Fan controller control parameter prediction method and system and edge device

Abstract

According to the fan controller control parameter prediction method and system and the edge device, when the number n of samples is smaller than or equal to the set number m, the XGBoost model is established, the XGBoost model is trained through the n samples, and the trained XGBoost model serves as the prediction model; when the number n of the samples is increased to be larger than the set number m, a DNN model is established, the newly-increased n-m samples are used for training the DNN model, and the feature importance weight output by the XGBoost model serves as the initial attention weight of the DNN model; and taking the trained DNN model as a prediction model. Therefore, according to the fan controller control parameter prediction method and system and the edge device, different models are established according to the number of samples, data overfitting is avoided, limitation of a fixed model is avoided, and the technical problem that overfitting is prone to occurring when the fixed model is used in the prior art is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of parameter prediction, and in particular, relates to a method, system and edge device for predicting control parameters of a wind turbine controller. Background Art

[0002] Traditional industrial fan variable frequency controllers rely on the experience of embedded software engineers for parameter debugging. Manual adjustment of multiple coupling parameters (such as current loop, carrier frequency, motor phase resistance, motor phase inductance, and pre-charge time) is required. This leads to the following core issues:

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

[0004] (2) Strong 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 capabilities: Traditional methods do not use historical data for modeling and are unable to optimize prediction accuracy through data accumulation.

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

[0007] Parameter debugging of industrial fan variable frequency controllers directly affects the operating efficiency and stability of fan equipment. Traditional methods rely on manual experience or a single model and are difficult to address the following challenges:

[0008] (1) Multi-parameter coupling: Parameters such as the current loop, carrier frequency, motor phase resistance, and motor phase inductance affect each other and require coordinated optimization;

[0009] (2) Dynamic changes in data scale: The amount of data is limited in the early stage, and the data continues to grow in the later stage, requiring adaptive adjustment of the model;

[0010] (3) Real-time requirements: Industrial scenarios require low-latency responses, and traditional models are large in deployment size and slow inference speed.

[0011] At present, the following schemes are usually used to adjust the controller parameters:

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

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

[0015] Therefore, the shortcomings of the prior art 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 cost, poor adaptability, and cannot handle new wind turbine controllers.

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

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

[0020] (5) Static DNN model: Severe overfitting of small data, no integration of lightweight technology, and high deployment cost. Summary of the Invention

[0021] The present invention provides a method for predicting control parameters of a fan controller, which solves the technical problem in the prior art that overfitting is prone to occur when using a fixed model.

[0022] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0023] The fan controller control parameter prediction method includes:

[0024] (1) Model establishment 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, wherein 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 a prediction model;

[0026] When the number of samples n increases to greater than the set number m, a DNN model is established, wherein 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, wherein the feature importance weights output by the XGBoost model serve as the initial attention weights of the DNN model; and the trained DNN model is used as a prediction model;

[0027] (2) Model calling process:

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

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

[0030] Normalize each input parameter;

[0031] Use shared feature extraction layers to extract common features for each input parameter;

[0032] The extracted common features are assigned to different prediction branch networks, each of which is used to predict an output parameter;

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

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

[0035] When the number of samples n≤q, the initial weight of the loss function is a static weight;

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

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

[0038] In some embodiments of the present application, the physical parameters of the fan include the rated speed, rated power, and pole pair number of the fan;

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

[0040] In some embodiments of the present 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, optimized by TensorRT and quantized by FP16.

[0042] The fan controller control parameter prediction system includes:

[0043] A dynamic modeling module, which is used to: when the number of samples n is less than or equal to the set number m, establish an XGBoost model, the input parameters of the XGBoost model include the physical parameters of the fan, and the output parameters of the XGBoost model include the control parameters of the fan and the feature importance weights of each control parameter; use n samples to train the XGBoost model, and use the trained XGBoost model as a prediction model; when the number of samples n increases to greater than the set number m, establish a DNN model, the input parameters of the DNN model include the physical parameters of the fan and the feature importance weights output by the XGBoost model, and the output parameters of the DNN model include the control parameters of the fan; use newly added nm samples to train the DNN model, wherein the feature importance weights output by the XGBoost model serve as the initial attention weights of the DNN model; and use the trained DNN model as a prediction model;

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

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

[0046] Normalize each input parameter;

[0047] Use shared feature extraction layers to extract common features for each input parameter;

[0048] The extracted common features are assigned to different prediction branch networks, each of which is used to predict an output parameter;

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

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

[0051] During the training process of the XGBoost model,

[0052] When the number of samples n≤q, the initial weight of the loss function is a static weight;

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

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

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

[0056] Export the trained XGBoost model as a prediction model through ONNX Runtime;

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

[0058] The edge device includes the wind turbine controller control parameter prediction system.

[0059] Compared with the prior art, the advantages and positive effects of the present invention are as follows: the fan controller control parameter prediction method, system and edge device of the present invention, when the number of samples n ≤ the set number m, establishes an XGBoost model, uses n samples to train the XGBoost model, and uses the trained XGBoost model as the prediction model; when the number of samples n increases to be greater than the set number m, establishes a DNN model, uses the newly added nm samples to train the DNN model, wherein the feature importance weights output by the XGBoost model are used as the initial attention weights of the DNN model; and uses the trained DNN model as the prediction model. Therefore, the fan controller control parameter prediction method, system and edge device of the present invention establish different models according to the number of samples, avoid data overfitting, avoid the limitations of fixed models, and solve the technical problem of the prior art that fixed models are prone to overfitting.

[0060] Other features and advantages of the present invention will become more apparent after reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0062] Figure 1 This is a flow chart of an embodiment of the method for predicting control parameters of a fan controller proposed in the present invention;

[0063] Figure 2 yes Figure 1 A flowchart of an embodiment of some steps in FIG.

[0064] Figure 3 This is a flow chart of another embodiment of the method for predicting control parameters of a fan controller proposed in the present invention;

[0065] Figure 4 yes Figure 3 A flowchart of an embodiment of some steps in FIG.

[0066] Figure 5 This is a graph of the loss of the multi-output model training based on motor parameters;

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

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

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

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

[0071] Figure 10 It is a structural block diagram of an embodiment of the wind turbine controller control parameter prediction system proposed by the present invention. DETAILED DESCRIPTION

[0072] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present 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 the present invention, terms such as "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer" indicating directions or positional relationships are based on the directions 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 component described must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present 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 the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0075] Example 1

[0076] The fan controller control parameter prediction method of this embodiment mainly includes the following steps, see Figure 1 shown.

[0077] Step S11: Model establishment 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 fan, and the output parameters of the XGBoost model include the control parameters of the fan 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 a prediction model.

[0079] When the number of samples n increases to 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 fan and the feature importance weights output by the XGBoost model. The output parameters of the DNN model include the control parameters of the fan. 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.

[0080] The feature importance weights provided by the XGBoost model are used to initialize the weights 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 been used to train the XGBoost model. Therefore, the remaining nm samples, that is, the newly added nm samples, are used to train the DNN model.

[0082] Step S12: Model calling process.

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

[0084] The prediction model is called, and the actual physical parameters of the fan are input into the prediction model, and the prediction model outputs the predicted actual control parameters of the fan.

[0085] Therefore, when the number of samples n ≤ the set number m, n samples are used to train the XGBoost model, 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 newly added nm samples are used to train the DNN model, 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] In the fan controller control parameter prediction method of this embodiment, when the sample number n ≤ the set number m, an XGBoost model is established, the XGBoost model is trained using n samples, and the trained XGBoost model is used as the prediction model; when the sample number n increases to greater than the set number m, a DNN model is established, the DNN model is trained using the newly added nm samples, wherein the feature importance weights output by the XGBoost model serve as the initial attention weights of the DNN model; and the trained DNN model is used as the prediction model. Therefore, the fan controller control parameter prediction method of this embodiment establishes different models based on the sample number, avoids data overfitting, avoids the limitations of fixed models, and solves the technical problem of overfitting easily occurring when using fixed models in the prior art.

[0087] During the training process, the input parameters of the XGBoost model also include annotation labels (label parameters), such as the startup status and operating status of the fan, to improve the accuracy of the trained XGBoost model.

[0088] In the DNN model, the ReLU activation function is used to solve the gradient disappearance problem and improve the nonlinear modeling capability.

[0089] In some embodiments of the present application, the training process of the XGBoost model and the DNN model specifically includes the following steps: Figure 2 shown.

[0090] Step S21: normalize each input parameter.

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

[0092] Step S23: assigning the extracted common features to different prediction branch networks, each prediction branch network is used to predict an output parameter.

[0093] Step S24: Design a loss function for each prediction branch network and optimize the loss function.

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

[0095] Through the design 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 the present application, during the training process of the XGBoost model,

[0097] When the number of samples n≤q, the initial weight of the loss function is a static weight.

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

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

[0100] Through the above design, we can select more appropriate weights for the loss function of the XGBoost model, which is convenient for the training of the XGBoost model.

[0101] In some embodiments of the present application, during the training process of the DNN model, the initial weights of the loss function are adjusted based on the SHAP value, thereby selecting a relatively appropriate initial weight for the loss function of the DNN model, thereby facilitating the training of the DNN model.

[0102] In some embodiments of the present application, q=1000, m=45000.

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

[0104] When the number of samples n≤1000, the initial weight of the loss function is a static weight.

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

[0106] When the number of samples n of historical data increases to >45,000, the newly added n-45,000 samples are used to train the DNN model; during the training process of the DNN model, the initial weight of the loss function is adjusted based on the SHAP value.

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

[0108] Dynamic adjustment based on error ratio: Applicable to adaptive data offset when data distribution changes.

[0109] SHAP value-based adjustment: Combined with SHAP value optimization (when feature interpretability is required), it is suitable for high-precision requirements and feature interpretability requirements.

[0110] The weight of the loss function is dynamically adjusted based on the error ratio. This refers to the dynamic adjustment based on the error ratio of output parameters such as the current loop proportional / integral coefficient, the speed loop proportional / integral coefficient, the carrier frequency, and the pre-charge time. By amplifying the weight of parameters with high error, the dimensionality difference between the proportional coefficient and the pre-charge time is resolved.

[0111] The initial weights of the loss function are adjusted based on SHAP values. This means that the initial weights are adjusted based on the SHAP values ​​of the quantified features. First, the SHAP values ​​of the input features (i.e., input parameters) on each output parameter are extracted from the XGBoost model. The relative influence of the input features on the 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 SHAP value extraction object is the XGBoost model. During the XGBoost training process, the global SHAP matrix is ​​calculated and stored, which is used to initialize the DNN model and then obtain the SHAP value.

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

[0114] Therefore, the model can simultaneously predict multiple control parameters, avoid step-by-step optimization error accumulation, and achieve end-to-end multi-output regression.

[0115] Control loop parameters, including 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, weak magnetic control threshold, etc.

[0117] Phase resistance and phase inductance are inherent physical parameters of the motor. However, during 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, pre-charge time, etc., avoiding the accumulation of step-by-step optimization errors.

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

[0120] In some embodiments of the present 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 through ONNX Runtime, compressing the model size by 40% (e.g., 200MB → 120MB).

[0122] In some embodiments of the present application, the trained DNN model is used as a prediction model, and TensorRT optimization and FP16 quantization are used to achieve model lightweighting and save deployment space.

[0123] Use TensorRT optimization and FP16 quantization on the trained DNN model to reduce model size and computing 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 perform real-time inference.

[0125] Next, combine Figure 3 and Figure 4 , specifically explain the specific workflow of the fan controller control parameter prediction method of this application.

[0126] (1) Dynamic modeling in stages.

[0127] Data collection stage: Data collection and data labeling are performed through the industrial fan inverter host computer system.

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

[0129] Data growth phase (data volume > 45,000): When the data volume threshold is triggered or the validation set loss does not decrease for three consecutive times, switch to the deep neural network model (DNN). 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 data volume refers to the number of samples. One sample (one piece of data) = 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 distribution: This solves the training imbalance problem caused by different dimensions and numerical ranges of multiple output parameters, and achieves error balance by dynamically adjusting the loss weights of different parameters.

[0133] The implementation process of multi-output end-to-end optimization technology, that is, the training process of the model, such as Figure 4 shown.

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

[0135] (22) Common feature extraction: All input features are extracted through a shared feature extraction layer (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 parameter group and performs specialized predictions 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 data volume is ≤1000, static weight (quick start) is selected. It is suitable for initial deployment or when the importance of parameters is clear, and it is simple to implement and has low computational overhead.

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

[0140] Dynamic weight allocation strategy (based on SHAP value quantification and feature impact adjustment): During the deployment phase, when the data volume is greater than 45,000, combined with SHAP value optimization (when feature interpretability is required), it is suitable for high-precision requirements and feature interpretability needs.

[0141] (25) Joint training and gradient coordination parallelism.

[0142] Joint training: propagates and 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, realizing a single forward propagation to complete multi-task prediction.

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

[0144] Joint training is similar to a multi-task pipeline. The core idea is one-time calculation and multiple outputs, which is a single forward propagation output.

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

[0146] In the training process, the parallelism of "joint training" and "gradient collaboration" is reflected in the fact that they jointly constitute 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 sum of the weighted losses of all output parameters and determine whether it has not decreased for ten consecutive rounds. Ten consecutive rounds refers to 10 complete epochs in the training process.

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

[0149] (3) Lightweight deployment.

[0150] Model lightweighting: Export the trained XGBoost model as a prediction model through ONNX Runtime. Use TensorRT optimization and FP16 quantization on the trained DNN model.

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

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

[0153] Next, we will explain the stage-by-stage model comparison and input feature importance analysis.

[0154] Figure 5 The 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 key role of the core parameters of the motor in the model training efficiency. Among them, the pole logarithm feature shows a significant convergence acceleration effect at the beginning of training through a high SHAP value (92.5±1.2), and the rated power feature reduces the loss value of the validation set by 27.8% (p<0.01) after multimodal fusion. A dynamic architecture switching strategy is further adopted to adaptively adjust the model complexity, achieving a 38% compression in training time while maintaining prediction accuracy. Experiments show 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 was compared from the output parameters, and the performance difference between the XGBoost model and the DNN model was systematically evaluated. The selected control parameters covered core functional dimensions such as carrier frequency, weak magnetic control threshold, and pre-charge timing. Among them, the DNN model showed significant advantages in the prediction of nonlinear strong coupling parameters (represented by the D-axis proportional coefficient KP), and the determination coefficient R 2 The XGBoost model achieved a relative error of less than 15% for the prediction of timing-sensitive parameters (such as pre-charge time). Thanks to an innovative end-to-end multi-output prediction framework, the two models were jointly optimized to effectively reduce the coupling interference between parameters, improving overall prediction stability by 42%. Experiments demonstrate that this framework significantly expands the boundaries of parameter prediction for complex motor control systems while maintaining model interpretability.

[0156] Figure 7 The global impact of some input features on all output parameters was quantified in the paper; the data showed that the pole logarithm feature became the most dominant input with a global variance explanation rate of 92% (based on the weighted SHAP value), affecting more output parameters. This can guide embedded software engineers to prioritize the calibration of high-importance features (such as pole logarithm), improve debugging efficiency, and accelerate model multi-parameter prediction.

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

[0158] Figure 9 This is a dynamic model switching decision boundary diagram. It shows that when the data volume is ≤45,000, the XGBoost model prioritizes real-time performance. When the data volume is ≥45,000, the DNN model is activated to capture complex coupling relationships. Complexity is strongly correlated with data volume (R = 0.82), reflecting the characteristics of real-world industrial data. This dynamic model switching decision improves debugging efficiency by 65% ​​compared to a fixed model solution.

[0159] The wind turbine controller control parameter prediction method of this embodiment adopts a dynamic modeling strategy, that is, adaptively switching 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 proposed in this application uses phased modeling based on data scale (dynamically switching between the XGBoost model and the DNN model) and an end-to-end training process for multi-output regression (including the design of a weighted MSE loss function). The XGBoost model's gain algorithm is more robust, while the DNN model's structure is lightweight and adaptable to medium-sized data scales.

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

[0162] (1) Dynamic model switching mechanism: When the sample data volume is ≤45,000, XGBoost parallelization is used for rapid modeling; when the data volume is greater than 45,000, ReLU-DNN deep fitting is used 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; multiple parameters are optimized simultaneously to reduce the error accumulation of step-by-step prediction.

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

[0165] The fan 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) Accuracy improvement: The average error of multiple parameters is reduced from 25% in manual debugging to below 10%; the verification loss in the XGBoost stage is 0.2, and it is further reduced to 0.1 in the DNN stage.

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

[0169] Example 2

[0170] Based on the design of the fan controller control parameter prediction method of the above embodiment 1, this embodiment 2 proposes a fan controller control parameter prediction system, including: a dynamic modeling module, a calling module, etc. Figure 10 shown.

[0171] A dynamic modeling module is used to: when the number of samples n is less than or equal to the set number m, establish an XGBoost model, the input parameters of the XGBoost model include the physical parameters of the fan, and the output parameters of the XGBoost model include the control parameters of the fan and the feature importance weights of each control parameter; use n samples to train the XGBoost model, and use the trained XGBoost model as a prediction model; when the number of samples n increases to greater than the set number m, establish a DNN model, the input parameters of the DNN model include the physical parameters of the fan and the feature importance weights output by the XGBoost model, and the output parameters of the DNN model include the control parameters of the fan; use the newly added nm samples to train the DNN model, wherein the feature importance weights output by the XGBoost model are used as the initial attention weights of the DNN model; and use the trained DNN model as a prediction model.

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

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

[0174] Normalize each input parameter;

[0175] Use shared feature extraction layers to extract common features for each input parameter;

[0176] The extracted common features are assigned to different prediction branch networks, each of which is used to predict an output parameter;

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

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

[0179] During the training process of the XGBoost model,

[0180] When the number of samples n≤q, the initial weight of the loss function is a static weight;

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

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

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

[0184] Export the trained XGBoost model as a prediction model through ONNX Runtime;

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

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

[0187] The fan controller control parameter prediction system of this embodiment establishes an XGBoost model when the sample number n ≤ the set number m, uses n samples to train the XGBoost model, and uses the trained XGBoost model as the prediction model. When the sample number n increases to greater than the set number m, a DNN model is established and the newly added nm samples are used to train the DNN model, wherein the feature importance weights output by the XGBoost model serve as the initial attention weights of the DNN model; and the trained DNN model is used as the prediction model. Therefore, the fan controller control parameter prediction system of this embodiment establishes different models based on the sample number, avoids data overfitting, avoids the limitations of fixed models, and solves the technical problem of overfitting easily occurring when using fixed models in the prior art.

[0188] The fan controller control parameter prediction system of this embodiment is a multi-parameter intelligent prediction system for industrial fan variable frequency controllers, including 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 fan controller control parameter prediction system of the above-mentioned embodiment 2, this embodiment 3 proposes an edge device, including: the fan controller control parameter prediction system.

[0191] The edge device of this embodiment is designed with a fan controller control parameter prediction system to implement a fan controller control parameter prediction method, establish different models according to the number of samples, avoid data overfitting, avoid the limitations of fixed models, and solve the technical problem of overfitting that is prone to occur when using fixed models in the prior art.

[0192] A fan controller control parameter prediction system is designed in the industrial edge device, 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, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting control parameters of a fan controller, characterized by: include: (1) Model establishment and training process: When the number of samples n is less than or equal to the set number m, an XGBoost model is established, wherein 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; Using n samples to train the XGBoost model, and using the trained XGBoost model as a prediction model; When the number of samples n increases to greater than the set number m, a DNN model is established, wherein 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, wherein the feature importance weights output by the XGBoost model serve as the initial attention weights of the DNN model; Using the trained DNN model as a prediction model; (2) Model calling process: The actual physical parameters of the fan are obtained and input into the prediction model to obtain the predicted actual control parameters of the fan.

2. The method for predicting control parameters of a fan controller according to claim 1, wherein: The training process of the XGBoost model and the DNN model specifically includes: Normalize each input parameter; Use shared feature extraction layers to extract common features for each input parameter; The extracted common features are assigned to different prediction branch networks, each of which is used to predict an output parameter; For each prediction branch network, a loss function is designed and optimized.

3. The method for predicting control parameters of a fan controller according to claim 2, characterized in that: During the training process of the XGBoost model, When the number of samples n≤q, the initial weight of the loss function is a static weight; When the number of samples n∈(q,m], the weight of the loss function is dynamically adjusted according to the error ratio; where q is a preset fixed value; During the training process of the DNN model, the initial weights of the loss function are adjusted based on the SHAP values.

4. The method for predicting control parameters of a fan controller according to claim 1, wherein: Physical parameters of the fan, including rated speed, rated power, and pole pairs; The control parameters of the fan include control loop parameters, carrier frequency, and pre-charging time.

5. The method for predicting control parameters of a wind turbine controller according to any one of claims 1 to 4, characterized in that: Export the trained XGBoost model as a prediction model through ONNX Runtime; The trained DNN model is used as the prediction model, optimized by TensorRT and quantized by FP16.

6. Fan controller control parameter prediction system, characterized by: include: A dynamic modeling module is configured to establish an XGBoost model when the number of samples n is less than or equal to the set number m, wherein 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 the control parameters; The XGBoost model is trained using n samples, and the trained XGBoost model is used as a prediction model. When the number of samples n increases to greater than a set number m, a DNN model is established, wherein 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, 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 a prediction model. The calling module is used to obtain the actual physical parameters of the fan and input them into the prediction model to obtain the predicted actual control parameters of the fan.

7. The wind turbine controller control parameter prediction system according to claim 6, characterized in that: The dynamic modeling module is specifically used to: Normalize each input parameter; Use shared feature extraction layers to extract common features for each input parameter; The extracted common features are assigned to different prediction branch networks, each of which is used to predict an output parameter; For each prediction branch network, a loss function is designed and optimized.

8. The wind turbine controller control parameter prediction system according to claim 7, characterized in that: The dynamic modeling module is specifically used to: During the training process of the XGBoost model, When the number of samples n≤q, the initial weight of the loss function is a static weight; When the number of samples n∈(q,m], the weight of the loss function is dynamically adjusted according to the error ratio; where q is a preset fixed value; During the training process of the DNN model, the initial weights of the loss function are adjusted based on the SHAP values.

9. The wind turbine controller control parameter prediction system according to any one of claims 6 to 8, characterized in that: The wind turbine controller control parameter prediction system further includes a lightweight deployment module, which is used to: Export the trained XGBoost model as a prediction model through ONNX Runtime; The trained DNN model is used as the prediction model, optimized by TensorRT and quantized by FP16.

10. Edge device, characterized by: The wind turbine controller comprises a control parameter prediction system according to any one of claims 6 to 9.

Images