A dynamic prediction method for a coal-fired power plant SCR denitration system based on a parameter identification mechanism-data hybrid model

Abstract

The application provides a dynamic prediction method for a coal-fired power plant SCR denitration system based on a parameter identification mechanism-data hybrid model, and relates to the technical field of intelligent modeling and prediction. x The mechanism prediction value of the NO x The application establishes an SCR dynamic mechanism model, completes parameter identification based on historical operation data, obtains the mechanism prediction value of the NO x Concentration and ammonia escape amount, constructs a series model of liquid neural network and long short-term memory network to learn the mechanism model prediction deviation, weights and fuses the mechanism prediction value and the prediction deviation, forms a hybrid model with physical consistency and dynamic adaptability, and realizes high-precision and stable prediction of the SCR outlet NO x Emission and ammonia escape amount under variable load conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent modeling and prediction technology, and in particular to a dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification. Background Technology

[0002] The current energy structure is rapidly transitioning towards a clean and low-carbon model, with the proportion of new energy installed capacity continuing to rise. The coupling relationship between power generation, grid, load, and storage is becoming increasingly complex, and coal-fired power units are gradually transforming from traditional baseload power sources to ancillary service power sources such as peak shaving and frequency regulation. Under peak and frequency changing conditions, unit loads often fluctuate at high frequencies within a wide range, causing drastic nonlinear changes in flue gas parameters (such as flow rate, temperature, and pollutant concentration). This dynamic condition leads to frequent instantaneous exceedances of pollutant emissions from traditional SCR denitrification systems; energy consumption runaway forces the system to add excessive amounts of reducing agents (such as liquid ammonia and urea) for extended periods, resulting in environmental pollution and serious resource waste. Therefore, it is necessary to establish an intelligent predictive model for SCR denitrification systems adapted to peak and frequency changing conditions to provide reliable guidance for subsequent intelligent optimization control systems.

[0003] Currently, SCR denitrification systems are widely used in coal-fired power plants. However, because the process is easily affected by boiler load fluctuations and dynamic changes in flue gas parameters, NO2 levels can rise. x Excessive emissions and increased ammonia slip. Existing methods for predicting SCR denitrification systems mainly fall into two categories: one is based on kinetic models constructed from the physicochemical laws of the SCR reaction, but these models contain multiple unknown parameters (such as reaction rate constants and activation energies) that are difficult to measure directly through experiments, leading to inaccurate parameter identification and decreased prediction accuracy under varying load conditions; the other is data-driven modeling, which directly maps the input-output relationship based on algorithms such as neural networks and machine learning, but purely data-driven models lack physical meaning support, have poor generalization ability, are insufficiently adaptable to unseen varying load scenarios, and cannot explain the physical logic of the prediction results; in addition, varying load processes exhibit both transient fluctuations and long-term operational dependence, making it difficult for a single model to simultaneously consider both transient response speed and long-term prediction stability. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification. This method constructs a physically consistent and dynamically adaptable hybrid model of the SCR denitrification system through "mechanism model parameter identification, deviation learning data model, and weighted fusion prediction," thereby achieving NO... x High-precision and stable prediction of ammonia escape under variable load conditions.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification, comprising:

[0007] A SCR kinetic mechanism model was established based on ammonia adsorption reaction, ammonia desorption reaction and SCR redox reaction to describe the spatiotemporal evolution of ammonia coverage, ammonia concentration and nitrogen oxide concentration in the catalyst layer, thus obtaining the SCR kinetic mechanism model.

[0008] The SCR kinetic mechanism model is discretized using the forward Euler method, and the parameters of the SCR kinetic mechanism model are identified based on historical operating data using the least squares method. The historical operating data is then input into the identified SCR kinetic mechanism model to obtain the SCR outlet NO. x Mechanistic predictions of concentration and ammonia slip;

[0009] Based on the historical operating data, a cascaded model of liquid neural network and long short-term memory network is constructed. The operating input in the historical operating data and the mechanism prediction value at multiple historical moments are used as inputs, and the difference between the mechanism prediction value and the actual measurement value is used as outputs. Transient deviation characteristics and long-term deviation characteristics are learned to obtain the prediction deviation.

[0010] A mechanism-data hybrid model is constructed using the predicted mechanism value as the baseline and the predicted deviation as the compensation value. The predicted mechanism value and the predicted deviation are weighted and fused to output NO under variable load conditions. x Dynamic prediction results of emission concentration and ammonia slip.

[0011] Preferably, the SCR kinetic mechanism model includes a set of kinetic equations for ammonia adsorption, ammonia desorption, ammonia oxidation, and standard reduction reactions. These kinetic equations include dynamic equations for NH3 surface coverage, NH3 concentration, and NO concentration. x The concentration dynamic equation and the corresponding NH3 adsorption rate, NH3 desorption rate, NH3 oxidation rate and NO x The rate expression for the reduction reaction includes the pre-exponential factor for NH3 adsorption rate, the pre-exponential factor for NH3 adsorption activation energy, the pre-exponential factor for NH3 desorption rate, the pre-exponential factor for NH3 desorption activation energy, the pre-exponential factor for NH3 oxidation rate, the NH3 oxidation activation energy, and NO. x Reduction rate pre-exponential factor, NO x The reduction activation energy was determined, taking into account the effects of flue gas velocity, catalyst axial coordinate, reaction temperature, ideal gas constant, and catalyst NH3 adsorption capacity.

[0012] Preferably, the discretization method of the forward Euler method includes:

[0013] By setting a time step in the time dimension and an axial discrete step in the catalyst axis dimension, the continuous kinetic equation of the SCR kinetic mechanism model is converted into a discrete recursive form. Under given initial state conditions, the mechanism prediction value is obtained by iteratively calculating segment by segment along the catalyst axis.

[0014] Preferably, when identifying the parameters of the SCR kinetic mechanism model using the least squares method, historical operating data from the actual continuous operation of a coal-fired power plant is selected. The historical operating data covers the operating conditions from 50% of the unit load to full load and the flue gas temperature from 300°C to 400°C. The historical operating data is divided into a training set, a validation set, and a test set in a 7:2:1 ratio. An objective function is constructed with the goal of minimizing the sum of squared errors between the predicted and actual measured values. The parameters are iteratively updated using the Levenberg-Marquardt algorithm until the parameter update threshold and the objective function descent threshold are met. The parameters of the SCR kinetic mechanism model include: pre-exponential factor for NH3 adsorption rate, NH3 adsorption activation energy, pre-exponential factor for NH3 desorption rate, NH3 desorption activation energy, pre-exponential factor for NH3 oxidation rate, NH3 oxidation activation energy, and NO... x Reduction rate pre-exponential factor and NO x Reduction activation energy.

[0015] Preferably, the liquid neural network uses neuron membrane potential as a state variable. The update of neuron membrane potential is jointly determined by membrane capacitance, leakage conductance, leakage potential, sensory input, and synaptic input. The sensory input is obtained by weighting the running input in the historical running data with sensory input weights and input activation functions. The synaptic input is obtained by weighting the inter-neuron synaptic weights and synaptic activation functions. The hidden state output of the liquid neural network is obtained by solving the membrane potential ordinary differential equation.

[0016] Preferably, the Long Short-Term Memory (LSTM) network includes a forget gate, an input gate, candidate cell states, cell states, and an output gate; the forget gate updates the historical memory retention ratio of the cell state based on the current input and the hidden state at the previous time step; the input gate writes the current input into the cell state to form the updated cell state; the output gate generates the current hidden state output from the updated cell state to extract the long-term bias feature.

[0017] Preferably, the cascaded model is trained with the difference between the predicted value and the actual measured value of the mechanism as the training objective; the loss function of the cascaded model includes a prediction deviation fitting error term, a regularization term for constraining the model parameter size, and a physical consistency penalty term related to the ammonia injection ratio, which is used to penalize the deviation prediction trend that violates the physical laws of the SCR kinetic mechanism model.

[0018] Preferably, the weighted fusion method of the mechanism-data hybrid model is to separately apply NO... x Mechanism prediction value and NO x The weighted sum of prediction biases yields NO. x The dynamic prediction results are obtained by weighting and summing the predicted values ​​of ammonia escape mechanism and the prediction deviation of ammonia escape, respectively. The weighting coefficients are adaptively determined based on the error of the predicted value of the mechanism and the prediction error of the series model deviation under the corresponding operating conditions in the historical operating data.

[0019] Preferably, before constructing the cascade model, the historical operating data is preprocessed. The preprocessing includes missing value handling, unified sampling frequency, smoothing and denoising, and standardization. A historical window length is set, and the operating input within the historical window length and the corresponding mechanism prediction value are used to form the time-series input of the cascade model.

[0020] Preferably, the mechanism prediction value includes SCR outlet NO. x Mechanism prediction values ​​and SCR outlet ammonia slip mechanism prediction values, the prediction deviations including NO x The prediction bias and ammonia slip prediction bias are used as the basis for the mechanism-data hybrid model to form the corresponding dynamic prediction result by superimposing the prediction bias.

[0021] The present invention discloses the following technical effects:

[0022] This invention introduces a mechanism modeling strategy of "forward Euler discretization + least squares parameter identification," enabling the SCR kinetic mechanism model to automatically calibrate unmeasurable parameters such as the pre-exponential factor of key reaction rates and activation energy based on historical operating data. This fundamentally overcomes the shortcomings of traditional mechanism models, such as uncertain parameter sources and difficulty in reflecting actual operating conditions. The identified mechanism model, while maintaining the integrity of its physical structure, possesses higher applicability and interpretability, enabling NO... x The basic prediction process for ammonia escape has a solid physical consistency basis.

[0023] This invention constructs a data-driven structure that connects a liquid neural network and a long short-term memory network. Using mechanistic predictions as the benchmark and prediction bias as the learning object, the data-driven model learns only the shortcomings of the mechanistic model under rapid fluctuations and long-term accumulation conditions, avoiding the physical omissions and insufficient generalization caused by the full-scale fitting of pure data models. The liquid neural network represents transient dynamics, while the long short-term memory network extracts long-term dependencies, giving bias learning clear physical boundaries and stronger robustness, significantly improving the model's adaptability in complex operating scenarios.

[0024] The weighted fusion mechanism of mechanistic predictions and prediction biases proposed in this invention enables the hybrid model to achieve adaptive balance under dynamic operating conditions such as load changes, ammonia injection ratio adjustments, and flue gas temperature variations. When the mechanistic model is more reliable, its contribution is strengthened; when the bias model captures significant system deviations, the compensation effect is enhanced, thus overcoming the instability of traditional methods under dynamic conditions. Ultimately, this invention significantly improves the SCR outlet NOo. x The dynamic prediction accuracy of concentration and ammonia escape enables coal-fired units to achieve stable compliance and optimized ammonia injection while supporting peak shaving. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of the SCR denitrification system provided in an embodiment of the present invention;

[0028] Figure 3 A flowchart of least squares parameter identification provided for embodiments of the present invention;

[0029] Figure 4 This is an overall framework diagram of the hybrid prediction method provided in the embodiments of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] The purpose of this invention is to provide a dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification. By integrating the stability of an identifiable mechanism model with the flexibility of data-driven deviation compensation, a dynamic prediction system for SCR suitable for complex operating conditions is constructed, which significantly improves prediction accuracy and adaptability to operating conditions while ensuring physical rationality.

[0032] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification, comprising:

[0034] Step 100: Establish an SCR kinetic mechanism model based on ammonia adsorption reaction, ammonia desorption reaction and SCR redox reaction to describe the spatiotemporal evolution of ammonia coverage, ammonia concentration and nitrogen oxide concentration in the catalyst layer, and obtain the SCR kinetic mechanism model;

[0035] Step 200: Discretize the SCR dynamic mechanism model using the forward Euler method, and identify the parameters of the SCR dynamic mechanism model based on historical operating data using the least squares method. Input the historical operating data into the identified SCR dynamic mechanism model to obtain the SCR outlet NO. x Mechanistic predictions of concentration and ammonia slip;

[0036] Step 300: Construct a cascaded model of liquid neural network and long short-term memory network based on historical operation data. The model takes the operation input in the historical operation data and the mechanism prediction values ​​at multiple historical moments as input, and the difference between the mechanism prediction value and the actual measurement value as output. The model learns transient deviation characteristics and long-term deviation characteristics to obtain the prediction deviation.

[0037] Step 400: Construct a mechanism-data hybrid model using the mechanism prediction value as the baseline and the prediction deviation as the compensation value. Weight the mechanism prediction value and the prediction deviation to output the NO under varying load conditions. x Dynamic prediction results of emission concentration and ammonia slip.

[0038] like Figures 2 to 4 As shown, the technical approach of this embodiment is as follows:

[0039] Step 1: Based on ammonia adsorption, ammonia desorption, and SCR redox reaction, describe the spatiotemporal evolution of ammonia coverage, ammonia concentration, and nitrogen oxide concentration in the catalyst layer, and establish an SCR kinetic mechanism model.

[0040] First, a mechanistic model was established based on the reaction kinetics of the SCR denitrification system. The model includes ammonia surface coverage, ammonia concentration, and NO. x The dynamic equation for concentration:

[0041] (1)

[0042] (2)

[0043] (3)

[0044] in,

[0045] (4)

[0046] (5)

[0047] (6)

[0048] (7)

[0049] This model comprehensively considers the effects of temperature, concentration, and coverage on the reaction rate, and can accurately characterize the transient dynamics of the SCR denitrification system.

[0050] Step 2: Discretize the model using the forward Euler method. Based on historical data, use the least squares method to identify the parameters of the SCR kinetic mechanism model and establish a prediction model for nitrogen oxides and ammonia escape. The specific method is as follows:

[0051] The model discretized using the forward Euler method is as follows:

[0052] (8)

[0053] (9)

[0054] (10)

[0055] NO x Concentration prediction:

[0056] (11)

[0058] Ammonia escape prediction:

[0059] (12)

[0061] The identified parameter set is β= , The input vector is used, and the initial conditions of all state variables are set by actual operating condition measurements, with preset parameter constraints. Historical data from continuous operation of the power plant are selected, covering operating conditions from 50% load to full load, with flue gas temperatures ranging from 300 to 400°C, ensuring the data includes the main characteristic operating conditions of the nonlinear model. The dataset is divided into a 7:2:1 ratio for training, validation, and test sets. The objective function is constructed with the goal of minimizing the sum of squared errors between the model's predicted and actual values:

[0062] (13)

[0063] in, , For model NO x and predicted ammonia escape amount, , NO x The true value of ammonia slip is obtained by updating parameters using the Levenberg-Marquardt algorithm and solving iteratively. .

[0064] First, set the initial parameter value β0, the parameter update threshold ɛ1, the objective function descent threshold ɛ2, the damping factor λ, and the maximum number of iterations I; then substitute the current parameter β0. k Through axial piecewise iterative calculation Construct the error vector For β k The corresponding predicted output matrix; calculate the Jacobian matrix J. k Solve for the parameter update amount Δβ k Iterative adjustment of Δβ k Finally, the parameters are verified and optimized. See the appendix for detailed operating steps. Figure 2 The flowchart for updating parameters using the least squares method is shown.

[0065] Step 3: Construct an LNN+LSTM cascaded model based on historical operating data to compensate for the bias in the mechanism prediction model.

[0066] The essence of a liquid neural network is to simulate the electrophysiological characteristics of biological neurons. The logic of neuron state updates is the rate of change of membrane potential over time, which is jointly determined by the input signal, synaptic interactions, and membrane properties. This is achieved by solving ordinary differential equations (ODEs). The model's output is consistent with the neuron's state, directly reflecting the neuron's dynamic response. Its membrane potential update is as follows:

[0067] (14)

[0068] in, It is a film capacitor. It is leakage conductivity. It is leakage potential. This is the membrane potential at the previous moment. The total input numerator is the sum of the sensory input numerator and the synaptic input numerator, and the total input denominator is the sum of the sensory input denominator and the synaptic input denominator. The formula is as follows:

[0069] (15)

[0070] (16)

[0071] in, These are the sensory input weight, the input activation function, and the inversion potential corresponding to the sensory input, respectively. These are the synaptic weights between neurons, the synaptic activation function, and the inversion potential corresponding to the synapse, respectively.

[0072] The core of a long short-term memory (LSTM) neural network is to simulate the dynamic management of "memory." For time-series data of SCR denitrification reactions, it is necessary to selectively remember important long-term information (such as the accumulation trend of ammonia slip under sustained low load), forget irrelevant short-term noise (such as instantaneous fluctuations in sensors), and update key new information at the current moment (such as NO during a sudden increase in load). x (Sudden changes in concentration), and long-term memories are stored through a cellular state. The neuronal renewal mechanism is as follows:

[0073] (17)

[0074] (18)

[0075] (19)

[0076] (20)

[0077] in, , These are the forget gate, input gate, candidate cell state, and cell state, respectively. These represent the current input and the hidden state from the previous moment, respectively. It is the sigmoid function. These are the input weights, hidden state weights, and bias, respectively. The output gate controls the extraction of information from the cell state to the hidden state.

[0078] (twenty one)

[0079] (twenty two)

[0080] in, For output gate, The cell state is compressed to (-1,1) to avoid excessively large output values.

[0081] The goal of data-driven models is to learn the predicted values ​​of mechanistic models. Compared with the true value The deviation between them Construct a cascaded data model. First, use an LNN to capture transient dynamics. Then, feed the hidden states of the LNN along with the input into an LSTM to learn long-term dependencies, and output... .

[0082] The specific steps are as follows: First, preprocess the data collected by the factory's DCS system, including handling missing values, unifying the sampling frequency, smoothing and denoising, and standardization; input the data and select the historical window length; construct and initialize a hybrid LNN and LSTM neural network model; train the model and define the training objective function as shown in the following formula:

[0083] (twenty one)

[0084] A regularization term is introduced to improve the model's generalization ability and convergence stability, while a physical consistency penalty term is introduced to penalize predicted trends that violate physical laws. This represents all trainable parameters of the model, including the weights and biases of the LNN layer, LSTM layer, and output layer. The model represents the The output of the time-of-flight deviation prediction; These are regularization weight coefficients. The results are predicted by the mechanism and data fusion model, where R is the ammonia injection ratio. These are the physical constraint weighting coefficients, used to balance physical consistency and prediction accuracy.

[0085] Step 4: Weighted fusion of the prediction results from the mechanism model and the prediction bias from the data model, and output the actual prediction results of the system.

[0086] Based on the denitrification reaction kinetic equation and parameter identification results, the NO at the SCR system outlet was analyzed. x The concentration and ammonia slip are predicted to obtain the mechanism prediction output. and The data-driven model, composed of a concatenated LNN-LSTM network, takes historical input sequences and mechanistic model predictions as inputs and outputs the mechanistic model prediction bias. and By weighted and fused the results of the two prediction branches, the final output of the system is obtained as follows:

[0087] (twenty two)

[0088] (twenty three)

[0089] in, , is a dynamic weighting coefficient that reflects the reliability of the mechanistic model and the data model under the current operating conditions. When the prediction error of the mechanistic model is small and the confidence level of the data model is low... When the value approaches 0, the system mainly relies on the output of the mechanistic model; when the error of the mechanistic model is large but the data model performs well... When the value approaches 1, the system relies more on the correction results of the data-driven model. This weighted fusion strategy enables the present invention to adaptively adjust the prediction source according to changes in operating conditions, balancing stability and accuracy.

[0090] The beneficial effects of this invention are as follows:

[0091] This invention is based on the collaborative concept of "mechanism modeling to ensure physical consistency and data-driven compensation and correction of prediction biases," and targets NO2 in SCR denitrification systems of coal-fired power plants under variable load conditions. x To address the challenge of predicting emissions and ammonia slip, a coupled architecture of "discrete mechanism model + LNN-LSTM hybrid compensation model" is constructed. The core principle is as follows: Starting from the physicochemical essence of SCR denitrification, the mechanism model is based on NH3 adsorption and desorption kinetics and the SCR redox reaction laws. This is achieved by constructing a model that includes NH3 surface coverage, NO... x The dynamic equations (partial differential equations) for the concentrations of NH3 and other components characterize the continuous changes in the reaction along the catalyst axis over time. Since continuous partial differential equations cannot be directly adapted to discrete sampling data from power plants, a forward Euler method is used to discretize them: the time dimension is split into discrete time intervals Δt, and the spatial dimension is divided into uniform segments Δz along the catalyst axis length L, with each segment serving as an independent computational unit. The current state is iteratively calculated using the "previous time / previous segment state" formula, transforming the complex continuous model into a real-time solvable algebraic equation while preserving the physical logic of the reaction. To address the difficulty in determining unknown parameters in the mechanistic model, the least squares method is introduced: a linearized observation equation is constructed based on historical operating data, aiming to minimize the sum of squared errors between model predictions and measured values. This solves for the unknown parameter vector, giving the mechanistic model basic predictive capabilities. Considering the susceptibility of mechanistic models to transient fluctuations and long-term cumulative biases under varying loads, a hybrid LNN-LSTM data-driven model is designed: the LNN subnetwork captures short-term transient biases through fully connected layers, while the LSTM subnetwork learns long-term operational dependencies using memory units; the input includes the mechanistic model's input parameters and historical operational information, which can smooth measurement noise, compensate for the mechanistic model's lag, and output the mechanistic model's prediction bias. Finally, through the superposition logic of "mechanistic model output + data-driven bias compensation," a balance between physical rationality and prediction accuracy is achieved, meeting the need for accurate prediction under varying load conditions.

[0092] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0093] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification, characterized in that, include: A SCR kinetic mechanism model was established based on ammonia adsorption reaction, ammonia desorption reaction and SCR redox reaction to describe the spatiotemporal evolution of ammonia coverage, ammonia concentration and nitrogen oxide concentration in the catalyst layer, thus obtaining the SCR kinetic mechanism model. The SCR kinetic mechanism model is discretized using the forward Euler method, and the parameters of the SCR kinetic mechanism model are identified based on historical operating data using the least squares method. The historical operating data is then input into the identified SCR kinetic mechanism model to obtain the SCR outlet NO. x Mechanistic predictions of concentration and ammonia slip; Based on the historical operating data, a cascaded model of liquid neural network and long short-term memory network is constructed. The operating input in the historical operating data and the mechanism prediction value at multiple historical moments are used as inputs, and the difference between the mechanism prediction value and the actual measurement value is used as outputs. Transient deviation characteristics and long-term deviation characteristics are learned to obtain the prediction deviation. A mechanism-data hybrid model is constructed using the predicted mechanism value as the baseline and the predicted deviation as the compensation value. The predicted mechanism value and the predicted deviation are weighted and fused to output NO under variable load conditions. x Dynamic prediction results of emission concentration and ammonia slip.

2. The dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification as described in claim 1, characterized in that, The SCR kinetic mechanism model includes a set of kinetic equations for ammonia adsorption, ammonia desorption, ammonia oxidation, and standard reduction reactions. These equations include dynamic equations for NH3 surface coverage, NH3 concentration, and NO concentration. x The concentration dynamic equation and the corresponding NH3 adsorption rate, NH3 desorption rate, NH3 oxidation rate and NO x The rate expression for the reduction reaction includes the pre-exponential factor for NH3 adsorption rate, the pre-exponential factor for NH3 adsorption activation energy, the pre-exponential factor for NH3 desorption rate, the pre-exponential factor for NH3 desorption activation energy, the pre-exponential factor for NH3 oxidation rate, the NH3 oxidation activation energy, and NO. x Reduction rate pre-exponential factor, NO x The reduction activation energy was determined, taking into account the effects of flue gas velocity, catalyst axial coordinate, reaction temperature, ideal gas constant, and catalyst NH3 adsorption capacity.

3. The dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification as described in claim 1, characterized in that, The discretization methods of the forward Euler method include: By setting a time step in the time dimension and an axial discrete step in the catalyst axis dimension, the continuous kinetic equation of the SCR kinetic mechanism model is converted into a discrete recursive form. Under given initial state conditions, the mechanism prediction value is obtained by iteratively calculating segment by segment along the catalyst axis.

4. The dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification as described in claim 1, characterized in that, When the least squares method is used to identify the parameters of the SCR kinetic mechanism model, the historical operating data of the actual continuous operation of the coal-fired power plant is selected. The operating conditions of the historical operating data cover the unit load from 50% to full load and the flue gas temperature from 300 degrees Celsius to 400 degrees Celsius. The historical operating data is divided into training set, validation set and test set in a ratio of 7:2:

1. The objective function is constructed with the goal of minimizing the sum of squared errors between the mechanism prediction value and the actual measurement value. The parameters are iteratively updated based on the Levenberg-Marquardt algorithm until the parameter update threshold and the objective function descent threshold are met. The parameters of the SCR kinetic mechanism model include: pre-exponential factor of NH3 adsorption rate, NH3 adsorption activation energy, pre-exponential factor of NH3 desorption rate, NH3 desorption activation energy, pre-exponential factor of NH3 oxidation rate, NH3 oxidation activation energy, and NO. x Reduction rate pre-exponential factor and NO x Reduction activation energy.

5. The dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification as described in claim 1, characterized in that, The liquid neural network uses neuron membrane potential as a state variable. The update of neuron membrane potential is jointly determined by membrane capacitance, leakage conductance, leakage potential, sensory input, and synaptic input. The sensory input is obtained by weighting the running input in the historical running data with sensory input weights and input activation functions. The synaptic input is obtained by weighting the inter-neuron synaptic weights and synaptic activation functions. The hidden state output of the liquid neural network is obtained by solving the membrane potential ordinary differential equation.

6. The dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification as described in claim 1, characterized in that, The Long Short-Term Memory (LSTM) network includes a forget gate, an input gate, candidate cell states, cell states, and an output gate. The forget gate updates the historical memory retention ratio of the cell state based on the current input and the hidden state at the previous time step. The input gate writes the current input into the cell state to form the updated cell state. The output gate generates the current hidden state output from the updated cell state to extract the long-term bias feature.

7. The dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification as described in claim 1, characterized in that, The cascaded model is trained with the difference between the predicted value and the actual measured value of the mechanism as the training objective; the loss function of the cascaded model includes a prediction deviation fitting error term, a regularization term for constraining the model parameter size, and a physical consistency penalty term related to the ammonia injection ratio, which is used to penalize the deviation prediction trend that violates the physical laws of the SCR kinetic mechanism model.

8. The dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification as described in claim 1, characterized in that, The weighted fusion method of the mechanism-data hybrid model is to separately apply NO... x Mechanism prediction value and NO x The weighted sum of prediction biases yields NO. x The dynamic prediction results are obtained by weighting and summing the predicted values ​​of ammonia escape mechanism and the prediction deviation of ammonia escape, respectively. The weighting coefficients are adaptively determined based on the error of the mechanism prediction value and the error of the series model deviation prediction under the corresponding operating conditions in the historical operating data.

9. The dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification as described in claim 1, characterized in that, Before constructing the cascaded model, the historical operating data is preprocessed. The preprocessing includes missing value handling, unified sampling frequency, smoothing and denoising, and standardization. The historical window length is set, and the operating input within the historical window length and the corresponding mechanism prediction value are used to form the time-series input of the cascaded model.

10. The dynamic prediction method for SCR denitrification systems in coal-fired power plants based on a mechanism-data hybrid model with parameter identification as described in claim 1, characterized in that, The mechanism prediction value includes SCR export NO. x Mechanism prediction values ​​and SCR outlet ammonia slip mechanism prediction values, the prediction deviations including NO x The prediction bias and ammonia slip prediction bias are used as the basis for the mechanism-data hybrid model to form the corresponding dynamic prediction result by superimposing the prediction bias.

Images