Fan controller control parameter prediction method and system, and edge device

Abstract

The fan controller control parameter prediction method, system and edge device of the application, when the sample quantity n is less than or equal to the set quantity m, an XGBoost model is established, the XGBoost model is trained by using n samples, and the trained XGBoost model is used as a prediction model; when the sample quantity n increases to more than the set quantity m, a DNN model is established, the DNN model is trained by using n-m newly added samples, wherein the feature importance weight output by the XGBoost model is used as the initial attention weight of the DNN model; and the trained DNN model is used as a prediction model. Therefore, the fan controller control parameter prediction method, system and edge device of the application establish different models according to the sample quantity, avoid data overfitting, avoid the limitation of fixed models, and solve the technical problem that overfitting is prone to occur in the prior art using fixed models.

Description

Technical Field

[0001] This invention belongs to the field of parameter prediction technology, specifically, it relates to a method, system and edge device for predicting control parameters of a wind turbine controller. Background Technology

[0002] Traditional industrial fan frequency converters rely on the experience of embedded software engineers for parameter tuning, requiring manual adjustment of multiple coupling parameters (such as current loop, carrier frequency, motor phase resistance, motor phase inductance, pre-charge time, etc.), which presents the following core problems:

[0003] (1) Inefficient: Manual debugging requires repeated trial and error, taking several hours to several days, and it is difficult to cover multi-parameter nonlinear coupling relationships.

[0004] (2) High dependence on experience: The debugging results are affected by the subjective experience of embedded software engineers and lack standardization, resulting in poor parameter consistency in different scenarios.

[0005] (3) Lack of data-driven capability: Traditional methods do not utilize historical data for modeling and cannot optimize prediction accuracy through data accumulation.

[0006] (4) Poor deployment flexibility: Existing automation solutions do not integrate lightweight technology and are difficult to adapt to edge computing scenarios.

[0007] The parameter tuning of the frequency converter for industrial fans directly affects the operating efficiency and stability of the fan equipment. Traditional methods rely on manual experience or a single model, which is insufficient to address the following challenges:

[0008] (1) Multi-parameter coupling: Parameters such as current loop, carrier frequency, motor phase resistance, and motor phase inductance affect each other and need to be optimized in a coordinated manner;

[0009] (2) Dynamic changes in data scale: The initial data volume is limited, but the data continues to grow in the later stage, requiring the model to adapt and adjust accordingly;

[0010] (3) Real-time requirements: Industrial scenarios require low-latency response, while traditional models have large deployment volume and slow inference speed.

[0011] The following methods are commonly used to adjust controller parameters:

[0012] (1) Linear regression extended with multiple outputs: Using MultiOutputRegressor to wrap the linear model cannot capture non-linear relationships, and the prediction error is >30%.

[0013] (2) Static DNN: Directly applying DNN for multi-output prediction, resulting in overfitting on small data and a validation set loss fluctuation of >50%.

[0014] (3) Traditional PID tuning: Single-variable PID parameter adjustment is only applicable to a single parameter, and manual intervention is required for multiple objectives.

[0015] Therefore, the disadvantages of the existing technology are as follows:

[0016] (1) Manual trial and error method: The debugging cycle is long, it cannot cover complex nonlinear relationships, and the results are inconsistent.

[0017] (2) Rule-based expert system: The rule base has high maintenance costs, poor adaptability, and cannot handle new wind turbine controllers.

[0018] (3) Single-objective optimization algorithms (such as GA, PSO): only optimize a single parameter, ignore the synergistic effect of multiple parameters, resulting in suboptimal system.

[0019] (4) Traditional machine learning models (such as SVM, RF): multiple outputs require multiple modeling, resulting in computational redundancy; they are not adapted to dynamic changes in data scale.

[0020] (5) Static DNN model: It suffers from severe overfitting on small data, lacks lightweight technology, and has high deployment costs. Summary of the Invention

[0021] This invention provides a method for predicting control parameters of a wind turbine controller, which solves the technical problem of overfitting that easily occurs when using a fixed model in the prior art.

[0022] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution:

[0023] Methods for predicting control parameters for wind turbine controllers include:

[0024] (1) Model building and training process:

[0025] When the number of samples n is less than or equal to the set number m, an XGBoost model is established. The input parameters of the XGBoost model include the physical parameters of the wind turbine, and the output parameters of the XGBoost model include the control parameters of the wind turbine and the feature importance weights of each control parameter. The XGBoost model is trained using n samples, and the trained XGBoost model is used as the prediction model.

[0026] When the number of samples n increases to a value greater than the set number m, a DNN model is established. The input parameters of the DNN model include the physical parameters of the wind turbine and the feature importance weights output by the XGBoost model. The output parameters of the DNN model include the control parameters of the wind turbine. The DNN model is trained using the newly added nm samples, wherein the feature importance weights output by the XGBoost model are used as the initial attention weights of the DNN model. The trained DNN model is then used as the prediction model.

[0027] (2) Model invocation process:

[0028] The actual physical parameters of the wind turbine are obtained and input into the prediction model to obtain the predicted actual control parameters of the wind turbine.

[0029] In some embodiments of this application, the training process of the XGBoost model and the DNN model specifically includes:

[0030] Normalize each input parameter;

[0031] A shared feature extraction layer is used to extract common features for each input parameter;

[0032] The extracted general features are assigned to different prediction branch networks, and each prediction branch network is used to predict one output parameter;

[0033] For each prediction branch network, a loss function is designed and optimized.

[0034] In some embodiments of this application, during the training process of the XGBoost model,

[0035] When the number of samples n≤q, the initial weights of the loss function are static weights;

[0036] When the number of samples n∈(q, m], the weights of the loss function are dynamically adjusted according to the error ratio; where q is a preset fixed value;

[0037] During the training of a DNN model, the initial weights of the loss function are adjusted based on the SHAP value.

[0038] In some embodiments of this application, the physical parameters of the fan include the fan's rated speed, rated power, and number of pole pairs;

[0039] The control parameters of the wind turbine include control loop parameters, carrier frequency, and pre-charging time.

[0040] In some embodiments of this application, the trained XGBoost model is used as a prediction model and exported through ONNXRuntime;

[0041] The trained DNN model is used as the prediction model, and then optimized using TensorRT and quantized using FP16.

[0042] The wind turbine controller control parameter prediction system includes:

[0043] A dynamic modeling module is used for: when the number of samples n is less than or equal to a set number m, establishing an XGBoost model, the input parameters of which include the physical parameters of the wind turbine, and the output parameters of which include the control parameters of the wind turbine and the feature importance weights of each control parameter; training the XGBoost model with n samples, and using the trained XGBoost model as a prediction model; when the number of samples n increases to greater than the set number m, establishing a DNN model, the input parameters of which include the physical parameters of the wind turbine and the feature importance weights output by the XGBoost model, and the output parameters of the DNN model include the control parameters of the wind turbine; training the DNN model with the newly added nm samples, wherein the feature importance weights output by the XGBoost model are used as the initial attention weights of the DNN model; and using the trained DNN model as a prediction model.

[0044] The calling module is used to: obtain the actual physical parameters of the wind turbine and input them into the prediction model to obtain the predicted actual control parameters of the wind turbine.

[0045] In some embodiments of this application, the dynamic modeling module is specifically used for:

[0046] Normalize each input parameter;

[0047] A shared feature extraction layer is used to extract common features for each input parameter;

[0048] The extracted general features are assigned to different prediction branch networks, and each prediction branch network is used to predict one output parameter;

[0049] For each prediction branch network, a loss function is designed and optimized.

[0050] In some embodiments of this application, the dynamic modeling module is specifically used for:

[0051] During the training of the XGBoost model,

[0052] When the number of samples n≤q, the initial weights of the loss function are static weights;

[0053] When the number of samples n∈(q, m], the weights of the loss function are dynamically adjusted according to the error ratio; where q is a preset fixed value;

[0054] During the training of a DNN model, the initial weights of the loss function are adjusted based on the SHAP value.

[0055] In some embodiments of this application, the wind turbine controller control parameter prediction system further includes: a lightweight deployment module, which is used for:

[0056] The trained XGBoost model is used as a prediction model and exported via ONNX Runtime;

[0057] The trained DNN model is used as the prediction model, and then optimized using TensorRT and quantized using FP16.

[0058] Edge devices, including the aforementioned wind turbine controller control parameter prediction system.

[0059] Compared with existing technologies, the advantages and positive effects of this invention are as follows: When the number of samples n ≤ a set number m, an XGBoost model is established, trained using n samples, and the trained XGBoost model is used as the prediction model. When the number of samples n increases to greater than the set number m, a DNN model is established, trained using the newly added nm samples, wherein the feature importance weights output by the XGBoost model are used as the initial attention weights of the DNN model; the trained DNN model is used as the prediction model. Therefore, the wind turbine controller control parameter prediction method, system, and edge device of this invention establish different models according to the number of samples, avoiding data overfitting and the limitations of fixed models, thus solving the technical problem of overfitting that easily occurs when using fixed models in existing technologies.

[0060] Other features and advantages of the present invention will become clearer after reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a flowchart of an embodiment of the wind turbine controller control parameter prediction method proposed in this invention;

[0063] Figure 2 yes Figure 1 A flowchart of one embodiment of the steps in the middle section;

[0064] Figure 3 This is a flowchart of another embodiment of the wind turbine controller control parameter prediction method proposed in this invention;

[0065] Figure 4 yes Figure 3 A flowchart of one embodiment of the steps in the middle section;

[0066] Figure 5 This is a training loss curve for a multi-output model based on motor parameters;

[0067] Figure 6 This is a comparison chart of the accuracy of multi-output regression prediction (partial parameters);

[0068] Figure 7 It is a feature importance analysis diagram;

[0069] Figure 8 This is a performance comparison chart of edge deployment using a full-parameter model;

[0070] Figure 9 It is a dynamic model switching decision boundary diagram;

[0071] Figure 10 This is a structural block diagram of an embodiment of the wind turbine controller control parameter prediction system proposed in this invention. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0073] It should be noted that in the description of this invention, the terms "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," which indicate directional or positional relationships, are based on the directional or positional relationships shown in the accompanying drawings. These are merely for ease of description and do not indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0074] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0075] Example 1

[0076] The wind turbine controller control parameter prediction method in this embodiment mainly includes the following steps, see below. Figure 1 As shown.

[0077] Step S11: Model building and training process.

[0078] When the number of samples n is less than or equal to the set number m, an XGBoost model is established. The input parameters of the XGBoost model include the physical parameters of the wind turbine, and the output parameters of the XGBoost model include the control parameters of the wind turbine and the feature importance weights of each control parameter. The XGBoost model is trained using n samples, and the trained XGBoost model is used as the prediction model.

[0079] When the number of samples n increases to a value greater than the set number m, a DNN model (deep neural network model) is established. The input parameters of the DNN model include the physical parameters of the wind turbine and the feature importance weights output by the XGBoost model. The output parameters of the DNN model include the control parameters of the wind turbine. The DNN model is trained using the newly added nm samples, where the feature importance weights output by the XGBoost model are used as the initial attention weights of the DNN model. The trained DNN model is then used as the prediction model.

[0080] The feature importance weights provided by the XGBoost model are used for the weight initialization of the DNN model input layer, that is, as the initial weights of the DNN model input layer.

[0081] When the number of samples n > the set number m, among the n samples, the first m samples have already been used to train the XGBoost model. Therefore, the remaining nm samples, i.e. the newly added nm samples, are used to train the DNN model.

[0082] Step S12: Model invocation process.

[0083] The actual physical parameters of the wind turbine are obtained and input into the prediction model to obtain the predicted actual control parameters of the wind turbine.

[0084] The prediction model is invoked, and the actual physical parameters of the wind turbine are input into the prediction model. The prediction model then outputs the predicted actual control parameters of the wind turbine.

[0085] Therefore, when the number of samples n ≤ the set number m, the XGBoost model is trained using n samples, and the trained XGBoost model is used to predict the actual control parameters of the wind turbine; when the number of samples n > the set number m, the DNN model is trained using the newly added nm samples, and the trained DNN model is used to predict the actual control parameters of the wind turbine; this solves the problem of data overfitting caused by using a fixed model and improves training efficiency.

[0086] The wind turbine controller control parameter prediction method of this embodiment establishes an XGBoost model when the number of samples n ≤ a set number m, trains the XGBoost model using n samples, and uses the trained XGBoost model as the prediction model. When the number of samples n increases to greater than the set number m, a DNN model is established, trains the DNN model using the newly added nm samples, and uses the feature importance weights output by the XGBoost model as the initial attention weights of the DNN model. The trained DNN model is then used as the prediction model. Therefore, the wind turbine controller control parameter prediction method of this embodiment establishes different models according to the number of samples, avoiding data overfitting and the limitations of a fixed model, thus solving the technical problem of overfitting that easily occurs when using a fixed model in the prior art.

[0087] During training, the input parameters of the XGBoost model also include labels (label parameters), such as the start-up status and running status of the wind turbine, in order to improve the accuracy of the trained XGBoost model.

[0088] In DNN models, the ReLU activation function is used to solve the gradient vanishing problem and improve nonlinear modeling capabilities.

[0089] In some embodiments of this application, the training process of the XGBoost model and the DNN model specifically includes the following steps, see [link to relevant documentation]. Figure 2 As shown.

[0090] Step S21: Normalize each input parameter.

[0091] Step S22: Extract common features for each input parameter using a shared feature extraction layer.

[0092] Step S23: Assign the extracted general features to different prediction branch networks, with each prediction branch network used to predict one output parameter.

[0093] Step S24: For each prediction branch network, design a loss function and optimize the loss function.

[0094] For each prediction branch network, a loss function is designed, and the predicted value is calculated using the forward propagation algorithm. The weighted sum of losses for all output parameters is calculated, and the weights of each loss function are updated using the backpropagation algorithm.

[0095] By designing steps S21 to S24, a loss function is designed and updated for each prediction branch network to optimize the loss function and train a qualified model.

[0096] In some embodiments of this application, during the training process of the XGBoost model,

[0097] When the number of samples n ≤ q, the initial weights of the loss function are static weights.

[0098] When the number of samples n∈(q, m], the weights of the loss function are dynamically adjusted according to the error ratio.

[0099] Where q is a preset fixed value, 0 < q < m.

[0100] The above design allows for the selection of appropriate weights for the loss function of the XGBoost model, facilitating its training.

[0101] In some embodiments of this application, during the training of the DNN model, the initial weights of the loss function are adjusted based on the SHAP value. This allows for the selection of suitable initial weights for the loss function of the DNN model, facilitating its training.

[0102] In some embodiments of this application, q = 1000 and m = 45000.

[0103] Therefore, during the training process of the XGBoost model,

[0104] When the number of samples n ≤ 1000, the initial weights of the loss function are static weights.

[0105] When the number of samples n∈(1000, 45000], the weights of the loss function are dynamically adjusted according to the error ratio.

[0106] When the number of historical data samples n increases to >45000, the DNN model is trained using the newly added n-45000 samples; during the training of the DNN model, the initial weights of the loss function are adjusted based on the SHAP value.

[0107] Static weights (quick start): Suitable for initial deployments or when parameter importance is clear, simple to implement, and low computational overhead. Static weights are pre-set fixed values.

[0108] Dynamically adjusts based on error ratio: suitable for adaptive data offset when data distribution changes.

[0109] SHAP value-based adjustment: Combining SHAP value optimization (when feature interpretability is required) is suitable for high accuracy requirements and feature interpretability needs.

[0110] The weights of the loss function are dynamically adjusted based on the error proportions. This means that the weights are dynamically adjusted according to the error proportions of output parameters such as the proportional / integral coefficients of the current loop, the proportional / integral coefficients of the speed loop, the carrier frequency, and the pre-charge time. By amplifying the weights of parameters with high error rates, issues such as the dimensional differences between the proportional coefficient and the pre-charge time are addressed.

[0111] The initial weights of the loss function are adjusted based on SHAP values, meaning the initial weights are adjusted based on the quantified influence of features according to SHAP values. First, the SHAP values ​​of input features (i.e., input parameters) on each output parameter are extracted from the XGBoost model. The relative influence of input features on output parameters is calculated using a normalization formula. During dynamic adjustment, the loss weights of parameters associated with features with high SHAP values ​​are amplified.

[0112] The object of SHAP value extraction is the XGBoost model. During the training process of XGBoost, the global SHAP matrix is ​​calculated and stored, which is used to initialize and inherit the DNN model, and then the DNN model obtains the SHAP value.

[0113] In some embodiments of this application, the physical parameters of the wind turbine include its rated speed, rated power, number of pole pairs, etc. The control parameters of the wind turbine include control loop parameters, carrier frequency, pre-charging time, etc.

[0114] Therefore, the model can predict multiple control parameters simultaneously, avoiding the accumulation of errors in step-by-step optimization and achieving end-to-end multi-output regression.

[0115] Control loop parameters include current loop proportional coefficient, current loop integral coefficient, speed loop proportional coefficient, and speed loop integral coefficient.

[0116] The control parameters of the fan may also include: motor phase resistance, motor phase inductance, field weakening control threshold, etc.

[0117] Phase resistance and phase inductance are inherent physical parameters of a motor, but in software debugging, the control algorithm can be optimized by adjusting the parameter values ​​used in the control algorithm (rather than changing the physical parameters themselves).

[0118] Therefore, the model can achieve end-to-end multi-output regression: synchronously predict multiple parameters such as current loop, speed loop, carrier frequency, motor phase resistance, motor phase inductance, and pre-charge time, avoiding the accumulation of errors in step-by-step optimization.

[0119] The predicted values ​​output by the XGBoost model are regression prediction results of multiple wind turbine parameters (XGBRegressor's multi-output regression task), such as proportional (P) / integral (I) coefficients, phase resistance, phase inductance, etc.

[0120] In some embodiments of this application, the trained XGBoost model is used as a prediction model and exported through ONNXRuntime to achieve model lightweighting and save deployment space.

[0121] Export the trained XGBoost model via ONNX Runtime, compressing the model size by 40% (e.g., 200MB → 120MB).

[0122] In some embodiments of this application, the trained DNN model is used as the prediction model, and the model is lightweighted by TensorRT optimization and FP16 quantization, thus saving deployment space.

[0123] The trained DNN model is optimized using TensorRT and FP16 quantization to reduce model size and computational latency; for example, the inference latency is reduced from 30ms to 10ms.

[0124] Using ONNX / TensorRT, industrial-grade models can be deployed across platforms and inferred in real time.

[0125] Below, in conjunction with Figure 3 and Figure 4 This paper will explain the specific workflow of the wind turbine controller control parameter prediction method of this application.

[0126] (1) Phased dynamic modeling.

[0127] Data acquisition phase: Data acquisition and annotation are performed through the host computer system of the industrial fan frequency converter.

[0128] Initial stage (data volume ≤ 45,000 records): XGBoost+XGBRegressor is used to leverage its parallel computing advantages and quickly fit the multi-output nonlinear relationship through the Gradient Boosting algorithm to create a single prediction model.

[0129] Data growth phase (data volume > 45,000 records): When the data volume threshold is triggered or the validation set loss does not decrease for 3 consecutive times, switch to a deep neural network model (DNN) to ensure that the input feature dimensions of the XGBoost model and the DNN model are consistent, and use the feature importance output by the XGBoost model as the initial attention weight of the DNN model.

[0130] The amount of data is the number of samples. One sample (one data point) = one complete set of input features (fixed feature dimensions, such as rated speed, rated power, etc.) + one corresponding set of output parameters (such as KP, KI, carrier frequency, phase resistance, etc.).

[0131] (2) Multi-output end-to-end optimization.

[0132] Weighted MSE loss function design and weight allocation: This addresses the training imbalance problem caused by differences in the units and numerical ranges of multiple output parameters. Error balance is achieved by dynamically adjusting the loss weights of different parameters.

[0133] The implementation process of multi-output end-to-end optimization technology, i.e., the model training process, is as follows: Figure 4 As shown.

[0134] (21) Parameter normalization: Standardize each input parameter.

[0135] (22) General feature extraction: All input features are extracted through a shared feature extraction layer (a shared neural network layer, such as a fully connected layer).

[0136] (23) Branch output layer: The output of the shared feature extraction layer is distributed to different prediction branch networks. Each prediction branch network is responsible for the prediction of a specific set of parameters and performs specialized prediction for different parameter characteristics.

[0137] (24) Loss function design.

[0138] Static weight allocation strategy (manually set based on prior knowledge): In the initial stage, when the amount of data is ≤1000, static weight (quick start) is selected. It is suitable for initial deployment or when the importance of parameters is clear. It is simple to implement and has low computational overhead.

[0139] Dynamic weight allocation strategy (based on dynamic adjustment of error ratio): When the number of data items is ∈ (1000, 45000), dynamic adjustment of error ratio is enabled. This is suitable for adaptive data offset when the data distribution changes.

[0140] Dynamic weight allocation strategy (adjustment based on SHAP value quantification of feature influence): During the deployment phase, when the data volume is >45000, it is combined with SHAP value optimization (when feature interpretability is required), which is suitable for high accuracy requirements and feature interpretability needs.

[0141] (25) Joint training and gradient collaborative parallel training.

[0142] Joint training: The propagation calculates the predicted values ​​of all parameters. A single batch of data (referring to a small batch of sample sets) simultaneously calculates the predicted values ​​of all parameters, enabling multi-task prediction to be completed with a single forward propagation.

[0143] Gradient collaboration: The gradient of the shared layer is updated by the weighted average of all branches, and the gradient of the shared layer is updated by the weighted average of the gradients of each branch; backpropagation updates the shared layer.

[0144] Joint training is similar to a multi-task pipeline. The core idea is to perform a single computation and output multiple paths, which is a single forward propagation output.

[0145] The core of gradient coordination is to prevent a certain parameter from dominating the model and causing the prediction of other parameters to fail. The goal is to improve system stability and reduce multi-parameter prediction error.

[0146] In the training process, the parallelism of "joint training" and "gradient collaboration" is reflected in the fact that they together form a closed-loop training unit, rather than being completely synchronized in time. They are completed within the same cycle, characterized by the synchronous completion of multi-task prediction and shared gradient updates in a single iteration.

[0147] (26) Monitor the weighted sum of losses for all output parameters; determine if the loss has not decreased for ten consecutive epochs. Ten consecutive epochs refer to 10 complete epochs in the training process.

[0148] If the weighted sum of losses does not decrease for 10 consecutive epochs, training is stopped; this is an early stopping strategy.

[0149] (3) Lightweight deployment.

[0150] Model lightweighting: The trained XGBoost model is used as the prediction model and exported via ONNX Runtime. The trained DNN model is optimized using TensorRT and quantized using FP16.

[0151] Edge deployment and continuous learning: After edge deployment through the industrial fan inverter host computer system, local real-time prediction is performed through the host computer, with a rapid response time; subsequently, the model is continuously updated incrementally, and redeployed after fine-tuning the model.

[0152] Edge deployment is achieved through ONNX / TensorRT, and dynamic model updates are supported, enabling lightweight and continuous learning.

[0153] The following section explains the comparison of phased models and the analysis of the importance of input features.

[0154] Figure 5 This paper presents a comparison of the loss convergence curves of the XGBoost model and the Deep Neural Network (DNN) model under different data scales, revealing the crucial role of core motor parameters in model training efficiency. Specifically, the extreme logarithmic feature, with its high SHAP value (92.5±1.2), exhibits a significant convergence acceleration effect in the early stages of training, while the rated power feature, after multimodal fusion, reduces the validation set loss by 27.8% (p<0.01). Furthermore, a dynamic architecture switching strategy is employed, adaptively adjusting model complexity to achieve a 38% reduction in training time while maintaining prediction accuracy. Experiments demonstrate that this composite optimization strategy enables the motor parameter modeling task to achieve a breakthrough balance between accuracy and efficiency in limited data scenarios.

[0155] Figure 6The prediction accuracy of six key motor control parameters selected from the output parameters was compared to systematically evaluate the performance differences between the XGBoost model and the DNN model. The selected control parameters cover core functional dimensions such as carrier frequency, field weakening control threshold, and pre-charge timing. The DNN model showed a significant advantage in predicting nonlinear strongly coupled parameters (represented by the D-axis proportional coefficient KP), with a determination coefficient R0. 2 The accuracy exceeds 0.9 (within the 95% confidence interval), while the XGBoost model maintains a relative error of <15% in predicting time-sensitive parameters (such as precharge time). Thanks to the innovative end-to-end multi-output prediction framework, the two models effectively reduce the coupling interference between parameters through joint optimization, improving the overall prediction stability by 42%. Experiments show that this framework significantly expands the parameter prediction boundary of complex motor control systems while maintaining model interpretability.

[0156] Figure 7 The data quantifies the global impact of some input features on all output parameters. Data shows that the extreme logarithmic feature, with a global variance explanation rate of 92% (weighted based on SHAP values), is the most dominant input, influencing many output parameters. This can guide embedded software engineers to prioritize the calibration of highly important features (such as extreme logarithmic features), improve debugging efficiency, and accelerate the prediction of multiple parameters of the model.

[0157] Figure 8 The performance of different frameworks was compared from a deployment perspective. In terms of memory optimization, the ONNX format reduces model size by 40% (215MB → 128MB), significantly reducing storage pressure at the edge. Regarding inference performance, the DNN-TensorRT solution exhibits a low latency of 13.7ms, fully meeting the <20ms real-time requirements of industrial control. Its throughput of 195req / s, combined with a multi-device heterogeneous parallel mechanism, can support the collaborative debugging needs of complex systems. A comparison of deployment strategies shows that the XGBoost model achieves inference acceleration through ONNX Runtime, while the DNN model, compiled with TensorRT, performs better in terms of GPU resource utilization and memory bandwidth optimization. Together, they constitute the Pareto optimal solution for edge deployment.

[0158] Figure 9 This is a boundary graph for dynamic model switching decisions. The graph shows that when the data volume is ≤45000, the XGBoost model prioritizes real-time performance; when the data volume is ≥45000, the DNN model is activated to capture complex coupling relationships. Complexity is strongly correlated with data volume (R = 0.82), reflecting the characteristics of actual industrial data. Compared to a fixed model approach, this dynamic model switching decision improves debugging efficiency by 65%.

[0159] The wind turbine controller control parameter prediction method in this embodiment adopts a dynamic modeling strategy, that is, it adaptively switches between the XGBoost model (small data volume) and the DNN model (large data volume) based on the data scale.

[0160] The wind turbine controller control parameter prediction method of this application performs staged modeling based on data scale (dynamic switching between XGBoost model and DNN model) and conducts an end-to-end training process of multi-output regression (including the design of weighted MSE loss function). Among them, the XGBoost model gain algorithm is more robust, and the DNN model has a lightweight structure and is suitable for medium data scale.

[0161] The wind turbine controller control parameter prediction method of this application has the following characteristics:

[0162] (1) Dynamic model switching mechanism: When the amount of sample data is ≤45,000, XGBoost parallelization is used for fast modeling; when the amount of data is >45,000, ReLU-DNN deep fitting is switched to avoid the limitations of static models.

[0163] (2) Multi-output end-to-end optimization: The weighted MSE loss function assigns weights to different parameters to achieve error balance; multi-parameter optimization is performed simultaneously to reduce the accumulation of errors in step-by-step prediction.

[0164] (3) Lightweight deployment technology: ONNX achieves cross-platform compatibility (supports Windows / Linux embedded systems); TensorRT compresses model volume through layer fusion and quantization to adapt to edge devices.

[0165] The wind turbine controller control parameter prediction method of this application has the following technical advantages:

[0166] (1) Efficiency improvement: Manual debugging takes an average of 8 hours, while model prediction takes less than 1 minute; the multi-output end-to-end framework saves 70% of training time compared to traditional step-by-step optimization.

[0167] (2) Improved accuracy: The average error of multiple parameters was reduced from 25% in manual debugging to below 10%; the verification loss in the XGBoost stage was 0.2, and it was further reduced to 0.1 in the DNN stage.

[0168] (3) Deployment advantages: ONNX model size is reduced by 40%, TensorRT inference latency is <10ms; supports real-time response of edge devices, and memory usage is reduced by 50%.

[0169] Example 2

[0170] Based on the design of the wind turbine controller control parameter prediction method in Embodiment 1 above, Embodiment 2 proposes a wind turbine controller control parameter prediction system, including: a dynamic modeling module, a calling module, etc., see [link / reference]. Figure 10 As shown.

[0171] The dynamic modeling module is used for the following purposes: When the number of samples n is less than or equal to a set number m, an XGBoost model is built. The input parameters of the XGBoost model include the physical parameters of the wind turbine, and the output parameters include the control parameters of the wind turbine and the feature importance weights of each control parameter. The XGBoost model is trained using n samples, and the trained XGBoost model is used as the prediction model. When the number of samples n increases to more than the set number m, a DNN model is built. The input parameters of the DNN model include the physical parameters of the wind turbine and the feature importance weights output by the XGBoost model, and the output parameters of the DNN model include the control parameters of the wind turbine. The DNN model is trained using the newly added nm samples, where the feature importance weights output by the XGBoost model are used as the initial attention weights of the DNN model. The trained DNN model is used as the prediction model.

[0172] The calling module is used to: obtain the actual physical parameters of the wind turbine and input them into the prediction model to obtain the predicted actual control parameters of the wind turbine.

[0173] In some embodiments of this application, the dynamic modeling module is specifically used for:

[0174] Normalize each input parameter;

[0175] A shared feature extraction layer is used to extract common features for each input parameter;

[0176] The extracted general features are assigned to different prediction branch networks, and each prediction branch network is used to predict one output parameter;

[0177] For each prediction branch network, a loss function is designed and optimized.

[0178] In some embodiments of this application, the dynamic modeling module is specifically used for:

[0179] During the training of the XGBoost model,

[0180] When the number of samples n≤q, the initial weights of the loss function are static weights;

[0181] When the number of samples n∈(q, m], the weights of the loss function are dynamically adjusted according to the error ratio; where q is a preset fixed value.

[0182] During the training of a DNN model, the initial weights of the loss function are adjusted based on the SHAP value.

[0183] In some embodiments of this application, the wind turbine controller control parameter prediction system further includes: a lightweight deployment module, which is used for:

[0184] The trained XGBoost model is used as a prediction model and exported via ONNX Runtime;

[0185] The trained DNN model is used as the prediction model, and then optimized using TensorRT and quantized using FP16.

[0186] The specific working process of the wind turbine controller control parameter prediction system has been detailed in the above-mentioned wind turbine controller parameter prediction method, and will not be repeated here.

[0187] In this embodiment of the wind turbine controller control parameter prediction system, when the number of samples n ≤ a set number m, an XGBoost model is established, trained using n samples, and the trained XGBoost model is used as the prediction model. When the number of samples n increases to greater than the set number m, a DNN model is established, trained using the newly added nm samples, wherein the feature importance weights output by the XGBoost model are used as the initial attention weights of the DNN model; the trained DNN model is used as the prediction model. Therefore, this embodiment of the wind turbine controller control parameter prediction system establishes different models according to the number of samples, avoiding data overfitting and the limitations of a fixed model, thus solving the technical problem of overfitting that easily occurs when using a fixed model in the prior art.

[0188] The wind turbine controller control parameter prediction system in this embodiment is a multi-parameter intelligent prediction system for industrial wind turbine frequency converters. It includes a data preprocessing module, a dynamic modeling module, a lightweight deployment module, etc., and integrates ONNX / TensorRT model compression and acceleration technology.

[0189] Example 3

[0190] Based on the design of the wind turbine controller control parameter prediction system in Embodiment 2 above, Embodiment 3 proposes an edge device, including: the wind turbine controller control parameter prediction system.

[0191] The edge device in this embodiment implements a wind turbine controller control parameter prediction system to predict wind turbine controller control parameters. It establishes different models based on the number of samples to avoid data overfitting and the limitations of fixed models, thus solving the technical problem of overfitting that easily occurs when using fixed models in the prior art.

[0192] The industrial edge device is designed with a fan controller control parameter prediction system, which can perform real-time parameter prediction and debugging.

[0193] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of predicting a control parameter of a fan controller, the method comprising: include: ​ (1) Model building and training process: When the number of samples n is less than or equal to the set number m, an XGBoost model is established. The input parameters of the XGBoost model include the physical parameters of the wind turbine, and the output parameters of the XGBoost model include the control parameters of the wind turbine and the feature importance weights of each control parameter. The XGBoost model is trained using n samples, and the trained XGBoost model is used as the prediction model. When the number of samples n increases to a value greater than the set number m, a DNN model is established. The input parameters of the DNN model include the physical parameters of the wind turbine and the feature importance weights output by the XGBoost model. The output parameters of the DNN model include the control parameters of the wind turbine. The DNN model is trained using the newly added nm samples, wherein the feature importance weights output by the XGBoost model are used as the initial attention weights of the DNN model. The trained DNN model is then used as the prediction model. (2) Model invocation process: The actual physical parameters of the wind turbine are obtained and input into the prediction model to obtain the predicted actual control parameters of the wind turbine. The training process of the XGBoost model and the DNN model specifically includes: Normalize each input parameter; A shared feature extraction layer is used to extract common features for each input parameter; The extracted general features are assigned to different prediction branch networks, and each prediction branch network is used to predict one output parameter; For each prediction branch network, a loss function is designed and optimized. During the training of the XGBoost model, When the number of samples n ≤ q, the initial weights of the loss function are static weights; When the number of samples n∈(q, m], the weights of the loss function are dynamically adjusted according to the error ratio; where q is a preset fixed value; During the training of a DNN model, the initial weights of the loss function are adjusted based on the SHAP value; The object for SHAP value extraction is the XGBoost model. During the training process of the XGBoost model, the global SHAP matrix is ​​calculated and stored.

2. The method for predicting control parameters of a wind turbine controller according to claim 1, characterized in that: The physical parameters of the fan include its rated speed, rated power, and number of pole pairs; The control parameters of the wind turbine include control loop parameters, carrier frequency, and pre-charging time.

3. The method for predicting control parameters of a wind turbine controller according to any one of claims 1 to 2, characterized in that: The trained XGBoost model is used as a prediction model and exported via ONNX Runtime; The trained DNN model is used as the prediction model, and then optimized using TensorRT and quantized using FP16.

4. A fan controller control parameter prediction system characterized by: include: The dynamic modeling module is used to: establish an XGBoost model when the number of samples n is less than or equal to a set number m. The input parameters of the XGBoost model include the physical parameters of the wind turbine, and the output parameters of the XGBoost model include the control parameters of the wind turbine and the feature importance weights of each control parameter. The XGBoost model is trained using n samples, and the trained XGBoost model is used as the prediction model. When the number of samples n increases to a value greater than a set number m, a DNN model is established. The input parameters of the DNN model include the physical parameters of the wind turbine and the feature importance weights output by the XGBoost model. The output parameters of the DNN model include the control parameters of the wind turbine. The DNN model is trained using an additional nm samples, wherein the feature importance weights output by the XGBoost model are used as the initial attention weights of the DNN model. The trained DNN model is used as the prediction model. The calling module is used to: obtain the actual physical parameters of the wind turbine and input them into the prediction model to obtain the predicted actual control parameters of the wind turbine; The dynamic modeling module is specifically used for: Normalize each input parameter; A shared feature extraction layer is used to extract common features for each input parameter; The extracted general features are assigned to different prediction branch networks, and each prediction branch network is used to predict one output parameter; For each prediction branch network, a loss function is designed and optimized. The dynamic modeling module is specifically used for: During the training of the XGBoost model, When the number of samples n ≤ q, the initial weights of the loss function are static weights; When the number of samples n∈(q, m], the weights of the loss function are dynamically adjusted according to the error ratio; where q is a preset fixed value; During the training of a DNN model, the initial weights of the loss function are adjusted based on the SHAP value; The object for SHAP value extraction is the XGBoost model. During the training process of the XGBoost model, the global SHAP matrix is ​​calculated and stored.

5. The wind turbine controller control parameter prediction system according to claim 4, characterized in that: The wind turbine controller control parameter prediction system further includes: a lightweight deployment module, which is used for: The trained XGBoost model is used as a prediction model and exported via ONNX Runtime; The trained DNN model is used as the prediction model, and then optimized using TensorRT and quantized using FP16.

6. An edge device, characterized in that: Includes a wind turbine controller control parameter prediction system as described in any one of claims 4 to 5.

Images