A regularization CNN-GRU control method for physical mechanism of a thermal power plant excitation system

Abstract

This invention provides a regularized CNN-GRU control method based on the physical mechanism of the excitation system in thermal power plants, belonging to the field of excitation control technology. The method mainly includes: constructing an input feature vector and normalizing it; mapping the normalized time-series features to a multi-branch dilated convolutional operator layer composed of multiple sets of parallel operators to generate a global spatial feature vector; feeding the global spatial feature vector into a GRU network embedding a physical consistency gating unit; embedding the second-order rotor motion equation as a gradient constraint factor into the loss function; and linearly superimposing the discretized PID main path output with the network output, generating the final excitation voltage command through a limiting operator. This invention can force the control command to return to the feasible region of physical laws under small sample and strong random disturbance environments, improving the transient damping support efficiency and adaptive control accuracy of the excitation circuit.

Description

Technical Field

[0001] This invention mainly relates to the field of excitation control technology, specifically a regularized CNN-GRU control method based on the physical mechanism of the excitation system in thermal power plants. Background Technology

[0002] With the deepening of the construction of new power systems, thermal power units frequently participate in deep peak shaving and operate under low-load conditions for extended periods. In low-load regions, the internal electromagnetic characteristics of synchronous generators exhibit strong nonlinearity and time-varying behavior, with the magnetic circuit saturation level drifting drastically with the operating point. Traditional excitation regulators are mainly based on proportional-integral-derivative (PID) control operators with fixed parameters; however, PID operators are essentially linearized products for a specific operating point. When the system experiences large-scale fluctuations across operating conditions or encounters short-circuit faults, fixed-parameter PID often fails to provide sufficient damping ratios, leading to low-frequency oscillations or even loss of transient stability in the unit.

[0003] To overcome the limitations of fixed-parameter control, existing technologies attempt to introduce numerical mapping models for nonlinear compensation of excitation regulation. However, black-box mapping structures based purely on data fitting lack semantic awareness of the electromagnetic transient mechanisms inside synchronous generators. In practical engineering applications, when such models encounter atypical disturbances such as grid short circuits or asynchronous grid connection, the generated regulation sequences often produce invalid oscillations that violate the physical semantics of the oscillation equation, resulting in severe distortion of the control output in the physical dimension, and even inducing safety risks such as excitation winding overvoltage or rotor loss of synchronization.

[0004] Therefore, how to maintain the powerful multi-condition adaptive adjustment capability of the nonlinear mapping operator while forcibly anchoring the generator rotor motion mechanism as a "hard constraint" within the control logic, so that the adjustment pulse strictly follows the physical law mapping on the entire time axis, has become a key bottleneck that urgently needs to be overcome in the transformation of modern thermal power plant excitation control systems towards intelligence and high safety. Summary of the Invention

[0005] To address the shortcomings of current technologies, this invention combines existing technologies and, based on practical applications, provides a regularized CNN-GRU control method oriented towards the physical mechanism of thermal power plant excitation systems. This method explicitly embeds generator kinematic differential operators into a nonlinear time-series mapping structure, aiming to solve the core technical bottlenecks of insufficient transient damping and control law distortion in the deep peak-shaving process of thermal power units under the background of large-scale new energy grid connection.

[0006] The technical solution of the present invention is as follows:

[0007] A regularized CNN-GRU control method based on the physical mechanism of the excitation system in thermal power plants includes the following steps:

[0008] Step 1: Data acquisition. Real-time acquisition of the synchronous generator state sequence. Based on the generator terminal voltage, active power, angular velocity deviation and rotor angle, construct the input feature vector and normalize the input feature vector.

[0009] Step 2: Multi-branch dilated convolutional CNN spatial feature extraction, the normalized temporal features are mapped to a multi-branch dilated convolutional operator layer composed of multiple sets of parallel operators, and a global spatial feature vector is generated by connecting the operators;

[0010] Step 3: GRU temporal evolution mapping with physical consistency gate PC-Gate, which feeds the global spatial feature vector into the GRU network with embedded physical consistency gate unit to correct the physical validity of the GRU temporal evolution state;

[0011] Step 4: Construct a physical mechanism regularization loss function, embed the second-order rotor motion equation as a gradient constraint factor into the loss function, and dynamically correct the hidden layer state of the network according to the real-time physical mechanism residual to achieve regularization adjustment of the model weights;

[0012] Step 5: Dual-channel signal superposition and execution. The output of the discretized PID main path and the network output are linearly superimposed, and the final excitation voltage command is generated by the limiting operator.

[0013] Furthermore, in step 1, the input feature vector is mapped to using the maxima-mina mapping operator. The interval mitigates the risk of gradient vanishing and eliminates the influence of dimensions on gradient calculation.

[0014] Furthermore, in step 2, the multi-branch dilated convolutional operator layer dynamically constructs multi-scale spatiotemporal receptive fields by configuring differentiated dilation rates, and performs causal zero-filling operations on each convolutional branch.

[0015] Furthermore, the output tensor of the k-th branch dilated convolution operator layer The expression is:

[0016]

[0017] in, Indicates the first The void ratio of the branch; L represents the temporal index of the convolution kernel; L represents the kernel length. for Branch weights The corresponding weight matrix; represents the bias vector, m represents the number of convolution output channels, and n represents the number of feature dimensions; The expression is , This represents the input feature vector.

[0018] Furthermore, the specific evolutionary logic in step 3 is as follows:

[0019] Step 3.1: Update the door ;

[0020] Step 3.2: Reset the door ;

[0021] Step 3.3: Generate Physical Consistency Gating Activation Values ;

[0022] Step 3.4: Calculate the candidate hidden state ;

[0023] Step 3.5: Calculate the final hidden state ;

[0024] In the formula, For the Sigmoid operator, the formula is: , For input scalars, It is a natural constant; Here, is the Hadamard product operator, representing element-wise multiplication; tanh is the activation function operator, with the formula: ; These represent the weight matrices for the update gate, reset gate, and candidate hidden state, respectively. These represent the bias vectors for the update gate, reset gate, and candidate state corresponding gating, respectively. Represents physical information mapping parameters; Represents the real-time physical mechanism residual; ; This represents the generated global spatial feature vector.

[0025] Furthermore, the physical consistency gate PC-Gate uses real-time physical mechanism residuals. The driver, expressed as follows:

[0026]

[0027] in, Indicates the moment of inertia coefficient of the generator rotor; Indicates the damping coefficient; Indicates the current The mechanical power input to the prime mover at any given moment; Indicates the current The electromagnetic power output by the generator at any given time; This indicates the discrete sampling period of the controller. This indicates the deviation in angular velocity.

[0028] Furthermore, in step 4, the total loss function is defined. Mean square error term With physical regularization term The weighted sum is expressed as follows:

[0029]

[0030] Mean square error term The expression is as follows:

[0031]

[0032] in, Indicates the physical constraint weight coefficient; N represents the total number of training samples; Indicates the sample index; Indicates the first The expert PID algorithm reference output for each sample; This indicates samples generated by network secondary pathways. Residual compensation output.

[0033] Furthermore, the physical regularization term The continuously differentiable operator based on the synchronous generator rotor motion equation is constructed as follows:

[0034]

[0035] Indicates the moment of inertia coefficient of the generator rotor; Indicates the damping coefficient; Indicates the current The mechanical power input to the prime mover at any given moment; Indicates the current The electromagnetic power output by the generator at any given time; This indicates the discrete sampling period of the controller. Indicates angular velocity deviation. Indicates the rotor angle.

[0036] Furthermore, in step 5, the output of the discretized PID main path is linearly superimposed with the network output, and then subjected to a limiting operator. Generate final excitation voltage command ;

[0037] in, These represent the maximum and minimum amplitude limits of the excitation regulator output, respectively.

[0038] Furthermore, the excitation voltage command is sent to the drive and pulse trigger circuit module, which converts the digital control quantity into a phase-shift trigger pulse with a specific conduction angle to drive the thyristor rectifier unit to perform power conversion and ultimately regulate the excitation voltage of the synchronous generator set.

[0039] The beneficial effects of this invention are:

[0040] To address the problem of excitation voltage instability caused by the strong nonlinear characteristics of synchronous generators under deep peak-shaving conditions, this invention implements a topology architecture that coordinates a "proportional-integral-differential main path" and a "physical mechanism regularization network auxiliary path." The auxiliary path employs an improved CNN-GRU network structure, combining a convolutional neural network (CNN) and a gated recurrent unit (GRU). A multi-branch dilated convolutional CNN extracts spatial coupling information of generator operating characteristics, and a Physical Consistency Gate (PC-Gate) unit is used to correct the physical validity of the GRU's temporal evolution state. By embedding the second-order rotor motion equation as a gradient constraint factor into the loss function, and based on the real-time physical mechanism residual... The invention dynamically corrects the hidden layer state of the network to achieve regularized adjustment of model weights. Under small sample and strong random disturbance environments, it can force the control command to return to the feasible region of physical laws, thereby improving the transient damping support efficiency and adaptive control accuracy of the excitation circuit. Attached Figure Description

[0041] Figure 1 This is the overall logic flow diagram of the algorithm in this invention;

[0042] Step 1 describes the real-time operating input vector of the synchronous generator. Specifically, this includes terminal voltage. Electromagnetic power Angular velocity deviation and rotor angle The original four-dimensional signal was normalized; step 2 describes the parallel extraction of spatial features by the three-branch dilated convolutional layer, processing transient, oscillatory, and global features respectively; step 3 is the core temporal mapping layer, demonstrating how the Physical Consistency Gating (PC-Gate) intervenes in the GRU state; step 4 illustrates the total loss function in offline training. Physical constraints on the weights; step 5 demonstrates the generation of the final excitation command after the dual-channel superposition. The execution process.

[0043] Figure 2 This is a CNN-GRU cascade diagram of the present invention;

[0044] This figure depicts the tensor evolution topology of the neural network, where the input layer receives the normalized terminal voltage. Electromagnetic power Angular velocity deviation and rotor angle A four-dimensional time-series signal is then fed in parallel into three signals with different void ratios. The system employs a one-dimensional convolutional branch; each branch extracts multi-scale features through receptive fields of different spans, and generates a high-dimensional fused global spatial feature vector through a concatenation operator (Concat). The core processing mechanism is based on an embedded physical consistency gating activation value. The GRU, corresponding to Figure 1 Step 3 in the diagram illustrates the physical mechanism of the residual. As an independent external stimulus signal, it is directly injected into the activation function of the PC-Gate. In the middle, and then the candidate hidden state. Element-by-element modulation is performed; the final structure maps the timing mapping result to the linear output layer to generate a compensation voltage signal.

[0045] Figure 3 This is the gradient feedback loop diagram based on physical mechanism regularization of the present invention;

[0046] The diagram shows Figure 1 Step 4 involves the logic for adjusting network parameters using the loss function; the input to the feedback loop is the network's current parameters. The predicted compensation value generated below; the logical connection lies in the construction of two parallel loss evaluation paths: the first path is based on the mean squared error term of the reference label. The second step involves substituting the network output into the second-order rotor motion equation of the synchronous generator. The physical regularization term generated later The feedback output is a hybrid gradient signal synthesized through an automatic differentiation mechanism. This signal corrects the weights of the convolution kernel and the gate unit through the backpropagation algorithm, forcing the model space to converge to a manifold region that conforms to the physical laws of the power system.

[0047] Figure 4 This is a hardware closed-loop control topology diagram of the thermal power plant excitation system of the present invention;

[0048] This diagram illustrates the collaborative interaction link between the algorithm logic and the industrial control hardware; the input terminal captures the pre-processed terminal voltage via a voltage / current transformer array. Electromagnetic power Angular velocity deviation and rotor angle The analog signal stream is quantized into a digital timing sequence by an analog-to-digital converter (ADC) module, and then sent to a digital processing unit with DSP / FPGA as its core. Figure 1 The superposition logic in step 5 is shown in the diagram as a parallel output mechanism of the PID unit and the CNN-GRU model, which ultimately generates the excitation voltage command. The command is sent to the drive and pulse trigger circuit module, which converts the digital control quantity into a phase-shift trigger pulse with a specific conduction angle to drive the "thyristor rectifier unit" to perform power conversion, and finally adjust the excitation voltage of the synchronous generator set to form a complete hardware closed-loop control chain. Detailed Implementation

[0049] The present invention will be further described in conjunction with the accompanying drawings and 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.

[0050] This embodiment provides a regularized CNN-GRU control method based on the physical mechanism of the excitation system in a thermal power plant. This embodiment takes a 600MW large-scale synchronous generator unit in a thermal power plant as the research object; it utilizes a dual-channel residual compensation structure (such as...) Figure 1 The discrete PID main path and network auxiliary path parallel architecture shown realizes precise control of excitation voltage under all operating conditions; the detailed steps are as follows.

[0051] Step 1: Real-time acquisition of synchronous generator state sequences. Constructing input feature vectors. Among them, the number of feature dimensions subscript This indicates that time-indexing is used as the input feature vector. Including terminal voltage Active power Angular velocity deviation and rotor angle Terminal voltage The unit is volts. Active power The unit is watts. Angular velocity deviation The unit is Rotor angle The unit is Set the data usage period to [period]. The input feature vector is transformed by the maxima-mina mapping operator. Mapped to The interval mitigates the risk of gradient vanishing and eliminates the influence of dimensions on gradient calculation.

[0052] Step 2: Multi-branch dilated convolution spatial feature extraction. The normalized temporal features are mapped to a multi-branch dilated convolution operator layer composed of multiple sets of parallel operators. This operator layer is configured with differentiated dilation rates. To achieve dynamic construction of multi-scale spatiotemporal receptive fields, and to eliminate feature map stride mismatch caused by different expansion intervals, causal zero-padding is performed on each convolutional branch to ensure strict alignment of temporal causal relationships.

[0053] like Figure 2 As shown, the normalized sequence is fed into a dilated convolutional layer with three parallel branches; the kernel length is set. Number of output channels Branches 1, 2, and 3 respectively adopt void ratio These are used to extract electromagnetic transient features with shorter, normal, and longer periods, respectively.

[0054] To ensure consistent output tensor dimensions under different voiding rates, each branch performs a causal zero-filling operation. Branch output Calculate according to the formula:

[0055]

[0056] Among them, the Branch void ratio Number of convolution output channels Convolutional kernel temporal index The unit is the number of kernels; kernel length Dimensionless; for Branch weights The corresponding weight matrix and bias vector ; Linear rectifier unit The expression is , dimensionless.

[0057] Generate global spatial feature vectors using connection operators. ;in, , dimensionless.

[0058] Step 3: GRU temporal evolution mapping with Physical Consistency Gate (PC-Gate). For example... Figure 2 As shown, The data is fed into a GRU network with an embedded physical consistency gating unit; the specific steps are as follows:

[0059] Step 3.1: Update the door ;

[0060] Step 3.2: Reset the door The reset and update gates are jointly responsible for filtering historical information.

[0061] Step 3.3: Generate Physical Consistency Gating Activation Values ;

[0062] Step 3.4: Calculate the candidate hidden state ;

[0063] Step 3.5: Calculate the final hidden state ;pass The modulation effect occurs when the system operating state violates the oscillation equation (such as residuals). When (increase), The value of decreases, thereby inhibiting the network's exploration of the non-physical understanding space;

[0064] Among them, the update door Reset the door Physical consistency gating activation value Candidate hidden state Ultimately hidden state Dimensionless; hidden layer state dimension The unit is one; For the Sigmoid operator, the formula is: , For input scalars, It is a natural constant, dimensionless; Hadamard's product operator represents element-wise multiplication and is dimensionless; tanh is the activation function operator, with the formula: Dimensionless; weight matrices for update gate, reset gate, and candidate states. Dimensionless; bias vectors corresponding to the update gate, reset gate, and candidate state gating. Dimensionless; physical information mapping parameters Dimensionless; Real-time physical mechanism residual Dimensionless; , dimensionless.

[0065] Preferably, the PC-Gate is composed of real-time physical mechanism residuals. The driving force, its mathematical expression is:

[0066]

[0067] Among them, the generator rotor moment of inertia coefficient The unit is Damping coefficient The unit is ;current Mechanical power input to the prime mover at any time The unit is watts. ;current Electromagnetic power output by the generator at any time Discrete sampling period of the controller The unit is seconds. .

[0068] Step 4: Construct the physical mechanism regularization loss function; define the total loss function. Mean square error term With physical regularization term Weighted sum:

[0069]

[0070] in, , Dimensionless; total number of training samples The unit is individuals; sample index The unit is "one"; the first The expert PID algorithm reference output for each sample The unit is volts. Samples generated by network secondary pathways Residual compensation output The unit is volts. .

[0071] Physical regularization term Based on the construction of a continuously differentiable operator according to the synchronous generator rotor motion equation, its analytical mapping expression is as follows:

[0072]

[0073] Among them, rotor angular acceleration Characterizing the second-order dynamic change of the synchronizing machine's power angle over time, in units of The second derivative of the rotor angle with respect to time is the angular acceleration. This is the second-order rotor motion equation, which constitutes the mechanical-electromagnetic energy mismatch term of the rotor motion; As a soft constraint term, the parameter tuning gradient of the computational network is introduced. By calculating the partial derivative of this term with respect to the internal weight matrix of the model, the mapping logic of the control operator is forced to converge within a manifold space that conforms to the unit's power angle stability criterion and the law of energy conservation. In this embodiment, the generator rotor moment of inertia coefficient... The unit is Damping coefficient The unit is ;current Mechanical power input to the prime mover at any time The unit is watts. ;current Electromagnetic power output by the generator at any time Discrete sampling period of the controller The unit is seconds. Physical weighting coefficient The weights of each layer are updated through backpropagation, forcing the model to follow Newton's second law.

[0074] Step 5: Dual-channel signal superposition and execution; such as Figure 4 As shown, the output of the discretized PID main path is... With network output Linear superposition, after amplitude limiting operator Generate final excitation voltage command Set maximum amplitude Minimum amplitude The unit is volts. The synthesized signal is applied to the thyristor rectifier unit via the drive circuit to achieve closed-loop excitation regulation of the synchronous generator.

Claims

1. A regularized CNN-GRU control method based on the physical mechanism of excitation systems in thermal power plants, characterized in that, Includes the following steps: Step 1: Data acquisition. Real-time acquisition of the synchronous generator state sequence. Based on the generator terminal voltage, active power, angular velocity deviation and rotor angle, construct the input feature vector and normalize the input feature vector. Step 2: Multi-branch dilated convolutional CNN spatial feature extraction, the normalized temporal features are mapped to a multi-branch dilated convolutional operator layer composed of multiple sets of parallel operators, and a global spatial feature vector is generated by connecting the operators; Step 3: GRU temporal evolution mapping with physical consistency gate PC-Gate, which feeds the global spatial feature vector into the GRU network with embedded physical consistency gate unit to correct the physical validity of the GRU temporal evolution state; Step 4: Construct a physical mechanism regularization loss function, embed the second-order rotor motion equation as a gradient constraint factor into the loss function, and dynamically correct the hidden layer state of the network according to the real-time physical mechanism residual to achieve regularization adjustment of the model weights; Step 5: Dual-channel signal superposition and execution. The output of the discretized PID main path and the network output are linearly superimposed, and the final excitation voltage command is generated by the limiting operator.

2. The regularized CNN-GRU control method for the physical mechanism of thermal power plant excitation systems according to claim 1, characterized in that, In step 1, the input feature vector is mapped to the minimum-maximum mapping operator. The interval mitigates the risk of gradient vanishing and eliminates the influence of dimensions on gradient calculation.

3. The regularized CNN-GRU control method for the physical mechanism of thermal power plant excitation systems according to claim 1, characterized in that, In step 2, the multi-branch dilated convolutional operator layer dynamically constructs multi-scale spatiotemporal receptive fields by configuring differentiated dilation rates, and performs causal zero-filling operations on each convolutional branch.

4. The regularized CNN-GRU control method for the physical mechanism of thermal power plant excitation systems according to claim 3, characterized in that, The output tensor of the k-th branch dilated convolution operator layer The expression is: ； in, Indicates the first The void ratio of the branch; L represents the temporal index of the convolution kernel; L represents the kernel length. for Branch weights The corresponding weight matrix; represents the bias vector, m represents the number of convolution output channels, and n represents the number of feature dimensions; The expression is , This represents the input feature vector.

5. The regularized CNN-GRU control method for the physical mechanism of thermal power plant excitation systems according to claim 1, characterized in that, The specific evolutionary logic in step 3 is as follows: Step 3.1: Update the door ; Step 3.2: Reset the door ; Step 3.3: Generate Physical Consistency Gating Activation Values ; Step 3.4: Calculate the candidate hidden state ; Step 3.5: Calculate the final hidden state ; In the formula, For the Sigmoid operator, the formula is: , For input scalars, It is a natural constant; Here, is the Hadamard product operator, representing element-wise multiplication; tanh is the activation function operator, with the formula: ; These represent the weight matrices for the update gate, reset gate, and candidate hidden state, respectively. These represent the bias vectors for the update gate, reset gate, and candidate state corresponding gating, respectively. Represents physical information mapping parameters; Represents the real-time physical mechanism residual; ; This represents the generated global spatial feature vector.

6. The regularized CNN-GRU control method for the physical mechanism of thermal power plant excitation systems according to claim 5, characterized in that, Physical consistency gated PC-Gate uses real-time physical mechanism residuals The driver, expressed as follows: ； in, Indicates the moment of inertia coefficient of the generator rotor; Indicates the damping coefficient; Indicates the current The mechanical power input to the prime mover at any given moment; Indicates the current The electromagnetic power output by the generator at any given time; This represents the discrete sampling period of the controller. This indicates the deviation in angular velocity.

7. The regularized CNN-GRU control method for the physical mechanism of thermal power plant excitation systems according to claim 1, characterized in that, In step 4, the total loss function is defined. Mean square error term With physical regularization term The weighted sum is expressed as follows: ； Mean square error term The expression is as follows: ； in, Indicates the physical constraint weight coefficient; N represents the total number of training samples; Indicates the sample index; Indicates the first The expert PID algorithm reference output for each sample; This indicates samples generated by network secondary pathways. Residual compensation output.

8. The regularized CNN-GRU control method for the physical mechanism of thermal power plant excitation systems according to claim 7, characterized in that, Physical regularization term The continuously differentiable operator based on the synchronous generator rotor motion equation is constructed as follows: ； Indicates the moment of inertia coefficient of the generator rotor; Indicates the damping coefficient; Indicates the current The mechanical power input to the prime mover at any given moment; Indicates the current The electromagnetic power output by the generator at any given time; This represents the discrete sampling period of the controller. Indicates angular velocity deviation. Indicates the rotor angle.

9. The regularized CNN-GRU control method for the physical mechanism of thermal power plant excitation systems according to claim 1, characterized in that, In step 5, the output of the discretized PID main path is linearly superimposed with the network output, and then subjected to a limiting operator. Generate final excitation voltage command ; in, These represent the maximum and minimum amplitude limits of the excitation regulator output, respectively.

10. The regularized CNN-GRU control method for the physical mechanism of thermal power plant excitation systems according to claim 9, characterized in that, Excitation voltage command The signal is fed into the drive and pulse trigger circuit module, which converts the digital control signal into a phase-shift trigger pulse with a specific conduction angle. This pulse drives the thyristor rectifier unit to perform power conversion and ultimately regulates the excitation voltage of the synchronous generator set.

Images