Multi-dimensional dynamic safety domain construction method for wind generating set

Abstract

The invention discloses a multi-dimensional dynamic security domain construction method for a wind generating set, and the method comprises the steps: building a multi-dimensional dynamic security domain containing a plurality of key variables, such as wind speed, load and rotating speed, through a deep neural network technology of sharing weight in combination with a confidence level theory; the method comprises the following steps: firstly, carrying out abnormal value detection, filtering smoothing and normalization preprocessing on collected fan operation data; then, a deep neural network based on shared weight is adopted, multiple groups of two-dimensional safety surfaces such as wind speed-load, wind speed-rotating speed, load-rotating speed and the like are constructed according to the number of related variables, and the two-dimensional safety surfaces are combined into a complete multi-dimensional safety domain through a weighted fusion algorithm; further introducing a confidence level concept, calculating a confidence level from a state point to a security domain boundary by adopting an exponential decay function, and dividing the security domain into a security region, a critical security region and a dangerous region; and finally, designing a sliding time window and a time attenuation weight mechanism to realize adaptive dynamic updating of the security domain. According to the method, accurate physical modeling is not needed, the safe operation boundary of the fan under different working conditions can be accurately described, real-time adjustment can be carried out in a self-adaptive mode along with the system state and the environment change, reliable constraint conditions are provided for fan safety control, and the method is particularly suitable for safe operation monitoring and control of a large wind generating set under complex wind conditions. The invention provides an effective method for solving the problems that the safety boundary of the fan is difficult to describe and the multi-parameter coupling is difficult to operate, and the method has wide engineering application value.

Description

Technical Field

[0001] The present invention relates to the field of wind farm power generation safety, and in particular to a data-driven method for constructing a multi-dimensional dynamic safety domain for a wind turbine generator set. Background Art

[0002] As a key form of clean energy, wind power generation continues to see increasing installed capacity and unit power. By the end of 2023, global installed wind power capacity will have reached approximately 1,021 GW, meeting approximately 10% of global electricity demand. However, with the expansion of wind turbines and the increasing complexity of their operating environments, the wind power industry faces unprecedented safety challenges. Modern large-scale wind turbines often operate in harsh natural environments, and their operating conditions are inevitably affected by a variety of complex factors, such as the long-term effects of multiple loads, the randomness and unevenness of wind speeds, and the impact of extreme operating conditions.

[0003] Existing wind turbine control methods, primarily based on mathematical models and design parameters, struggle to effectively address complex operating conditions and extreme environments. Existing technologies typically describe wind turbine operating boundaries using the following approaches: 1) static safety domains based on physical models, which struggle to adapt to the dynamic changes in the wind turbine operating environment; 2) single-parameter or two-dimensional safety boundaries, which fail to fully reflect the coupling relationships between key parameters such as wind speed, load, and rotational speed; and 3) empirically based fixed threshold methods, which lack the ability to fine-tune system state classification and dynamic adjustment.

[0004] These methods have obvious shortcomings in application: first, the accuracy and real-time performance of the dynamic safety domain construction are insufficient, and the existing models are difficult to fully reflect the dynamic characteristics of wind turbines; second, the safety domain model usually adopts a fixed or simplified form and cannot be dynamically updated in a high-dimensional nonlinear state space; third, it is difficult to simultaneously consider the coupling relationship between key variables such as wind speed, load, and rotation speed.

[0005] Therefore, there is an urgent need to develop a method that can dynamically construct a multi-dimensional safety domain of wind turbines to provide more accurate safe operation boundaries and provide a reliable basis for the safety control of wind turbines. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for constructing a multidimensional dynamic safety domain for a wind turbine generator set. By using the deep neural network technology with shared weights and combining it with confidence theory, a multidimensional dynamic three-layer model safety domain including multiple key variables such as wind speed, load and rotational speed is established, thereby achieving an accurate description and real-time update of the safe operation boundary of the wind turbine.

[0007] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions:

[0008] A model predictive control method based on distributed neighbor communication in a gridded wind farm includes the following steps:

[0009] S1. Data acquisition and preprocessing: Collect wind turbine operating data under different operating conditions, including wind speed, load, and rotational speed, and perform outlier detection, filtering, smoothing, and normalization on the data; the outlier detection uses the 3σ criterion combined with the boxplot method; the filtering and smoothing uses a second-order Butterworth low-pass filter; the normalization uses the minimum-maximum normalization method.

[0010] S2. Construction of two-dimensional safety surfaces and fusion of multi-dimensional safety domains: Using a deep neural network based on shared weights, multiple sets of two-dimensional safety surfaces such as wind speed-load, wind speed-rotation speed, and load-rotation speed are constructed according to the number of variables involved; a fusion mechanism is established to fuse the safety surfaces into multi-dimensional safety domains.

[0011] S3. Confidence level calculation and safety domain stratification: Use an exponential decay function to calculate the confidence level from the state point to the safety domain boundary, divide the safety domain into a safe area, a critical safety area, and a dangerous area; and establish a three-layer model of the safety domain.

[0012] S4. Establish a security domain adaptive update mechanism: Dynamically update the security domain based on newly collected operational data. Set a sliding time window and assign time-decaying weights to the data within the window. Adaptively update the fusion weight coefficients based on the rolling data within the sliding window. Simultaneously, establish a model-triggered full retraining mechanism.

[0013] In step S1, for the dataset , calculate the mean and standard deviation , define the outlier detection threshold:

[0014]

[0015] For the outlier judgment criteria: or ,but For outliers, a second-order Butterworth low-pass filter is used for filtering and smoothing:

[0016]

[0017] in is the cutoff angular frequency ( , ), is the damping coefficient ( ). For the normalization of key parameters such as wind speed v, load l and speed w, the minimum-maximum normalization method is used:

[0018] .

[0019] In step S2, the deep neural network adopts a shared weight network structure, and the network model is expressed as:

[0020] .

[0021] in is the input feature vector, is the weight matrix, is the bias vector, For the activation function (ReLU: ).

[0022] For each set of two-dimensional safety surfaces , the training loss function is designed as:

[0023]

[0024] in is a label (0 for unsafe, 1 for safe), is the predicted value, is the regularization coefficient.

[0025] The two-dimensional security surface is combined into a complete multi-dimensional security domain through the fusion algorithm. Let the multi-dimensional state point , the two-dimensional safety surfaces are 、 、 …. The fusion function is defined as:

[0026]

[0027] in 、 、 ...is the weight coefficient, satisfying , and the initial state satisfies = = ...; Define the judgment criteria: If Points is inside the security domain, otherwise it is outside the security domain, where is the safety threshold, initially set to .

[0028] In step S3, the confidence level from the state point to the safety domain boundary is calculated, and the safety domain is divided into a safe area, a critical safety area, and a dangerous area; , calculate its distance to the boundary of the safe zone: .in , is a point on the boundary of the safety region; the confidence level is calculated using an exponential decay function: .in is a scale parameter, which is 10% of the radius of the safety zone; for the safety zone stratification criteria: definition, safety zone: ; Critical safety area: ; Dangerous areas: .

[0029] In step S4, the security domain is dynamically updated based on the newly collected operating data. Set the sliding time window , window length Hours; assign time-decay weights to data within the window: in is the attenuation coefficient, is the data timestamp;

[0030] Time decay weight The weighted loss function applied to model training: , making the model pay more attention to recent data. It is also used in the prediction error calculation: , enhance the evaluation of recent performance. The adaptive update formula of fusion weight coefficient is:

[0031]

[0032] in is the learning rate ( ), 、 、 are weighted prediction errors for each dimension, is the average weighted forecast error;

[0033] In step S4, the full retraining condition is triggered for the model. , then retrain the entire model, where is the weighted average prediction error within the sliding window ( ), is the retraining threshold (initial set to 0.2).

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] A method for constructing a multi-dimensional safety domain for wind turbines based on deep neural networks is proposed, which can accurately describe the safe operating boundaries of wind turbines under different operating conditions and provide reliable constraints for control strategies; the innovative use of a two-dimensional safety surface combination method to construct a multi-dimensional safety domain can effectively solve the coupling problem of multiple parameters in wind turbine operation and improve computational efficiency; the introduction of confidence levels and safety domain stratification can achieve a refined division of system states and provide a quantitative basis for risk assessment; an adaptive update mechanism is designed to enable the safety domain to be adjusted in real time with changes in system states and environment, significantly improving the adaptability of the model under complex wind conditions; the data-driven modeling method does not require accurate physical modeling and has a wider applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is an overall flow chart of the method for constructing a multi-dimensional safety domain for a wind turbine generator set according to the present invention.

[0037] Figure 2 Schematic diagram of the neural network structure built for the safety domain (including the adaptive update part and the projection of the three-layer model of the safety domain).

[0038] Figure 3 It is a scatter plot of multi-dimensional security domain (taking three-dimensional security domain as an example).

[0039] Figure 4 The following is a schematic diagram of the multi-dimensional security domain after algorithm fusion (taking the three-dimensional security domain as an example). DETAILED DESCRIPTION

[0040] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.

[0041] Example: As shown in the attached Figure 1-4The present invention describes a method for constructing a dynamic safety domain for a wind turbine. Using shared-weight deep neural network technology and combined with confidence theory, a multidimensional dynamic three-layer model safety domain is established, encompassing multiple key variables such as wind speed, load, and rotational speed. This allows for a precise description and real-time update of the wind turbine's safe operating boundaries. This example uses the construction of a three-dimensional safety domain as an example. A simulation model of a 2MW wind turbine is constructed using Bladed software. A test protocol is designed encompassing normal operating conditions, extreme operating conditions, and fault conditions: a) Normal operating conditions: wind speed range 3-25 m / s, covering the full turbine power curve; b) Extreme operating conditions: extreme wind speed, gusts (amplitude 8 m / s), wind shear (index 0.2), and turbulence (intensity 16%); c) Fault conditions: pitch system failure and yaw system failure. Key data is collected: wind speed (v), blade load (l) (including blade root bending moment and thrust), and rotational speed (w). The sampling frequency is 50 Hz, and a total of 100 hours of operating data is acquired.

[0042] Including steps:

[0043] S1. Preprocessing of acquired data:

[0044] S11, perform outlier detection, using the improved The criterion is combined with the box plot method to identify outliers. For the data set , calculate the mean and standard deviation : , Calculate the quartiles for the box plot 、 、 and interquartile range 1. The comprehensive threshold is set as: , The outlier judgment criteria are: or ,but For data points identified as outliers, local interpolation is used to replace them: hour, .

[0045] S12. Perform data smoothing and design a second-order Butterworth low-pass filter to eliminate high-frequency noise. The frequency domain transfer function is:

[0046]

[0047] in , is the cutoff frequency, is the damping coefficient. Time domain discretization (using bilinear transformation): , where the coefficient calculation formula is:

[0048]

[0049] S13. Perform data normalization processing on wind speed, load and rotation speed to unify the data dimensions:

[0050]

[0051] in and are the minimum and maximum values ​​of the wind speed data, 、 、 、 Same thing.

[0052] To ensure the robustness of normalization, the actual value is:

[0053]

[0054] in ,right and Use the same treatment.

[0055] S2. Construction of two-dimensional security plane and integration of multi-dimensional security domains:

[0056] S21. For three key variables (wind speed, load, and rotation speed), a deep neural network (DNN) is used to construct a two-dimensional safety surface. First, the wind speed-load safety surface is constructed. First, the rotation speed is fixed to a specific value set. (Select n=9, evenly distributed in the speed range), for each speed value , construct the corresponding wind speed-load two-dimensional safety surface Secondly, determine the DNN network structure: the input layer contains 2 neurons, corresponding to the normalized wind speed and load ; Hidden layer 1 contains 32 neurons, activation function ReLU: The hidden layer 2 contains 16 neurons, and the activation function is ReLU; the output layer contains 1 neuron, and the activation function is Sigmoid: Then, design the forward propagation calculation process: the input is ; The output of hidden layer 1 is ; The output of hidden layer 2 is ; Output layer output: ,in, Finally, the objective function is trained using a binary cross entropy loss function with regularization:

[0057]

[0058] in is the sample size is the true label (0 means unsafe, 1 means safe), is the network prediction output, is the regularization coefficient, Represents the Frobenius norm of the weight matrix.

[0059] S22: For step S2, establish the wind speed-rotation speed safety surface and the load-rotation speed safety surface, and use the same network structure and training method as the wind speed-load safety surface. For each fixed load value, , build the corresponding wind speed-rotation speed safety surface To avoid repeated training and maintain consistency, a parameter sharing strategy is adopted: 1) Sharing the underlying feature extraction network parameters and ;2) Customize high-level network parameters for each security plane 、 、 、 .

[0060] S23. For step S2, the training implementation details are as follows: 1) Dataset partitioning: 70% training set, 15% validation set, and 15% test set. 2) Batch size: 64. 3) Adam optimizer with the following parameters: Learning rate: , momentum decay rate: , second-order moment decay rate: , value stability constant: 4) Training epochs: 200. 5) Early stopping strategy: Stop training when there is no improvement in performance on the validation set for 10 consecutive epochs. 6) Learning rate scheduling: When the validation set loss does not improve for 5 consecutive epochs, halve the learning rate.

[0061] S24, parameter update rule of the training process, for the parameters θ (including weight W and bias b), each step is updated:

[0062]

[0063] S25. Combine the constructed two-dimensional safety surfaces into a complete three-dimensional safety domain. First, use boundary interpolation. For each set of two-dimensional safety surfaces, use the cubic spline interpolation method to achieve smooth connection between different discrete points. Given a discrete speed point Wind speed-load safety surface , for any speed , using cubic spline interpolation to calculate the wind speed-load safety surface :

[0064]

[0065] in, are the interpolation coefficients, for cubic splines:

[0066] like ,but:

[0067]

[0068] Secondly, ensure the continuity of each safety surface at the intersection. Apply constraints at the intersection of two-dimensional safety surfaces: 1) Value continuity: . 2) First-order derivative is continuous: . 3) The second-order derivative is continuous: ∂ .

[0069] S26, using the domain fusion algorithm, calculate the projection value of the point on the wind speed-load safety surface under specific speed w, load l, and wind speed w , Projection value on wind speed-rotation speed safety surface , Projection value on load-speed safety surface . Then calculate the average safety score .like ≥ 0.5, then the decision point (v, l, w) is in the safety domain. The weighted fusion method is used to integrate the decision results of each two-dimensional safety domain:

[0070]

[0071] in Score the safety of the fusion. 、 、 is the weight coefficient of each dimension, satisfying . Initialize the weights to equal weights:

[0072] At the same time, define the boundary judgment criteria: Points It is within the security zone, otherwise it is outside the security zone. The threshold is initially set to . Points outside the safe zone are directly removed.

[0073] S3. Calculate confidence levels and perform security domain stratification

[0074] S31. Calculate the confidence level. Define the security domain boundary. ,in is the state point, Score safety; calculate status points To the security domain boundary The shortest distance: To improve computational efficiency, the gradient descent method is used for approximate calculation:

[0075]

[0076] in Score safety At the point The gradient at:

[0077]

[0078] Sign convention for distance calculation: If In the security domain, ;like Outside the security zone, . The confidence level is calculated using an exponential decay function:

[0079]

[0080] in is a scale parameter, and its value is 10% of the "radius" of the safety zone:

[0081]

[0082] in is a set of uniformly sampled state points that satisfies .

[0083] S32. Stratify security domains based on confidence levels. Based on confidence level thresholds, divide the state space into three layers: , Critical safety area: , Dangerous areas: . Conduct statistical analysis on historical operating data and combine expert experience to determine the stratification threshold. The specific calculation is as follows:

[0084]

[0085] in is the safety zone threshold, is the danger zone threshold is the average confidence level of the state points in the safe area, is the standard deviation. Based on this experimental data, we can calculate , .

[0086] S4. Design of security domain adaptive update mechanism

[0087] S41, keep the network structure unchanged, initialize the new network parameters to the current model parameters. With a lower learning rate ( ) Fine-tune the model and reduce the number of training rounds. Set the sliding time window:

[0088]

[0089] Where T is the window length, which is 1 hour. The timestamp of the data point.

[0090] Set the weight reduction strategy. Assign time-decaying weights to the data in the window:

[0091]

[0092] in is the attenuation coefficient, initially set to , so that the data weight at the beginning of the window is about 13.5% of the data weight at the end. The weighted loss function is designed as: .

[0093] S42. Design an error feedback adjustment strategy. Define the prediction error:

[0094]

[0095] in Status point The true security status (0 or 1), The time decay weight of the data point. Set the error threshold , when detected When , the model is completely retrained: 1) merge historical data with new data, and decay weights according to time Assign the importance of training samples. 2) Reinitialize the network parameters. 3) Use the weighted full dataset for training, and the loss function becomes:

[0096]

[0097] in Ensure that the model pays more attention to recent data patterns and adapts to changes in wind turbine operating conditions.

[0098] S43, perform adaptive parameter update. First, calculate the weighted prediction error of each dimension:

[0099]

[0100] Update the fusion weight coefficient based on the weighted prediction error:

[0101]

[0102] in is the learning rate, which is set to 0.05. Then the weights are normalized:

[0103]

[0104] The above steps complete the construction of the three-dimensional dynamic safety domain of the case wind turbine.

Claims

1. A method for constructing a multi-dimensional dynamic safety domain of a wind turbine generator set, characterized in that: The following steps are involved: S1. Data acquisition and preprocessing: Collect wind turbine operating data, including key parameters such as wind speed, load, and speed, and perform outlier detection, filtering, smoothing, and normalization on the data; S2. Two-dimensional safety surface construction and multi-dimensional safety domain fusion: Using a deep neural network based on shared weights, we construct multiple two-dimensional safety surfaces, such as wind speed-load, wind speed-rotation speed, and load-rotation speed, based on the number of variables involved. We also establish a fusion mechanism to fuse the safety surfaces into multi-dimensional safety domains. S3. Confidence level calculation and safety domain stratification: Use an exponential decay function to calculate the confidence level from the state point to the safety domain boundary, divide the safety domain into a safe area, a critical safety area, and a dangerous area; and establish a three-layer model of the safety domain. S4. Establish a security domain adaptive update mechanism: dynamically update the security domain based on newly collected operational data; Set a sliding time window and assign time-decay weights to the data within the window; Based on the sliding window rolling data, the fusion weight coefficient is adaptively updated; at the same time, a model triggering full retraining mechanism is established.

2. The method for constructing a multi-dimensional dynamic safety domain of a wind turbine generator set according to claim 1, characterized in that: The outlier detection in step S1 adopts the 3σ criterion combined with the boxplot method, and the detection criterion is: or ,but is an outlier, where is the sample mean, is the sample standard deviation.

3. The method for constructing a multi-dimensional dynamic safety domain of a wind turbine generator set according to claim 1, characterized in that: The filtering and smoothing in step S1 adopts a second-order Butterworth low-pass filter, and the transfer function is: ,in , is the cutoff frequency, is the damping coefficient. The normalization process uses the minimum-maximum normalization method, and the normalization formula is:

4. The method for constructing a multi-dimensional dynamic safety domain of a wind turbine generator set according to claim 1, characterized in that: In step S2, the two-dimensional security surface is constructed using a shared weight deep neural network. The network structure includes an input layer, multiple hidden layers and an output layer. The output layer represents the security of the state point. is the input feature vector, is the weight matrix, is the bias vector, is the activation function, and the network model is expressed as:

5. The method for constructing a multi-dimensional dynamic safety domain of a wind turbine generator set according to claim 4, characterized in that: The training loss function of the deep neural network is:

6. The method for constructing a multi-dimensional dynamic safety domain of a wind turbine generator set according to claim 1, characterized in that: In step S2, Score the safety of the fusion. 、 、 ,...are the weight coefficients of each dimension, and satisfy ,Multi-dimensional security domain fusion adopts weighted fusion method:

7. The method for constructing a multi-dimensional dynamic safety domain of a wind turbine generator set according to claim 1, characterized in that: In step S3, is the shortest distance from the state point to the security domain boundary, is a scale parameter, which is set to 10% of the radius of the safety zone. The confidence level is calculated using an exponential decay function:

8. The method for constructing a multi-dimensional dynamic safety domain of a wind turbine generator set according to claim 1, characterized in that: In step S4, the security domain adaptive update includes: (a) Setting a sliding time window , window length Hour; (b) Assign time-decay weights to the data in the window: ,in is the attenuation coefficient, is the data timestamp; (c) Decay the time weight Applied to: weighted loss function for training, weighted prediction error calculation, and fusion weight coefficient update; (d) Adaptive update of fusion weight coefficients (when the prediction error exceeds the preset threshold, the model is completely retrained):

Images