Lightweight tcn broadband oscillation suppression method for power excitation system
By using a lightweight causal dilated convolutional network and a frequency-adaptive channel attention mechanism, the problems of narrow bandwidth and difficulty in implementing complex deep learning models in traditional PSS are solved, and fast and effective broadband oscillation suppression is achieved in the excitation system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 四川华电珙县发电有限公司
- Filing Date
- 2026-04-03
- Publication Date
- 2026-06-19
AI Technical Summary
The traditional PSS of existing power systems lacks sufficient frequency band adaptability, making it difficult to provide effective positive damping over a wide frequency range. Furthermore, complex deep learning models are difficult to implement in real-time closed-loop control in excitation controllers with limited computing resources.
A lightweight causal dilated convolutional network (TCN) combined with a frequency-adaptive channel attention mechanism is adopted. The multi-scale temporal features of the excitation system are extracted through causal dilated convolution, and the additional damping control signal is generated by dynamically allocating weights using the frequency-adaptive channel attention layer.
It achieves low computational complexity, fast response speed, and wide frequency band adaptation in broadband oscillation suppression, and can effectively suppress broadband oscillations of generator sets under complex power grid conditions.
Smart Images
Figure CN122246702A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system stability control technology, specifically a lightweight TCN broadband oscillation suppression method for power excitation systems. Background Technology
[0002] Low-frequency oscillations in power systems have long been one of the main factors threatening the safe and stable operation of the power grid. In order to suppress such oscillations, synchronous generators in thermal power plants are generally equipped with power system stabilizers (PSS). Traditional PSS usually adopts a linear control strategy based on lead-lag compensation links, and its parameters are tuned according to a specific single operating point.
[0003] However, with the rapid development of wind power, photovoltaic power, and high-voltage direct current transmission technologies, the dynamic processes of modern power systems exhibit significant "wideband oscillation" characteristics. Their oscillation spectrum has broken through the conventional low-frequency band limitations, extending to subsynchronous oscillations (SSO) and even higher frequency bands. Therefore, it can be clearly inferred that existing control technologies face severe adaptability bottlenecks under such operating conditions.
[0004] (1) The traditional PSS has insufficient frequency band adaptability; the traditional PSS is essentially a linear filter with fixed phase compensation characteristics, which makes it difficult to provide effective positive damping in a wide frequency band at the same time; the parameters optimized for low-frequency oscillations often show a lack of phase compensation or even negative damping (reverse modulation) in the high-frequency band, which may deteriorate the wideband stability of the system.
[0005] (2) Existing intelligent algorithms struggle to balance performance and real-time performance. Although deep learning control methods based on Long Short-Term Memory (LSTM) networks or Transformers theoretically possess strong nonlinear fitting capabilities, they face significant engineering application obstacles: LSTM employs serial computation, resulting in high inference latency; while Transformers have high parallelism, the computational complexity of their self-attention mechanism increases quadratically with sequence length, and the number of parameters is enormous, making it difficult to achieve millisecond-level real-time closed-loop control in excitation controllers (usually DSPs or embedded FPGAs) with limited computational resources; furthermore, ordinary neural networks lack strict temporal causal constraints, which can easily introduce future information leakage, leading to control failure.
[0006] Therefore, there is an urgent need to develop an excitation control algorithm that has both wideband feature extraction capabilities and meets the requirements of lightweight and real-time performance of embedded systems, in order to fill the gap in the field of wideband oscillation suppression in existing technologies. Summary of the Invention
[0007] To address the shortcomings of existing technologies, this invention provides a lightweight TCN broadband oscillation suppression method for power excitation systems. This method combines the efficient feature extraction capability of causal dilated convolution with the frequency selection capability of channel attention, solving the technical challenges of narrow bandwidth in traditional PSS and the difficulty in deploying complex deep learning models.
[0008] To achieve the above objectives, the present invention employs the following technical solution:
[0009] A lightweight TCN broadband oscillation suppression method for electric excitation systems includes the following steps:
[0010] Step 1: Construct a broadband oscillation suppression network model for the excitation system. The model sequentially includes an input preprocessing layer, a lightweight causal TCN feature extraction layer, a frequency adaptive channel attention layer, and an output mapping layer.
[0011] Step 2: Collect real-time operating status data of the synchronous generator, construct a timing input vector, and preprocess the timing input vector;
[0012] Step 3: Input the preprocessed temporal input vector into the lightweight causal TCN feature extraction layer, extract multi-scale temporal features through causal dilated convolution operation, and output feature tensors;
[0013] Step 4: Input the feature tensor into the frequency adaptive channel attention layer, compress the sequence dimension through global pooling to obtain channel statistics, use the fully connected layer network to learn the dependencies between channels to generate channel weight vectors, and fuse the channel weight vectors with the feature tensor to obtain a weighted feature vector.
[0014] Step 5: Input the weighted feature vector into the output mapping layer to calculate the additional damping control signal;
[0015] Step 6: Superimpose the additional damping control signal onto the voltage reference setting value of the excitation regulator to adjust the excitation current of the generator.
[0016] Furthermore, the preprocessing described in step 2 specifically includes: using a sliding window to extract data segments and performing online normalization processing.
[0017] Furthermore, the lightweight causal TCN feature extraction layer consists of several cascaded causal dilated residual blocks, used to extract multi-scale temporal features of the excitation system state variables; the frequency adaptive channel attention layer dynamically allocates weights according to the importance of the feature channels.
[0018] Furthermore, the fully connected layer network in step 4 comprises two fully connected layers, and the fusion operation is to multiply the channel weight vector with the feature tensor channel by channel.
[0019] Furthermore, in step 6, the trainable parameters of the network model are fine-tuned and updated online using the gradient descent method based on the damping objective function.
[0020] Furthermore, the causal dilation residual block adopts a depthwise separable convolutional architecture, which includes sequentially cascaded channel-wise convolutional sub-operations and pointwise convolutional sub-operations.
[0021] Furthermore, the calculation process of the frequency-adaptive channel attention layer in step 4 includes the following:
[0022] Step 4.1: Perform global information compression. Compress the feature tensor along the time dimension using global average pooling to obtain channel statistics.
[0023] Step 4.2: Perform channel dependency learning, using two fully connected layers and activation functions to learn the dependencies between channels and generate channel weight vectors;
[0024] Step 4.3: Perform feature recalibration by multiplying the channel weight vector with the original feature tensor channel by channel to obtain the frequency-weighted feature vector.
[0025] Furthermore, the normalization process described in step 2 employs an online Z-score standardization method, dynamically adjusting the input data based on the mean and standard deviation within the sliding time window.
[0026] Furthermore, the operating status data includes the change in generator active power and the deviation of generator rotor angular velocity.
[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0028] This invention first designs a lightweight temporal convolutional network (TCN) with causal dilated convolutional layers to extract multi-scale temporal features of excitation operation data using exponentially increasing dilation coefficients, while strictly ensuring the temporal causality of control signals. Secondly, it introduces a frequency-adaptive channel attention mechanism to dynamically and adaptively allocate feature channel weights based on real-time oscillation modes, capturing oscillations from low to high frequencies. Finally, a fully connected layer maps the features to additional damping control commands. This invention features low computational cost, fast response speed, and wide bandwidth adaptability, exhibiting excellent broadband oscillation suppression performance for generator units under complex power grid conditions. Attached Figure Description
[0029] Appendix Figure 1 This is a flowchart of the present invention;
[0030] Appendix Figure 2This is a schematic diagram of the lightweight causal TCN feature extraction layer;
[0031] Appendix Figure 3 This is a schematic diagram of the structure of the frequency-adaptive channel attention layer;
[0032] Appendix Figure 4 This is a structural diagram of the coordinated operation of the excitation system and control algorithm in this invention. Detailed Implementation
[0033] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.
[0034] like Figure 1-4 As shown in the figure, a lightweight TCN broadband oscillation suppression method for power excitation systems is disclosed. The figure illustrates the complete closed-loop process from data acquisition, preprocessing, network feature extraction, attention weighting, control output to online optimization from top to bottom. Specifically, it includes: firstly, real-time acquisition of generator active power variation. and rotor angular velocity deviation Constructing time-series input vectors Subsequently, after online Z-score normalization preprocessing, the data is input into a lightweight causal TCN to extract feature tensors containing different time scales. Then, a weight vector is generated through a frequency-adaptive channel attention mechanism. Recalibrate the features to obtain Finally, the additional damping control signal is output via mapping through the fully connected layer. And based on the comprehensive objective function Gradient-based online fine-tuning of network parameters.
[0035] Figure 2 This is a schematic diagram of the lightweight causal TCN feature extraction layer; the diagram shows the details of the first layer from left to right. The internal structure and data flow of the residual block of the layered causal dilation convolution; the specific process is as follows: input tensor First, causal zero-padding is applied to ensure that the convolutional window only covers historical data; then, a depthwise separable convolutional layer is entered, utilizing an inflation coefficient that grows exponentially with the number of layers. Expand the receptive field; after layer normalization, ReLU activation function and Dropout layer processing, connect the original input through the residual connection path. Add them together to output the feature tensor. (i.e., the input to the next layer), thus constructing a deep network and preventing gradient vanishing; it should be noted that when this residual block is the last layer of the network, its output is a feature tensor that extracts multi-scale temporal dependencies. It also constitutes Figure 3 Input feature tensor ;
[0036] Figure 3 This is a schematic diagram of the structure of the frequency-adaptive channel attention layer; the diagram shows the frequency-adaptive channel attention layer from left to right. Figure 2 Output feature tensor The process of frequency-weighted processing includes: inputting the feature tensor. First, it is compressed into channel statistics through global average pooling. Then it enters the excitation module, passing through a dimensionality-reduced fully connected layer (compression ratio of...). The system employs ReLU activation function, a fully connected layer of increased dimension, and Sigmoid activation function to learn the non-linear dependencies between channels and generate normalized channel weight vectors. Finally, the weights are multiplied channel by channel. Acting on the original feature Output the frequency adaptive feature tensor after channel weight recalibration. The tensor While preserving the original timing information, the amplitude of the characteristic channels of the key frequency bands was enhanced according to the current dominant oscillation mode;
[0037] Figure 4 This is a structural diagram of the coordinated operation of the excitation system and control algorithm in this invention; the controller described in this invention (such as...) Figure 2 , Figure 3 The network topology described is embedded in the synchronous generator excitation closed-loop control loop using the embedded logic: that is, the controller collects the state variables of the synchronous generator in real time. , The additional damping control signal is obtained after processing by the internal neural network. The signal at the addition point is related to the voltage reference value. The summation is used as a new reference voltage input to the automatic voltage regulator (AVR); the AVR adjusts the excitation power unit voltage based on the new reference voltage value. and current This achieves the goal of suppressing broadband oscillations in the power grid.
[0038] The specific steps are as follows:
[0039] Step 1: Construct a broadband oscillation suppression network model for the excitation system. The model sequentially includes an input preprocessing layer, a lightweight causal TCN feature extraction layer, a frequency-adaptive channel attention layer, and an output mapping layer. The lightweight causal TCN feature extraction layer consists of several cascaded causal dilated residual blocks, used to extract multi-scale temporal features of the excitation system's state variables. The frequency-adaptive channel attention layer dynamically assigns weights based on the importance of the feature channels. The specific method is as follows:
[0040] Step 1.1: Design the input layer, defining the network input as... ;in, express The input matrix at any given time is dimensionless; 2 represents the change in generator active power. and generator rotor angular velocity deviation Two feature channels, , The unit is per unit (pu); sliding window length The unit is the number of sampling points;
[0041] Step 1.2: As Figure 2 As shown, the construction is made of The cascaded causal dilated residual blocks form the backbone network; for the first... Each residual block has an expansion coefficient set to... That is, the dilation coefficients of layers 1, 2, and 3 are 1, 2, and 4, respectively; the kernel size is set to... The number of output channels per residual block is set to... The internal structure of the residual block consists of a causal filling layer, a depthwise separable convolutional layer, a layer normalization function, a ReLU activation function, and a dropout layer.
[0042] Step 1.3: As Figure 3 As shown, an SE module is connected after the TCN backbone network; the compression ratio is set. This module is used to recalibrate based on the importance of the features. The weight of each channel;
[0043] Step 1.4: Use a fully connected layer with input dimension . The output dimension is 1, and the output value is the additional damping control signal, where, , ;
[0044] Step 2: Collect real-time operating status data of the synchronous generator and construct a time-series input vector. The operating status data includes the change in generator active power. and generator rotor angular velocity deviation For the time-series input vector Perform sliding window capture and online normalization; among them, Input data matrix at time step Sliding window length The unit is the number of sampling points; change in generator active power. The unit is per unit (pu), which is the ratio of the actual value to the selected reference value; generator rotor angular velocity deviation. The unit is per unit (pu), specifically:
[0045] Step 2.1: The excitation controller uses a sampling period Real-time acquisition of generator terminal data; calculation of changes in generator active power. and generator rotor angular velocity deviation Among them, the change in generator active power The unit is per unit (pu); generator rotor angular velocity deviation The unit is per unit (pu);
[0046] Step 2.2: Construct a length of The first-in-first-out queue performs online Z-score normalization on the data in the queue. The calculation formula is as follows:
[0047]
[0048] Among them, the original input data Normalized data Mean within the sliding window Standard deviation within the sliding window ; To prevent the division of tiny positive numbers with a denominator of zero, the fraction is dimensionless;
[0049] Step 3: Input the processed temporal input vector into the lightweight causal TCN feature extraction layer; in each causal dilated residual block, perform causal dilated convolution operation by setting an exponentially growing dilation coefficient. Expanding the receptive field without revealing future information, and outputting feature tensors containing information at different time scales. Among them, the coefficient of thermal expansion Dimensionless; characteristic tensor Number of feature channels Dimensionless, sliding window length Dimensionless, specifically:
[0050] Step 3.1: In the first... In each residual block, a causal dilatation convolution operation is performed; for the input sequence Its output The calculation is as follows:
[0051]
[0052] Among them, network layer index ;No. Layer input sequence vector ; Output value at time kernel size , dimensionless; the first Convolutional kernel weight parameters Dimensionless; coefficient of thermal expansion Dimensionless; For the current layer at time The convolution output value is dimensionless;
[0053] Step 3.2: Pad the left side of the sequence before convolution. One zero; this ensures that the convolution kernel is at time 1. , , Data from historical moments must strictly meet physical feasibility requirements.
[0054] Step 4: Convert the feature tensor Input the frequency-adaptive channel attention layer; calculate channel statistics by compressing the spatial dimension of features through global average pooling. By utilizing the non-linear dependencies between channels learned from two fully connected layers, channel weight vectors are generated. The channel weight vector is combined with the original feature tensor. Perform channel-by-channel multiplication to obtain the frequency-weighted eigenvector. Among them, channel statistics Number of feature channels Dimensionless; channel weight vector And the range of element values is Dimensionless; frequency-weighted eigenvectors Dimensionless, specifically:
[0055] Step 4.1: Perform global average pooling along the time dimension to calculate channel statistics. ;
[0056]
[0057] Among them, the Descriptors of each channel Dimensionless; , is a set of positive integers, and represents the number of feature channels. Dimensionless; sliding window length Dimensionless; characteristic tensor of the first Channel 1 Element value of time step ;
[0058] Step 4.2: Capture channel dependencies through two fully connected layers and generate channel weight vectors. :
[0059]
[0060] Among them, the channel weight vector Dimensionless; dimensionality-reduced weight matrix Dimensionless; Upgraded weight matrix Number of feature channels Dimensionless; The scaling factor is dimensionless. It is the ReLU activation function; Use the Sigmoid activation function;
[0061] Step 4.3: Perform feature recalibration; convert the channel weight vector With the original feature tensor Perform channel-by-channel multiplication to obtain the frequency-weighted eigenvector. ;in, Dimensionless; , dimensionless.
[0062] Step 5: Calculate the frequency-weighted feature vector. The input to the output mapping layer is used to calculate the current additional damping control signal through linear transformation. ;in, Apply damping control signal at all times The unit is per unit (pu), specifically.
[0063] Step 6: Obtain the additional damping control signal As a supplementary instruction, it is superimposed on the voltage reference setpoint used in the automatic voltage regulation channel within the excitation regulator to drive the excitation power unit to adjust the generator's excitation current; simultaneously, according to the selected damping objective function... The gradient descent method is used to fine-tune and update the trainable parameters in the neural network online; whereby the objective function value... Dimensionless, specifically:
[0064] Step 6.1: Execute closed-loop control, such as... Figure 4 As shown, The AVR reference voltage summation point sent to the excitation system drives the exciter to adjust the generator terminal voltage;
[0065] Step 6.2: Perform online updates and define the comprehensive objective function appropriately. :
[0066]
[0067] Among them, the active power deviation weighting coefficient Dimensionless; weighted coefficient for rotational speed deviation Dimensionless; control output energy weighting coefficient It is dimensionless; therefore, when the system oscillation exceeds a given threshold, reasonable calculation is required. Then, the gradient descent method is used to fine-tune the output layer weights online and in real time so that the model can adapt well to the current working conditions.
[0068] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.
Claims
1. A lightweight TCN broadband oscillation suppression method for electric excitation systems, characterized in that, Includes the following steps: Step 1: Construct a broadband oscillation suppression network model for the excitation system. The model sequentially includes an input preprocessing layer, a lightweight causal TCN feature extraction layer, a frequency adaptive channel attention layer, and an output mapping layer. Step 2: Collect real-time operating status data of the synchronous generator, construct a timing input vector, and preprocess the timing input vector; Step 3: Input the preprocessed temporal input vector into the lightweight causal TCN feature extraction layer, extract multi-scale temporal features through causal dilated convolution operation, and output feature tensors; Step 4: Input the feature tensor into the frequency adaptive channel attention layer, compress the sequence dimension through global pooling to obtain channel statistics, use the fully connected layer network to learn the dependencies between channels to generate channel weight vectors, and fuse the channel weight vectors with the feature tensor to obtain a weighted feature vector. Step 5: Input the weighted feature vector into the output mapping layer to calculate the additional damping control signal; Step 6: Superimpose the additional damping control signal onto the voltage reference setting value of the excitation regulator to adjust the excitation current of the generator.
2. The lightweight TCN broadband oscillation suppression method for electric excitation systems according to claim 1, characterized in that, The preprocessing described in step 2 specifically includes: using a sliding window to extract data segments and performing online normalization.
3. The lightweight TCN broadband oscillation suppression method for electric excitation systems according to claim 1, characterized in that, The lightweight causal TCN feature extraction layer consists of several cascaded causal dilated residual blocks, used to extract multi-scale temporal features of the excitation system state variables; The frequency-adaptive channel attention layer dynamically assigns weights based on the importance of the feature channels.
4. The lightweight TCN broadband oscillation suppression method for electric excitation systems according to claim 1, characterized in that, The fully connected layer network in step 4 consists of two fully connected layers, and the fusion operation is to multiply the channel weight vector with the feature tensor channel by channel.
5. A lightweight TCN broadband oscillation suppression method for power excitation systems according to claim 1, characterized in that, In step 6, the trainable parameters of the network model are fine-tuned and updated online using the gradient descent method according to the damping objective function.
6. A lightweight TCN broadband oscillation suppression method for electric excitation systems according to claim 3, characterized in that, The causal dilated residual block employs a depthwise separable convolutional architecture, which includes sequentially cascaded channel-wise convolutional sub-operations and pointwise convolutional sub-operations.
7. A lightweight TCN broadband oscillation suppression method for power excitation systems according to claim 1, characterized in that, The calculation process of the frequency-adaptive channel attention layer in step 4 includes the following: Step 4.1: Perform global information compression. Compress the feature tensor along the time dimension using global average pooling to obtain channel statistics. Step 4.2: Perform channel dependency learning, using two fully connected layers and activation functions to learn the dependencies between channels and generate channel weight vectors; Step 4.3: Perform feature recalibration by multiplying the channel weight vector with the original feature tensor channel by channel to obtain the frequency-weighted feature vector.
8. A lightweight TCN broadband oscillation suppression method for power excitation systems according to claim 2, characterized in that, The normalization process described in step 2 uses an online Z-score standardization method, where the mean and standard deviation within a sliding time window are dynamically adjusted to adjust the input data.
9. A lightweight TCN broadband oscillation suppression method for power excitation systems according to claim 1, characterized in that, The operating status data includes the change in generator active power and the deviation of generator rotor angular velocity.