltaic power generation systems in the prior art; propose a spatiotemporal neural network calculation method based on the distribution characteristics of the wind farm flow field, and construct a spatiotemporal neural network model that can accurately describe the distribution characteristics of the wind farm flow field according to the spatial position relationship between the wind tower, wind turbine and wind speed meter and the spatiotemporal correlation between wind speed and wind direction; and propose a training strategy for the spatiotemporal neural network model for the constructed spatiotemporal neural network model. The constructed spatiotemporal neural network model is used to realize the accurate calculation of the spatiotemporal distribution characteristics of the wind farm flow field.

[0006] In order to solve the above technical problems, the present invention provides a spatiotemporal neural network calculation method based on the flow field distribution characteristics of a wind farm, comprising the following steps: S1. Using the communication system to obtain the location data of the wind tower, wind turbine and anemometer, as well as the wind speed and direction data at the corresponding locations at a set time resolution; S2. Based on the spatial position relationship between the wind tower, wind turbine and anemometer and the spatiotemporal correlation between wind speed and direction, a spatiotemporal neural network calculation model of the wind farm flow field is constructed; S21, determining the structure of the spatiotemporal three-dimensional neural network model according to the spatial position relationship between the wind farm wind tower, wind turbines and anemometer and the spatiotemporal correlation between wind speed and wind direction; S22. Based on the spatiotemporal three-dimensional neural network model structure, a spatiotemporal three-dimensional neural network calculation model of the wind farm flow field is constructed; S3. Based on the spatiotemporal three-dimensional neural network calculation model of the wind farm flow field, a training strategy for the spatiotemporal three-dimensional neural network calculation model is proposed.

[0007] Preferably, the spatiotemporal three-dimensional neural network model structure includes a plurality of neurons, the input and output of each neuron have time-varying characteristics and spatial correlation characteristics, and each neuron is connected to a plurality of neurons around it.

[0008] Preferably, the input of the neuron is specifically expressed as: ; In the formula, Represents the coordinate position in the spatiotemporal neural network model as The neurons in Wind speed and direction information input at all times; express The coordinate position of the moment flowing from the directly connected neuron is The wind speed and direction information of the neurons; among them, Quantity indication The time from the coordinate position is The neuron inflow coordinate position is The wind speed and direction information of the neuron, and the meanings of other wind speed and direction components are similar; The neurons directly connected to the coordinate position are The connection weights between neurons are expressed in The wind speed and direction information output by the neurons directly connected at the moment is The degree of influence of wind speed and direction information input by the neurons, The value range of is between [0,1]; Indicated in The time coordinate position is The neuron and its coordinate position are The connection weight of the neuron, and the meaning of other connection weight components are similar; here each neuron is directly connected to the 6 neurons around it, so the number of wind speed and direction information input by each neuron is 6, and the number of connection weights between each neuron and the surrounding neurons is also 6; The spatial distance of directly connected neurons to the coordinate position is The greater the spatial distance, the smaller the influence of the directly connected neuron on the neuron; Indicated in The wind speed and direction information output by the neurons directly connected at the moment is The influence of wind speed and direction information input by the neurons is The offset value at time.

[0009] Preferably, the spatiotemporal stereo neural network calculation model is specifically expressed as: ; In the formula, The coordinate position in the spatiotemporal neural network model is The neurons in The wind speed and direction information output at all times, Represents the coordinate position in the spatiotemporal neural network model as The neurons in Wind speed and direction information input at all times.

[0010] Preferably, the training strategy of the spatiotemporal stereo neural network calculation model includes a calculation strategy of the spatiotemporal stereo neural network model weight adjustment amount, which is specifically expressed as: ; In the formula, for The time coordinate position is The actual measured value of wind speed and direction at the location; is the learning rate of the weight, and its value range is [0,1]; is the neuron connection weight in The amount of adjustment at the moment.

[0011] Preferably, a spatiotemporal stereo neural network weight update formula is obtained according to the adjustment amount of the spatiotemporal stereo neural network weight.

[0012] As a preferred embodiment, the spatiotemporal stereo neural network weight update formula is specifically expressed as: ; In the formula, It is the momentum coefficient in the weight adjustment process. The value range of the momentum coefficient is [0,1].

[0013] Preferably, the training strategy of the spatiotemporal neural network calculation model also includes: predicting by calculating the spatiotemporal neural network model of the wind farm flow field and comparing it with the actual measurement results. If the difference between the two meets the requirements, the connection weights of the current wind farm flow field spatiotemporal neural network model are saved; if the difference between the two does not meet the requirements, the connection weights of each neuron in the spatiotemporal neural network model of the wind farm flow field are adjusted according to the calculation strategy of the spatiotemporal stereoscopic neural network model weight adjustment amount and the spatiotemporal stereoscopic neural network weight update formula, and then the spatiotemporal distribution characteristics of the wind farm flow field are recalculated until the difference between the calculated value of the spatiotemporal stereoscopic neural network model and the actual measurement value meets the requirements or the model training times are reached.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This scheme constructs a spatiotemporal neural network calculation model of the wind farm flow field based on the spatial position relationship between the wind tower, wind turbine and anemometer and the spatiotemporal correlation between wind speed and direction.

[0015] Most existing neural network models are two-dimensional models, which cannot accurately describe the three-dimensional spatiotemporal distribution characteristics of wind farm flow fields. For this reason, the present invention proposes a spatiotemporal three-dimensional neural network model for calculating the spatiotemporal distribution characteristics of wind farm flow fields based on the three-dimensional spatiotemporal distribution characteristics of wind farm flow fields. The spatiotemporal three-dimensional neural network model takes into account both the temporal correlation between neuronal data transmission and the spatial distance relationship between neurons. Based on the structure of the spatiotemporal three-dimensional neural network model, a mathematical calculation formula that takes into account the spatiotemporal distribution characteristics of neurons is proposed, which provides a new calculation method for the accurate calculation of the spatiotemporal distribution characteristics of wind farm flow fields.

[0016] 2. This scheme is based on the spatiotemporal three-dimensional neural network calculation model of the wind farm flow field, proposes a training strategy for the spatiotemporal three-dimensional neural network calculation model, and realizes the convergence of the spatiotemporal three-dimensional neural network calculation model.

[0017] The single neuron connection weights of the wind farm flow field spatiotemporal neural network calculation model include time parameters and space parameters. Conventional neuron weight adjustment strategies are not suitable for the adjustment of neuron weights in the spatiotemporal neural network model. To this end, the present invention proposes a spatiotemporal neural network model weight adjustment strategy based on the spatiotemporal neural network model structure and neuron input-output characteristics. In the spatiotemporal neural network model weight adjustment strategy, the influence of neuron time factors and spatial position relationships on weight changes is fully considered, and a neuron weight adjustment mathematical model including spatiotemporal factors is constructed to achieve rapid convergence of neuron weights, and the correctness of neuron weight convergence is verified through examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the system flow of the present invention; Figure 2 It is a schematic diagram of the spatiotemporal neural network computing model of the present invention; Figure 3 is a schematic diagram of the device of the present invention; Figure 4 This is a layout diagram of the virtual wind farm units of the present invention; Figure 5 It is a schematic diagram of the calculation results of the spatiotemporal distribution of the flow field in a wind farm according to the present invention; Figure 6 It is a schematic diagram of the arrangement of units of a wind farm of the present invention; Figure 7 This is a schematic diagram of flow field distribution characteristics of a wind farm at a wind speed of 8.5 m / s and a wind direction of 270° in an embodiment of the present invention; Figure 8 It is a schematic diagram comparing the calculated wind speed and the measured wind speed at the position of the fourth row of wind turbines calculated by the spatiotemporal neural network model under the condition of an incoming wind speed of 8.5 m / s and a wind direction of 270° according to an embodiment of the present invention; Fig. 9 This is a schematic diagram of flow field distribution characteristics of a wind farm at a wind speed of 8.5 m / s and a wind direction of 222° in an embodiment of the present invention; Fig.10 This is a schematic diagram comparing the calculated wind speed and the measured wind speed at the position of the 8th row of wind turbines calculated by the spatiotemporal neural network model under the conditions of an incoming wind speed of 8.5 m / s and a wind direction of 222° according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] Example 1: Figure 1 As shown, the spatiotemporal three-dimensional neural network calculation method based on the wind farm flow field distribution characteristics includes: (1) Obtain wind speed and direction data from wind towers, wind turbines and anemometers. The wind speed and direction data collection time resolution is 1 minute. The dedicated communication network for wind farms usually uses optical fiber communication, and can also use 5G or other communication networks for communication.

[0020] (2) Based on the spatial position relationship between the wind tower, wind turbines and anemometer and the spatiotemporal correlation between wind speed and direction, a spatiotemporal neural network model of the wind farm flow field is constructed. The construction process of the spatiotemporal neural network model is as follows: 1) According to the spatial position relationship between the wind farm wind tower, wind turbines and anemometer and the spatiotemporal correlation between wind speed and direction, the structure of the spatiotemporal neural network model is determined as follows: Figure 2 As shown. Figure 2In the example, each neuron represents a position point in space, the input of each neuron represents the wind speed and direction flowing into the position, and the output of each neuron represents the wind speed and direction flowing out of the position point. The reason why neurons are used to construct a spatiotemporal neural network model is that neurons can accurately describe the nonlinear mapping relationship between the output data of other neurons and the output data of this neuron, which corresponds to the influence of the wind speed and direction at other positions in the wind farm flow field on the wind speed and direction at a certain position point. In order to accurately calculate the spatiotemporal distribution characteristics of the wind farm flow field, the present invention uses neurons to construct a spatiotemporal neural network model, thereby realizing the accurate calculation of the spatiotemporal distribution characteristics of the wind farm flow field.

[0021] 2) By Figure 2 From the structure of the spatiotemporal neural network model, we can see that the wind speed and direction flowing into each neuron is related to the distance of the directly connected neurons and the time-varying characteristics of the wind speed and direction. Figure 2 It can be seen that the number of neurons directly connected to each neuron is at least 6 (upper, lower, front, back, left and right), and neurons at more positions can also be considered as needed (such as upper left, lower, upper front, lower front and other positions). For the convenience of explaining the problem, the present invention chooses to select 6 neurons directly connected to each neuron for calculation. Therefore, the wind speed and direction data flowing into the neuron comes from the 6 neurons directly connected to it; at the same time, the wind speed and direction information output from the neuron also flows to the 6 neurons directly connected to it. According to the correlation between the wind speed and direction at the position of each neuron and the wind speed and direction of other directly connected neurons, the input expression of each neuron is constructed as follows: (1) Represents the coordinate position in the spatiotemporal neural network model as The neurons in All wind speed and direction values ​​flowing in at all times. express The coordinate position of the moment flowing from the directly connected neuron is The wind speed and direction information of the neurons; among them, Quantity indication The time from the coordinate position is The neuron inflow coordinate position is The wind speed and direction information of the neuron, and the meanings of other wind speed and direction components are similar. The neurons directly connected to the coordinate position are The connection weights between neurons are expressed in The wind speed and direction information output by the neurons directly connected at the moment is The degree of influence of wind speed and direction information input by the neurons, The value range of is between [0, 1]; Indicated in The time coordinate position is The neuron and its coordinate position are The connection weight of the neuron, and the meaning of other connection weight components are similar. Each neuron is directly connected to the 6 neurons around it, so the number of wind speed and direction information input by each neuron is 6, and the number of connection weights between each neuron and the surrounding neurons is also 6. The neurons in the spatiotemporal neural network model are The offset value at time.

[0022] The spatial distance of directly connected neurons to the coordinate position is The greater the spatial distance, the smaller the influence of the directly connected neuron on the neuron.

[0023] 3) In order to accurately describe the The present invention adopts the hyperbolic tangent function to describe the quantitative relationship between the input and output of neurons in the spatiotemporal neural network. The transfer function expression between the input and output of neurons in the spatiotemporal neural network is shown in formula (2).

[0024] (2) The coordinate position in the spatiotemporal neural network model is The neurons in The wind speed and direction information output at all times, Represents the coordinate position in the spatiotemporal neural network model as The neurons in Wind speed and direction information input at all times.

[0025] (3) Based on the constructed spatiotemporal neural network calculation model of wind farm flow field, a strategy for adjusting the weights of the spatiotemporal neural network model is proposed to achieve rapid convergence of the spatiotemporal neural network calculation model.

[0026] 1) For each neuron in the spatiotemporal neural network model, when the difference between the wind speed and direction value calculated by the neuron at a certain moment and the actually measured wind speed and direction value is greater than a given threshold, the weight connected to the neuron needs to be adjusted so that the difference between the wind speed and direction value calculated by the neuron and the actually measured wind speed and direction value is less than the given threshold; 2) According to the adjustment principle of neural network weights, the adjustment amount of neuron weights is related to factors such as the size of the neuron calculation error, the derivative of the neuron transfer function with respect to the weight, the spatial distance between neurons, the neuron input and the weight learning rate. Therefore, the neuron weight adjustment formula can be obtained as shown in formula (3).

[0027] (3) In formula (3), Indicates the coordinate position is The neurons in The weight adjustment amount at the moment; is the actual measured wind speed and direction value; is the learning rate of the weight, and the value range of the learning rate is between [0, 1]; For neurons in The input at the moment; the derivative of the neuron transfer function with respect to the weight is .

[0028] 3) The number of other neurons directly connected to the neuron is 6. According to the coordinate positions of other neurons, The positional relationship of the neurons can be obtained by calculating the connection weight adjustment formula of the neurons as shown in formula (4).

[0029] (4) In formula (4), for The time coordinate position is The actual measured value of wind speed and direction at the location; is the learning rate of the weight, ranging from [0, 1]; is the neuron connection weight in The amount of adjustment at the moment.

[0030] 4) After calculating the size of the neuron weight adjustment, the connection weights of each neuron can be updated according to formula (5). In the process of updating the neuron weight, in order to balance the quantitative relationship between the original weight and the weight adjustment, the momentum coefficient is introduced The updating formula of the spatiotemporal stereo neural network weights is shown in formula (5).

[0031] (5) In formula (5), is the momentum coefficient in the weight adjustment process, and the value range of the momentum coefficient is [0, 1].

[0032] The single neuron connection weights of the wind farm flow field spatiotemporal neural network calculation model include time parameters and space parameters. Conventional neuron weight adjustment strategies are not suitable for the adjustment of neuron weights in the spatiotemporal neural network model. To this end, the present invention proposes a spatiotemporal neural network model weight adjustment strategy based on the spatiotemporal neural network model structure and neuron input-output characteristics. In the spatiotemporal neural network model weight adjustment strategy, the influence of neuron time factors and spatial position relationships on weight changes is fully considered, and a neuron weight adjustment mathematical model including spatiotemporal factors is constructed to achieve rapid convergence of neuron weights, and the correctness of neuron weight convergence is verified through examples.

[0033] On the other hand, the present application also proposes a wind farm flow field spatiotemporal distribution characteristics calculation device based on a spatiotemporal neural network, which is composed of a data acquisition and transmission system, a wind farm flow field spatiotemporal neural network model, a software operating system and related supporting software, hardware equipment and other components.

[0034] The data acquisition and transmission system is used to obtain the location data of the wind tower, wind turbine and anemometer, as well as the wind speed and direction data at the location; The spatiotemporal neural network model of wind farm flow field is used to propose a spatiotemporal neural network calculation model of wind farm flow field based on the spatial position relationship between wind towers, wind turbines and anemometers and the spatiotemporal correlation between wind speed and direction; Software operating systems and related supporting software and hardware equipment are relatively common existing carriers and will not be elaborated on in detail here.

[0035] In addition, the present application also proposes a method for constructing a wind farm flow field spatiotemporal distribution characteristic calculation device based on spatiotemporal neural network. The wind farm flow field spatiotemporal distribution characteristic calculation device based on spatiotemporal neural network is composed of a data acquisition and transmission system, a wind farm flow field spatiotemporal neural network model, a software operating system and related supporting software, hardware equipment and other components. The structure of the wind farm flow field spatiotemporal distribution characteristic calculation device based on spatiotemporal neural network is as follows: Figure 3 As shown, the specific steps are as follows: 1) Use the dedicated communication network of the wind farm or the 5G network to transmit the wind speed and direction data in the SCADA system of the wind turbine, the wind speed and direction data of the wind tower, and the wind speed and direction data of the anemometer to the support platform where the spatiotemporal neural network model of the wind farm flow field is located; 2) The wind farm flow field spatiotemporal neural network model calculates the spatiotemporal distribution characteristics of the wind farm flow field based on the historical wind speed and direction data on the support platform and the real-time wind speed and direction data transmitted; 3) Calculate the difference between the calculated value of the wind farm flow field spatiotemporal stereo neural network model and the actual measured value. If the difference meets the requirements, save the connection weight of the current wind farm flow field spatiotemporal stereo neural network model. If the difference between the calculated value of the spatiotemporal stereo neural network model and the actual measured value does not meet the requirements, adjust the connection weight of each neuron in the wind farm flow field spatiotemporal stereo neural network model according to formulas (3), (4) and (5), and then recalculate the spatiotemporal distribution characteristics of the wind farm flow field until the difference between the model calculated value and the actual measured value meets the requirements or reaches the number of model training times.

[0036] Embodiment 2: In the example analysis, two examples are used to verify the correctness and feasibility of the present invention.

[0037] The first example is a virtual wind farm, which contains 9 wind turbines. According to the Energy Industry Standard of the People's Republic of China, Technical Specifications for Micro-site Selection of Wind Farm Projects, NB / T 10103-2018, the row spacing of wind turbines should not be less than 3 times the diameter of the wind rotor, and the column spacing should not be less than 5 times the diameter of the wind rotor. Therefore, in this example, the north-south spacing (row spacing) between wind turbines is set to 4 times the diameter of the wind rotor (4D), and the east-west spacing (column spacing) is set to 7 times the diameter of the wind rotor (7D). Figure 4 shown.

[0038] Figure 5 This is the wind farm flow field distribution characteristic diagram calculated by the present invention. Figure 5 From the calculation results, it can be seen that when the incoming wind speed passes through the wind turbine, the wind speed decreases to form a wake wind speed. As the wake wind speed propagates downstream, the wake wind speed gradually recovers, which is consistent with the flow field distribution characteristics of the actual wind farm, indicating that the method proposed in this patent can correctly calculate the flow field distribution characteristics of the wind farm.

[0039] Figure 6 , Figure 7 , Figure 8 , Fig. 9 and Fig.10 The following are the unit arrangement and calculation results of a wind farm, where Figure 6 This is the layout of the turbines in a wind farm. The spacing between the turbines along the 270° direction (east-west direction) is 7 times the rotor diameter (7D), and the spacing along the 222° direction (northeast direction) is 9.4 times the rotor diameter (9.4D).

[0040] Figure 7 The wind field distribution characteristics of the wind farm calculated by the spatiotemporal neural network model under the wind speed of 8.5 m / s and wind direction of 270°. Figure 7It can be seen that the wind speed distribution at the first row of wind turbines is 8.5 m / s, which is consistent with the actual distribution of the wind farm flow field, because the first row of wind turbines is not disturbed by the wake of other wind turbines, and the wind speed at its location is the incoming wind speed (wind speed 8.5 m / s). For the second row and other rows of wind turbines, due to the influence of the wake of the previous wind turbines, their flow field speed is significantly lower than the incoming wind speed, which is consistent with the actual distribution characteristics of the wind farm flow field, indicating that the spatiotemporal neural network model can correctly calculate the spatiotemporal distribution characteristics of the wind farm flow field. Figure 8 Comparison between the calculated wind speed and the measured wind speed at the location of the 4th row of wind turbines calculated by the spatiotemporal neural network model when the incoming wind speed is 8.5m / s and the wind direction is 270°. Figure 8 It can be seen that the maximum error between the calculated wind speed and the measured wind speed of the space-time stereo neural network model is 0.102 m / s, indicating that the space-time stereo neural network model can accurately calculate the flow field distribution characteristics of the wind farm.

[0041] Fig. 9 The wind farm flow field distribution characteristics under wind speed of 8.5 m / s and wind direction of 222° are shown in Figure 2. Fig. 9 It can be seen that when the incoming wind direction is 222°, the wind turbines in the first column and the last row are not affected by the wake of other wind turbines, and the wind speed at their location is the incoming wind speed (wind speed 8.5 m / s). For wind turbines at other locations, due to the influence of the wake of the previous wind turbines, their flow field speed is significantly lower than the incoming wind speed, which is consistent with the real distribution characteristics of the wind farm flow field, indicating that the spatiotemporal three-dimensional neural network model can correctly calculate the spatiotemporal distribution characteristics of the wind farm flow field. Fig.10 The eighth row is calculated by the spatiotemporal neural network model under the wind speed of 8.5 m / s and the wind direction of 222° (see Figure 6 ) Comparison of the calculated wind speed and measured wind speed at the wind turbine location. Fig.10 It can be seen that the maximum error between the calculated wind speed and the measured wind speed of the space-time stereo neural network model is 0.126m / s, which further proves that the space-time stereo neural network model can accurately calculate the flow field distribution characteristics of the wind farm.

Claims

1. A spatiotemporal neural network calculation method based on the flow field distribution characteristics of a wind farm, characterized in that: The steps include: S1. Using the communication system to obtain the location data of the wind tower, wind turbine and anemometer, as well as the wind speed and direction data at the corresponding locations at a set time resolution; S2. Based on the spatial position relationship between the wind tower, wind turbine and anemometer and the spatiotemporal correlation between wind speed and direction, a spatiotemporal neural network calculation model of the wind farm flow field is constructed; S21, determining the structure of the spatiotemporal three-dimensional neural network model according to the spatial position relationship between the wind farm wind tower, wind turbines and anemometer and the spatiotemporal correlation between wind speed and wind direction; S22. Based on the spatiotemporal three-dimensional neural network model structure, a spatiotemporal three-dimensional neural network calculation model of the wind farm flow field is constructed; S3. Based on the spatiotemporal three-dimensional neural network calculation model of the wind farm flow field, a training strategy for the spatiotemporal three-dimensional neural network calculation model is proposed.

2. The spatiotemporal neural network calculation method based on the wind farm flow field distribution characteristics according to claim 1 is characterized in that: The spatiotemporal three-dimensional neural network model structure includes a plurality of neurons, the input and output of each neuron have time-varying characteristics and spatial correlation characteristics, and each neuron is connected to a plurality of neurons around it.

3. The spatiotemporal neural network calculation method based on the wind farm flow field distribution characteristics according to claim 2 is characterized in that: The input of the neuron is specifically expressed as: ； In the formula, Represents the coordinate position in the spatiotemporal neural network model as The neurons in Wind speed and direction information input at all times; express The coordinate position of the moment flowing from the directly connected neuron is The wind speed and direction information of the neurons; among them, Quantity indication The time from the coordinate position is The neuron inflow coordinate position is The wind speed and direction information of the neuron, and the meanings of other wind speed and direction components are similar; The neurons directly connected to the coordinate position are The connection weights between neurons are expressed in The wind speed and direction information output by the neurons directly connected at the moment is The degree of influence of wind speed and direction information input by the neurons; Indicated in The time coordinate position is The neuron and its coordinate position are The connection weight of the neuron, and the meaning of other connection weight components are similar; The spatial distance of directly connected neurons to the coordinate position is The degree of influence of neurons; Indicated in The wind speed and direction information output by the neurons directly connected at the moment is The influence of wind speed and direction information input by the neurons is The offset value at time.

4. The spatiotemporal neural network calculation method based on the wind farm flow field distribution characteristics according to claim 3 is characterized in that: The spatiotemporal stereo neural network calculation model is specifically expressed as: ； In the formula, The coordinate position in the spatiotemporal neural network model is The neurons in The wind speed and direction information output at all times, Represents the coordinate position in the spatiotemporal neural network model as The neurons in Wind speed and direction information input at all times.

5. The spatiotemporal neural network calculation method based on the wind farm flow field distribution characteristics according to claim 1 is characterized in that: The training strategy of the spatiotemporal stereo neural network computing model includes a computing strategy for the spatiotemporal stereo neural network model weight adjustment amount, which is specifically expressed as: ； In the formula, for The time coordinate position is The actual measured value of wind speed and direction at the location; is the learning rate of the weight, ranging from [0,1]. is the neuron connection weight in The amount of adjustment at the moment.

6. The spatiotemporal neural network calculation method based on the wind farm flow field distribution characteristics according to claim 5 is characterized in that: According to the adjustment amount of the space-time stereo neural network weight, the space-time stereo neural network weight update formula is obtained.

7. The spatiotemporal neural network calculation method based on the wind farm flow field distribution characteristics according to claim 6 is characterized in that: The weight update formula of the spatiotemporal stereo neural network is specifically expressed as: ； In the formula, It is the momentum coefficient in the weight adjustment process. The value range of the momentum coefficient is [0,1].

8. The spatiotemporal neural network calculation method based on the wind farm flow field distribution characteristics according to claim 1 is characterized in that: The training strategy of the spatiotemporal neural network calculation model also includes: predicting by calculating the spatiotemporal neural network model of the wind farm flow field and comparing it with the actual measurement results. If the difference between the two meets the requirements, the connection weights of the current wind farm flow field spatiotemporal neural network model are saved; if the difference between the two does not meet the requirements, the connection weights of each neuron in the spatiotemporal neural network model of the wind farm flow field are adjusted according to the calculation strategy of the spatiotemporal stereoscopic neural network model weight adjustment amount and the spatiotemporal stereoscopic neural network weight update formula, and then the spatiotemporal distribution characteristics of the wind farm flow field are recalculated until the difference between the calculated value of the spatiotemporal stereoscopic neural network model and the actual measurement value meets the requirements or the model training times are reached.

Images