Wind power plant control system

Abstract

The invention relates to the technical field of wind power plant control, and discloses a wind power plant control system. The system comprises a data acquisition module, a feature extraction module, a decision instruction generation module, an optimization model construction module and a hierarchical control execution module. The data acquisition module acquires data through a multi-source sensor; the feature extraction module performs multi-modal feature extraction by using a deep belief network; the decision instruction generation module generates a regulation and control instruction by means of a pre-trained collaborative decision model; the optimization model construction module constructs a multi-target dynamic optimization model, and adopts a decomposition coordination optimization algorithm to adjust operation parameters; the hierarchical control execution module executes a regulation and control instruction through a hierarchical control architecture, and the decision layer, the coordination layer and the execution layer respectively adopt corresponding algorithms to realize global regulation and control, local parameter adaptation and accurate tracking of power and pitch angles. The system can improve the power generation efficiency of the wind power plant, balance the equipment loss, enhance the cooperative operation capability with the power grid, and guarantee the stable and efficient operation of the wind power plant.

Description

Technical Field

[0001] This invention relates to the field of wind farm control technology, specifically a wind farm control system. Background Technology

[0002] With the continuous growth of global demand for clean energy, wind power, as a sustainable and clean energy source, is playing an increasingly important role in electricity supply. As wind farms continue to expand in scale, their operation and control face numerous complex challenges.

[0003] The operating environment of wind farms is extremely complex. Wind resources are highly random and intermittent, with wind speed and direction constantly changing, and wind conditions varying greatly across different regions. The diversity of meteorological conditions, such as heavy rain, sandstorms, and low temperatures, not only affects the power generation efficiency of wind turbines but can also damage the equipment. Currently, existing monitoring equipment such as anemometers, vibration sensors, and temperature sensors often suffer from low data accuracy and inconsistent data acquisition frequencies, leading to errors in the acquired real-time operating data of wind farms and making it difficult to comprehensively and accurately reflect the actual operating status of the wind farm.

[0004] The inherent characteristics of wind turbine equipment also increase the difficulty of control. Wind turbines have a complex structure, comprising numerous components such as blades, gearboxes, and generators. These components are interconnected, and a failure in any one component can affect the operation of the entire unit. Furthermore, the equipment operates in harsh environments for extended periods, inevitably leading to wear and fatigue, making equipment condition monitoring and fault prediction difficult. Traditional equipment condition monitoring methods often rely on data from single sensors, failing to comprehensively consider the impact of multiple factors on equipment status. This makes it difficult to accurately assess the equipment's health condition, hindering timely and effective maintenance measures and increasing downtime and repair costs.

[0005] From the perspective of grid connection, the large-scale integration of wind farms into the grid poses a significant challenge to grid stability. Fluctuations in wind farm output power cause grid voltage fluctuations and frequency changes, affecting power quality. When wind farm power fluctuations are large, the grid voltage may exceed the allowable range, affecting the normal operation of other electrical equipment in the grid. Existing wind farm control strategies are insufficient in coordinating the relationship between wind farms and the grid, making it difficult to precisely control wind turbines according to real-time grid demand, and failing to effectively balance the relationship between power generation efficiency and grid stability.

[0006] In terms of overall wind farm operation and management, existing control systems lack global optimization capabilities. Most wind farms adopt a decentralized control approach, with each wind turbine operating independently, lacking a unified coordination mechanism and failing to fully utilize wind resources to maximize power generation efficiency. When optimizing wind turbine operating parameters, only a single objective, such as maximizing power generation, is often considered, while other important factors such as equipment wear and tear and maintenance costs are ignored, leading to shortened equipment lifespan and increased operating costs.

[0007] To address these challenges, there is an urgent need for an advanced wind farm control system that can comprehensively process multi-source data to achieve accurate perception of the wind farm's operating status; generate reasonable control commands through intelligent decision-making; and use optimization algorithms to globally optimize the operating parameters of wind turbine units to improve power generation efficiency, reduce equipment losses, and enhance the coordinated operation capability between the wind farm and the power grid, thereby ensuring the safe, stable, and efficient operation of the wind farm. Summary of the Invention

[0008] The purpose of this invention is to provide a wind farm control system to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a wind farm control system, the system comprising: Data acquisition module: used to collect real-time operating data of the wind farm through multi-source sensors; Feature extraction module: Based on deep belief network, multimodal feature extraction is performed on the real-time running data to generate comprehensive state features; Decision instruction generation module: Inputs the comprehensive state features into the pre-trained collaborative decision-making model to generate wind turbine control instructions; Optimization model construction module: Constructs a multi-objective dynamic optimization model based on the control instructions. The multi-objective dynamic optimization model takes maximizing power generation efficiency and balancing equipment losses as optimization objectives, and uses a decomposition and coordination optimization algorithm to globally adjust the operating parameters of the wind turbine. The hierarchical control execution module outputs the optimal control strategy based on the multi-objective dynamic optimization model and realizes the distributed execution of control commands through a hierarchical control architecture. The hierarchical control architecture includes a decision layer, a coordination layer, and an execution layer. The decision layer generates a global control sequence based on the control commands, the coordination layer uses a rolling time window optimization algorithm to dynamically adapt local control parameters, and the execution layer uses a fuzzy adaptive control algorithm to achieve accurate tracking of wind turbine power and pitch angle.

[0010] Preferably, the acquisition of real-time wind farm operation data through multi-source sensors includes: Multi-source sensors include anemometers, vibration sensors, temperature sensors, current and voltage sensors, and weather radar; Spatiotemporal alignment of anemometer data and meteorological radar data is performed to construct a three-dimensional wind field distribution map; multi-scale wavelet transform is performed on vibration sensor data and temperature sensor data to generate equipment state feature sequences. A dual-channel feature fusion network is constructed. The first channel uses a convolutional autoencoder to extract the spatial features of the three-dimensional wind field distribution map, and the second channel uses a long short-term memory network to extract the temporal features of the equipment state feature sequence. The spatial and temporal features are fused through a cross-modal attention mechanism to generate a joint feature matrix. Dynamic correlation modeling is performed on the joint feature matrix based on a gated cyclic unit to output comprehensive state features including wind speed distribution, equipment health status, and grid load demand.

[0011] Preferably, the collaborative decision-making model adopts a hierarchical graph convolutional network structure, and generates multi-objective control instructions based on a dynamic weight allocation mechanism; the hierarchical graph convolutional network structure includes: Construct a wind farm-grid interaction diagram. The nodes in the diagram include wind turbine nodes, grid nodes, meteorological nodes, and equipment nodes. The node attributes include power output, operating temperature, and vibration amplitude. A two-stage attention mechanism is adopted. In the first stage, the interaction weights between wind turbine nodes and adjacent nodes are calculated through spatial graph convolutional layers. In the second stage, the importance of historical operating states is filtered through temporal graph convolutional layers. The node features are iteratively updated based on the multi-head graph attention module, and each attention head integrates node attributes and external environment parameters. The training process is stabilized through residual connection and batch normalization mechanism, and finally the control command containing power grid constraints and equipment protection is output.

[0012] Preferably, the decomposition coordination optimization algorithm integrates dynamic weight adjustment and constraint relaxation strategies, including: The parameter optimization problem is modeled as a multi-objective mixed integer nonlinear programming problem, with decision variables including discrete unit start-up and shutdown variables and continuous power regulation variables; The slack problem is initialized and the Pareto front approximation solution is calculated. A dynamic weight adjustment mechanism is used to update the weight coefficients of the objective function according to the real-time power grid demand. In the decomposition phase, the global problem is divided into multiple single-objective sub-problems based on the objective space decomposition technique; in the coordination phase, slack variables are introduced to balance the conflict constraints between the sub-problems. A parallel solver is used to iteratively optimize the subproblem. In each iteration, the local solution is updated and the global solution is synchronized through coordination variables.

[0013] Preferably, the rolling time window optimization algorithm dynamically adapts local control parameters by: A time-varying prediction model is constructed, and the dynamic equations of the wind turbine are discretized into a state-space model. The state-space model includes wind speed prediction error, power tracking error, and pitch angle differential term. Design a sliding window optimization objective function, which includes a power fluctuation suppression term, a mechanical stress smoothing term, and a grid frequency response term.

[0014] Preferably, the execution layer, based on a fuzzy adaptive control algorithm, achieves precise tracking of wind turbine power and pitch angle, including: Design a two-layer fuzzy inference system that maps power error and pitch angle error to fuzzy rule activation degree; construct an adaptive membership function adjustment mechanism to dynamically scale the coverage of the fuzzy set according to the error change rate.

[0015] Preferably, the convolutional autoencoder employs a pyramid pooling structure to accelerate feature extraction, including: The three-dimensional wind field distribution map is divided into a multi-resolution grid, and each grid cell stores the wind speed gradient and turbulence intensity statistics. In the encoding stage, dilated convolutional kernels are used to expand the receptive field, and in the decoding stage, deconvolutional layers are used to restore spatial details. A channel attention module is introduced to reweight the feature maps.

[0016] Preferably, the spatial graph convolutional layer employs a relative attribute encoding mechanism, including: Define the relative attribute vector between wind turbine nodes and adjacent nodes, including electrical distance, power transmission efficiency, and phase difference; The relative attribute vectors are mapped to the bias parameters of the graph convolution kernel through a nonlinear transformation layer; The bias parameters are superimposed on the standard graph convolution operation.

[0017] Preferably, the dynamic weight adjustment mechanism is implemented based on an online learning strategy, including: Collect historical optimization process data of objective function values ​​and constraint violation degrees as training samples; Construct a radial basis function network to fit the mapping relationship between the weight coefficients and the system state; The network parameters are updated online using incremental gradient descent, and the weight distribution of the objective function is adjusted in real time. When a sudden change in grid demand is detected, an emergency weight reset operation is triggered.

[0018] Preferably, the interval analysis method is implemented through orthogonal decomposition, including: Wind speed uncertainty is modeled as an interval variable, with upper and lower bounds determined by the standard deviation of the prediction error; the interval variable is orthogonally decomposed to separate deterministic and random components.

[0019] Compared with the prior art, the beneficial effects of the present invention are: In data acquisition and processing, the system collects data from multiple sources, including anemometers, vibration sensors, temperature sensors, current and voltage sensors, and weather radar. The anemometer and weather radar data are spatiotemporally aligned to construct a three-dimensional wind field distribution map. Vibration and temperature sensor data are processed using multi-scale wavelet transform to generate equipment state feature sequences. Then, advanced technologies such as dual-channel feature fusion networks and cross-modal attention mechanisms are employed to deeply fuse spatial and temporal features, ultimately outputting comprehensive state features including wind speed distribution, equipment health status, and grid load demand. This process makes the wind farm operation data acquired by the system more comprehensive and accurate, accurately reflecting the actual operating conditions of the wind farm and providing a solid and reliable data foundation for subsequent decision-making and control. Compared to traditional single or simple data acquisition and processing methods, this represents a qualitative leap in data completeness and accuracy.

[0020] From the perspective of decision command generation, the collaborative decision-making model adopts a hierarchical graph convolutional network structure and a dynamic weight allocation mechanism. By constructing a wind farm-grid interaction graph, and utilizing a two-stage attention mechanism and a multi-head graph attention module, it comprehensively considers multiple factors such as grid constraints and equipment protection to generate control commands. This approach can fully explore the complex relationships between various parts of the wind farm and with the grid, resulting in more scientific and reasonable control commands. It can effectively balance multiple objectives such as power generation, equipment protection, and grid stability, avoiding the limitations of traditional decision-making methods that only focus on a single objective, and improving the overall coordination and safety of wind farm operation.

[0021] The optimization model building module transforms control commands into a multi-objective dynamic optimization model, aiming to maximize power generation efficiency and balance equipment losses, employing a decomposition-coordination optimization algorithm. This algorithm integrates dynamic weight adjustment and constraint relaxation strategies, modeling the complex parameter optimization problem as a multi-objective mixed-integer nonlinear programming problem. It appropriately handles sub-problems at different stages and iteratively optimizes through a parallel solver, enabling global adjustment of wind turbine operating parameters under various operating conditions. This not only improves power generation efficiency, allowing wind turbines to better adapt to changes in wind resources, but also balances equipment losses, extends equipment lifespan, and reduces maintenance costs, achieving a dual improvement in economic benefits and equipment reliability.

[0022] In terms of executing control commands, the hierarchical control architecture of the hierarchical control execution module plays a crucial role. The decision-making layer generates a global control sequence, providing direction for overall operation; the coordination layer employs a rolling time window optimization algorithm to dynamically adapt local control parameters based on real-time conditions, enhancing the system's real-time response capability; and the execution layer uses a fuzzy adaptive control algorithm to accurately track wind turbine power and pitch angle, ensuring the stability and accuracy of wind turbine operation. This entire hierarchical control architecture enables efficient and accurate execution of control commands, ensuring stable operation of the wind farm under various complex conditions and significantly improving the reliability and stability of wind farm operation.

[0023] Furthermore, the application of innovative technologies such as the pyramid pooling structure of the convolutional autoencoder, the relative attribute encoding mechanism of the spatial graph convolutional layer, and the dynamic weight adjustment mechanism based on an online learning strategy further enhances the performance and efficiency of each module of the system. The synergistic effect of these technologies enables the wind farm control system of this invention to excel in improving power generation efficiency, ensuring stable equipment operation, and enhancing grid adaptability, providing strong support for the intelligent and efficient development of wind farms, and possessing broad application prospects and significant economic and social benefits. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the working principle of the wind farm control system described in this invention; Figure 2 Schematic diagram for data acquisition and integrated status feature generation; Figure 3 This is a schematic diagram illustrating the working principle of the fuzzy adaptive control algorithm at the execution layer. Figure 4 This is a flowchart illustrating the workflow of the dynamic weight adjustment mechanism. Detailed Implementation

[0025] 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.

[0026] Please see Figures 1-4 This invention provides a wind farm control system, the overall implementation of which is as follows: Data Acquisition Module: This module collects real-time operational data from the wind farm using multi-source sensors. These sensors include anemometers, vibration sensors, temperature sensors, current and voltage sensors, and weather radar. These sensors gather various types of information about the wind farm's operation from different perspectives, providing a data foundation for subsequent analysis.

[0027] Feature extraction module: Based on deep belief networks, multimodal feature extraction is performed on the collected real-time operational data to generate comprehensive state features. Deep belief networks can effectively process multi-source data, uncover potential relationships between data, and transform different types of operational data into representative comprehensive features, providing more valuable information for subsequent decision-making.

[0028] The decision command generation module inputs the generated comprehensive state features into a pre-trained collaborative decision-making model. This model, trained on a large amount of data, is capable of generating wind turbine control commands based on the comprehensive state features. These commands will be used to adjust the operating status of the wind turbines to adapt to the actual conditions of the wind farm.

[0029] The optimization model construction module constructs a multi-objective dynamic optimization model based on the control instructions output by the decision instruction generation module. This model aims to maximize power generation efficiency and balance equipment losses, employing a decomposition and coordination optimization algorithm to globally adjust the wind turbine's operating parameters. In this way, power generation efficiency is improved while equipment losses are effectively controlled, ensuring the long-term stable operation of the wind turbine.

[0030] The hierarchical control execution module, based on the optimal control strategy output by a multi-objective dynamic optimization model, implements distributed execution of control commands through a hierarchical control architecture. This architecture comprises a decision-making layer, a coordination layer, and an execution layer. The decision-making layer generates a global control sequence based on the control commands, formulating a macro-level plan for the entire system's operation. The coordination layer uses a rolling time window optimization algorithm to dynamically adapt local control parameters, enabling the system to flexibly adjust according to real-time conditions. The execution layer uses a fuzzy adaptive control algorithm to accurately track wind turbine power and pitch angle, ensuring the wind turbine operates stably according to the expected control strategy.

[0031] The implementation of the present invention will be further described below with reference to Examples 1 to 5.

[0032] Example 1: In this embodiment, the application of real-time data acquisition and convolutional autoencoder in feature extraction in the data acquisition module is further illustrated.

[0033] In the data acquisition module, an anemometer monitors wind speed and direction in real time, while the weather radar acquires meteorological information about the wind field from a more macroscopic perspective. After collecting anemometer and weather radar data, these two types of data are spatiotemporally aligned to construct a three-dimensional wind field distribution map, thus presenting the actual situation of the wind field more intuitively and comprehensively. Vibration and temperature sensors collect vibration and temperature data from the equipment, respectively, providing a basis for equipment condition monitoring. Multi-scale wavelet transform is performed on the vibration and temperature sensor data, enabling data analysis at different time and frequency scales to generate equipment condition feature sequences and uncover potential information about the equipment's operating status.

[0034] To extract features more efficiently, a dual-channel feature fusion network is constructed. The first channel employs a convolutional autoencoder with a pyramid pooling structure to accelerate feature extraction. Specifically, the 3D wind field distribution map is divided into multi-resolution grids, with each grid cell storing wind speed gradients and turbulence intensity statistics. During the encoding stage, dilated convolutional kernels are used to expand the receptive field, allowing for the acquisition of a wider range of information without increasing the number of parameters, thus improving the accuracy of feature extraction. In the decoding stage, deconvolutional layers are used to restore spatial details, enabling the extracted features to more accurately reflect the characteristics of the original data. Simultaneously, a channel attention module is introduced to reweight the feature maps. This module automatically learns the importance of features from different channels, highlighting key information and suppressing redundant information. The second channel uses a long short-term memory (LSTM) network to extract temporal features from the equipment state feature sequence. LSM excels at processing time-series data and can effectively capture long-term dependencies in the data. Finally, a cross-modal attention mechanism is used to fuse spatial and temporal features to generate a joint feature matrix. Based on gated recurrent units, dynamic correlation modeling is performed on the joint feature matrix, outputting comprehensive state features including wind speed distribution, equipment health status, and grid load demand. Through this series of operations, we can more comprehensively and deeply extract useful information from the data, providing strong support for subsequent decision-making and control.

[0035] Example 2: The collaborative decision-making model employs a hierarchical graph convolutional network structure and generates multi-objective control commands based on a dynamic weight allocation mechanism. First, a wind farm-grid interaction graph is constructed, with nodes including wind turbine nodes, grid nodes, meteorological nodes, and equipment nodes. Each node possesses rich attributes, such as power output, operating temperature, and vibration amplitude, reflecting the node's real-time status.

[0036] During model operation, a two-stage attention mechanism is employed. The first stage calculates the interaction weights between wind turbine nodes and their neighboring nodes using a spatial graph convolutional layer. This spatial graph convolutional layer employs a relative attribute encoding mechanism, specifically defining relative attribute vectors between wind turbine nodes and their neighboring nodes. These vectors contain information such as electrical distance, power transmission efficiency, and phase difference. These relative attributes more accurately describe the relationships between nodes. A nonlinear transformation layer maps the relative attribute vectors to bias parameters of the graph convolutional kernel, and then these bias parameters are superimposed on the standard graph convolution operation. This allows the spatial graph convolutional layer to better capture local structural information between nodes, improving the accuracy of calculating interaction weights. The second stage uses a temporal graph convolutional layer to filter the importance of historical operating states, fully utilizing useful information from historical data to provide a reference for current decision-making.

[0037] The model iteratively updates node features using a multi-head graph attention module, with each attention head fusing node attributes and external environment parameters. This multi-head structure allows for learning node features from different perspectives, enhancing the model's learning ability and adaptability. Simultaneously, residual connections and batch normalization mechanisms stabilize the training process. Residual connections effectively address the vanishing gradient problem in deep network training, enabling smoother training. Batch normalization accelerates model convergence and improves training efficiency. Ultimately, this collaborative decision-making model outputs control commands that incorporate grid constraints and equipment protection, ensuring that wind turbines meet grid demands while guaranteeing safe and stable operation.

[0038] Example 3: The decomposition-coordination optimization algorithm is used to solve parameter optimization problems in multi-objective dynamic optimization models. This algorithm integrates dynamic weight adjustment and constraint relaxation strategies. First, the parameter optimization problem is modeled as a multi-objective mixed-integer nonlinear programming problem, with decision variables including discrete turbine start-up / shutdown variables and continuous power regulation variables. The discrete turbine start-up / shutdown variables determine whether the wind turbines are put into operation, while the continuous power regulation variables control the power generation of the wind turbines. This setting of decision variables allows for more flexible regulation of the wind turbines.

[0039] The slack problem is initialized, and the Pareto front approximation solution is calculated. The Pareto front approximation solution provides a set of solutions that achieve equilibrium between different objectives, serving as a reference for subsequent optimization. A dynamic weight adjustment mechanism is adopted to update the objective function weight coefficients according to real-time grid demand. This mechanism is implemented based on an online learning strategy. Specifically, historical objective function values ​​and constraint violation degrees are collected as training samples to construct a radial basis function network (RBF network) to fit the mapping relationship between the weight coefficients and the system state. The RDF network can effectively approximate complex nonlinear functions, establishing a connection between the weight coefficients and the actual operating state of the system. Incremental gradient descent is used to update the network parameters online, adjusting the objective function weight allocation in real time. When a sudden change in grid demand is detected, an emergency weight reset operation is triggered, enabling the system to quickly adapt to changes in grid demand.

[0040] In the decomposition phase, the global problem is divided into multiple single-objective sub-problems based on the objective space decomposition technique, which simplifies the solution difficulty. In the coordination phase, slack variables are introduced to balance conflicting constraints between sub-problems. A parallel solver is used to iteratively optimize the sub-problems, updating the local solution in each iteration and synchronizing the global solution through coordination variables. In this way, multi-objective optimization problems can be solved efficiently while ensuring global optimality, enabling global adjustment of wind turbine operating parameters, improving power generation efficiency, and balancing equipment losses.

[0041] Example 4: In order to achieve accurate dynamic adaptation of local control parameters during wind farm operation, rolling time window optimization algorithms and interval analysis methods are required.

[0042] The core of the rolling time window optimization algorithm lies in constructing a time-varying prediction model. The operating state of a wind turbine is influenced by various factors, making its dynamic equations quite complex. The wind turbine's dynamic equations are discretized into a state-space model, which includes wind speed prediction error, power point tracking error, and pitch angle differential terms. The wind speed prediction error is denoted as... This is the actual wind speed. With predicted wind speed The difference, i.e. It reflects the degree of accuracy deviation in wind speed prediction and has a direct impact on the power output of wind turbines. Power point tracking error, denoted as [missing value], is [missing value]. This is the actual power generation of the wind turbine. With target power The difference, i.e. This error reflects the difference between the actual power generation of the wind turbine and the expected result. The differential term of the pitch angle is... Indicated, used to describe the pitch angle The rate of change of the pitch angle, which is the derivative of the pitch angle with respect to time, reflects the dynamic trend of the pitch angle. These three parameters combined comprehensively characterize the operating state of the wind turbine at a given moment.

[0043] Based on the aforementioned state-space model, a sliding window optimization objective function is designed. This function includes a power fluctuation suppression term, a mechanical stress smoothing term, and a grid frequency response term. The power fluctuation suppression term can be expressed as follows: ,in This represents the number of sampling points within the sliding window. It is the first The power tracking error at each sampling time. By optimizing the power fluctuation suppression term, the fluctuation of wind turbine power generation can be effectively reduced, ensuring the stability of power quality. The mechanical stress smoothing term is assumed to use... This indicates that it is related to the stress changes of various components of the wind turbine, and can generally be obtained through calculation and processing of relevant stress parameters, such as blade stress and tower stress. Optimizing this item can reduce the mechanical stress during the operation of the wind turbine and extend the service life of the equipment. The grid frequency response item is denoted as... It is related to the power grid frequency. And it is related to the wind turbine's ability to respond to changes in grid frequency, for example, it can be expressed as ,in It is the first The power grid frequency at each sampling time. This refers to the grid's set frequency. Optimizing this factor allows wind turbines to better respond to changes in the grid frequency, maintaining stable grid operation. The entire sliding window optimization objective function... The sum of these three terms, namely .

[0044] Because wind speed is uncertain, interval analysis is used to handle this factor more accurately. The uncertainty of wind speed is modeled as an interval variable, with its upper and lower bounds determined by the standard deviation of the prediction error. Let the standard deviation of the wind speed prediction error be... Then the wind speed range variable lower limit upper limit ,in This represents the mean of the wind speed forecast. Orthogonal decomposition is performed on the interval variable to separate the deterministic and random components. Orthogonal decomposition can use mathematical transformations to decompose the wind speed interval variable into a deterministic part. and randomness This allows for a more precise consideration of the impact of wind speed uncertainty during the optimization process. In the rolling time window optimization algorithm, local control parameters, such as power adjustment parameters and pitch angle adjustment parameters, are continuously adjusted based on the decomposed wind speed components. This enables the wind farm control system to better adapt to complex and variable wind speed environments, achieving precise dynamic adaptation of local control parameters and improving the overall operating efficiency and stability of the wind farm.

[0045] Example 5: The execution layer of the wind farm control system is responsible for translating control commands into actual control actions to achieve precise tracking of wind turbine power and pitch angle. This process is mainly achieved through fuzzy adaptive control algorithms.

[0046] The foundation of the fuzzy adaptive control algorithm is the design of a two-layer fuzzy inference system. This system uses power error and pitch angle error as inputs. The power error, as mentioned earlier, is the actual power generated by the wind turbine. With target power The difference The pitch angle error is the actual pitch angle. With the desired pitch angle The difference is denoted as ,Right now These two error values ​​reflect the deviation between the current operating state of the wind turbine and the ideal state.

[0047] The two-layer fuzzy inference system maps these errors to fuzzy rule activations. In the first layer of fuzzy inference, the power error is... and pitch angle error Each error is transformed into a fuzzy linguistic variable according to its respective fuzzification rule, such as "positive large," "positive small," "zero," "negative small," and "negative large." Each fuzzy linguistic variable corresponds to a specific fuzzy set, and the degree to which a specific error value belongs to each fuzzy set is determined by a membership function. Taking power error as an example, let's assume its membership function is... , The actual value of the power error, when When it is within a certain range, It has different values ​​under different fuzzy sets, and the range of values ​​is... Similarly, the pitch angle error also has its corresponding membership function. , This represents the actual value of the pitch angle error. Then, based on a pre-defined fuzzy rule base, fuzzy logic operations are used to obtain preliminary control output fuzzy values. For example, if the power error is "positive large" and the pitch angle error is "positive small", a corresponding preliminary control action fuzzy value may be derived based on the rule base.

[0048] In the second layer of fuzzy inference, the output of the first layer is further processed, taking into account more factors, such as the error change rate, to obtain a more accurate fuzzy rule activation degree. The error change rate reflects the trend of error change and has important guiding significance for adjusting the control strategy. For example, the power error change rate... , is the derivative of power error with respect to time, which indicates how fast the power error changes; the rate of change of pitch angle error. Similarly.

[0049] To enable the fuzzy inference system to better adapt to various conditions during wind turbine operation, an adaptive membership function adjustment mechanism is constructed. This mechanism dynamically scales the coverage of the fuzzy set based on the error change rate. When the power error change rate... A larger value indicates a rapid change in power error. In this case, appropriately expanding the coverage range of the fuzzy set for power error makes the control action more flexible and able to respond quickly to system changes. For example, the original power error range corresponding to the "Zhengda" fuzzy set was... ,when When it is large, its range can be expanded to , For pitch angle error, when the rate of change of pitch angle error... A smaller value indicates that the pitch angle error changes slowly, thus reducing the coverage range of the fuzzy set of pitch angle error and improving control accuracy. For example, the original pitch angle error range corresponding to the "positive small" fuzzy set was... ,when When it is small, it can be reduced to , .

[0050] In actual operation, based on power error and pitch angle error A two-layer fuzzy inference system is used to generate control signals, and the control effect is continuously optimized through an adaptive membership function adjustment mechanism. When the wind turbine power is lower than the target power and the power error rate of change is large, the fuzzy inference system generates a corresponding pitch angle adjustment signal according to rules. At the same time, the adaptive membership function adjustment mechanism expands the relevant fuzzy set range, enabling the pitch angle to be adjusted more quickly, thereby improving power generation. In this way, the power output of the wind turbine is ensured to be stable, and the pitch angle can be adjusted quickly and accurately according to the actual situation, thereby improving the power generation efficiency and operational stability of the wind turbine and ensuring the reliable operation of the wind farm.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wind farm control system, characterized in that, include: Data acquisition module: Used to collect real-time operating data of the wind farm through multi-source sensors; Feature extraction module: Based on deep belief network, multimodal feature extraction is performed on the real-time running data to generate comprehensive state features; Decision instruction generation module: Inputs the comprehensive state features into the pre-trained collaborative decision-making model to generate wind turbine control instructions; Optimization model construction module: Constructs a multi-objective dynamic optimization model based on the control instructions. The multi-objective dynamic optimization model takes maximizing power generation efficiency and balancing equipment losses as optimization objectives, and uses a decomposition and coordination optimization algorithm to globally adjust the operating parameters of the wind turbine. The hierarchical control execution module outputs the optimal control strategy based on the multi-objective dynamic optimization model and realizes the distributed execution of control commands through a hierarchical control architecture. The hierarchical control architecture includes a decision layer, a coordination layer, and an execution layer. The decision layer generates a global control sequence based on the control commands, the coordination layer uses a rolling time window optimization algorithm to dynamically adapt local control parameters, and the execution layer uses a fuzzy adaptive control algorithm to achieve accurate tracking of wind turbine power and pitch angle.

2. The wind farm control system according to claim 1, characterized in that, The acquisition of real-time wind farm operation data through multi-source sensors includes: Multi-source sensors include anemometers, vibration sensors, temperature sensors, current and voltage sensors, and weather radar; Spatiotemporal alignment of anemometer data and meteorological radar data is performed to construct a three-dimensional wind field distribution map; multi-scale wavelet transform is performed on vibration sensor data and temperature sensor data to generate equipment state feature sequences. A dual-channel feature fusion network is constructed. The first channel uses a convolutional autoencoder to extract the spatial features of the three-dimensional wind field distribution map, and the second channel uses a long short-term memory network to extract the temporal features of the equipment state feature sequence. The spatial and temporal features are fused through a cross-modal attention mechanism to generate a joint feature matrix. Dynamic correlation modeling is performed on the joint feature matrix based on a gated cyclic unit to output comprehensive state features including wind speed distribution, equipment health status, and grid load demand.

3. The wind farm control system according to claim 1, characterized in that, The collaborative decision-making model employs a hierarchical graph convolutional network structure and generates multi-objective control commands based on a dynamic weight allocation mechanism; the hierarchical graph convolutional network structure includes: Construct a wind farm-grid interaction diagram. The nodes in the diagram include wind turbine nodes, grid nodes, meteorological nodes, and equipment nodes. The node attributes include power output, operating temperature, and vibration amplitude. A two-stage attention mechanism is adopted. In the first stage, the interaction weights between wind turbine nodes and adjacent nodes are calculated through spatial graph convolutional layers. In the second stage, the importance of historical operating states is filtered through temporal graph convolutional layers. The node features are iteratively updated based on the multi-head graph attention module, and each attention head integrates node attributes and external environment parameters. The training process is stabilized through residual connection and batch normalization mechanism, and finally the control command containing power grid constraints and equipment protection is output.

4. The wind farm control system according to claim 1, characterized in that, The decomposition coordination optimization algorithm integrates dynamic weight adjustment and constraint relaxation strategies, including: The parameter optimization problem is modeled as a multi-objective mixed integer nonlinear programming problem, with decision variables including discrete unit start-up and shutdown variables and continuous power regulation variables; The slack problem is initialized and the Pareto front approximation solution is calculated. A dynamic weight adjustment mechanism is used to update the weight coefficients of the objective function according to the real-time power grid demand. In the decomposition phase, the global problem is divided into multiple single-objective sub-problems based on the objective space decomposition technique; in the coordination phase, slack variables are introduced to balance the conflict constraints between the sub-problems. A parallel solver is used to iteratively optimize the subproblem. In each iteration, the local solution is updated and the global solution is synchronized through coordination variables.

5. The wind farm control system according to claim 1, characterized in that, The rolling time window optimization algorithm dynamically adapts local control parameters, including: A time-varying prediction model is constructed, and the dynamic equations of the wind turbine are discretized into a state-space model. The state-space model includes wind speed prediction error, power tracking error, and pitch angle differential term. Design a sliding window optimization objective function, which includes a power fluctuation suppression term, a mechanical stress smoothing term, and a grid frequency response term.

6. The wind farm control system according to claim 1, characterized in that, The execution layer, based on a fuzzy adaptive control algorithm, achieves precise tracking of wind turbine power and pitch angle, including: Design a two-layer fuzzy inference system that maps power error and pitch angle error to fuzzy rule activation degree; construct an adaptive membership function adjustment mechanism to dynamically scale the coverage of the fuzzy set according to the error change rate.

7. The wind farm control system according to claim 2, characterized in that, The convolutional autoencoder employs a pyramid pooling structure to accelerate feature extraction, including: The three-dimensional wind field distribution map is divided into a multi-resolution grid, and each grid cell stores the wind speed gradient and turbulence intensity statistics. In the encoding stage, dilated convolutional kernels are used to expand the receptive field, and in the decoding stage, deconvolutional layers are used to restore spatial details. A channel attention module is introduced to reweight the feature maps.

8. The wind farm control system according to claim 3, characterized in that, The spatial graph convolutional layer employs a relative attribute encoding mechanism, including: Define the relative attribute vector between wind turbine nodes and adjacent nodes, including electrical distance, power transmission efficiency, and phase difference; The relative attribute vectors are mapped to the bias parameters of the graph convolution kernel through a nonlinear transformation layer; The bias parameters are superimposed on the standard graph convolution operation.

9. The wind farm control system according to claim 4, characterized in that, The dynamic weight adjustment mechanism is implemented based on an online learning strategy and includes: Collect historical optimization process data of objective function values ​​and constraint violation degrees as training samples; Construct a radial basis function network to fit the mapping relationship between the weight coefficients and the system state; The network parameters are updated online using incremental gradient descent, and the weight distribution of the objective function is adjusted in real time. When a sudden change in grid demand is detected, an emergency weight reset operation is triggered.

10. The wind farm control system according to claim 5, characterized in that, The interval analysis method is implemented through orthogonal decomposition, including: Wind speed uncertainty is modeled as an interval variable, with upper and lower bounds determined by the standard deviation of the prediction error; the interval variable is orthogonally decomposed to separate deterministic and random components.

Images