CNN-LSTM deep learning-based wind speed dynamic correction method and system

Abstract

The invention discloses a wind speed dynamic correction method and system based on CNN-LSTM deep learning, and relates to the technical field of wind speed correction, and the method comprises the steps: receiving actual measurement sensor data, carrying out the sliding window processing of the actual measurement sensor data, and obtaining time sequence input sample data, the environment parameters comprise temperature data, humidity data, air pressure data and pitch angle data; inputting the time sequence input sample data into a pre-trained multi-scale mixed correction model, outputting to obtain a standardized correction amount, and performing anti-standardization processing based on the standardized correction amount to obtain a corrected wind speed value; and the accuracy of wind power prediction and fan performance evaluation is improved.

Description

Technical Field

[0001] This invention relates to the field of wind speed correction technology, specifically a method and system for dynamic wind speed correction based on CNN-LSTM deep learning. Background Technology

[0002] Wind speed is a core parameter for wind farm power prediction, wind turbine performance evaluation, and grid dispatch. However, in practical applications, wind speed sensors are susceptible to interference from environmental factors such as temperature, humidity, air pressure, wake effects, and geographical location, causing the measured values ​​collected by the SCADA system to deviate from the true wind speed. For example, increased bearing friction due to low temperatures, sensor drift caused by high humidity, and wake effects leading to wind speed deviations in the SCADA system can cause systematic deviations in wind speed data of up to 10% to 15%. Such errors significantly reduce the accuracy of wind power prediction, thereby affecting power generation planning and grid stability control. According to industry statistics, for every 1% increase in wind speed data error, the annual power generation loss of a wind farm can reach 0.8% to 1.2%. In addition, most mainstream wind turbine manufacturers currently use the main control coefficient correction method to calibrate SCADA wind speed: by preset correction coefficients (usually linear scaling factors or offsets) in the wind turbine main control system, the measured value (Vx) of the anemometer behind the nacelle is compared and calibrated with the reference wind speed (Vy) in front of the hub. However, with the increasing size of wind turbines and their widespread application in complex terrain wind fields, current mainstream wind speed data correction methods have certain limitations. While deep learning has shown advantages in time-series data processing with technological advancements, it faces challenges in intelligent wind speed correction. These challenges include: significant differences in the dimensions and frequencies of change of parameters such as wind speed, temperature, and humidity, which weakens the model's learning ability when features are directly concatenated; and the strong non-stationarity of wind speed fluctuations, requiring the simultaneous capture of local abrupt changes (such as gusts) and long-term trends (such as gradual changes in air pressure). Existing LSTM (Long Short-Term Memory) models are insufficient for extracting short-term high-frequency features, while CNN (Convolutional Neural Network) models struggle to model long-term dependencies, resulting in low accuracy in wind power prediction and wind turbine performance evaluation. Summary of the Invention

[0003] To address the shortcomings mentioned in the background section, the present invention aims to provide a method and system for dynamic wind speed correction based on CNN-LSTM deep learning.

[0004] Firstly, the objective of this invention can be achieved through the following technical solution: a dynamic wind speed correction method based on CNN-LSTM deep learning, the method comprising the following steps: Receive measured sensor data, process the measured sensor data through a sliding window to obtain time-series input sample data, wherein the sensor stream data includes environmental parameters and disturbed wind speed data, and the environmental parameters include temperature data, humidity data, air pressure data and blade pitch angle data; The time-series input sample data is input into a pre-trained multi-scale hybrid correction model, and the standardized correction value is output. Based on the standardized correction value, the de-standardization process is performed to obtain the corrected wind speed value. The pre-trained multi-scale hybrid correction model is executed as follows: Environmental parameters from time-series input sample data are extracted using a recurrent neural network to obtain long-period environmental coupled time-series features; Disturbed wind speed data is extracted using a convolutional neural network to obtain local wind speed features; Long-period environmental coupled temporal features and local wind speed features are fused and stitched together. The fused features are then processed through a fully connected layer to obtain a standardized correction value.

[0005] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the multi-scale hybrid correction model is a dual-channel hybrid architecture, including a recurrent neural network and a convolutional neural network, wherein the recurrent neural network is a bidirectional LSTM layer and the convolutional neural network is a three-level 1D convolutional layer.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: a training process for the multi-scale hybrid correction model, comprising: The basic aerodynamic model is trained using a training dataset that has been enhanced with physical constraints and labeled, and then the final model is obtained by combining it with a deep learning algorithm.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the physical constraints being icing condition simulation and pitch angle turbulence effect, wherein the icing condition simulation is to apply a preset ratio of attenuation to the reference wind speed when the temperature is below a preset temperature threshold; and the pitch angle turbulence effect is to add Gaussian noise when the pitch angle is greater than a preset pitch angle threshold.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the expression for the multi-scale hybrid correction model is as follows: Input data definition: Suppose that after the measured sensor data is processed by a sliding window, two types of time-series input sample data are obtained, including environmental parameter time-series data and disturbed wind speed time-series data, as follows: Environmental parameter time series data E = {et, f} ∈ R T×Fe Disturbed wind speed time series data V ={vt}∈R T×1 LSTM long-period feature extraction of environmental parameter channels: Bidirectional LSTM layers extract long-term environmental coupling temporal features from environmental parameters: H LSTM =BiLSTM(E;W lstm ,b lstm )∈R T×Dlstm CNN local feature extraction of the disturbed wind speed channel: A three-level 1D convolutional layer extracts local wind speed features from disturbed wind speeds: First-level convolution: C1 =σ(Conv1D(V;W c1 ,b c1 ,s1))∈R T×K1 Second-level convolution: C2 =σ(Conv1D(C1;W c2 ,b c2 ,s2))∈R T×K2 Third-level convolution: H CNN =σ(Conv1D(C2;W c3 ,b c3 ,s3))∈R T×K3 Feature fusion and fully connected layer output: Feature fusion and stitching: F fusion =Concat( HLSTM H CNN )∈R T×(Dlstm+K3) The fully connected layer outputs a normalized correction value: ΔV std =Linear(F fusion ;W fc ,b fc )∈R T×1 Destandardization yields corrected wind speed: Vcorr =ΔV std ⋅σ V +μ V Where E is the environmental parameter time series data matrix, containing four types of parameters: temperature, humidity, air pressure, and propeller pitch angle; et,f is the value of the environmental parameter of type f at time step t; T is the time step size of the time series input sample; Fe is the number of features of the environmental parameter; V is the disturbed wind speed time series data vector; vt is the disturbed wind speed value at time step t; H LSTM W represents the long-period environmental coupling temporal feature matrix output by the bidirectional LSTM layer, where BiLSTM(·) is the bidirectional long short-term memory network operation. lstm Let b be the weight matrix of the bidirectional LSTM layer. lstm D is the bias vector of the bidirectional LSTM layer. lstm C1, C2, and C3 are the local feature maps output by the three 1D convolutional layers, respectively. Conv1D(·) is the one-dimensional convolution operation. W c1 W c2 W c3 These are the kernel weight matrices for each of the three 1D convolutional layers, b c1 b c2 b c3 s1, s2, and s3 are the bias vectors of the three-level 1D convolutional layers, respectively; s1, s2, and s3 are the convolution strides of the three-level 1D convolutional layers, respectively; K1, K2, and K3 are the number of feature map channels output by the three-level 1D convolutional layers, respectively; σ(·) is the activation function; H CNN F is the local wind speed feature vector output by the final three-level 1D convolutional layer. fusion The feature matrix is ​​the result of fusing long-period environmental features and local wind speed features. Concat(·) is the feature concatenation operation. ΔV std The normalized correction value is the output of the fully connected layer, Linear(·) is the operation of the fully connected layer, and W is the normalized correction value. fc Let b be the weight matrix of the fully connected layer. fc Here, Vcorr is the bias vector of the fully connected layer, Vcorr is the final corrected wind speed value, and σ is the bias vector of the fully connected layer. V μ represents the standard deviation of the original wind speed data. V This is the mean of the original wind speed data.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: during the process of performing de-standardization processing based on the standardized correction amount to obtain the corrected wind speed value, if the data triggers a disaster response mechanism, the method automatically switches to the backup filtering channel and starts the data verification process.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the determination criteria of the disaster response mechanism include at least one of the mutation magnitude of the standardized correction amount and the confidence index.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the recurrent neural network may also select a GRU gated recurrent unit, and the convolutional neural network may select a Wavelet convolutional layer.

[0012] Secondly, in order to achieve the above objectives, this invention discloses a wind speed dynamic correction system based on CNN-LSTM deep learning, comprising: The data processing module is used to receive measured sensor data, process the measured sensor data through a sliding window, and obtain time-series input sample data. The sensor stream data includes environmental parameters and disturbed wind speed data. The environmental parameters include temperature data, humidity data, air pressure data, and blade pitch angle data. The wind speed correction module is used to input time-series input sample data into a pre-trained multi-scale hybrid correction model, output a standardized correction value, and perform destandardization based on the standardized correction value to obtain the corrected wind speed value. The pre-trained multi-scale hybrid correction model is executed as follows: Environmental parameters from time-series input sample data are extracted using a recurrent neural network to obtain long-period environmental coupled time-series features; Disturbed wind speed data is extracted using a convolutional neural network to obtain local wind speed features; Long-period environmental coupled temporal features and local wind speed features are fused and stitched together. The fused features are then processed through a fully connected layer to obtain a standardized correction value.

[0013] In another aspect of the present invention, in order to achieve the above-mentioned objective, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor. When the processor loads and executes the computer program, it employs the wind speed dynamic correction method based on CNN-LSTM deep learning as described above.

[0014] The beneficial effects of this invention are: This invention effectively separates the coupling interference between environmental parameters and wind speed signals by constructing a dual-channel hybrid model, thereby improving data reliability under complex operating conditions. It embeds wind turbine aerodynamics into the data generation process, enabling the model to learn physical rules such as icing attenuation and turbulence compensation during the training phase, maintaining stable correction performance even in harsh environments. Using a dual-threshold catastrophe response mechanism, it identifies sensor anomalies within milliseconds by real-time monitoring of correction abrupt changes and confidence indices, significantly reducing the risk of control malfunctions caused by data distortion. By improving wind speed measurement accuracy, it directly optimizes the accuracy of wind power prediction and unit control efficiency, thus enhancing the accuracy of wind power prediction and wind turbine performance evaluation. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the core logic of the disaster response of this invention; Figure 3 This is a schematic diagram of the training progress of the model of the present invention; Figure 4 This is a schematic diagram comparing the time series data after the model of this invention has been corrected; Figure 5 This is a schematic diagram of the scatter point density distribution of the present invention; Figure 6 This is a schematic diagram illustrating the evaluation of the model correction effect of the present invention; Figure 7 This is a schematic diagram of the system structure of the present invention; Figure 8 This is a schematic diagram of wind measurement data before the application of the technology of this invention; Figure 9 This is a schematic diagram of wind measurement data after the application of the technology of this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: like Figure 1 As shown, a dynamic wind speed correction method based on CNN-LSTM deep learning is proposed, which includes the following steps: S101: Receive measured sensor data, process the measured sensor data through a sliding window to obtain time-series input sample data, wherein the sensor stream data includes environmental parameters and disturbed wind speed data, and the environmental parameters include temperature data, humidity data, air pressure data and blade pitch angle data. Disturbed wind speed data are SCADA disturbed wind speeds (derived from the text format "[B,C,D,E,F]"). Specifically, in this embodiment, the scenario is set as follows: A wind farm is located in a complex terrain area (a mountainous wind farm in a certain province). The wind speed sensors deployed on-site (wind speeds derived from the SCADA system) are affected by environmental factors, leading to systematic deviations and poor consistency of the wind power curve. Specific issues: ① Environmental interference: Low temperature (< -25℃) causes increased bearing friction, which reduces the wind speed measurement value.

[0018] ② Data error: The timing asynchrony between the wind speed sensor (1Hz sampling) and the environmental sensor (temperature 0.2Hz, humidity 0.1Hz) results in a 15% to 20% time domain misalignment.

[0019] ③ Actual impact: When the turbulence intensity is greater than 0.2, the difference in mountainous terrain leads to a large error in the traditional fixed coefficient correction. Furthermore, the use of a unified main control wind speed correction coefficient results in a large error in wind power prediction and a poor consistency coefficient of the group's wind power curve assessment.

[0020] ④ Boundary conditions: Wind farms need to cope with icing conditions (-30℃) and high aerodynamic efficiency loss conditions with a pitch angle >10°.

[0021] S102: Input the time-series input sample data into the pre-trained multi-scale hybrid correction model, output the standardized correction amount, and perform de-standardization based on the standardized correction amount to obtain the corrected wind speed value; The pre-trained multi-scale hybrid correction model is executed as follows: Environmental parameters from time-series input sample data are extracted using a recurrent neural network to obtain long-period environmental coupled time-series features; Disturbed wind speed data is extracted using a convolutional neural network to obtain local wind speed features; Long-period environmental coupled temporal features and local wind speed features are fused and stitched together. The fused features are then processed through a fully connected layer to obtain a standardized correction value.

[0022] The multi-scale hybrid correction model is a dual-channel hybrid architecture, comprising a recurrent neural network (RNN) and a convolutional neural network (CNN). The RNN uses bidirectional LSTM layers, and the CNN uses three-level 1D convolutional layers. Furthermore, alternative network structures are provided to expand applicability. Specifically, in the dual-channel model, GRU (Gated Recurrent Unit) layers are used instead of bidirectional LSTM layers to reduce the number of parameters by 30%, making it suitable for edge computing; or the three-level 1D convolutional layers are replaced with wavelet convolutional layers to enhance high-frequency noise capture and improve model flexibility.

[0023] Specifically, the core advantages and performance comparisons of the three models are as follows: (a) Original model: Dual-channel CNN-LSTM (bidirectional LSTM + three-level 1D convolution) Table 1. Core advantages of dual-channel CNN-LSTM and its effect on dynamic wind speed correction. Core advantages The effect corresponding to this application (dynamic wind speed correction) Dual advantages of long-term dependence and local features: Bidirectional LSTM can simultaneously capture the bidirectional dependence of time series data "past→present" and "present→future", and three-level 1D convolution can extract local fluctuation features (such as gusts and small turbulence) in a layered manner. Adaptable to multiple wind farm scenarios: It can handle the long-term impact of cross-time environmental parameters (such as diurnal temperature difference and air pressure changes) on wind speed, and accurately correct local deviations in SCADA data caused by sensor fluctuations. The time-series trend matching degree between the corrected data and the actual field value is over 92% (stable period) and over 88% (turbulent period). High accuracy and stability: The gating mechanism of bidirectional LSTM (forget gate, input gate) can filter redundant information in environmental parameters, and the three-level 1D convolution reduces noise interference through "convolution-pooling" to avoid overfitting to a single feature. Small fluctuations in correction error: Within the full range of wind speeds from 0 to 25 m / s, the average absolute error remains stable at 0.3 to 0.5 m / s, and the relative error is ≤8%, with no sudden increase in error under extreme scenarios (such as strong gusts). The computational complexity is moderate: no special complex operators (such as wavelet transform) are introduced, and the number of parameters is controlled between 500,000 and 800,000 (adjusted according to the time step), making it suitable for conventional edge servers or cloud computing power in wind farms. High deployment feasibility: Inference time for a single sample (24-hour time series data) is ≤0.5s, which can meet the dual requirements of wind farms for "real-time correction + hourly data backtracking" without the need for additional hardware upgrades. (ii) Alternative Solution 1: GRU replaces bidirectional LSTM (dual-channel CNN-GRU) Table 2. Core advantages and dynamic wind speed correction effect of dual-channel CNN-GRU Core advantages The effect corresponding to this application (dynamic wind speed correction) Lightweight and deployment-friendly: GRU combines the "cell state" and "hidden state" of LSTM, retaining only the reset gate and update gate, reducing the number of parameters by 30% to 40% compared to bidirectional LSTM. Improved edge computing adaptability: The number of parameters is reduced to 350,000 to 550,000, and the single-sample inference time is shortened to less than 0.3 seconds. It can be deployed at the edge box of wind farm tower base (computing power ≤ 2 TOPS). Faster training convergence: Reduced gating computation and slower gradient decay during backpropagation shorten the model training cycle by 20%–30%. High efficiency in project implementation: For fresh air field data, model fine-tuning only takes 2-3 hours (compared to 4-5 hours for the original model). Good robustness in low-computing-power scenarios: In environments with limited computing power, it avoids inference interruptions caused by insufficient computing resources. Available for extreme deployment scenarios: suitable for remote wind farms (without a stable cloud connection), allowing for real-time local correction. limitation 1. Weak long-term dependency capture: GRU lacks independent cell states, resulting in decreased processing capability for environmental couplings spanning more than 24 hours (such as the impact of continuous rainy weather on wind speed), and the average correction error rises to 0.6–0.8 m / s; 2. Decreased temporal trend matching accuracy: Due to the lack of bidirectional information (GRU is mostly unidirectional), the accuracy of correction for wind speed "peak lag" (such as wind speed drop after a gust) decreases, and the true value matching accuracy drops below 85%. (III) Alternative Solution 2: Wavelet Convolution replaces Three-Level 1D Convolution (Wavelet-CNN + Bidirectional LSTM) Table 3. Core advantages of Wavelet-CNN + Bidirectional LSTM and their effect on dynamic wind speed correction. Core advantages The effect corresponding to this application (dynamic wind speed correction) Strong high-frequency noise capture capability: Wavelet convolution, through "multi-scale wavelet decomposition," can separate different frequency components in wind speed data (such as turbulent noise of 0.1–1 Hz and sensor electromagnetic interference of 1–10 Hz), accurately locating high-frequency interference sources. Excellent correction for extreme turbulence scenarios: In turbulent wind fields (such as mountainous areas and coastlines), it can effectively filter instantaneous high-frequency noise, and the average absolute error after correction is reduced to 0.2-0.4 m / s, which is 10%-15% better than the original model. High interpretability of frequency characteristics: Wavelet decomposition can output the contribution of each frequency band, making it easy to trace the source of error (e.g., deviation in a certain frequency band is caused only by sensor interference). It has good fault diagnosis assistance: it can locate the "high-frequency drift" fault of SCADA sensors through frequency characteristics, providing data support for equipment maintenance. limitation 1. Increased computational complexity: Wavelet convolution requires additional convolution operations using wavelet basis functions (such as db4 and sym5), increasing computational cost by 40%–60% compared to three-level 1D convolution, extending single-sample inference time to over 1.2 seconds; 2. Redundancy in stable scenarios: During periods of stable wind speed (such as windless periods in plain wind farms), high-frequency noise is minimal, making the multi-scale decomposition of Wavelet convolution redundant, resulting in no difference in correction accuracy compared to the original model (and even a slight increase in error due to over-decomposition); 3. High hardware requirements: Requires deployment on high-performance GPU servers (computing power ≥ 10 TOPS), making it unsuitable for conventional edge devices in wind farms. (iv) Applicable Scenarios for the Three Models Table 4 Comparison of Applicable Scenarios for the Three Models Model Original model (CNN - Bidirectional LSTM) Alternative Solution 1 (CNN-GRU) Alternative Solution 2 (Wavelet-CNN-LSTM) Applicable wind farm scenarios All scenes (plains / mountains / coastal areas) Low computing power + stable scenarios High turbulence + high computing power scenarios In actual operation, wind speed is affected by both "long-term environmental coupling" (such as diurnal temperature difference and continuous rainfall) and "short-term local fluctuations" (such as gusts and small-scale turbulence), and the scenarios vary greatly (stable wind farms in plains and frequent turbulence in mountainous areas).

[0024] The proposed "dual-channel CNN-bidirectional LSTM model" is a general, efficient, and stable method for dynamic wind speed correction, applicable to most wind farm deployment environments (under non-extreme computing power conditions) and their diverse operating scenarios (including stable and turbulent periods). In this model, bidirectional LSTM captures long-term temporal dependencies, while three-level 1D convolutions handle regular local fluctuations. The combination of these two methods allows the model to maintain stable correction accuracy (absolute error ≤ 0.5 m / s) across all scenarios without requiring specific optimization for particular scenarios (such as extreme turbulence). In contrast, Alternative Solution 1 (GRU) lacks accuracy in long-period modeling, and Alternative Solution 2 (wavelet method) exhibits redundancy in stable scenarios, both failing to meet the requirement of "universality across all scenarios."

[0025] The "dynamic correction" process requires real-time updates of correction parameters based on SCADA data and backtracking historical data (e.g., calibration using 72-hour data), thus placing high demands on the model's temporal consistency. Bidirectional LSTM, through bidirectional temporal modeling, ensures a coherent correction logic from "history → present → future," avoiding the "peak lag" error caused by unidirectional structures (such as the commonly used unidirectional GRU). Meanwhile, three-level 1D convolution, through hierarchical feature extraction, effectively matches the dynamic changes in wind speed from macroscopic trends to microscopic fluctuations, making the correction results more closely match the temporal patterns of actual field measurements. Other alternative methods cannot simultaneously meet both the core requirements of "temporal consistency" and "dynamic adaptability."

[0026] The training process of the multi-scale hybrid correction model includes: The basic aerodynamic model is trained using a training dataset that has been enhanced with physical constraints and labeled, and then the final model is obtained by combining it with a deep learning algorithm.

[0027] The physical constraints are icing condition simulation and pitch angle turbulence effect. The icing condition simulation is to apply a preset ratio of attenuation to the reference wind speed when the temperature is lower than a preset temperature threshold. The pitch angle turbulence effect is to add Gaussian noise when the pitch angle is greater than a preset pitch angle threshold.

[0028] Specifically, using the physical laws governing wind turbine operation as constraints, an aerodynamic model is used to simulate abnormal conditions such as icing attenuation and pitch angle turbulence. Physical feature labels are dynamically injected into the training data, and then a multi-channel hybrid correction model is constructed using deep learning algorithms. In other words, a real-time correction framework that integrates data-driven approaches and physical laws is established using the aerodynamic characteristics of the wind turbine as boundary conditions. Among them, the simulation of low-temperature icing conditions of wind turbines is carried out. That is, based on the negative correlation between bearing friction coefficient and temperature, when the temperature is below the physical threshold (such as -25℃), an attenuation compensation (such as 20% attenuation) that conforms to the aerodynamic law is applied to the reference wind speed. By dynamically adjusting the attenuation rate, the real wind speed under icing conditions is accurately restored, ensuring the robustness of the model in extreme low-temperature environments.

[0029] Simulates high-pitch-angle aerodynamic loss conditions. Specifically, based on the physical law that blade aerodynamic efficiency decreases with increasing pitch angle, when the pitch angle exceeds a critical value (e.g., >10°), a Gaussian noise model conforming to hydrodynamic characteristics is injected into the disturbed wind speed to effectively characterize the measurement deviation caused by the decrease in aerodynamic efficiency and improve the correction accuracy under complex conditions.

[0030] The expression for the multi-scale hybrid correction model is as follows: Input data definition: Suppose that after the measured sensor data is processed by a sliding window, two types of time-series input sample data are obtained, including environmental parameter time-series data and disturbed wind speed time-series data, as follows: Environmental parameter time series data E = {et, f} ∈ R T×Fe Disturbed wind speed time series data V ={v t}∈R T×1 LSTM long-period feature extraction of environmental parameter channels: Bidirectional LSTM layers extract long-term environmental coupling temporal features from environmental parameters: H LSTM =BiLSTM(E;W lstm ,b lstm )∈R T×Dlstm CNN local feature extraction of the disturbed wind speed channel: A three-level 1D convolutional layer extracts local wind speed features from disturbed wind speeds: First-level convolution: C1 =σ(Conv1D(V;W c1 ,b c1 ,s1))∈R T×K1 Second-level convolution: C2 =σ(Conv1D(C1;W c2 ,b c2 ,s2))∈R T×K2 Third-level convolution: H CNN =σ(Conv1D(C2;W c3 ,b c3 ,s3))∈R T×K3 Feature fusion and fully connected layer output: Feature fusion and stitching: F fusion =Concat( HLSTM H CNN )∈R T×(Dlstm+K3) The fully connected layer outputs a normalized correction value: ΔV std =Linear(F fusion ;W fc ,b fc )∈R T×1 Destandardization yields corrected wind speed: Vcorr =ΔV std ⋅σ V+μ V Where E is the environmental parameter time series data matrix, containing four types of parameters: temperature, humidity, air pressure, and propeller pitch angle; et,f is the value of the environmental parameter of type f at time step t; T is the time step size of the time series input sample; Fe is the number of features of the environmental parameter; V is the disturbed wind speed time series data vector; vt is the disturbed wind speed value at time step t; H LSTM W represents the long-period environmental coupling temporal feature matrix output by the bidirectional LSTM layer, where BiLSTM(·) is the bidirectional long short-term memory network operation. lstm Let b be the weight matrix of the bidirectional LSTM layer. lstm D is the bias vector of the bidirectional LSTM layer. lstm C1, C2, and C3 are the local feature maps output by the three 1D convolutional layers, respectively. Conv1D(·) is the one-dimensional convolution operation. W c1 W c2 W c3 These are the kernel weight matrices for each of the three 1D convolutional layers, b c1 b c2 b c3 s1, s2, and s3 are the bias vectors of the three-level 1D convolutional layers, respectively; s1, s2, and s3 are the convolution strides of the three-level 1D convolutional layers, respectively; K1, K2, and K3 are the number of feature map channels output by the three-level 1D convolutional layers, respectively; σ(·) is the activation function; H CNN F is the local wind speed feature vector output by the final three-level 1D convolutional layer. fusion The feature matrix is ​​the result of fusing long-period environmental features and local wind speed features. Concat(·) is the feature concatenation operation. ΔV std The normalized correction value is the output of the fully connected layer, Linear(·) is the operation of the fully connected layer, and W is the normalized correction value. fc Let b be the weight matrix of the fully connected layer. fc Here, Vcorr is the bias vector of the fully connected layer, Vcorr is the final corrected wind speed value, and σ is the bias vector of the fully connected layer. V μ represents the standard deviation of the original wind speed data. V This is the mean of the original wind speed data.

[0031] Specifically, the design of the multi-scale hybrid correction model employs a dual-channel deep learning architecture to achieve multi-scale feature hierarchical extraction. Specifically, an environmental parameter channel (input temperature, humidity, air pressure, and propeller pitch angle) is constructed using a bidirectional LSTM layer (64 units) to capture long-term environmental coupled temporal features. A wind speed feature channel (input disturbed wind speed) extracts local turbulence and sensor noise patterns through a three-level 1D convolutional layer. These two channels work together to address the temporal misalignment (15%–20%) and dimensional differences in multi-source heterogeneous data. After concatenating the dual-channel feature vectors, the model is processed by a fully connected layer and a standardized correction value is output via regression. This achieves end-to-end mapping learning of "original measurement value → environmental parameter → corrected wind speed," avoiding the limitations of relying on manually designed compensation formulas.

[0032] During the de-standardization process based on the standardized correction value to obtain the corrected wind speed value, if the data triggers the disaster response mechanism, it will automatically switch to the backup filtering channel and initiate the data verification process. The judgment criteria for the disaster response mechanism include at least one of the abrupt change magnitude of the standardized correction value and confidence level indicators.

[0033] Specifically, a millisecond-level disaster response mechanism is deployed to ensure system security. This involves a dual-threshold judgment logic based on a change in correction magnitude (ΔV > 3σ) and the model confidence probability (p < 0.8), triggering a switch to the Kalman filter backup channel within 5ms and initiating a moving median verification process to quickly block the propagation chain of abnormal data. The 3σ threshold can be replaced with Mahalanobis distance detection. Furthermore, it enhances cross-scenario generalization capabilities. Specifically, it uses physical constraint data generation technology to cover extreme operating conditions (temperature range of -30℃ to 50℃, humidity range of 0% to 100%), enabling the model to adapt to complex terrains (such as mountainous or coastal wind fields) and different turbine models (such as 3MW and 5MW units), achieving a parameter transfer mechanism without the need for repeated modeling. Inversely derived from the power curve using the icing attenuation model: (Cp∈[0.35,0.45]) In the formula: V ice Wind speed at which blades ic up; P measured This refers to the fan power. A represents the swept area of ​​the wind turbine, in meters (m²). 2 ; ρ is the air density, with units of kg / m³. 3 ; C p This represents the wind energy utilization coefficient.

[0034] Specifically, the present invention will be further illustrated below through embodiments: Application effects and technology verification Accuracy improvement effect: The systematic deviation of the corrected wind speed data is significantly reduced. (1) Measurement accuracy: Under high pitch angle conditions (10°), the environmental channel calculates the temperature-humidity-pressure coupling equation to reduce the error caused by the decrease in aerodynamic efficiency; the measured RMSE is stable within 0.7m / s.

[0035] (2) Comparison results: Figure 4 (Comparison of corrected time series) shows that the temporal correlation between the original disturbed wind speed and the corrected wind speed is improved by 40%.

[0036] Enhanced environmental adaptability: The system efficiently copes with extreme conditions. Under icing conditions, the physical attenuation compensation mechanism ensures that the low temperature deviation does not exceed 2.5 m / s.

[0037] Cross-scenario generalization: It overcomes the terrain dependence of existing models and is applicable to both plateau wind fields (air pressure 860hPa) and coastal wind fields (1020hPa).

[0038] (3) System security upgrade: Disaster response trigger: When a sensor suddenly fails, the system restores data reliability within 10ms to avoid pitch control malfunctions (single-time loss control) caused by signal transmission delay.

[0039] Case verification: In field testing, the disaster response mechanism blocked the propagation chain of abnormal data, ensuring the continuous operation of the wind turbine (refer to the core logic of disaster response).

[0040] (4) Release of economic benefits: Improved accuracy of wind power prediction: As mentioned in the reference text, a 1% reduction in wind speed error can reduce annual wind power loss by 0.8% to 1.2%; this implementation can not only correct real-time wind measurement data, but also optimize the prediction model, thereby reducing annual power generation loss.

[0041] Other benefits include reduced unplanned downtime (such as avoiding malfunctions caused by sensor failures) and extended equipment lifespan.

[0042] like Figure 4 As shown, the comparison between the "original wind speed measured by lidar" curve and the "wind speed after correction by the CNN-LSTM model" curve reveals that the corrected wind speed data closely matches the temporal trend of the actual wind speed, effectively improving the systematic bias problem of the original measured data. The original data generally exhibits a bias characteristic of "higher peaks and lower valleys," while the corrected curve closely matches the fluctuation pattern of the actual wind speed throughout, fully verifying the model's accurate correction capability for wind speed data.

[0043] like Figure 5As shown, the corrected scatter plots exhibit a highly clustered distribution along the theoretical power curve, with a concentrated and uniform density. Under the same wind speed conditions, the power fluctuation range is significantly narrowed, indicating that the CNN-LSTM model effectively filters out random biases caused by sensor noise and environmental interference, and accurately compensates for system errors. This concentrated scatter plot density distribution verifies the consistency and reliability of the corrected data, ensuring that the wind measurement data can truly match the actual operating state of the wind turbine.

[0044] like Figure 8 , Figure 9 As shown, based on dynamic correction of wind measurement data, the technology was applied as follows: Figure 8 (Data collected for a full year) As shown, after the application of the technology, as follows: Figure 9 As shown in the data collected over a full year, it is evident that the discrete wind measurement data has significantly improved after the application of the technology.

[0045] The CNN-LSTM multi-scale hybrid correction model is used to filter out random biases caused by sensor noise and environmental interference, and to compensate for systematic errors, resulting in a highly clustered distribution of scatter points. Throughout the entire wind speed range, the corrected scatter points are closely distributed along the theoretical power curve, significantly narrowing the power fluctuation range under the same wind speed conditions, thereby improving the accuracy of the correspondence between wind measurement data and actual wind turbine power.

[0046] The reliability of the corrected data is improved, and the concentrated distribution of scatter points indicates data consistency, making it directly usable for wind power curve calibration. The uniform clustering of scatter points across the entire wind speed range verifies that the model possesses stable correction capabilities under different wind speed conditions.

[0047] Example 2: To achieve the above objective, such as Figure 7 As shown, based on Embodiment 1, this invention discloses a wind speed dynamic correction system based on CNN-LSTM deep learning, comprising: Data processing module 11 is used to receive measured sensor data, process the measured sensor data through a sliding window to obtain time-series input sample data, wherein the sensor stream data includes environmental parameters and disturbed wind speed data, and the environmental parameters include temperature data, humidity data, air pressure data and blade pitch angle data. The wind speed correction module 12 is used to input the time-series input sample data into the pre-trained multi-scale hybrid correction model, output the standardized correction amount, and perform de-standardization based on the standardized correction amount to obtain the corrected wind speed value. The pre-trained multi-scale hybrid correction model is executed as follows: Environmental parameters from time-series input sample data are extracted using a recurrent neural network to obtain long-period environmental coupled time-series features; Disturbed wind speed data is extracted using a convolutional neural network to obtain local wind speed features; Long-period environmental coupled temporal features and local wind speed features are fused and stitched together. The fused features are then processed through a fully connected layer to obtain a standardized correction value.

[0048] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0049] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0050] In the description of this specification, references to terms such as "an embodiment," "example," and "specific example" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0051] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A method for dynamic wind speed correction based on CNN-LSTM deep learning, characterized in that, The method includes the following steps: Receive measured sensor data, process the measured sensor data through a sliding window to obtain time-series input sample data, wherein the sensor stream data includes environmental parameters and disturbed wind speed data, and the environmental parameters include temperature data, humidity data, air pressure data and blade pitch angle data; The time-series input sample data is input into a pre-trained multi-scale hybrid correction model, and the standardized correction value is output. Based on the standardized correction value, the de-standardization process is performed to obtain the corrected wind speed value. The pre-trained multi-scale hybrid correction model is executed as follows: Environmental parameters from time-series input sample data are extracted using a recurrent neural network to obtain long-period environmental coupled time-series features; Disturbed wind speed data is extracted using a convolutional neural network to obtain local wind speed features; Long-period environmental coupled temporal features and local wind speed features are fused and stitched together. The fused features are then processed through a fully connected layer to obtain a standardized correction value.

2. The wind speed dynamic correction method based on CNN-LSTM deep learning according to claim 1, characterized in that, The multi-scale hybrid correction model is a dual-channel hybrid architecture, including a recurrent neural network and a convolutional neural network. The recurrent neural network is a bidirectional LSTM layer, and the convolutional neural network is a three-level 1D convolutional layer.

3. The wind speed dynamic correction method based on CNN-LSTM deep learning according to claim 2, characterized in that, The training process of the multi-scale hybrid correction model includes: The basic aerodynamic model is trained by enhancing the labeled training dataset with physical constraints, and then the multi-scale hybrid correction model is finally trained by combining deep learning algorithms.

4. The wind speed dynamic correction method based on CNN-LSTM deep learning according to claim 3, characterized in that, The physical constraints are icing condition simulation and pitch angle turbulence effect. The icing condition simulation involves applying a preset ratio of attenuation to the reference wind speed when the temperature is below a preset temperature threshold. The pitch angle turbulence effect involves adding Gaussian noise when the pitch angle is greater than a preset pitch angle threshold.

5. The wind speed dynamic correction method based on CNN-LSTM deep learning according to claim 4, characterized in that, The expression for the multi-scale hybrid correction model is as follows: Input data definition: Suppose that after the measured sensor data is processed by a sliding window, two types of time-series input sample data are obtained, including environmental parameter time-series data and disturbed wind speed time-series data, as follows: Environmental parameter time series data E = {et, f} ∈ R T×Fe Disturbed wind speed time series data V ={v t }∈R T×1 LSTM long-period feature extraction of environmental parameter channels: Bidirectional LSTM layers extract long-term environmental coupling temporal features from environmental parameters: H LSTM =BiLSTM(E;W lstm ,b lstm )∈R T×Dlstm CNN local feature extraction of the disturbed wind speed channel: A three-level 1D convolutional layer extracts local wind speed features from disturbed wind speeds: First-level convolution: C1 =σ(Conv1D(V;W). c1 ,b c1 ,s1))∈R T×K1 Second-level convolution: C2 =σ(Conv1D(C1;W c2 ,b c2 ,s2))∈R T×K2 Third-level convolution: H CNN =σ(Conv1D(C2;W c3 ,b c3 ,s3))∈R T×K3 Feature fusion and fully connected layer output: Feature fusion and stitching: F fusion =Concat( HLSTM ,H CNN )∈R T×(Dlstm+K3) The fully connected layer outputs a normalized correction value: ΔV std =Linear(F fusion ;W fc ,b fc )∈R T×1 Destandardization yields corrected wind speed: Vcorr =ΔV std ⋅σ V +μ V Where E is the environmental parameter time series data matrix, containing four types of parameters: temperature, humidity, air pressure, and propeller pitch angle; et,f is the value of the environmental parameter of type f at time step t; T is the time step size of the time series input sample; Fe is the number of features of the environmental parameter; V is the disturbed wind speed time series data vector; vt is the disturbed wind speed value at time step t; H LSTM W represents the long-period environmental coupling temporal feature matrix output by the bidirectional LSTM layer, where BiLSTM(·) is the bidirectional long short-term memory network operation. lstm Let b be the weight matrix of the bidirectional LSTM layer. lstm D is the bias vector of the bidirectional LSTM layer. lstm C1, C2, and C3 are the local feature maps output by the three 1D convolutional layers, respectively. Conv1D(·) is the one-dimensional convolution operation. W c1 W c2 W c3 These are the kernel weight matrices for each of the three 1D convolutional layers, b c1 b c2 b c3 s1, s2, and s3 are the bias vectors of the three-level 1D convolutional layers, respectively; s1, s2, and s3 are the convolution strides of the three-level 1D convolutional layers, respectively; K1, K2, and K3 are the number of feature map channels output by the three-level 1D convolutional layers, respectively; σ(·) is the activation function; H CNN F is the local wind speed feature vector output by the final three-level 1D convolutional layer. fusion The feature matrix is ​​the result of fusing long-period environmental features and local wind speed features. Concat(·) is the feature concatenation operation. ΔV std The normalized correction value is the output of the fully connected layer, Linear(·) is the operation of the fully connected layer, and W is the normalized correction value. fc Let b be the weight matrix of the fully connected layer. fc Here, Vcorr is the bias vector of the fully connected layer, Vcorr is the final corrected wind speed value, and σ is the bias vector of the fully connected layer. V μ represents the standard deviation of the original wind speed data. V This is the mean of the original wind speed data.

6. The wind speed dynamic correction method based on CNN-LSTM deep learning according to claim 1, characterized in that, If the data triggers a disaster response mechanism during the process of de-standardizing based on the standardized correction amount to obtain the corrected wind speed value, the system will automatically switch to the backup filtering channel and start the data verification process.

7. The wind speed dynamic correction method based on CNN-LSTM deep learning according to claim 6, characterized in that, The criteria for determining the disaster response mechanism include at least one of the following: the magnitude of the mutation in the standardized correction amount and the confidence index.

8. The wind speed dynamic correction method based on CNN-LSTM deep learning according to claim 5, characterized in that, The recurrent neural network may also select a GRU gated recurrent unit, and the convolutional neural network may select a Wavelet convolutional layer.

9. A wind speed dynamic correction system based on CNN-LSTM deep learning, employing the wind speed dynamic correction method based on CNN-LSTM deep learning as described in any one of claims 1 to 8, characterized in that, include: The data processing module is used to receive measured sensor data, process the measured sensor data through a sliding window, and obtain time-series input sample data. The sensor stream data includes environmental parameters and disturbed wind speed data. The environmental parameters include temperature data, humidity data, air pressure data, and blade pitch angle data. The wind speed correction module is used to input time-series input sample data into a pre-trained multi-scale hybrid correction model, output a standardized correction value, and perform destandardization based on the standardized correction value to obtain the corrected wind speed value. The pre-trained multi-scale hybrid correction model is executed as follows: Environmental parameters from time-series input sample data are extracted using a recurrent neural network to obtain long-period environmental coupled time-series features; Disturbed wind speed data is extracted using a convolutional neural network to obtain local wind speed features; Long-period environmental coupled temporal features and local wind speed features are fused and stitched together. The fused features are then processed through a fully connected layer to obtain a standardized correction value.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs the wind speed dynamic correction method based on CNN-LSTM deep learning as described in any one of claims 1 to 8.

Images