A method, system, electronic device, and storage medium for wind farm regulation based on neural networks.

Abstract

This invention belongs to the field of wind power generation technology and provides a wind farm regulation method, system, electronic device, and storage medium based on neural networks. The method includes: raw dataset acquisition, regulation command evaluation and classification, neural network model training, and wind farm regulation model regulation. This invention evaluates the regulation commands generated by the preset regulation strategy by using two adjacent sets of data in the raw dataset, realizing real-time analysis of the preset regulation strategy. The generated dataset provides a foundation for subsequent model training based on historical effects. By using labeled data to train the pre-built neural network model, the judgment logic of the command is optimized, improving the model's dynamic regulation capability for different wavebands, the degree of development of historical regulation data experience, and the stability of system operation.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and in particular to a wind farm control method, system, electronic device and storage medium based on neural networks. Background Technology

[0002] Existing wind power regulation technology is centered on primary frequency regulation and inertial response, employing hierarchical control based on grid frequency variation indicators, combined with turbine self-regulation and energy storage. Real-time data collection of voltage, current, and grid frequency at the wind farm's grid connection point is used, with frequency variation and frequency variation rate as the core indicators for selecting the regulation mode. When the frequency variation exceeds the dead zone, primary frequency regulation is executed, power commands are calculated using formulas, and the power of individual wind turbines is controlled via the energy management platform. When the frequency variation rate exceeds the threshold, inertial response is superimposed, or adaptive operating conditions are applied. When frequency fluctuations are small, turbine self-regulation is used; when fluctuations are excessive, energy storage compensation is activated.

[0003] However, existing technologies only classify operating conditions using fixed thresholds, which can only distinguish simple scenarios and cannot accurately match the different levels of disturbance requirements of the power grid; the regulation amount of existing technologies relies entirely on fixed mathematical formulas for calculation, and the formula parameters are mostly preset fixed values, which cannot dynamically adapt to real-time operating conditions; existing technologies only focus on whether the regulation is executed, without establishing a quantitative evaluation system for the regulation effect, and cannot optimize the subsequent regulation logic based on historical effects, and do not make full use of the experience in historical regulation data. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a wind farm regulation method, system, electronic device and storage medium based on neural network, which solves the problems of poor dynamic regulation capability, fixed regulation logic and inability to learn from historical data experience in the prior art.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A wind farm regulation method based on neural networks includes:

[0007] The target wind farm is regulated using a preset control strategy, and the operation data of the grid connection point and the regulation instructions generated by the control strategy are collected during the regulation process to obtain the original dataset.

[0008] The control instructions are evaluated and classified based on the operational data at adjacent times in the original dataset to obtain a labeled dataset; the labeled dataset includes: timestamp, operational data, control instructions, evaluation level, and operating condition status;

[0009] The pre-constructed neural network model is trained using the labeled dataset to obtain a wind farm control model;

[0010] The newly collected operating data from the grid connection point is input into the wind farm control model to generate instructions, thereby obtaining optimized control instructions, which are then updated to the target wind farm.

[0011] Preferably, a preset control strategy is used to regulate the target wind farm, and the operating data of the grid connection point and the regulation commands generated by the control strategy are collected during the regulation process to obtain the original dataset, including:

[0012] The operating data of the grid connection point is collected; the operating data includes: electrical parameters, frequency parameters, and status parameters; the electrical parameters include: voltage data and current data; the frequency parameters include: actual grid frequency, frequency change, and frequency change rate; the status parameters include: initial power of the wind farm, rated power, wind turbine speed, pitch angle, energy storage margin, and wind turbine power upsizing margin.

[0013] Set the primary frequency modulation dead zone, mild disturbance threshold, moderate disturbance threshold, and severe disturbance threshold;

[0014] When the frequency change is greater than the primary frequency regulation dead zone, the state of the target wind farm is determined to be a frequency regulation trigger state.

[0015] After determining the frequency modulation trigger state, if the frequency change rate is less than or equal to the mild disturbance threshold, mild disturbance control is applied to the target wind farm; if the frequency change rate is between the mild disturbance threshold and the moderate disturbance threshold, moderate disturbance control is applied to the target wind farm; and if the frequency change rate is greater than the severe disturbance threshold, severe disturbance control is applied to the target wind farm.

[0016] Preferably, a preset control strategy is used to regulate the target wind farm, and the operating data of the grid connection point and the regulation commands generated by the control strategy are collected during the regulation process to obtain the original dataset, including:

[0017] The adjustment direction is determined based on the sign of the frequency change. If the sign is positive, the adjustment direction is to reduce the power; otherwise, it is to increase the power.

[0018] When performing mild disturbance control, the primary frequency regulation adjustment and virtual inertia adjustment are calculated based on the operating data to obtain the total mild disturbance regulation. This total mild disturbance regulation is then updated to the wind turbines of the target wind farm according to the adjustment direction. The expression for the primary frequency regulation is: The expression for the virtual inertia adjustment is: ;in, , These are the primary frequency modulation adjustment amount and the virtual inertia adjustment amount, respectively; This refers to the rated power of the wind farm. The rated frequency of the power grid; This refers to the actual frequency of the power grid. This is a primary frequency modulation dead zone; This refers to the vortex rate of a single frequency modulation (FM). The equivalent inertial time constant of the wind farm; The rate of change of frequency;

[0019] When performing moderate disturbance control, the wind turbine speed is reduced or the pitch angle is narrowed to obtain reserve power. The total adjustment amount of the mild disturbance and the reserve power are integrated to obtain the total adjustment amount of the moderate disturbance. If the energy storage margin is greater than or equal to 10%, the output power of the wind turbine is adjusted to the target power using the energy storage module in the target wind farm. If the energy storage margin is less than 10%, the power of the wind turbine is adjusted until the output power of the wind turbine is adjusted to the target power.

[0020] Preferably, a preset control strategy is used to regulate the target wind farm, and the operating data of the grid connection point and the regulation commands generated by the control strategy are collected during the regulation process to obtain the original dataset, including:

[0021] During severe disturbance regulation, the extreme adjustment amount of primary frequency regulation is calculated based on the operating data, and an emergency compensation amount for energy storage is set. The extreme adjustment amount of primary frequency regulation and the emergency compensation amount for energy storage are integrated to obtain the total adjustment amount for severe disturbance. If the adjustment direction is power increase and the wind turbine unit is unrestricted load, the inertia regulation capability of each wind turbine in the wind turbine unit is calculated, and the total adjustment amount for severe disturbance is allocated to each wind turbine according to the calculated inertia regulation capability to obtain an active power command. If the wind turbine unit is at full load, the emergency compensation amount for energy storage is provided using the energy storage module, and the extreme adjustment amount of primary frequency regulation is allocated to the wind turbine unit. The expression for the extreme adjustment amount of primary frequency regulation is: The expression for the active power command is: ;in, , These are the primary frequency modulation extreme adjustment amount and the active power command, respectively; It is the sum of the primary frequency regulation extreme adjustment amount and the energy storage emergency compensation amount;

[0022] The control commands during the mild disturbance control, moderate disturbance control, and severe disturbance control processes are collected, and the control commands and the operating data are integrated to obtain the raw data.

[0023] Preferably, the control instructions are evaluated and classified based on the operational data at adjacent time points in the original dataset to obtain a labeled dataset, including:

[0024] Construct an accuracy evaluation formula; the expression of the accuracy evaluation formula is: ;in, Power deviation rate; , These are the actual power and the target power, respectively.

[0025] Construct a velocity evaluation formula; the expression of the velocity evaluation formula is: ;in, For response time; , These are the time when the power target is met and the time when the instruction is issued, respectively.

[0026] Construct a stability evaluation formula; the expression of the stability evaluation formula is: ;in, For frequency recovery deviation; This is the average frequency value of the power grid when it transitions from transient fluctuations to steady-state equilibrium.

[0027] Construct a supporting evaluation formula; the expression of the supporting evaluation formula is: ;in, It is the generalized inertia constant; This refers to the per-unit value of the energy released or stored by the fan rotor during the control process; for The per-unit value of the power grid frequency at any given time; For time intervals;

[0028] The accuracy evaluation formula, the velocity evaluation formula, the stability evaluation formula, and the support evaluation formula are used to evaluate each group of data in the original dataset to obtain an evaluation value;

[0029] The evaluation values ​​are classified using a preset threshold judgment table and an evaluation level judgment table to obtain the label dataset; the label dataset includes: first to fourth levels.

[0030] Preferably, the pre-constructed neural network model is trained using the labeled dataset to obtain a wind farm control model, which includes:

[0031] Construct an input layer; the data types of the input layer include: the operating data, operating condition classification identifiers, historical control instructions, and the tag dataset;

[0032] A condition-aware dual-attention layer is constructed; the condition-aware dual-attention layer includes: a time-step attention layer and a parameter attention layer connected to each other; the expression of the time-step attention layer is: The expression for the parameter attention layer is: ;in, Attention weights for each time step; Let be the weight matrix for the attention at the given time step; For the first Continuously run data vectors; For the first Real-time operating condition classification labels; Indicates feature splicing; The bias term for the attention at the given time step; Represents the normalization function; For attention weights; The attention weight matrix is ​​a parameter. for Time of the first The weighted values ​​of each feature; To assess the weighting of grades; For attention bias terms; It is a non-linear activation function;

[0033] Construct a domain-specific LSTM layer; the domain-specific LSTM layer includes: sequentially connected embedding units, three parallel LSTM units, and a hidden state fusion unit; the expression of the embedding unit includes: , , The expression for the hidden state fusion unit is: ; To embed feature vectors, The value range is 1, 2, 3. Embedded in the electrical domain, For frequency domain embedding, Embedded for state domain; for Voltage data at any given time; for Current data at any given time; for The actual frequency of the power grid at any given time; for Change in frequency over time; for Rate of change of frequency at any given time; for Electric field power data at any given time; for Constant fan speed; for Pitch angle at any moment; for Energy storage margin at all times; for Real-time historical power adjustment commands; for Real-time historical speed adjustment command; For the first Domain operating condition coding; For the first Domain LSTM hidden state; For the first Domain LSTM output gate; This represents the element-wise product operation; For the first Domain LSTM cell state; It is the hyperbolic tangent function; For the first Domain embedding feature vectors;

[0034] Construct an instruction generation layer; the expression for the instruction generation layer is: ;in, ; For constraints; Generate intermediate feature vectors for the instruction layer; To correct the linear unit; Generate layer weight matrices for instructions; This is the overall hidden state of the merged system; Generate layer bias terms for instructions;

[0035] The neural network model is obtained by fusing the input layer, the condition-aware dual-attention layer, the domain-specific LSTM layer, and the instruction generation layer; wherein the input layer, the condition-aware dual-attention layer, the domain-specific LSTM layer, and the instruction generation layer are connected in sequence.

[0036] Preferably, the pre-constructed neural network model is trained using the labeled dataset to obtain a wind farm control model, including:

[0037] Construct the total loss function; the expression for the total loss function is: ;in, For the first The evaluation level of each sample; To assess the weighting of the grades, when When the value is 1 or 2 Greater than 0, when When the value is 3 or 4 Less than 0; This represents the total loss value of the model. The total number of training samples; The first generated for the model Individual sample control instructions; For the first The actual regulatory instructions for each sample;

[0038] The total loss function is used to iteratively train the neural network model to obtain the wind farm control model.

[0039] Preferably, a wind farm control system based on a neural network includes:

[0040] The data acquisition module is used to regulate the target wind farm using a preset control strategy, and to collect the operation data of the grid connection point and the regulation instructions generated by the control strategy during the regulation process to obtain the raw dataset.

[0041] The data labeling module is used to evaluate and classify the control instructions based on the operating data at adjacent times in the original dataset to obtain a labeled dataset; the labeled dataset includes: timestamp, operating data, control instructions, evaluation level, and operating condition status;

[0042] The model training module is used to train the pre-built neural network model using the labeled dataset to obtain the wind farm control model.

[0043] The optimization and control module is used to input the newly collected operating data from the grid connection point into the wind farm control model to generate instructions, obtain optimized control instructions, and update the optimized control instructions to the target wind farm.

[0044] Preferably, an electronic device includes: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned neural network-based wind farm control method.

[0045] Preferably, a non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the aforementioned neural network-based wind farm control method.

[0046] The present invention discloses the following technical effects:

[0047] This invention provides a wind farm regulation method, system, electronic device, and storage medium based on neural networks. By evaluating regulation commands generated by a preset regulation strategy using two adjacent sets of data in the original dataset, it solves the deficiency of existing technologies in not establishing a quantitative evaluation of regulation effects and realizes real-time analysis of preset regulation strategies. By training a pre-built neural network model using labeled data, it solves the problems of poor dynamic regulation capability, fixed regulation logic, and inability to learn from historical data experience in existing technologies, and realizes the learning of historical operation control data and optimization of regulation commands based on evaluation data. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A schematic diagram of a wind farm control process based on a neural network provided for an embodiment of the present invention;

[0050] Figure 2 This is a schematic diagram of the operation flow of the preset control strategy provided in the embodiments of the present invention;

[0051] Figure 3 This is a schematic diagram of the label dataset construction process provided in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the neural network model construction process provided in an embodiment of the present invention;

[0053] Figure 5 This is a schematic diagram of the model training process provided in an embodiment of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The purpose of this invention is to provide a wind farm regulation method, system, electronic device and storage medium based on neural networks, which solves the problems of poor dynamic regulation capability, fixed regulation logic and inability to learn from historical data experience in the prior art.

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Figure 1 A schematic diagram of a wind farm control process based on a neural network provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this invention provides a wind farm regulation method based on neural networks, comprising:

[0058] Step 100: Regulate the target wind farm using a preset control strategy, and collect the operation data of the grid connection point and the regulation instructions generated by the control strategy during the regulation process to obtain the original dataset;

[0059] Step 200: Evaluate and classify the control instructions based on the operational data at adjacent time points in the original dataset to obtain a label dataset; the label dataset includes: timestamp, operational data, control instructions, evaluation level, and operating condition status;

[0060] Step 300: Train the pre-built neural network model using the labeled dataset to obtain the wind farm control model;

[0061] Step 400: Input the newly collected operating data from the grid connection point into the wind farm control model to generate instructions, obtain optimized control instructions, and update the optimized control instructions to the target wind farm.

[0062] refer to Figure 2 The target wind farm is regulated using a preset control strategy, and the operation data of the grid connection point and the regulation commands generated by the control strategy are collected during the regulation process to obtain the original dataset, including:

[0063] Step 101: Collect the operating data of the grid connection point; the operating data includes: electrical parameters, frequency parameters, and status parameters; the electrical parameters include: voltage data and current data; the frequency parameters include: actual grid frequency, frequency change, and frequency change rate; the status parameters include: initial power of the wind farm, rated power, wind turbine speed, pitch angle, energy storage margin, and wind turbine power adjustment margin.

[0064] Step 102: Set the primary frequency modulation dead zone, mild disturbance threshold, moderate disturbance threshold, and severe disturbance threshold;

[0065] Step 103: When the frequency change is greater than the primary frequency regulation dead zone, the state of the target wind farm is determined to be the frequency regulation trigger state;

[0066] Step 104: After determining the frequency modulation trigger state, if the frequency change rate is less than or equal to the mild disturbance threshold, then mild disturbance control is performed on the target wind farm; if the frequency change rate is between the mild disturbance threshold and the moderate disturbance threshold, then moderate disturbance control is performed on the target wind farm; if the frequency change rate is greater than the severe disturbance threshold, then severe disturbance control is performed on the target wind farm.

[0067] Furthermore, the target wind farm is regulated using a preset control strategy, and the operation data of the grid connection point and the regulation commands generated by the control strategy are collected during the regulation process to obtain the original dataset, including:

[0068] The adjustment direction is determined based on the sign of the frequency change. If the sign is positive, the adjustment direction is to reduce the power; otherwise, it is to increase the power.

[0069] When performing mild disturbance control, the primary frequency regulation adjustment and virtual inertia adjustment are calculated based on the operating data to obtain the total mild disturbance regulation. This total mild disturbance regulation is then updated to the wind turbines of the target wind farm according to the adjustment direction. The expression for the primary frequency regulation is: The expression for the virtual inertia adjustment is: ;in, , These are the primary frequency modulation adjustment amount and the virtual inertia adjustment amount, respectively; This refers to the rated power of the wind farm. The rated frequency of the power grid; This refers to the actual frequency of the power grid. This is a primary frequency modulation dead zone; This refers to the vortex rate of a single frequency modulation (FM). The equivalent inertial time constant of the wind farm; The rate of change of frequency;

[0070] When performing moderate disturbance control, the wind turbine speed is reduced or the pitch angle is narrowed to obtain reserve power. The total adjustment amount of the mild disturbance and the reserve power are integrated to obtain the total adjustment amount of the moderate disturbance. If the energy storage margin is greater than or equal to 10%, the output power of the wind turbine is adjusted to the target power using the energy storage module in the target wind farm. If the energy storage margin is less than 10%, the power of the wind turbine is adjusted until the output power of the wind turbine is adjusted to the target power.

[0071] Specifically, a preset control strategy is used to regulate the target wind farm, and the operation data of the grid connection point and the regulation commands generated by the control strategy are collected during the regulation process to obtain the original dataset, including:

[0072] During severe disturbance regulation, the extreme adjustment amount of primary frequency regulation is calculated based on the operating data, and an emergency compensation amount for energy storage is set. The extreme adjustment amount of primary frequency regulation and the emergency compensation amount for energy storage are integrated to obtain the total adjustment amount for severe disturbance. If the adjustment direction is power increase and the wind turbine unit is unrestricted load, the inertia regulation capability of each wind turbine in the wind turbine unit is calculated, and the total adjustment amount for severe disturbance is allocated to each wind turbine according to the calculated inertia regulation capability to obtain an active power command. If the wind turbine unit is at full load, the emergency compensation amount for energy storage is provided using the energy storage module, and the extreme adjustment amount of primary frequency regulation is allocated to the wind turbine unit. The expression for the extreme adjustment amount of primary frequency regulation is: The expression for the active power command is: ;in, , These are the primary frequency modulation extreme adjustment amount and the active power command, respectively; It is the sum of the primary frequency regulation extreme adjustment amount and the energy storage emergency compensation amount;

[0073] The control commands during the mild disturbance control, moderate disturbance control, and severe disturbance control processes are collected, and the control commands and the operating data are integrated to obtain the raw data.

[0074] refer to Figure 3 The control instructions are evaluated and classified based on the operational data at adjacent time points in the original dataset to obtain a labeled dataset, including:

[0075] Step 201: Construct the accuracy evaluation formula; the expression of the accuracy evaluation formula is: ;in, Power deviation rate; , These are the actual power and the target power, respectively.

[0076] Step 202: Construct a velocity evaluation formula; the expression of the velocity evaluation formula is: ;in, For response time; , These are the time when the power target is met and the time when the instruction is issued, respectively.

[0077] Step 203: Construct a stability evaluation formula; the expression of the stability evaluation formula is: ;in, For frequency recovery deviation; This is the average frequency value of the power grid when it transitions from transient fluctuations to steady-state equilibrium.

[0078] Step 204: Construct the support evaluation formula; the expression of the support evaluation formula is: ;in, It is the generalized inertia constant; This refers to the per-unit value of the energy released or stored by the fan rotor during the control process; for The per-unit value of the power grid frequency at any given time; For time intervals;

[0079] Step 205: Evaluate each group of data in the original dataset using the accuracy evaluation formula, the velocity evaluation formula, the stability evaluation formula, and the support evaluation formula to obtain an evaluation value;

[0080] Step 206: Classify the evaluation values ​​using a preset threshold judgment table and evaluation level judgment table to obtain the label dataset; the label dataset includes: first to fourth levels.

[0081] refer to Figure 4 The pre-built neural network model is trained using the labeled dataset to obtain a wind farm control model, which includes:

[0082] Step 301: Construct the input layer; the data types of the input layer include: the operating data, operating condition classification identifiers, historical control instructions, and the tag dataset;

[0083] Step 302: Construct a condition-aware dual attention layer; the condition-aware dual attention layer includes: a time-step attention layer and a parameter attention layer connected to each other; the expression for the time-step attention layer is: The expression for the parameter attention layer is: ;in, Attention weights for each time step; Let be the weight matrix for the attention at the given time step; For the first Continuously run data vectors; For the first Real-time operating condition classification labels; Indicates feature splicing; The bias term for the attention at the given time step; Represents the normalization function; For attention weights; The attention weight matrix is ​​a parameter. for Time of the first The weighted values ​​of each feature; To assess the weighting of grades; For attention bias terms; It is a non-linear activation function;

[0084] Step 303: Construct a domain-specific LSTM layer; the domain-specific LSTM layer includes: sequentially connected embedding units, three parallel LSTM units, and a hidden state fusion unit; the expression of the embedding unit includes: , , The expression for the hidden state fusion unit is: ; To embed feature vectors, The value range is 1, 2, 3. Embedded in the electrical domain, For frequency domain embedding, Embedded for state domain; for Voltage data at any given time; for Current data at any given time; for The actual frequency of the power grid at any given time; for Change in frequency over time; for Rate of change of frequency at any given time; for Electric field power data at any given time; for Constant fan speed; for Pitch angle at any moment; for Energy storage margin at all times; for Real-time historical power adjustment commands; for Real-time historical speed adjustment command; For the first Domain operating condition coding; For the first Domain LSTM hidden state; For the first Domain LSTM output gate; This represents the element-wise product operation; For the first Domain LSTM cell state; It is the hyperbolic tangent function; For the first Domain embedding feature vectors;

[0085] Step 304: Construct the instruction generation layer; the expression for the instruction generation layer is: ;in, ; For constraints; Generate intermediate feature vectors for the instruction layer; To correct the linear unit; Generate layer weight matrices for instructions; This is the overall hidden state of the merged system; Generate layer bias terms for instructions;

[0086] Step 305: Fuse the input layer, the condition-aware dual-attention layer, the domain-specific LSTM layer, and the instruction generation layer to obtain the neural network model; wherein the input layer, the condition-aware dual-attention layer, the domain-specific LSTM layer, and the instruction generation layer are connected in sequence.

[0087] refer to Figure 5The pre-built neural network model is trained using the labeled dataset to obtain a wind farm control model, including:

[0088] Step 306: Construct the total loss function; the expression for the total loss function is: ;in, For the first The evaluation level of each sample; To assess the weighting of the grades, when When the value is 1 or 2 Greater than 0, when When the value is 3 or 4 Less than 0; This represents the total loss value of the model. The total number of training samples; The first generated for the model Individual sample control instructions; For the first The actual regulatory instructions for each sample;

[0089] Step 307: Iteratively train the neural network model using the total loss function to obtain the wind farm control model.

[0090] Specifically, real-time data collection at a fixed frequency at the wind farm grid connection point includes: electrical parameters: grid connection point voltage and current; frequency parameters: actual grid frequency, frequency deviation, and frequency change rate; and status parameters: initial power, rated power, turbine speed, pitch angle, energy storage margin, and turbine power upscaling margin. Primary frequency regulation dead zone. ,Right now Triggering primary frequency modulation; Level 3 frequency change rate threshold: mild disturbance threshold: Rotor kinetic energy control trigger; moderate disturbance threshold: Power reserve control triggered; severe disturbance threshold: Energy storage compensation is triggered.

[0091] Furthermore, if Directly superimposed with a frequency modulation adjustment; according to The threshold range into which the disturbance falls determines its intensity, specifically corresponding to the following: mild disturbance. Moderate disturbance Severe disturbance Combining The sign determines the direction of the disturbance; power is reduced during upward disturbances and increased during downward disturbances. The specific control strategy is as follows:

[0092] 1) Mild disturbance control:

[0093] Primary frequency modulation adjustment amount: Virtual inertia adjustment: The total adjustment is the sum of the primary frequency regulation and the virtual inertia regulation. The total adjustment is sent to the energy management platform, and the power adjustment is evenly distributed according to the current output of each wind turbine and then sent to each wind turbine for execution.

[0094] 2) Moderate disturbance control:

[0095] The secondary frequency regulation and virtual inertia regulation are calculated in the same way as in 1). Power reserve regulation: reserve 5% to 10% of the preset power by reducing the turbine speed or reducing the pitch angle; if the energy storage margin is... During upward disturbances, the wind turbine maintains its current power, while excess power is transferred to the energy storage module for storage until the wind farm's output power reaches the target power. During downward disturbances, the energy storage module is prioritized for discharge, while the wind turbine power is released for backup. The two work together to increase the power to the target power. If the energy storage margin is sufficient... When encountering an upward disturbance, directly reduce the fan power until the target power is reached. When encountering a downward disturbance, release the fan power for standby. If the standby is insufficient, check the fan power adjustment margin. Then increase the fan power until the target power is reached.

[0096] 3) Severe disturbance control:

[0097] By multiplying the primary frequency modulation adjustment by 1.2, we obtain the primary frequency modulation extreme adjustment: Set the emergency compensation amount for energy storage; the energy storage will be at full load during upward disturbances and at full load during downward disturbances. If the disturbance is frequency downward and the wind turbine has no load limit, calculate the inertia regulation capacity of each wind turbine and allocate the active power command to each wind turbine according to the regulation capacity.

[0098]

[0099] The command is directly sent to the wind turbine, while the energy storage is controlled to discharge at full load, with the two working in tandem. If there is a frequency upswing or downswing but the wind turbine is at full load, the energy storage is controlled to store or discharge electricity at full load to handle the emergency compensation portion of the energy storage. The remaining active power command is then sent to the wind turbine until the frequency change rate drops to a minimum. Next, we will switch to moderate disturbance control.

[0100] Preferably, this embodiment evaluates the actual execution effect of the control commands during the operation of the above control strategy in four aspects: power regulation accuracy, frequency response speed, frequency stability effect, and inertia support effect.

[0101] Power regulation accuracy: ;

[0102] Frequency response speed: ;

[0103] Frequency stabilization effect: ;

[0104] Inertia support effect: ;

[0105] In the formula, Power deviation rate; , These are the actual power and the target power, respectively. For response time; , These are the time when the power target is met and the time when the instruction is issued, respectively. For frequency recovery deviation; This is the average frequency value of the power grid when it transitions from transient fluctuations to steady-state equilibrium. It is the generalized inertia constant; This refers to the per-unit value of the energy released or stored by the fan rotor during the control process; for The per-unit value of the power grid frequency at any given time; For time intervals.

[0106] The four calculated indicators are compared one by one with their corresponding thresholds. The specific judgment rules can be found in Table 1.

[0107] Table 1

[0108]

[0109] Based on the number of times the above four indicators are met, the evaluation results are divided into four levels: Level 1 to Level 4.

[0110] Specifically, the model architecture used in this embodiment includes: Input layer: Normalizing the collected time-series running data, constructing a time-series window and calculating the difference between adjacent time points; converting the operating condition classification identifier into a binary representation using one-hot encoding; Normalizing the historical control commands; Mapping the evaluation level to weights for subsequent model training.

[0111] The condition-aware dual-attention layer includes a time-step attention layer and a parameter attention layer; the expression for the time-step attention layer is:

[0112]

[0113] The expression for the parameter attention layer is:

[0114]

[0115] In the formula, Attention weights for each time step; This is the weight matrix for attention at each time step; For the first Continuously run data vectors; For the first Real-time operating condition classification labels; Indicates feature splicing; The bias term for attention at time steps; Represents the normalization function; For attention weights; The attention weight matrix is ​​a parameter. for Time of the first The weighted values ​​of each feature; To assess the weighting of grades; For attention bias terms; It is a non-linear activation function.

[0116] The domain-specific LSTM layer includes: an embedding unit, three parallel LSTM units, and a hidden state fusion unit; the expression for the embedding unit includes:

[0117]

[0118]

[0119]

[0120] In the formula, To embed feature vectors, The value range is 1, 2, 3. Embedded in the electrical domain, For frequency domain embedding, Embedded for state domain; for Voltage data at any given time; for Current data at any given time; for The actual frequency of the power grid at any given time; for Change in frequency over time; for Rate of change of frequency at any given time; for Electric field power data at any given time; for Constant fan speed; for Pitch angle at any moment; for Energy storage margin at all times; for Real-time historical power adjustment commands; for Real-time historical speed adjustment command; For the first Domain operating condition coding.

[0121] The gating formula for a single LSTM sub-cell in an LSTM unit is as follows:

[0122] Forgotten Gate: ;

[0123] Input Gate: ;

[0124] Cell status update: ;

[0125] Cell state: ;

[0126] Output gate: ;

[0127] Hidden state: ;

[0128] In the formula, Use the Sigmoid activation function; Output for the forget gate; Here is the forget gate weight matrix; This represents the hidden state of the electrical domain at the previous moment; The electrical domain input features at the current moment; Forget gate bias term; For input gate output; The input gate weight matrix; For input gate bias terms; Candidate cell state; Update the weight matrix for cell state; Update the bias term for cell state; This represents the current state of the cell. This represents the cell state at the previous moment; Output gate output; This is the output gate weight matrix; This is the output gate bias term; Hide the current state.

[0129] The expression for the hidden state fusion unit is:

[0130]

[0131] In the formula, For the first Domain LSTM hidden state; For the first Domain LSTM output gate; This represents the element-wise product operation; For the first Domain LSTM cell state; It is the hyperbolic tangent function; For the first Domain embedding feature vectors.

[0132] Instruction generation layer, the expression is:

[0133]

[0134] In the formula, ; For constraints; Generate intermediate feature vectors for the instruction layer; To correct the linear unit; Generate layer weight matrices for instructions; This is the overall hidden state of the merged system; Generate layer bias terms for instructions.

[0135] Furthermore, construct the total loss function:

[0136]

[0137] In the formula, For the first The evaluation level of each sample; To assess the weighting of the grades, when When the value is 1 or 2 Greater than 0, when When the value is 3 or 4 Less than 0; This represents the total loss value of the model. The total number of training samples; The first generated for the model Individual sample control instructions; For the first The actual regulatory instructions for each sample;

[0138] The neural network model described above was trained using a labeled dataset and a total loss function, ultimately yielding a wind farm control model.

[0139] As an optional implementation, this embodiment also provides a wind farm control system based on a neural network, including:

[0140] The data acquisition module is used to regulate the target wind farm using a preset control strategy, and to collect the operation data of the grid connection point and the regulation instructions generated by the control strategy during the regulation process to obtain the raw dataset.

[0141] The data labeling module is used to evaluate and classify the control instructions based on the operating data at adjacent times in the original dataset to obtain a labeled dataset; the labeled dataset includes: timestamp, operating data, control instructions, evaluation level, and operating condition status;

[0142] The model training module is used to train the pre-built neural network model using the labeled dataset to obtain the wind farm control model.

[0143] The optimization and control module is used to input the newly collected operating data from the grid connection point into the wind farm control model to generate instructions, obtain optimized control instructions, and update the optimized control instructions to the target wind farm.

[0144] As an optional implementation, this embodiment also provides an electronic device, including: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to execute the aforementioned neural network-based wind farm regulation method.

[0145] As an optional implementation, this embodiment also provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the aforementioned neural network-based wind farm control method.

[0146] The beneficial effects of this invention are as follows:

[0147] (1) The present invention pre-sets a three-level disturbance threshold system, which combines trigger frequency modulation and graded disturbance to subdivide the operating conditions into mild, moderate and severe disturbances under the frequency modulation trigger state, covering the entire scenario from small fluctuations to extreme fluctuations, and avoiding control mismatch caused by coarse division of operating conditions. Differentiated control logic is designed for different disturbance levels to achieve precise matching of disturbance degree and control intensity.

[0148] (2) The present invention constructs an adaptive control model based on neural networks. The dual attention layer of working condition perception captures the time correlation and key parameter weights in the historical control data. The domain-specific LSTM layer further splits the operating data into electrical parameters, frequency parameters and state parameters, realizing deep adaptation to dynamic working conditions. The model can iteratively optimize the adjustment amount according to the historical control effect, rather than mechanically applying fixed formulas, reducing the adjustment deviation caused by fixed parameters.

[0149] (3) This invention uses accuracy evaluation formula, speed evaluation formula, stability evaluation formula, and support evaluation formula to quantify and score each set of control data as training labels for the neural network, enabling the model to autonomously learn the deep control logic corresponding to high-level evaluations. By using past control command data and effect evaluations as input, the model can reuse high-quality historical experience, improving the reliability of command generation.

[0150] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0151] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A wind farm control method based on neural networks, characterized in that, include: The target wind farm is regulated using a preset control strategy, and the operation data of the grid connection point and the regulation instructions generated by the control strategy are collected during the regulation process to obtain the original dataset. The control instructions are evaluated and classified based on the operational data at adjacent times in the original dataset to obtain a labeled dataset; the labeled dataset includes: timestamp, operational data, control instructions, evaluation level, and operating condition status; The pre-constructed neural network model is trained using the labeled dataset to obtain a wind farm control model; the neural network model includes: an input layer, a condition-aware dual attention layer, a domain-specific LSTM layer, and an instruction generation layer connected in sequence; The newly collected operating data from the grid connection point is input into the wind farm control model to generate instructions, thereby obtaining optimized control instructions, which are then updated to the target wind farm.

2. The wind farm regulation method based on neural networks according to claim 1, characterized in that, The target wind farm is regulated using a preset control strategy, and the operation data of the grid connection point and the regulation commands generated by the control strategy are collected during the regulation process to obtain the original dataset, including: The operating data of the grid connection point is collected; the operating data includes: electrical parameters, frequency parameters, and status parameters; the electrical parameters include: voltage data and current data; the frequency parameters include: actual grid frequency, frequency change, and frequency change rate; the status parameters include: initial power of the wind farm, rated power, wind turbine speed, pitch angle, energy storage margin, and wind turbine power upsizing margin. Set the primary frequency modulation dead zone, mild disturbance threshold, moderate disturbance threshold, and severe disturbance threshold; When the frequency change is greater than the primary frequency regulation dead zone, the state of the target wind farm is determined to be a frequency regulation trigger state. After determining the frequency modulation trigger state, if the frequency change rate is less than or equal to the mild disturbance threshold, mild disturbance control is applied to the target wind farm; if the frequency change rate is between the mild disturbance threshold and the moderate disturbance threshold, moderate disturbance control is applied to the target wind farm; and if the frequency change rate is greater than the severe disturbance threshold, severe disturbance control is applied to the target wind farm.

3. The wind farm regulation method based on neural networks according to claim 2, characterized in that, The target wind farm is regulated using a preset control strategy, and the operation data of the grid connection point and the regulation commands generated by the control strategy are collected during the regulation process to obtain the original dataset, including: The adjustment direction is determined based on the sign of the frequency change. If the sign is positive, the adjustment direction is to reduce the power; otherwise, it is to increase the power. When performing mild disturbance control, the primary frequency regulation adjustment and virtual inertia adjustment are calculated based on the operating data to obtain the total mild disturbance adjustment, and the total mild disturbance adjustment is updated to the wind turbines of the target wind farm according to the adjustment direction; the expression for the primary frequency regulation adjustment is: The expression for the virtual inertia adjustment is: ;in, , These are the primary frequency modulation adjustment amount and the virtual inertia adjustment amount, respectively; This refers to the rated power of the wind farm. The rated frequency of the power grid; This refers to the actual frequency of the power grid. This is a primary frequency modulation dead zone; This refers to the vortex rate of a single frequency modulation (FM). The equivalent inertial time constant of the wind farm; The rate of change of frequency; When performing moderate disturbance control, the wind turbine speed is reduced or the pitch angle is narrowed to obtain reserve power. The total adjustment amount of the mild disturbance and the reserve power are integrated to obtain the total adjustment amount of the moderate disturbance. If the energy storage margin is greater than or equal to 10%, the output power of the wind turbine is adjusted to the target power using the energy storage module in the target wind farm. If the energy storage margin is less than 10%, the power of the wind turbine is adjusted until the output power of the wind turbine is adjusted to the target power.

4. The wind farm regulation method based on neural networks according to claim 3, characterized in that, The target wind farm is regulated using a preset control strategy, and the operation data of the grid connection point and the regulation commands generated by the control strategy are collected during the regulation process to obtain the original dataset, including: During severe disturbance regulation, the extreme adjustment amount of primary frequency regulation is calculated based on the operating data, and an emergency compensation amount for energy storage is set. The extreme adjustment amount of primary frequency regulation and the emergency compensation amount for energy storage are integrated to obtain the total adjustment amount for severe disturbance. If the adjustment direction is power increase and the wind turbine unit is unrestricted load, the inertia regulation capability of each wind turbine in the wind turbine unit is calculated, and the total adjustment amount for severe disturbance is allocated to each wind turbine according to the calculated inertia regulation capability to obtain an active power command. If the wind turbine unit is at full load, the emergency compensation amount for energy storage is provided using the energy storage module, and the extreme adjustment amount of primary frequency regulation is allocated to the wind turbine unit. The expression for the extreme adjustment amount of primary frequency regulation is: The expression for the active power command is: ;in, , These are the primary frequency modulation extreme adjustment amount and the active power command, respectively; It is the sum of the primary frequency regulation extreme adjustment amount and the energy storage emergency compensation amount; The control commands during the mild disturbance control, moderate disturbance control, and severe disturbance control processes are collected, and the control commands and the operating data are integrated to obtain the raw data.

5. The wind farm regulation method based on neural networks according to claim 4, characterized in that, The control instructions are evaluated and classified based on the operational data at adjacent time points in the original dataset to obtain a labeled dataset, including: Construct an accuracy evaluation formula; the expression of the accuracy evaluation formula is: ;in, Power deviation rate; , These are the actual power and the target power, respectively. Construct a velocity evaluation formula; the expression of the velocity evaluation formula is: ;in, For response time; , These are the time when the power target is met and the time when the instruction is issued, respectively. Construct a stability evaluation formula; the expression of the stability evaluation formula is: ;in, For frequency recovery deviation; This is the average frequency value of the power grid when it transitions from transient fluctuations to steady-state equilibrium. Construct a supporting evaluation formula; the expression of the supporting evaluation formula is: ;in, It is the generalized inertia constant; This refers to the per-unit value of the energy released or stored by the fan rotor during the control process; for The per-unit value of the power grid frequency at any given time; For time intervals; The accuracy evaluation formula, the velocity evaluation formula, the stability evaluation formula, and the support evaluation formula are used to evaluate each group of data in the original dataset to obtain an evaluation value; The evaluation values ​​are classified using a preset threshold judgment table and an evaluation level judgment table to obtain the label dataset; the label dataset includes: first to fourth levels.

6. The wind farm regulation method based on neural networks according to claim 5, characterized in that, The pre-built neural network model is trained using the labeled dataset to obtain a wind farm control model, which includes: Construct the input layer; the data types of the input layer include: the operating data, operating condition classification identifiers, historical control instructions, and the label dataset; Construct the condition-aware dual attention layer; the condition-aware dual attention layer includes: a time-step attention layer and a parameter attention layer connected to each other; the expression of the time-step attention layer is: The expression for the parameter attention layer is: ;in, Attention weights for each time step; Let be the weight matrix for the attention at the given time step; For the first Continuously run data vectors; For the first Real-time operating condition classification labels; Indicates feature splicing; The bias term for the attention at the given time step; Represents the normalization function; For attention weights; The attention weight matrix is ​​a parameter. for Time of the first The weighted values ​​of each feature; To assess the weighting of the ratings; For attention bias terms; It is a non-linear activation function; Construct the domain-specific LSTM layer; the domain-specific LSTM layer includes: sequentially connected embedding units, three parallel LSTM units, and a hidden state fusion unit; the expression of the embedding unit includes: , , The expression for the hidden state fusion unit is: ; To embed feature vectors, The value range is 1, 2, 3. Embedded in the electrical domain, For frequency domain embedding, Embedded for state domains; for Voltage data at any given time; for Current data at any given time; for The actual frequency of the power grid at any given time; for Change in frequency over time; for Rate of change of frequency at any given time; for Electric field power data at any given time; for Constant fan speed; for Pitch angle at any moment; for Energy storage margin at all times; for Real-time historical power adjustment commands; for Real-time historical speed adjustment command; For the first Domain operating condition coding; For the first Domain LSTM hidden state; For the first Domain LSTM output gate; This represents the element-wise product operation; For the first Domain LSTM cell state; It is the hyperbolic tangent function; For the first Domain embedding feature vectors; Construct the instruction generation layer; the expression for the instruction generation layer is: ;in, ; For constraints; Generate intermediate feature vectors for the instruction layer; To correct the linear unit; Generate layer weight matrices for instructions; This is the overall hidden state of the merged system; Generate layer bias terms for instructions; The neural network model is obtained by fusing the input layer, the condition-aware dual-attention layer, the domain-specific LSTM layer, and the instruction generation layer.

7. The wind farm regulation method based on neural networks according to claim 6, characterized in that, The pre-built neural network model is trained using the labeled dataset to obtain a wind farm regulation model, including: Construct the total loss function; the expression for the total loss function is: ;in, For the first The evaluation level of each sample; To assess the weighting of the grades, when When the value is 1 or 2 Greater than 0, when When the value is 3 or 4 Less than 0; This represents the total loss value of the model. The total number of training samples; The first generated for the model Individual sample control instructions; For the first The actual regulatory instructions for each sample; The total loss function is used to iteratively train the neural network model to obtain the wind farm control model.

8. A wind farm control system based on neural networks, characterized in that, The system is used to implement the neural network-based wind farm control method according to claim 1, the system comprising: The data acquisition module is used to regulate the target wind farm using a preset control strategy, and to collect the operation data of the grid connection point and the regulation instructions generated by the control strategy during the regulation process to obtain the raw dataset. The data labeling module is used to evaluate and classify the control instructions based on the operating data at adjacent times in the original dataset to obtain a labeled dataset; the labeled dataset includes: timestamp, operating data, control instructions, evaluation level, and operating condition status; The model training module is used to train the pre-built neural network model using the labeled dataset to obtain the wind farm control model. The optimization and control module is used to input the newly collected operating data from the grid connection point into the wind farm control model to generate instructions, obtain optimized control instructions, and update the optimized control instructions to the target wind farm.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform a neural network-based wind farm control method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute a wind farm control method based on a neural network, as described in any one of claims 1 to 7.

Images