A method and system for identifying an inertia deviation of guide vanes of a hydraulic turbine
By constructing a dynamic neural warping network and utilizing deep learning and time series analysis, the inertial offset of the turbine guide vanes is identified and eliminated, solving the problem of insufficient control precision in existing technologies and achieving more efficient turbine control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- THREE GORGES JINSHAJIANG CHUANYUN HYDROPOWER DEV CO LTD
- Filing Date
- 2023-08-30
- Publication Date
- 2026-05-19
AI Technical Summary
The existing turbine governors fail to effectively handle the inertial offset of the guide vanes, resulting in insufficient control accuracy and affecting the stability of the grid connection frequency.
By employing deep artificial neural networks and dynamic time warping algorithms, a dynamic neural warping network is constructed to identify and eliminate guide vane inertial misalignment. The model is trained using time series data, features are extracted, and similarity is calculated to achieve accurate identification of inertial misalignment.
This improved the reliability and control accuracy of turbine guide vane inertial offset identification, ensuring stable turbine operation under different operating conditions and enhancing the control accuracy and frequency stability of the speed regulation system.
Smart Images
Figure CN117194996B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system equipment feature recognition technology, and in particular to a method and system for recognizing the inertial offset of turbine guide vanes. Background Technology
[0002] Guide vanes are an important component of a water turbine, typically located around the turbine runner. Their primary function is to control the direction and flow rate of water, thereby regulating the turbine's output power.
[0003] Guide vanes typically consist of a series of movable blades that can rotate or tilt to change the direction of water flow. The construction and adjustment methods of guide vanes vary depending on the type and design of the turbine. By appropriately adjusting the position of the guide vanes, the turbine can operate efficiently under different flow conditions and adapt to varying load requirements. The flexible adjustment of the guide vanes enables the turbine to operate stably under various operating conditions, improving its efficiency and reliability.
[0004] During turbine operation, guide vanes may be subjected to the impact force of water flow, hydraulic torque, and torque variations generated by friction with bearings. These torque variations affect the speed and rotational motion of the guide vanes. However, due to the rotational inertia of the guide vanes, they possess a certain resistance to these torque variations. This results in the sampled guide vane opening during turbine operation being a continuously changing value containing an inertial quantity. This inertial value also exists in the control system, causing feedback and control to sometimes deviate and lag, thus affecting the control accuracy of the entire speed regulation system and consequently the grid connection frequency. Existing turbine governor control of the guide vanes only considers the control dead zone, failing to address inertial offsets and hysteresis outside the dead zone. This results in turbine control being a perpetual control process, never reaching a stable (relative) state. Summary of the Invention
[0005] In view of the deficiencies in the prior art, the purpose of this invention is to provide a method and system for identifying the inertial drift of turbine guide vanes, aiming to solve the problem that the prior art lacks the ability to identify the inertial drift of turbine guide vanes.
[0006] The technical solution adopted by the present invention is as follows: In the first aspect, the present invention provides a method for identifying the inertial offset of a water turbine guide vane, comprising the following steps:
[0007] Collect relevant parameters of the turbine guide vanes under specific operating conditions and determine the time series data under the corresponding index parameters; the relevant parameters include: governor guide vane opening, governor opening setpoint, PID setpoint, primary frequency regulation action, AGC command, and unit frequency;
[0008] A dynamic neural warping network is constructed to train and learn time series data using a deep artificial neural network to extract features; at the same time, a dynamic time warping algorithm is used to align time series data and learn their similarities.
[0009] For real-time data, the time series similarity is calculated using the corresponding similarity metric to identify the inertial offset of the turbine guide vanes.
[0010] Furthermore, the time series is determined in the following manner:
[0011] During normal operation, a typical invalid guide vane inertial misalignment should meet the following conditions: Under grid-connected conditions, the primary frequency regulation of the speed control system starts to operate, while the AGC should not operate, the unit frequency should exceed the primary frequency regulation dead zone, and the governor PID setpoint starts to increase, but the final opening setpoint does not increase; under these conditions, time series data under specific operating conditions are obtained, and then normalized and interpolated.
[0012] Furthermore, before training a deep artificial neural network, it is necessary to extract the representation vector from the time series data.
[0013] The extracted representation vector of the time series is processed through a temporal convolutional network; the time series data is used as network input, a convolutional layer is built, residual connections are added, and finally the time dimension is compressed into a fixed-length representation vector through a global pooling layer.
[0014] Furthermore, the method for constructing the dynamic neural warping network is as follows: the input of the deep artificial neural network is the representation vector and the sequence length, and the output is the dynamic time warping distance between two time series; the hidden layer uses the ReLU activation function, and the output layer uses the linear activation function; by learning the dynamic neural warping network, the alignment relationship between time series is automatically learned, thereby realizing the similarity measurement and pattern matching tasks of time series.
[0015] Furthermore, the method for calculating the dynamic time warping distance between the two time series includes: matrixing the two time series, using Euclidean distance as a similarity metric, and calculating the dynamic time warping distance.
[0016] In a second aspect, the present invention provides a turbine guide vane inertial offset identification system, comprising:
[0017] The historical data acquisition module is used to collect relevant parameters of the turbine guide vanes under specific operating conditions; the relevant parameters include: governor guide vane opening, governor opening setpoint, PID setpoint, primary frequency regulation action, AGC command, and unit frequency.
[0018] The data processing module is used to determine the time series data under the corresponding indicator parameters based on relevant parameters, and then perform normalization and difference filling processing.
[0019] The Dynamic Neural Warping Network Construction Module is used to build a dynamic neural warping network, which trains and learns time series data through a deep artificial neural network to extract features; at the same time, it uses a dynamic time warping algorithm to align time series data and learn their similarities.
[0020] The identification and judgment module is used to calculate the time series similarity of real-time data using corresponding similarity metrics, thereby identifying the inertial offset of the turbine guide vanes.
[0021] Furthermore, it also includes:
[0022] The representation vector training module is used to process representation vectors through a temporal convolutional network. It takes time series data as network input, builds convolutional layers, adds residual connections, and finally compresses the time dimension into a fixed-length representation vector through a global pooling layer.
[0023] The beneficial effects of this invention are: This invention learns the inertial offset recognition of turbine guide vanes based on historical data of the turbine. The network structure and parameters can be adjusted according to specific data characteristics. Multiple experiments and optimizations are conducted to train the most suitable dynamic neural regularization network, so that the calculated inertial offset recognition of the guide vanes has high reliability. Attached Figure Description
[0024] Figure 1 This is a flowchart of the turbine guide vane inertial offset identification method disclosed in an embodiment of the present invention;
[0025] Figure 2 This is a structural diagram of the TCN deep neural network disclosed in an embodiment of the present invention;
[0026] Figure 3 This is a block diagram of the turbine guide vane inertial offset recognition system disclosed in an embodiment of the present invention. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in further detail below with reference to the accompanying drawings, but the embodiments of the present invention are not limited thereto.
[0028] Example 1:
[0029] This invention discloses a method for identifying the inertial offset of turbine guide vanes. It uses an artificial neural network model to fit the expression relationship of relevant time series data, thereby improving the performance of time series similarity measurement and pattern matching tasks, more accurately identifying characteristic operating conditions, and eliminating invalid inertial offsets.
[0030] Figure 1 The flowchart of the turbine guide vane inertial offset identification method provided in the embodiment of the present invention is as follows: Figure 1 As shown, it includes the following steps:
[0031] S101, collect relevant parameters of the turbine guide vanes under specific operating conditions, and determine the time series data under the corresponding index parameters; the relevant parameters include: governor guide vane opening, governor opening setpoint, PID setpoint, primary frequency regulation action, AGC command, and unit frequency.
[0032] Specifically, during normal operation, a typical invalid guide vane inertial misalignment should meet the following conditions: under grid-connected conditions, the primary frequency regulation of the speed control system starts to operate (at the same time, the AGC should not operate, eliminating AGC interference), the unit frequency should exceed the primary frequency regulation dead zone, and the governor PID setpoint starts to increase, but the final opening setpoint does not increase.
[0033] Taking maintenance period data as an example, data is collected using the method described above to obtain multiple sets of typical time series data. The retrieved N sets of data are then subjected to interpolation and normalization to be within the range [0,1]. Simultaneously, the data is divided into a training set N1 and a test set N2.
[0034] Normalization formula:
[0035]
[0036] In the above formula, y' is the normalized value, y is the actual value of the variable in the training set, and y max and y min These are the upper and lower limits of the variable in the speed control system, respectively.
[0037] interpolation filling:
[0038]
[0039] In the above formula, y k The value after the difference, y n and y n-1 The actual value of the variable in the training set.
[0040] S102, construct a dynamic neural warping network, train and learn time series data through a deep artificial neural network to extract features; at the same time, use a dynamic time warping algorithm to align time series data and learn their similarity.
[0041] Specifically, before model training, the time series needs to have its representation vector extracted; a Temporal Convolutional Network (TCN) is constructed to extract the representation vector.
[0042] Constructing a Temporal Convolutional Network (TCN) residual block: The residual block contains two dilated convolutional layers, two ReLU activation function layers, and two dropout layers. After weight normalization, the dilated convolutional layers are connected to the ReLU activation function layers. To prevent overfitting, a dropout layer is added, and this process is repeated. To ensure the input and output dimensions of the residual module are the same, a 1×1 convolution is added between the input and output. The kernel size of the dilated convolutional layers is set to z, and the expansion coefficient d between each layer grows exponentially to ensure a sufficiently large receptive field for the residual module.
[0043] Constructing a deep neural network for TCN, such as Figure 2 As shown, the temporal convolutional neural network (TCN) consists of multiple TCN residual modules and multiple fully connected network layers. The output of the previous residual block serves as the input to the next residual block, and the output of the last residual block serves as the input to the fully connected network layer. The number of neurons in the fully connected layer decreases sequentially, and finally, a fully connected network layer containing one neuron serves as the output layer of the model. The fully connected layers use the ReLU activation function, and the output layer uses a linear activation function.
[0044] The network weights are randomly initialized, and the size of a single training sample and the number of training epochs are specified. Finally, a global pooling layer compresses the time dimension into a fixed-length representation vector. This representation vector can serve as a high-level abstract feature of the time series data for subsequent time series analysis tasks.
[0045] Construct a dynamic neural regularization network (DNTW).
[0046] The Dynamic Time Warping (DTW) network comprises a Time Warping (DTW) computation module and a deep artificial neural network (or other AI algorithm). The deep AI network is used to train the data, but finding similarity between time series data requires comparing two sets of data. The comparison rules and metrics are implemented by the DTW computation module. DNTW can be intuitively understood as a deep AI network with DTW. The deep AI network consists of an input layer, multiple hidden layers, and an output layer. The output of one hidden layer serves as the input to the next hidden layer, and the output of the last hidden layer serves as the input to the output layer. The input to the deep AI network is the aforementioned representation vector and sequence length (sequence length refers to the length of the representation vector). The output is the Dynamic Time Warping (DTW) distance between two time series. The hidden layers use the ReLU activation function, and the output layer uses a linear activation function. Mean Squared Error (MSE) is the most commonly used regression loss function. MSE is the sum of the squared distances between the target variable and the predicted value. By learning the Dynamic Time Warping network, the alignment relationship between time series is automatically learned, thereby achieving tasks such as time series similarity measurement and pattern matching.
[0047] Dynamic time-warped distance calculation steps:
[0048] Input time series X and Y:
[0049] X = [x1, x2, x3, ..., x n ]
[0050] Y = [y1, y2, y3, ..., y m ]
[0051] Where n and m represent the lengths of the two time series.
[0052] Create a two-dimensional matrix.
[0053]
[0054] Initialize D[0,0] and other elements to positive infinity. Using dynamic programming (DP), recursively calculate each element in the matrix to obtain the dynamic normalized distance between the two time series. The recursive formula is as follows:
[0055] D[i,j]=distance(x i ,y j )+min(D[i-1,j],D[i,j-1],D[i-1,j])
[0056] Where, distance(x) i ,y jThe distance () represents the distance between the i-th element of sequence X and the j-th element of sequence Y. Here, Euclidean distance is used as the similarity measure.
[0057]
[0058] The calculated dynamic regularization distance between two time series is D[n,m].
[0059] Finally, the trained DNTW is evaluated on the test set. In actual computation, the network structure and parameters need to be adjusted according to the specific task and data characteristics, and multiple experiments and optimizations need to be conducted to find the most suitable dynamic neural regularization network, so as to effectively handle the similarity measurement and pattern matching tasks of time series data.
[0060] S103, for real-time data, uses the corresponding similarity metric to calculate the time series similarity, thereby identifying the inertial offset of the turbine guide vanes.
[0061] In actual operation, the system will analyze and compare the time series data in real time with reference to the model, identify the inertial offset of the turbine guide vanes, and eliminate the impact of inertial offset on turbine control.
[0062] The turbine guide vane inertial offset identification method of the present invention learns the turbine guide vane inertial offset identification based on historical data of the turbine. The network structure and parameters can be adjusted according to specific data characteristics. Multiple experiments and optimizations are carried out to train the most suitable dynamic neural regularization network, so that the calculated guide vane inertial offset identification has high reliability.
[0063] Example 2:
[0064] See Figure 3 This embodiment discloses a turbine guide vane inertial offset identification system, comprising:
[0065] The historical data acquisition module 200 is used to collect relevant parameters of the turbine guide vanes under specific operating conditions; the relevant parameters include: governor guide vane opening, governor opening setpoint, PID setpoint, primary frequency regulation action, AGC command, and unit frequency.
[0066] Data processing module 210 is used to determine the time series data under the corresponding indicator parameters based on relevant parameters, and then perform normalization and difference filling processing.
[0067] The dynamic neural warping network construction module 220 is used to construct a dynamic neural warping network, which trains and learns time series data through a deep artificial neural network to extract features; at the same time, it uses a dynamic time warping algorithm to align time series data and learn their similarity.
[0068] The identification and judgment module 240 is used to calculate the time series similarity for real-time data using corresponding similarity metrics, thereby identifying the inertial offset of the turbine guide vanes.
[0069] Furthermore, it also includes:
[0070] The representation vector training module 230 is used to process the representation vector through a temporal convolutional network. It takes time series data as network input, builds convolutional layers, adds residual connections, and finally compresses the time dimension into a fixed-length representation vector through a global pooling layer.
[0071] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for identifying the inertial offset of a water turbine guide vane, characterized in that, Includes the following steps: Collect relevant parameters of the turbine guide vanes under specific operating conditions and determine the time series data under the corresponding index parameters; the relevant parameters include: governor guide vane opening, governor opening setpoint, PID setpoint, primary frequency regulation action, AGC command, and unit frequency; Deep artificial neural networks are used to train and learn time series data to extract features; at the same time, dynamic time warping algorithm is used to align time series data and learn their similarity to construct a dynamic neural warping network. For real-time data, the time series similarity is calculated using the corresponding similarity metric to identify the inertial offset of the turbine guide vanes.
2. The method for identifying the inertial offset of turbine guide vanes according to claim 1, characterized in that, The time series was determined in the following manner: During normal operation, a typical invalid guide vane inertial misalignment should meet the following conditions: Under grid-connected conditions, the primary frequency regulation of the speed control system starts to operate, while the AGC should not operate, the unit frequency should exceed the primary frequency regulation dead zone, and the governor PID setpoint starts to increase, but the final opening setpoint does not increase; under these conditions, time series data under specific operating conditions are obtained, and then normalized and interpolated.
3. The method for identifying the inertial offset of turbine guide vanes according to claim 2, characterized in that, Before training a deep artificial neural network, it is necessary to extract the representation vector from the time series data. The extracted representation vector of the time series is processed through a temporal convolutional network; the time series data is used as network input, convolutional layers are built, residual connections are added, and finally the time dimension is compressed into a fixed-length representation vector through a global pooling layer.
4. The method for identifying the inertial offset of turbine guide vanes according to claim 3, characterized in that, The method for constructing the dynamic neural warping network is as follows: the input of the deep artificial neural network is the representation vector and the sequence length, and the output is the dynamic time warping distance between two time series; the hidden layer uses the ReLU activation function, and the output layer uses the linear activation function; by learning the dynamic neural warping network, the alignment relationship between time series is automatically learned, thereby realizing the similarity measurement and pattern matching tasks of time series.
5. The method for identifying the inertial offset of turbine guide vanes according to claim 4, characterized in that, The method for calculating the dynamic time warping distance between the two time series includes: matrixing the two time series, using Euclidean distance as a similarity metric, and calculating the dynamic time warping distance.
6. A system for identifying the inertial offset of a water turbine guide vane, characterized in that, include: The historical data acquisition module is used to collect relevant parameters of the turbine guide vanes under specific operating conditions; The relevant parameters include: governor guide vane opening, governor opening setpoint, PID setpoint, primary frequency regulation action, AGC command, and unit frequency. The data processing module is used to determine the time series data under the corresponding indicator parameters based on relevant parameters, and then perform normalization and difference filling processing. The Dynamic Neural Warping Network Construction Module is used to build a dynamic neural warping network, which trains and learns time series data through a deep artificial neural network to extract features; at the same time, it uses a dynamic time warping algorithm to align time series data and learn their similarities. The identification and judgment module is used to calculate the time series similarity of real-time data using corresponding similarity metrics, thereby identifying the inertial offset of the turbine guide vanes.
7. The turbine guide vane inertial offset identification system according to claim 6, characterized in that, Also includes: The representation vector training module is used to process representation vectors through a temporal convolutional network. It takes time series data as network input, builds convolutional layers, adds residual connections, and finally compresses the time dimension into a fixed-length representation vector through a global pooling layer.