CN-121689299-B - Wind power cluster wake optimizing method and device based on model predictive control

CN121689299BCN 121689299 BCN121689299 BCN 121689299BCN-121689299-B

Abstract

The application relates to the technical field of wind power generation predictive control and discloses a wind power cluster wake optimizing method and a device based on model predictive control, wherein the method comprises the steps of collecting operation state variables and meteorological environment information of each wind turbine generator set in a full farm; the method comprises the steps of inputting running state variables and meteorological environment information into a comprehensive prediction model, outputting a power generation sequence of each wind turbine generator in the whole field at N p moments in the future, constructing an objective function and constraint conditions, combining the power generation sequences at N p moments in the future, determining a yaw angle sequence of each wind turbine generator at N p moments in the future, controlling the wind turbine generator to execute yaw action according to a value at the first moment in the yaw angle sequence at N p moments in the future, and forming rolling closed-loop control according to a rolling time domain optimization mechanism. Therefore, the accuracy of a physical mechanism is reserved, the nonlinear fitting capability of deep learning is provided, the full-field power prediction accuracy can be improved, and the quick response of high-accuracy physical field description and cluster control is considered.

Inventors

ZHANG JI
GUO CHENGKE
SUN YONGCHAO
YUAN HAN
MEI NING
ZHU MENGGE
ZHENG XU

Assignees

中国海洋大学

Dates

Publication Date: 20260508
Application Date: 20260209

Claims (8)

1. The wind power cluster wake optimizing method based on model predictive control is characterized by comprising the following steps of: s10, data acquisition, namely collecting the running state variables and weather environment information of each wind turbine generator in the whole plant; S20, comprehensively predicting, namely inputting the running state variables and the meteorological environment information into a comprehensive prediction model, wherein the comprehensive prediction model comprises a PI-DANN model embedded with physical constraints of an N-S equation and a vortex quantity equation, a GAT model and a PI-Informer model introducing physical constraints based on a standard power curve in a loss function, generating an adjacency matrix of a wind power plant by using the PI-DANN model, determining the effective captured wind speed of each wind power plant by using the GAT model based on the adjacency matrix, and outputting a power generation sequence of each wind power plant in the whole field at the future N p moments by using the PI-Informer model based on the effective captured wind speed; S30, optimizing decision, namely constructing an objective function and constraint conditions, and determining a yaw angle sequence of each wind turbine set at the future N p moments by combining the power generation sequences at the future N p moments; S40, performing feedback, namely controlling the wind turbine generator to execute yaw action according to a value of a first moment in a yaw angle sequence of N p moments in the future, and forming rolling closed-loop control according to a rolling time domain optimization mechanism; In S20, the GAT model determines an effective captured wind speed of each wind turbine generator based on the adjacency matrix, including: Defining each wind turbine generator set as a node of the graph, and defining wake flow influence relation among the wind turbine generator sets as directed edges of the graph; introducing a multi-head attention mechanism, and automatically calculating the influence weight alpha ij of an upstream node j on a downstream node i: , Wherein, the Is a node feature vector; the term "vector" means a vector concatenation operation, The weight vector is a single-layer feedforward neural network, leakyReLU is a nonlinear activation function, and Softmax is a normalized exponential function; based on the distance and yaw angle between each wind turbine generator, adaptively distributing wake flow influence weight, and outputting physical wind speed; Carrying out weighted fusion on the physical wind speed and the measured wind speed of the cabin by using a Kalman filtering algorithm to obtain the effective captured wind speed of each wind turbine; The step of carrying out weighted fusion on the physical wind speed and the actually measured wind speed of the engine room by using a Kalman filtering algorithm to obtain the effective captured wind speed of each wind turbine generator, comprising the following steps: Defining the true effective capture wind speed at the moment k as a state variable Output values of the GAT model As a priori prediction, nacelle wind speed is defined as a measured variable The equations of state and the equations of measured variables are defined as follows: , Wherein N (0, Q) and N (0, R) respectively represent Gaussian distributions with a mean value of 0 and variances of Q and R; Transient flow field disturbances representing uncharacterized portions of the GAT model, which are process noise; For measuring noise, representing a measuring error of the wind meter including wake interference; , in the formula, For the effective captured wind speed estimated by kalman filtering at time k, Here as a priori state estimate, Is the kalman gain.
2. The wind power cluster wake optimization method based on model predictive control as claimed in claim 1, wherein the construction of the loss function of the PI-Informer model comprises: calculating the mean square error of the predicted value of the generating power of the wind turbine generator and the real monitoring power value in a single time step, and determining the mean value of the mean square error of all training samples as the data driving loss; Calculating the mean square error of the power generation power predicted value and the theoretical power value, and determining the mean square error average value of all training samples as physical prior loss; calculating counter-propagation gradient norms of the data driving loss and the physical prior loss in real time by using a gradient normalization algorithm, and dynamically adjusting data driving weights and physical prior weights according to gradient changes; The product of the data-driven loss and the data-driven weight, plus the product of the physical prior loss and the physical prior weight, is determined as a total loss.
3. The model predictive control-based wind power cluster wake optimization method as defined in claim 1, wherein S20 further comprises: and updating the output layer parameters of the comprehensive prediction model by utilizing the latest measured data, and keeping the parameters of the backbone network unchanged.
4. A wind power cluster wake optimization method based on model predictive control as defined in claim 3, wherein updating output layer parameters of said comprehensive predictive model with most recent measured data comprises: in a flow field correction loop, monitoring wake prediction errors of the PI-DANN model in real time And wind speed prediction error of the GAT model ; In the power correction loop, the power prediction error of the PI-Informer model is monitored in real time ; And when any one of the wake flow prediction error, the wind speed prediction error and the power prediction error continuously exceeds a preset threshold value, updating the output layer parameters of the corresponding model by using the latest actual measurement data.
5. The model predictive control-based wind power cluster wake optimization method as claimed in claim 2, wherein the theoretical power value of each wind turbine is determined by: Discrete design parameter points of each wind turbine are obtained, and a nonlinear fitting function of a partial load area is constructed by utilizing a piecewise cubic Hermite interpolation polynomial; When the input wind speed is smaller than the cut-in wind speed or larger than the cut-out wind speed, determining the theoretical power value as zero; When the input wind speed is greater than or equal to the cut-in wind speed and less than or equal to the rated wind speed, calculating a numerical value corresponding to the current input wind speed by utilizing the nonlinear fitting function, and determining the numerical value as a theoretical power value; and when the input wind speed is larger than the rated wind speed and smaller than or equal to the cut-out wind speed, determining the rated power of the wind turbine generator set as a theoretical power value.
6. The method for optimizing wake of a wind farm based on model predictive control of claim 1, wherein in S20, the generating an adjacency matrix of a wind farm using the PI-DANN model comprises: calculating wake envelope curves of each wind turbine generator by using the PI-DANN model; And determining whether a wake shielding relationship exists between any two wind motor groups according to the wake envelope curve, and further generating the adjacency matrix.
7. The wind power cluster wake optimization method based on model predictive control as claimed in any one of claims 1 to 6, wherein in S30, said constructing objective functions and constraints comprises: Maximizing the full-field total generated power in the prediction time domain N p ; The constraint conditions include: (1) The yaw angle of each wind turbine generator is limited in a safe range; (2) The variation of the yaw angle at adjacent moments is less than or equal to the single-step maximum allowable yaw rate; (3) The predicted power is within a reasonable interval that is non-negative and does not exceed the rated power.
8. Wind power cluster wake optimizing device based on model predictive control, which is characterized by comprising: the data acquisition module is configured to collect the running state variable and meteorological environment information of each wind turbine generator in the whole farm; The comprehensive prediction module is configured to input the running state variables and the meteorological environment information into a comprehensive prediction model, wherein the comprehensive prediction model comprises a PI-DANN model embedded with physical constraints of an N-S equation and a vortex quantity equation, a GAT model and a PI-Informer model introducing physical constraints based on a standard power curve in a loss function, an adjacency matrix of a wind power plant is generated by utilizing the PI-DANN model, an effective capture wind speed of each wind power plant is determined by the GAT model based on the adjacency matrix, and a power generation power sequence of each wind power plant in the whole field at N p moments is output by the PI-Informer model based on the effective capture wind speed; The optimization decision module is configured to construct an objective function and constraint conditions, and a yaw angle sequence of N p moments in the future of each wind turbine generator is determined by combining the power generation sequences of N p moments in the future; The execution feedback module is configured to control the wind turbine generator to execute yaw action according to a value of a first moment in a yaw angle sequence of N p moments in the future, and form rolling closed-loop control according to a rolling time domain optimization mechanism; wherein the GAT model determines an effective captured wind speed for each wind turbine based on the adjacency matrix, comprising: Defining each wind turbine generator set as a node of the graph, and defining wake flow influence relation among the wind turbine generator sets as directed edges of the graph; introducing a multi-head attention mechanism, and automatically calculating the influence weight alpha ij of an upstream node j on a downstream node i: , Wherein, the Is a node feature vector; the term "vector" means a vector concatenation operation, The weight vector is a single-layer feedforward neural network, leakyReLU is a nonlinear activation function, and Softmax is a normalized exponential function; based on the distance and yaw angle between each wind turbine generator, adaptively distributing wake flow influence weight, and outputting physical wind speed; Carrying out weighted fusion on the physical wind speed and the measured wind speed of the cabin by using a Kalman filtering algorithm to obtain the effective captured wind speed of each wind turbine; The step of carrying out weighted fusion on the physical wind speed and the actually measured wind speed of the engine room by using a Kalman filtering algorithm to obtain the effective captured wind speed of each wind turbine generator, comprising the following steps: Defining the true effective capture wind speed at the moment k as a state variable Output values of the GAT model As a priori prediction, nacelle wind speed is defined as a measured variable The equations of state and the equations of measured variables are defined as follows: , Wherein N (0, Q) and N (0, R) respectively represent Gaussian distributions with a mean value of 0 and variances of Q and R; Transient flow field disturbances representing uncharacterized portions of the GAT model, which are process noise; For measuring noise, representing a measuring error of the wind meter including wake interference; , in the formula, For the effective captured wind speed estimated by kalman filtering at time k, Here as a priori state estimate, Is the kalman gain.

Description

Wind power cluster wake optimizing method and device based on model predictive control Technical Field The application relates to the technical field of wind power generation predictive control, in particular to a wind power cluster wake optimizing method and device based on model predictive control. Background The core goal of the wind farm cluster control is to optimize the flow field distribution by cooperatively adjusting the control variables (such as yaw angle and pitch angle) of each machine set on the premise of ensuring the safe operation of the machine sets, thereby maximizing the total power generation of the whole farm. Along with the expansion of the installed scale and the increase of the layout density of the wind farm, the wake interference effect among units is remarkable, so that the overall benefit is difficult to be optimized by the traditional single machine control strategy. At present, wind farm cluster optimization control technologies are mainly divided into the following four types: (1) Based on the independent control method of the traditional MPPT, the method is the most commonly adopted control strategy of the commercial wind power plant at present. Each wind driven generator independently adjusts yaw and rotation speed only according to own sensor data (wind speed and wind direction) so as to track own maximum power point. However, as each unit operates independently, the upstream unit is always in a state of facing wind for capturing the maximum wind energy, and the generated strong wake can seriously interfere with the downstream unit, so that the power generation power of the downstream unit is reduced, the fatigue load of the downstream unit is increased, and the overall power generation efficiency of the whole plant is lower than the theoretical optimal value; (2) Collaborative optimization based on engineering wake model, namely describing attenuation and diffusion of wake speed by using a simplified analytical formula (such as a Jensen model or a Gaussian model), and searching the full-field optimal yaw angle by combining an optimization algorithm. Analytical models are usually based on static, linear simplifying assumptions, and it is difficult to accurately describe complex nonlinear turbulent mixing and dynamic deflection processes in actual flow fields. Under the actual working condition that the wind direction frequently fluctuates, the prediction deviation is larger, and the control strategy is very easy to fail; (3) Dynamic control based on high fidelity CFD, solving a fluid dynamics equation in real time by adopting a large vortex simulation (LES) or a Reynolds average equation (RANS), and performing control decision according to fine flow field information. This method theoretically has the highest control accuracy and can capture complex turbulence details, but has long calculation time. MPC control relies on rolling time domain optimization, requiring repeated deductions of flow field states for a future period to be completed in seconds or milliseconds. The computational overhead for solving the N-S equation by the CFD is extremely high, single computation often takes several hours, the real-time response period of the system is controlled far beyond, and real-time landing in engineering is difficult; (4) And the data driving control based on the traditional deep learning is to learn the time sequence mapping relation between wind speed and power through historical SCADA data by utilizing a long short term memory network (LSTM) or a standard transducer model, and embed the time sequence mapping relation into a Model Predictive Control (MPC) framework as a predictive model for rolling optimization. However, the method is excessively dependent on data quality, has poor generalization capability under extreme working conditions uncovered by data sparsity or training sets, lacks physical consistency constraint, and is easy to predict values violating physical common sense (such as violating a power curve). In addition, each unit is often regarded as an independent node by the existing time sequence model, and the influence of adjacent units in the wind power plant (namely, how the upstream wake flow is directionally transmitted to the downstream) is ignored, so that when the wake flow changes due to the change of the wind direction, the model prediction accuracy is greatly reduced. Meanwhile, the standard neural network architecture has higher computational complexity. When MPC needs to do long time sequence rolling prediction, the computing resource consumption increases in a quadratic way along with the increase of the prediction step length, so that the online reasoning speed is slower. It should be noted that the information disclosed in the above background section is only for enhancing understanding of the background of the application and thus may include information that does not form the prior art that is already known to those of ordinary skill in the a