Wind turbine full wind speed operation characteristic modeling method, system and electronic equipment

Abstract

This invention proposes a method, system, and electronic equipment for modeling the full-wind-speed operation characteristics of wind turbines. The method includes: collecting operational monitoring data of wind turbines; processing the operational monitoring data to form a sample set for modeling full-wind-speed operation characteristics; using a random forest algorithm to evaluate the importance of candidate features and select a subset of target features; using a recurrent neural network (RNN) to extract temporal dependency features from the subset of target features; using a convolutional neural network (CNN) to extract local pattern features from the subset of target features; and fusing the temporal dependency features and local pattern features to establish a model of the full-wind-speed operation characteristics of the wind turbine and output the predicted operation characteristics. By using the random forest algorithm to select the importance of multi-dimensional candidate features, and combining the ability of RNN to extract long-term temporal dependencies with the advantage of CNN in perceiving local fluctuation patterns, high-precision fusion modeling of the operation characteristics of wind turbines across the entire wind speed range is achieved.

Description

Technical Field

[0001] This invention relates to the fields of wind power generation and machine learning, and in particular to a method, system and electronic equipment for modeling the full wind speed operation characteristics of wind turbine units. Background Technology

[0002] Modeling the operational characteristics of wind turbines across the entire wind speed range is a crucial foundation for performance evaluation, condition monitoring, control optimization, and fault diagnosis. During actual operation, wind turbines experience several typical operating conditions, including power ramp-up, rated operation, and high-wind-speed power limitation, within the cut-in and cut-out wind speed range. The dynamic characteristics of each region differ significantly, and the coupling relationships between operating parameters exhibit strong nonlinearity and time-dependent behavior as operating conditions change. Establishing an accurate and robust operational characteristic model across the entire wind speed range remains a core technological challenge in the field of digital operation and maintenance of wind turbines.

[0003] Existing wind turbine operating characteristic modeling methods are mainly divided into two categories: mechanistic modeling and data-driven modeling. Mechanistic modeling methods rely on precise physical equations and parameter identification, requiring joint modeling of multiple physical processes such as aerodynamics, mechanical transmission, and electrical conversion. Parameter acquisition is difficult, model debugging cycles are long, and it is difficult to effectively characterize performance baseline drift caused by turbine aging and blade contamination, limiting its application in practical engineering scenarios. Data-driven modeling methods directly learn the input-output mapping relationship from historical operating data, requiring no precise physical priors, and have strong engineering applicability. However, existing data-driven methods still have significant shortcomings in model structure design. Specifically, in terms of model structure, existing methods mostly adopt a single neural network architecture. A single model exhibits unstable performance in cross-condition generalization. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a method, system, and electronic equipment for modeling the full wind speed operation characteristics of wind turbine units, in order to solve the problems in existing technologies.

[0005] This invention proposes a method for modeling the full-wind-speed operation characteristics of wind turbine units, including:

[0006] Collect operational monitoring data of wind turbine units across the entire wind speed range;

[0007] The operational monitoring data is preprocessed, divided into operating conditions, sliced ​​into time windows, and labeled with samples to form a full-wind-speed operational characteristic modeling sample set;

[0008] The importance of candidate features is evaluated using a random forest algorithm, and a subset of target features is obtained by selection. The candidate features are generated by feature engineering extension based on the features in the sample set of the full wind speed operation characteristics modeling.

[0009] A recurrent neural network (RNN) is used to extract the temporal dependency features from the target feature subset;

[0010] A convolutional neural network (CNN) is used to extract local pattern features from the target feature subset;

[0011] The time-dependent features and local pattern features are fused to establish a full-wind-speed operation characteristic model of the wind turbine and output the operation characteristic prediction results.

[0012] Furthermore, the operational monitoring data includes one or more of the following: wind speed, generator speed, impeller speed, blade pitch angle, yaw angle, active power, reactive power, torque, vibration, temperature, and unit operating status.

[0013] The preprocessing includes: abnormal operating condition removal, missing value imputation, time synchronization, normalization, and start / stop transition segment identification;

[0014] The candidate features include the product of rotational speed and pitch angle, the ratio of power to rotational speed, the difference between temperature measurement points, and the moving mean and moving standard deviation of each signal.

[0015] The predicted operating characteristics include one or more of the following: power characteristics, speed characteristics, load characteristics, temperature characteristics, and deviation residuals.

[0016] Furthermore, the step of using the random forest algorithm to evaluate the importance of candidate features and filter to obtain a subset of target features includes:

[0017] The importance of candidate features is evaluated using the random forest algorithm to obtain the normalized importance score of the candidate features;

[0018] Sort the features by importance score from highest to lowest to form a feature importance ranking list;

[0019] The candidate features in the feature importance ranking list are accumulated in descending order of importance score, and the feature set included when the cumulative importance reaches a preset total is used as the target feature subset.

[0020] Furthermore, if there are feature pairs in the target feature subset with a Pearson correlation coefficient higher than a preset threshold, then the feature pair with the lower importance score is removed from the target feature subset to update the target feature subset.

[0021] Furthermore, the RNN uses a Long Short-Term Memory (LSTM) network as its basic unit. The LSTM adopts a two-layer stacked design, wherein the first LSTM layer includes a first number of hidden units, the second LSTM layer includes a second number of hidden units, and a Dropout regularization layer is set between the first LSTM layer and the second LSTM layer; the first number is greater than the second number.

[0022] Furthermore, the CNN adopts a one-dimensional temporal convolutional structure with the time step as the convolution direction to perform sliding perception on the local morphological features in the input target feature subset; the one-dimensional temporal convolutional structure adopts a design of multi-scale convolutional kernels and multiple sets of convolutional branches, with different convolutional branches for different time steps to cover the fluctuation features of the running detection data at different time scales.

[0023] Furthermore, the feature fusion is achieved using an attention-weighted fusion mechanism.

[0024] Furthermore, the method also includes:

[0025] If new operation monitoring data is obtained, the parameters of the fully connected network of the output layer of the wind turbine full wind speed operation characteristic model are adjusted based on the operation monitoring data of normal operation within the first time period, and the parameters of the bottom feature extraction network of the wind turbine full wind speed operation characteristic model are fixed.

[0026] Based on the operational monitoring data during the second time period, the parameters of both the output layer fully connected network and the bottom layer feature extraction network of the wind turbine full-wind speed operation characteristic model are updated; wherein, the duration of the first time period is shorter than the duration of the second time period.

[0027] This invention also proposes a modeling system for the full wind speed operation characteristics of wind turbine units, comprising:

[0028] The data acquisition module is used to collect operational monitoring data of the wind turbine across the entire wind speed range;

[0029] The sample construction module is used to preprocess, divide, slice time windows, and label the operation monitoring data to form a full-wind speed operation characteristic modeling sample set.

[0030] The feature selection module is used to evaluate the importance of candidate features using the random forest algorithm and select a subset of target features; the candidate features are generated by feature engineering extension based on the features in the full wind speed operation characteristic modeling sample set;

[0031] The temporal feature extraction module is used to extract temporal dependency features from the target feature subset using a recurrent neural network (RNN).

[0032] The local pattern extraction module is used to extract local pattern features from the target feature subset using a convolutional neural network (CNN).

[0033] The fusion modeling module is used to perform feature fusion on the time-dependent features and local pattern features, establish a full-wind-speed operation characteristic model of the wind turbine, and output the operation characteristic prediction results.

[0034] This invention also proposes an electronic device, comprising:

[0035] One or more processors;

[0036] Memory, used to store one or more programs;

[0037] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-described method for modeling the full wind speed operation characteristics of wind turbine units.

[0038] This invention proposes a method, system, and electronic equipment for modeling the full-wind-speed operation characteristics of wind turbines. The method includes: collecting operational monitoring data of wind turbines across the entire wind speed range; preprocessing, dividing into operating conditions, slicing into time windows, and labeling samples of the operational monitoring data to form a full-wind-speed operation characteristic modeling sample set; using a random forest algorithm to evaluate the importance of candidate features and select a target feature subset; generating the candidate features through feature engineering based on features in the full-wind-speed operation characteristic modeling sample set; using a recurrent neural network (RNN) to extract temporal dependency features from the target feature subset; using a convolutional neural network (CNN) to extract local pattern features from the target feature subset; and fusing the temporal dependency features and local pattern features to establish a full-wind-speed operation characteristic model of the wind turbine and output the predicted operation characteristics. By using the random forest algorithm to screen the importance of multi-dimensional candidate features, and combining the ability of RNN to extract long-term temporal dependencies with the advantage of CNN in perceiving local fluctuation patterns, high-precision fusion modeling of the operating characteristics of wind turbines across the entire wind speed range is achieved. This helps to improve the problems of insufficient generalization ability of single architecture models under all operating conditions and the interference of redundant features on modeling accuracy, and provides a reliable operating characteristic modeling foundation for wind turbine performance evaluation, condition monitoring and intelligent operation and maintenance. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating a method for modeling the full wind speed operation characteristics of a wind turbine generator according to an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of the framework structure of a wind turbine full-wind-speed operation characteristic modeling system proposed in an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of the frame structure of an electronic device proposed in an embodiment of the present invention. Detailed Implementation

[0043] Various embodiments of this disclosure will be described more fully below. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0044] The terminology used in the various embodiments of this disclosure is for the purpose of describing particular embodiments only and is not intended to limit the various embodiments of this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this disclosure pertain. Terms (such as those defined in a generally used dictionary) are to be interpreted as having the same meaning as in the context of the relevant technical field and are not to be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this disclosure.

[0045] Example 1

[0046] Embodiment 1 of this invention discloses a method for modeling the full wind speed operation characteristics of wind turbine units, such as... Figure 1 As shown, it includes the following steps:

[0047] Step 101: Collect operational monitoring data of the wind turbine unit across the entire wind speed range; specifically, the operational monitoring data includes one or more of the following: wind speed, generator speed, rotor speed, pitch angle, yaw angle, active power, reactive power, torque, vibration, temperature, and unit operating status.

[0048] The data collection scope of the operation monitoring covers the core physical quantities of the wind turbine throughout its entire operating state. Wind speed data includes the wind speed at hub height measured by the nacelle anemometer and the reference wind speed at the meteorological tower, which is the primary input variable describing the unit's operating conditions and directly determines the operating segment of the unit. Generator speed and rotor speed (low-speed shaft speed) together reflect the operating state of the transmission system; their ratio corresponds to the gearbox transmission ratio and is an important input for power and load modeling. Pitch angle records the real-time pitch adjustment of each blade and is the core control variable for power regulation and load control in operating conditions above rated wind speed. Yaw angle reflects the unit's wind orientation; the existence of yaw error will lead to a reduction in the effective wind-catching area and introduce asymmetric aerodynamic loads. Active power and reactive power are the core output quantities of the unit during grid-connected operation, and accurate modeling of the power curve is the core objective of operating characteristic research; torque data, including low-speed shaft torque and generator electromagnetic torque, is a direct indicator for evaluating the mechanical load level of the transmission chain; vibration data comes from acceleration sensors in the nacelle, tower, and key parts of the transmission chain, reflecting the dynamic response characteristics of the unit structure; temperature data covers gearbox oil temperature, generator winding temperature, bearing temperature, and converter temperature, which are closely related to the thermal state of the unit; unit operating status signals (grid-connected / shutdown / fault / power restriction, etc.) are used for condition labeling and abnormal segment identification in the subsequent preprocessing stage.

[0049] The temporal resolution of the data is configured hierarchically according to signal type. SCADA (Supervisory Control And Data Acquisition) low-frequency data (power, speed, pitch angle, temperature, etc.) are typically acquired with a 1-minute acquisition cycle; for high-dynamic signals such as vibration, the system supports acquisition at second- or millisecond-level sampling frequencies, and uploads low-frequency feature quantities after feature extraction at the edge acquisition terminal to match the temporal granularity of SCADA data, facilitating time alignment and joint modeling of multi-source data.

[0050] The above systematic collection of multi-dimensional monitoring data of wind turbines across the entire wind speed operating range provides a complete data input covering multiple physical domains, including aerodynamics, mechanics, electrical systems, and thermodynamics, for subsequent feature selection and fusion modeling, laying a data foundation for high-precision modeling of the operating characteristics across the entire wind speed range.

[0051] Step 102: Preprocess the operation monitoring data, divide the operating conditions, slice the time window, and label the samples to form a full wind speed operation characteristic modeling sample set; specifically, the preprocessing includes: abnormal operating condition removal, missing value filling, time synchronization, normalization, and start-stop transition segment identification.

[0052] In this scheme, to eliminate the influence of different physical dimensions on model training, a minimum-maximum linear normalization method is used to map each feature quantity to the [0,1] interval. Let the original feature sequence be... Its normalized value The calculation formula is:

[0053]

[0054] in, and These represent the minimum and maximum values ​​of this feature in the full-speed operation monitoring data, respectively. The normalized data retains the distribution pattern of the original data and accelerates the gradient descent convergence process of the subsequent deep learning model.

[0055] This step is responsible for systematically preprocessing the raw operation monitoring data, dividing it into operating conditions and slicing it into time series samples, forming a standardized modeling sample set that can be directly used for feature selection and model training.

[0056] Preprocessing is the primary step in sample construction, aiming to eliminate various interference factors in the original data that affect the modeling quality. Abnormal operating condition removal identifies and excludes data from shutdowns, faults, power outages, maintenance, and start-up / shutdown transition periods, ensuring that the training samples only include data from the unit's normal steady-state operation. Specifically, this module reads the unit's operating status flags from the SCADA system, automatically identifies and removes off-grid operating periods; for start-up and shutdown transition periods, a 10-minute delay after the unit reaches grid-connected speed is used as the effective sample starting point, and a 10-minute delay before receiving the shutdown command is used as the effective sample ending point, avoiding contamination of steady-state operating characteristic modeling by transient transitions. Missing value imputation uses linear interpolation for short-term missing values ​​(continuous missing values ​​not exceeding 5 minutes), and directly removes the entire missing data segment for longer periods. Time synchronization resamples and aligns timestamps of signals with different sampling frequencies to a 1-minute resolution. Normalization performs minimum-maximum linear normalization on each feature quantity according to its physical range within the normal operating range, eliminating the adverse effects of different dimensions and magnitudes on model training.

[0057] The operating condition classification systematically segments the entire wind speed operating range according to the unit's control section. This module divides the unit's operating conditions into four typical ranges: the cut-in wind speed range (3 to 6 m / s), where the unit is in the low wind speed start-up and ramp-up phase, with power rapidly increasing with wind speed, and a strong nonlinear coupling relationship between speed and power; the wind speed below rated range (6 to 12 m / s), where the unit is in the maximum power point tracking control phase, with the pitch angle maintained at a fixed value corresponding to the optimal tip speed ratio, and power and speed mainly driven by wind speed; the range near rated wind speed (12 to 14 m / s), where the unit is in the power ramp-up phase and the transition zone between rated and rated operation, with the control system frequently switching pitch adjustment strategies, resulting in the greatest fluctuation in operating parameters; and the high wind speed range (14 m / s to cut-out wind speed), where the unit maintains power at the rated value through pitch adjustment, with the pitch angle becoming the most important control variable, and the load characteristics differ significantly from those in the low wind speed range. The samples for each working condition range are independently labeled, supporting the flexible choice between two strategies: segmented modeling of working conditions or unified modeling of all working conditions.

[0058] Time window slicing divides preprocessed continuous time-series data into fixed-length time-series sample segments. The default time window length is set to 30 time steps (i.e., 30 minutes), with a sliding step size of 5 time steps to ensure contextual continuity between adjacent samples. Each time window sample contains a multi-dimensional feature sequence of all time steps within the window and the corresponding modeling target value (such as the active power at the end of the window). For the segmented modeling strategy, only samples with the same operating condition label as the center of the time window are used to train the corresponding sub-model; for the unified modeling strategy, the operating condition category is included as an additional input feature in the samples, enabling the model to adaptively perceive operating condition segment information.

[0059] Through systematic preprocessing, working condition division, and time-series slicing, noise and non-steady-state interference in the original data were effectively removed, forming a high-quality modeling sample set with standardized structure and clear working condition annotations, providing a solid sample foundation for subsequent feature selection and deep learning model training.

[0060] Step 103: The importance of candidate features is evaluated using the random forest algorithm, and a subset of target features is obtained. Candidate features are generated by feature engineering based on features in the sample set of the full wind speed operation characteristics model. Specifically, candidate features include the product of speed and pitch angle, the ratio of power to speed, the difference between temperature measurement points, and the moving mean and moving standard deviation of each signal.

[0061] Step 103 uses the random forest algorithm to evaluate the importance of candidate features and selects a subset of target features. This includes: evaluating the importance of candidate features using the random forest algorithm to obtain normalized importance scores for the candidate features; sorting the candidate features by importance scores from high to low to form a feature importance ranking list; and accumulating the candidate features in the feature importance ranking list by importance scores from high to low, and using the accumulated importance to obtain the feature set included when the preset total is reached as the target feature subset.

[0062] This scheme uses Gini Importance as the feature importance evaluation metric for the Random Forest algorithm. Let the Random Forest consist of... The decision tree consists of several trees, for any feature Its importance score Defined as the sum of the reductions in Gini impurity caused by split nodes of this feature across all trees:

[0063]

[0064] in, For the first Using features in decision trees The set of nodes to be split; For nodes The reduction in Gini impurity before and after splitting can be replaced by the reduction in mean squared error for regression tasks. The feature importance scores, after normalization, form the final selection criterion.

[0065] Furthermore, if there are feature pairs in the target feature subset with a Pearson correlation coefficient higher than a preset threshold, then the feature pair with the lower importance score is removed from the target feature subset to update the target feature subset.

[0066] This step uses the random forest algorithm to perform data-driven importance assessment on all candidate input features. By objectively selecting and retaining a subset of core features that are strongly related to the modeling objective, redundant and noisy features are eliminated, thereby improving the training efficiency and generalization ability of the subsequent deep learning model.

[0067] The candidate feature set is constructed based on the original SCADA data, and further extended through feature engineering. Derived features built upon the original data include: the product of engine speed and pitch angle (approximately reflecting pitch control intensity), the power-to-engineering ratio (approximately electromagnetic torque, reflecting drivetrain load level), the difference between temperature measurement points (reflecting thermal equilibrium), and the moving mean and moving standard deviation of each signal (reflecting short-term trends and fluctuations). These derived features can extract more direct combined features representing the changing patterns of operational characteristics from the original data, helping to improve the discriminative power of random forest importance assessment.

[0068] Random forest importance evaluation uses Gini importance as the primary metric, statistically summarizing the contribution of each candidate feature to node splitting in the random forest ensemble classifier (or regressor). During training, the random forest consists of 500 decision trees. Each tree randomly selects the square root of the number of candidate features as a subset of candidate splitting features during training to effectively reduce interference from feature correlations. After training, the system outputs the normalized importance score for each feature, sorting them from highest to lowest to form a feature importance ranking list.

[0069] The feature selection strategy employs a cumulative importance threshold method. The system accumulates features according to their importance scores from high to low, and the feature set included when the cumulative importance reaches a preset total (the default setting is 90% of the total importance) is used as the target feature subset. This strategy automatically determines the optimal feature subset size while retaining most of the information, avoiding the subjectivity of manually setting the number of features. For the final selection results, the system additionally performs a correlation test. If there are highly correlated feature pairs with a Pearson correlation coefficient higher than 0.95 in the target feature subset, the one with lower importance is further removed to eliminate multicollinearity between features and improve the training stability of subsequent deep learning models.

[0070] In summary, by using the random forest algorithm to drive feature importance evaluation, objective and automated screening of multi-dimensional candidate features is achieved. This effectively eliminates the interference of redundant and weakly correlated features, reduces the input dimension and training complexity of subsequent RNN / CNN models, and improves the model's fitting accuracy and generalization ability to the changing patterns of real operating characteristics.

[0071] Step 104: Use a recurrent neural network (RNN) to extract temporal dependency features from the target feature subset;

[0072] This step uses a recurrent neural network (RNN) to deeply extract the time-dependent features in the target feature subset, capturing the long-term dynamic evolution of wind turbine operating parameters over time.

[0073] The RNN network architecture uses Long Short-Term Memory (LSTM) as its basic unit. LSTM, through a coordinated control mechanism of input, forget, and output gates, effectively overcomes the vanishing gradient problem in long-sequence training of standard RNNs. It can selectively remember and forget historical information within a 30-time-step input window, thereby accurately capturing the long-range dependencies between wind turbine operating parameters across multiple time steps. Typical long-range temporal dependencies include: the dynamic hysteresis characteristics of power response after wind speed rise, the impact of pitch control on power stability in subsequent time steps, and the thermal response hysteresis of temperature parameters to changes in mechanical load. These characteristics are key operational features that are difficult to accurately model using static or short-window models.

[0074] The LSTM network employs a two-layer stacked design: for example, the first LSTM layer contains 128 hidden units, responsible for extracting low-level local temporal patterns from the original temporal input; the second LSTM layer contains 64 hidden units, taking the hidden state sequence of the first layer as input, and further extracting high-level global temporal dependency features. A Dropout regularization layer (with a default Dropout rate of 0.2) is added between the two layers to suppress overfitting. The final hidden state vector of the second LSTM layer serves as the temporal feature extraction result and is output to the fusion modeling module for feature fusion. Furthermore, those skilled in the art will understand that the number of hidden units in the first and second LSTM layers can be adjusted according to the length of the input time window and the feature dimension, as long as the size of the hidden units in the first LSTM layer is greater than that in the second LSTM layer.

[0075] To address the unique characteristics of the full-wind-speed modeling scenario, the temporal feature extraction module also incorporates an adaptive working condition mechanism. The system concatenates the working condition segment codes (one-hot codes for four segments) into the input feature vector at each time step, enabling the LSTM to perceive the working condition environment at the current time step during temporal feature extraction. This allows for adaptive adjustment of the sensitivity to temporal patterns in different working condition segments within a unified modeling framework, thereby improving the targeting and accuracy of temporal feature extraction across the entire working condition range.

[0076] Therefore, by using deep temporal feature extraction through a dual-layer LSTM network, the dynamic evolution and cross-time step dependencies of wind turbine operating parameters across the entire wind speed range can be effectively captured. This provides a high-quality feature representation with rich temporal context information for the fusion modeling module, significantly improving the accuracy of operating characteristic prediction in response to dynamic operating condition changes.

[0077] Step 105: Use a convolutional neural network (CNN) to extract local pattern features from the target feature subset;

[0078] Specifically, RNN uses Long Short-Term Memory (LSTM) network as the basic unit. The LSTM adopts a two-layer stacked design, where the first LSTM layer includes a first number of hidden units and the second LSTM layer includes a second number of hidden units. A Dropout regularization layer is set between the first and second LSTM layers; the first number is greater than the second number.

[0079] CNN employs a one-dimensional temporal convolutional structure with the time step as the convolution direction, performing sliding perception on the local morphological features of the input target feature subset. The one-dimensional temporal convolutional structure uses a design of multi-scale convolutional kernels and multiple sets of convolutional branches. Different convolutional branches are for different time steps to cover the fluctuation features of the running detection data at different time scales.

[0080] This step uses a convolutional neural network (CNN) to perceive and extract local fluctuation features in the target feature subset, capturing short-term transient features and periodic fluctuation patterns within the time window of wind turbine operating parameters.

[0081] Specifically, the CNN network employs a one-dimensional temporal convolutional structure (1D-CNN), with the time step as the convolution direction, to perform sliding perception of local morphological features in the input temporal data. Compared to the advantage of LSTM in modeling global temporal dependencies, the convolutional kernel of 1D-CNN performs pattern matching on local combinations of feature values ​​within a fixed receptive field. This enables it to efficiently identify local morphological features such as short-term pulse fluctuations, periodic oscillations, and abrupt changes in feature slopes in operating parameters. These features have significant characterization value in scenarios such as instantaneous power fluctuations caused by wind speed turbulence, speed oscillations caused by pitch control, and periodic load pulses generated by the blade over-tower shadow effect.

[0082] The 1D-CNN network structure consists of three convolutional layers connected in series, for example, with 64, 128, and 64 kernels in each layer, and a uniform kernel length of 5 time steps. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function layer. The batch normalization layer accelerates training convergence and improves the model's robustness to changes in input distribution. A max-pooling layer (pooling window length of 2) is inserted between the first and second convolutional layers to downsample lower-level local features, increasing the receptive field of subsequent convolutional layers and reducing feature dimensionality. The output of the third convolutional layer is compressed into a fixed-dimensional feature vector using global average pooling and then output to the fusion modeling module, where it is fused with the temporal features extracted by LSTM.

[0083] The multi-scale convolutional kernel design further enhances the coverage of local pattern extraction. Based on the basic three-layer convolutional structure, this module sets up three sets of convolutional branches with kernel lengths of 3, 5, and 7 time steps in parallel, corresponding to local fluctuation patterns at three time scales: short (approximately 3 minutes), medium (approximately 5 minutes), and long (approximately 7 minutes), respectively. The output feature vectors of the three branches are concatenated and input into the subsequent fusion layer, enabling local pattern extraction to simultaneously cover the fluctuation features of operating parameters at different time scales, thus improving the model's ability to express multi-scale operational dynamics.

[0084] Therefore, by extracting local patterns through a multi-scale one-dimensional convolutional network, it is possible to effectively perceive the short-term transient fluctuations and periodic features within the time window of wind turbine operating parameters, making up for the shortcomings of LSTM in local morphological perception, providing a local morphological feature representation that complements the global time-series features for the fusion modeling module, and realizing a comprehensive perception of operating characteristics in multiple dimensions and at multiple scales.

[0085] Step 106: Perform feature fusion on time-dependent features and local pattern features to establish a full-wind-speed operation characteristic model of the wind turbine and output the operation characteristic prediction results.

[0086] Specifically, the predicted operating characteristics include one or more of the following: power characteristics, speed characteristics, load characteristics, temperature characteristics, and deviation residuals; in addition, feature fusion is achieved using an attention-weighted fusion mechanism.

[0087] Furthermore, this step also includes: if new operation monitoring data is obtained, adjusting the parameters of the fully connected network of the output layer of the wind turbine full-wind speed operation characteristic model based on the operation monitoring data of normal operation within the first time period, and fixing the parameters of the bottom feature extraction network of the wind turbine full-wind speed operation characteristic model; updating the parameters of both the fully connected network of the output layer and the bottom feature extraction network of the wind turbine full-wind speed operation characteristic model based on the operation monitoring data of normal operation within the second time period; wherein, the duration of the first time period is shorter than the duration of the second time period.

[0088] Specifically, this step deeply integrates the long-range time-series features output by the time-series feature extraction module with the local morphological features output by the local pattern extraction module to establish a prediction model for the full wind speed operation characteristics of the wind turbine, and outputs prediction results for multi-dimensional operation characteristics such as power characteristics, speed characteristics, load characteristics, temperature characteristics, and deviation residuals.

[0089] The feature fusion strategy employs an attention-weighted fusion mechanism. Compared to simple feature concatenation, the attention mechanism adaptively assigns fusion weights to RNN and CNN features based on the operating conditions of the current input sample. In operating conditions with more significant time-series dependence (such as power ramp-up zones below rated wind speed), RNN features are given higher weights; in operating conditions with more severe local fluctuations (such as rated wind speed zones under high turbulence), CNN features are given higher weights, achieving dynamic adaptation of the fusion strategy to operating conditions. The attention weights are generated through the following calculation process: the RNN and CNN feature vectors are concatenated and input into a two-layer fully connected attention network, outputting a normalized two-dimensional weight vector, which serves as the scaling coefficients for the RNN and CNN features, respectively. The weighted feature vectors are then summed to obtain the fused feature representation.

[0090] The multi-objective prediction output head is constructed based on a fused feature representation. The fused feature vector is sequentially transformed through two fully connected layers (dimensions 128 and 64, respectively, with ReLU activation function) and then connected to multiple independent output heads. Each output head corresponds to a type of operational characteristic prediction objective: the power prediction head outputs the predicted active power value at the current time step; the speed prediction head outputs the predicted generator speed value; the load prediction head outputs the predicted low-speed shaft torque value; the temperature prediction head outputs the predicted gearbox oil temperature value; and the deviation residual prediction head outputs the difference between the measured value and the theoretical expected value, used for state deviation detection. Multiple output heads share the same set of fused feature representations, achieving implicit modeling of the intrinsic relationships between multiple types of operational characteristics through parameter sharing, thus improving the overall consistency of each prediction objective.

[0091] The multi-objective loss function jointly optimizes the errors of each predictor during the model training phase. The total loss function consists of three terms: a power and state prediction error term, which is a weighted sum of the mean squared errors (MSEs) between the predicted and actual values ​​of each output head; a characteristic curve smoothness constraint term, which applies L2 regularization to the first-order differences of the model output values ​​within adjacent wind speed intervals to suppress discontinuous jumps in the prediction curve at the boundary of the operating condition transition; and a parameter regularization term, which applies L2 regularization to the weights of the fully connected layers to suppress overfitting. The weight coefficients of the three loss terms are automatically selected from the candidate parameter grid through cross-validation. The model is trained end-to-end using the Adam optimizer with an initial learning rate of 1×10⁻³, employing a cosine annealing learning rate decay strategy. The total training epochs are 200, and an early stopping mechanism automatically terminates training if the validation set loss does not decrease for 20 consecutive epochs.

[0092] Online calibration and incremental update mechanisms ensure the continued effectiveness of the model throughout the entire life cycle of the unit. As the unit operates for a longer period of time, factors such as blade contamination, gear wear, and control parameter adjustments can cause a gradual drift in the baseline of the unit's operating characteristics, leading to a decrease in the prediction accuracy of the trained model.

[0093] This module supports two model update strategies to address the gradual drift of unit operating characteristics over its lifespan:

[0094] Lightweight online calibration strategy: When new, short-term (e.g., one month) normal operation monitoring data is acquired, only the parameters of the fully connected network output layer of the wind turbine's full wind speed operation characteristic model are adjusted, while the parameters of the underlying feature extraction network (RNN and CNN parts) remain fixed. This strategy updates quickly and is suitable for fine-tuning.

[0095] Deep incremental update strategy: When a relatively long period (e.g., six months) of normal operation monitoring data has been accumulated, all network parameters, including the output layer and the bottom feature extraction network, are updated. This strategy adapts to a greater degree of feature drift and employs an experience replay mechanism to prevent the forgetting of historical knowledge.

[0096] Example 2

[0097] Embodiment 2 of the present invention also discloses a modeling system for the full wind speed operation characteristics of wind turbine units, such as... Figure 2 As shown, it includes:

[0098] The data acquisition module 201 is used to collect operational monitoring data of the wind turbine unit across the entire wind speed range;

[0099] The sample construction module 202 is used to preprocess, divide, slice time windows, and label samples of the operation monitoring data to form a full wind speed operation characteristic modeling sample set;

[0100] The feature selection module 203 is used to evaluate the importance of candidate features using the random forest algorithm and select a subset of target features; the candidate features are generated by feature engineering extension based on the features in the sample set of the full wind speed operation characteristics modeling.

[0101] The temporal feature extraction module 204 is used to extract temporal dependency features from the target feature subset using a recurrent neural network (RNN).

[0102] The local pattern extraction module 205 is used to extract local pattern features from a subset of target features using a convolutional neural network (CNN).

[0103] The fusion modeling module 206 is used to fuse time-dependent features and local pattern features to establish a full-wind-speed operation characteristic model of the wind turbine and output the operation characteristic prediction results.

[0104] Furthermore, the operational monitoring data includes one or more of the following: wind speed, generator speed, rotor speed, blade pitch angle, yaw angle, active power, reactive power, torque, vibration, temperature, and unit operating status.

[0105] Preprocessing includes: abnormal condition removal, missing value imputation, time synchronization, normalization, and start / stop transition segment identification;

[0106] Candidate features include the product of speed and pitch angle, the power-to-speed ratio, the difference between temperature measurement points, and the moving mean and moving standard deviation of each signal.

[0107] The predicted operating characteristics include one or more of the following: power characteristics, speed characteristics, load characteristics, temperature characteristics, and deviation residuals.

[0108] Furthermore, the random forest algorithm is used to evaluate the importance of candidate features and select a subset of target features, including:

[0109] The importance of candidate features is evaluated using the random forest algorithm, and the normalized importance scores of the candidate features are obtained.

[0110] Sort the features by importance score from highest to lowest to form a feature importance ranking list;

[0111] The candidate features in the feature importance ranking list are accumulated from high to low importance scores, and the feature set included when the cumulative importance reaches the preset total is used as the target feature subset.

[0112] Furthermore, if there are feature pairs in the target feature subset with a Pearson correlation coefficient higher than a preset threshold, then the feature pair with the lower importance score is removed from the target feature subset to update the target feature subset.

[0113] Furthermore, the RNN uses the Long Short-Term Memory (LSTM) network as the basic unit. The LSTM adopts a two-layer stacked design, where the first LSTM layer includes a first number of hidden units and the second LSTM layer includes a second number of hidden units. A Dropout regularization layer is set between the first and second LSTM layers; the first number is greater than the second number.

[0114] Furthermore, CNN employs a one-dimensional temporal convolutional structure with the time step as the convolution direction to perform sliding perception on the local morphological features in the input target feature subset. The one-dimensional temporal convolutional structure uses a multi-scale convolutional kernel and multiple sets of convolutional branches. Different convolutional branches are for different time steps to cover the fluctuation features of the running detection data at different time scales.

[0115] Furthermore, feature fusion is achieved using an attention-weighted fusion mechanism.

[0116] Furthermore, the fusion modeling module is also used for:

[0117] If new operation monitoring data is obtained, the parameters of the fully connected network of the output layer of the wind turbine full wind speed operation characteristic model are adjusted based on the operation monitoring data of normal operation within the first time period, and the parameters of the bottom feature extraction network of the wind turbine full wind speed operation characteristic model are fixed.

[0118] Based on the normal operation monitoring data during the second time period, the parameters of both the output layer fully connected network and the bottom layer feature extraction network of the wind turbine full wind speed operation characteristic model are updated; wherein, the duration of the first time period is shorter than the duration of the second time period.

[0119] Example 3

[0120] Embodiment 3 of the present invention also discloses an electronic device, such as Figure 3 As shown, it includes:

[0121] One or more processors 302;

[0122] Memory 303 is used to store one or more programs;

[0123] When one or more programs are executed by one or more processors 302, the one or more processors 302 implement the wind turbine full-wind-speed operation characteristic modeling method according to any one of Embodiment 1.

[0124] Specifically, such as Figure 3 As shown, a complete system may include electronic device 301 and acquisition device 305. Electronic device 301 may be an industrial control computer, GPU computing server, edge computing device, or cloud server, etc. Electronic device 301 may include, but is not limited to, processor 302 and memory 303. Those skilled in the art will understand that the illustration is merely an example of this modeling system 300 and does not constitute a limitation on the system. It may include more or fewer components than illustrated, or combine certain components, or use different components. For example, electronic device 301 and acquisition device 305 may be connected via industrial Ethernet, OPC-UA protocol, or wireless communication interface.

[0125] Processor 302 can be a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs), or other programmable logic devices. For training and inference tasks of RNN / CNN models, it is recommended to use a GPU processor with parallel floating-point computing capabilities to significantly improve model training speed and online inference efficiency.

[0126] The memory 303 can be an internal storage unit of the electronic device 301, such as a solid-state drive or RAM; or it can be an external storage device of the electronic device 301, such as network attached storage (NAS), a plug-in hard drive, or a flash memory card. Furthermore, the memory 303 can include both internal and external storage units of the electronic device 301. The memory 303 is used to store data required for the computer program 304 and system operation, including training sample sets, random forest feature importance parameters, LSTM network weights, CNN network weights, fusion layer attention weights, model validation and evaluation records, and online correction history logs.

[0127] The data acquisition device 305 is used to acquire multi-dimensional operational monitoring data of the wind turbine in real time. The device reads real-time sampled values ​​of operating parameters such as wind speed, power, rotational speed, pitch angle, and temperature from the turbine's SCADA system via a standard Modbus TCP / IP or OPC-UA interface. After timestamping and data format encapsulation, the data is transmitted to the electronic device 301 via the communication interface for reception and storage by the data acquisition module. This provides real-time data input for the online preprocessing and online characteristic prediction inference of the sample construction module.

[0128] Example 4

[0129] Embodiment 4 of the present invention also discloses a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the wind turbine full-wind-speed operation characteristic modeling method of any one of Embodiment 1.

[0130] This invention proposes a method, system, electronic device, and storage medium for modeling the full-wind-speed operation characteristics of wind turbines. The method includes: collecting operational monitoring data of wind turbines across the entire wind speed range; preprocessing, dividing into operating conditions, slicing into time windows, and labeling samples of the operational monitoring data to form a full-wind-speed operation characteristic modeling sample set; using a random forest algorithm to evaluate the importance of candidate features and select a target feature subset; generating candidate features through feature engineering based on features in the full-wind-speed operation characteristic modeling sample set; using a recurrent neural network (RNN) to extract temporal dependency features from the target feature subset; using a convolutional neural network (CNN) to extract local pattern features from the target feature subset; and fusing the temporal dependency features and local pattern features to establish a full-wind-speed operation characteristic model of the wind turbine and output the predicted operation characteristics. By using the random forest algorithm to screen the importance of multi-dimensional candidate features, and combining the ability of RNN to extract long-term temporal dependencies with the advantage of CNN in perceiving local fluctuation patterns, high-precision fusion modeling of the operating characteristics of wind turbines across the entire wind speed range is achieved. This helps to improve the problems of insufficient generalization ability of single architecture models under all operating conditions and the interference of redundant features on modeling accuracy, and provides a reliable operating characteristic modeling foundation for wind turbine performance evaluation, condition monitoring and intelligent operation and maintenance.

[0131] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0132] Those skilled in the art will understand that the modules in the apparatus of the implementation scenario can be distributed within the apparatus of the implementation scenario as described, or they can be located in one or more apparatuses different from this implementation scenario, with corresponding changes. The modules of the above-described implementation scenario can be combined into one module, or they can be further divided into multiple sub-modules.

[0133] The serial numbers of the present invention mentioned above are for descriptive purposes only and do not represent the superiority or inferiority of the implementation scenarios.

[0134] The above-disclosed examples are only a few specific implementation scenarios of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for modeling the full wind speed operation characteristics of a wind turbine, characterized in that, include: Collect operational monitoring data of wind turbine units across the entire wind speed range; The operational monitoring data is preprocessed, divided into operating conditions, sliced ​​into time windows, and labeled with samples to form a full-wind-speed operational characteristic modeling sample set; The importance of candidate features is evaluated using a random forest algorithm, and a subset of target features is obtained by selection. The candidate features are generated by feature engineering extension based on the features in the sample set of the full wind speed operation characteristics modeling. A recurrent neural network (RNN) is used to extract the temporal dependency features from the target feature subset; A convolutional neural network (CNN) is used to extract local pattern features from the target feature subset; The time-dependent features and local pattern features are fused to establish a full-wind-speed operation characteristic model of the wind turbine and output the operation characteristic prediction results.

2. The method as described in claim 1, characterized in that, The operational monitoring data includes one or more of the following: wind speed, generator speed, impeller speed, blade pitch angle, yaw angle, active power, reactive power, torque, vibration, temperature, and unit operating status. The preprocessing includes: abnormal operating condition removal, missing value imputation, time synchronization, normalization, and start / stop transition segment identification; The candidate features include the product of rotational speed and pitch angle, the ratio of power to rotational speed, the difference between temperature measurement points, and the moving mean and moving standard deviation of each signal. The predicted operating characteristics include one or more of the following: power characteristics, speed characteristics, load characteristics, temperature characteristics, and deviation residuals.

3. The method as described in claim 1, characterized in that, The process of using the random forest algorithm to evaluate the importance of candidate features and filter them to obtain a subset of target features includes: The importance of candidate features is evaluated using the random forest algorithm to obtain the normalized importance score of the candidate features; Sort the features by importance score from highest to lowest to form a feature importance ranking list; The candidate features in the feature importance ranking list are accumulated in descending order of importance score, and the feature set included when the cumulative importance reaches a preset total is used as the target feature subset.

4. The method as described in claim 3, characterized in that, If there are feature pairs in the target feature subset with a Pearson correlation coefficient higher than a preset threshold, then the feature pair with the lower importance score is removed from the target feature subset to update the target feature subset.

5. The method as described in claim 1, characterized in that, The RNN uses a Long Short-Term Memory (LSTM) network as its basic unit. The LSTM adopts a two-layer stacked design, wherein the first LSTM layer includes a first number of hidden units, the second LSTM layer includes a second number of hidden units, and a Dropout regularization layer is set between the first LSTM layer and the second LSTM layer; the first number is greater than the second number.

6. The method as described in claim 1, characterized in that, The CNN adopts a one-dimensional temporal convolutional structure with the time step as the convolution direction, and performs sliding perception on the local morphological features in the input target feature subset. The one-dimensional temporal convolutional structure adopts a design of multi-scale convolutional kernels and multiple sets of convolutional branches. Different convolutional branches are for different time steps to cover the fluctuation features of the running detection data at different time scales.

7. The method as described in claim 1, characterized in that, The feature fusion is achieved using an attention-weighted fusion mechanism.

8. The method as described in claim 1, characterized in that, Also includes: If new operation monitoring data is obtained, the parameters of the fully connected network of the output layer of the wind turbine full wind speed operation characteristic model are adjusted based on the operation monitoring data of normal operation within the first time period, and the parameters of the bottom feature extraction network of the wind turbine full wind speed operation characteristic model are fixed. Based on the operational monitoring data during the second time period, the parameters of both the output layer fully connected network and the bottom layer feature extraction network of the wind turbine full-wind speed operation characteristic model are updated; wherein, the duration of the first time period is shorter than the duration of the second time period.

9. A modeling system for the full wind speed operation characteristics of a wind turbine, characterized in that, include: The data acquisition module is used to collect operational monitoring data of the wind turbine across the entire wind speed range; The sample construction module is used to preprocess, divide, slice time windows, and label the operation monitoring data to form a full-wind speed operation characteristic modeling sample set. The feature selection module is used to evaluate the importance of candidate features using the random forest algorithm and select a subset of target features; the candidate features are generated by feature engineering extension based on the features in the full wind speed operation characteristic modeling sample set; The temporal feature extraction module is used to extract temporal dependency features from the target feature subset using a recurrent neural network (RNN). The local pattern extraction module is used to extract local pattern features from the target feature subset using a convolutional neural network (CNN). The fusion modeling module is used to perform feature fusion on the time-dependent features and local pattern features, establish a full-wind-speed operation characteristic model of the wind turbine, and output the operation characteristic prediction results.

10. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the wind turbine full-wind-speed operation characteristic modeling method according to any one of claims 1-8.

Images