Method for predicting flue gas nitrogen oxide and ammonia escape by fusing time derivative

Abstract

The application discloses a flue gas nitrogen oxide and ammonia escape synergistic prediction control method fusing time derivatives, and belongs to the field of flue gas pollutant emission control. The method comprises the following steps: constructing an ammonia escape prediction model feature set; collecting historical feature data; constructing a prediction model of a convolution-Transformer-time derivative hybrid structure, training the prediction model based on the historical feature data; and obtaining a nitrogen oxide emission concentration prediction sequence and an ammonia escape concentration prediction sequence of future time steps by using the trained prediction model; designing a reward function based on the nitrogen oxide emission concentration prediction sequence and the ammonia escape concentration prediction sequence; and solving an ammonia water flow control strategy by using a reinforcement learning algorithm based on the reward function and taking the ammonia water flow as a decision variable. The application solves the problem that existing methods do not consider the direct correlation between NO x removal and ammonia escape, and do not establish a synergistic control relationship between NO x emission and NH3 escape.

Description

Technical Field

[0001] This invention belongs to the field of flue gas pollutant emission control, and particularly relates to a method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on time derivatives. Background Technology

[0002] Waste incineration is an important way to reduce and harmlessly treat municipal solid waste, but it produces a large amount of nitrogen oxides (NOx) during combustion. x To meet environmental emission standards, ammonia water needs to be injected into the flue gas pollutant purification system for denitrification. However, excessive or mismatched ammonia injection can easily lead to the escape of unreacted ammonia (NH3), causing secondary pollution. The emission level directly affects the environmental performance and operational economy of the incineration system. To meet increasingly stringent environmental emission standards, waste incineration plants typically need to invest significant manpower and resources in the operation, regulation, and emission control of the denitrification system.

[0003] Current methods for controlling flue gas pollutant emissions mainly rely on mechanistic modeling or empirical regulation. While mechanistic models offer some theoretical guidance, they struggle to accurately describe complex multiphase combustion processes, and their parameter calibration is cumbersome and lacks applicability. Traditional empirical control methods, on the other hand, depend on human experience and are ill-suited to handling complex and fluctuating operating conditions. With the development of artificial intelligence, researchers have begun to explore data-driven approaches to control NO emissions in incineration systems. x With NH3 emission prediction and control. Existing solutions disclose a method using Transformer networks for boiler NO3 emission prediction and control. x The prediction methods currently only focus on the prediction level and do not optimize the control logic. For example, an existing solution discloses a "real-time online prediction and control method for ammonia slip concentration," but its model has obvious static characteristics and can only be tested and fitted at three fixed load points (high, medium, and low) of the unit, failing to adapt to continuous changes in incineration conditions. Existing technologies mostly predict single pollutants independently, without considering NO. x The direct correlation between NO removal and ammonia slip was not established. x The synergistic control relationship between emissions and NH3 escape limits the effectiveness of the model under complex dynamic conditions.

[0004] Therefore, constructing deep learning prediction models for NOx emission concentration and ammonia slip based on data from the DCS and CEMS systems of incineration plants, and realizing intelligent optimization control of the denitrification process, has become an important direction for improving the environmental and economic performance of incineration systems. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, this invention provides a synergistic predictive control method for flue gas nitrogen oxides and ammonia slip based on fusion of time derivatives, which solves the problem that existing methods do not consider NO...x The direct correlation between NO removal and ammonia slip was not established. x The issue of the synergistic control relationship between emissions and NH3 escape.

[0006] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a method for synergistic predictive control of flue gas nitrogen oxides and ammonia slip based on time derivatives, comprising: Operating parameters, nitrogen oxide emission concentration, and ammonia slip concentration are collected synchronously from the distributed control system and continuous emission monitoring system of the flue gas boiler, and a feature set of the ammonia slip prediction model is constructed based on the operating parameters, nitrogen oxide emission concentration, and ammonia slip concentration. Historical feature data was collected based on the feature set of the ammonia escape prediction model. A prediction model with a hybrid convolution-transformer-temporal derivative structure is constructed, and the prediction model is trained based on historical feature data. The trained prediction model is then used to obtain the predicted sequences of nitrogen oxide emission concentration and ammonia escape concentration for future time steps. A reward function is designed based on the predicted sequences of nitrogen oxide emission concentrations and ammonia escape concentrations at future time steps; Based on the reward function, with ammonia flow rate as the decision variable, a reinforcement learning algorithm is used to solve the ammonia flow rate control strategy.

[0007] The beneficial effects of this invention are as follows: The constructed prediction model with a convolution-Transformer-temporal derivative hybrid structure extracts local temporal features through one-dimensional convolution, captures long-range dependencies through Transformer, and optimizes dynamic trends by combining time derivative constraints. This effectively overcomes the lag and noise sensitivity problems of traditional BP and LSTM models under complex flue gas pollutant conditions. The prediction accuracy, stability, and convergence speed of nitrogen oxide emission concentration and ammonia escape are significantly better than existing methods. Introducing a time derivative error term into the loss function allows the model to not only fit the concentration values ​​but also simultaneously learn their rate of change, significantly improving the smoothness and temporal alignment of the prediction curve and avoiding control misjudgments caused by response delays. Explicitly fusing nitrogen oxide emission concentration and its historical sequence into the ammonia escape prediction input establishes a physical correlation mechanism between the two, breaking through the single-objective independent modeling paradigm and providing a reliable basis for the joint optimization of denitrification efficiency and ammonia escape suppression. By using high-precision prediction results as state input for reinforcement learning, an integrated control architecture of "prediction-decision-execution" is constructed to achieve early intervention based on future emission trends, effectively avoid overshoot and ammonia waste caused by monitoring lag, and reduce operating costs while ensuring emission compliance.

[0008] Furthermore, the construction of the ammonia escape prediction model feature set specifically includes: By using Gini importance, operating parameters with importance higher than the threshold are selected for nitrogen oxide emission concentration and ammonia slip concentration respectively, to obtain a high-importance input feature set; Features with absolute Pearson correlation coefficients greater than the correlation threshold are removed from the high-importance input feature set. At the same time, nitrogen oxide emission concentration and ammonia escape concentration are introduced based on the stoichiometric relationship to obtain the ammonia escape prediction model feature set.

[0009] The beneficial effects of the above-mentioned further scheme are as follows: by explicitly integrating the nitrogen oxide emission concentration and its historical sequence into the ammonia slip prediction input, a physical correlation mechanism between the two is established, breaking through the single-objective independent modeling paradigm, and providing a reliable basis for the joint optimization of denitrification efficiency and ammonia slip suppression.

[0010] Furthermore, the prediction model of the convolution-Transformer-temporal derivative hybrid structure includes: One-dimensional convolutional layers are used to extract local temporal features from historical feature data; Pooling layers are used to downsample and aggregate key features in the time dimension based on local temporal features, resulting in a simplified temporal feature sequence. The first Transformer encoder is used to capture short- to medium-range temporal dependencies between features based on the simplified temporal feature sequence through a multi-head attention mechanism, and combines layer normalization and residual connections to optimize feature propagation, outputting enhanced temporal features that fuse local correlations. The second Transformer encoder is used to mine long-range dependencies across time steps based on enhanced temporal features and through a multi-head attention mechanism, and outputs a final feature sequence that contains both local key features and long-range dependency information. A fully connected layer is used to map the final feature sequence to predicted sequences of nitrogen oxide emission concentration and ammonia escape concentration for future time steps.

[0011] The beneficial effects of the above-mentioned further scheme are as follows: The constructed prediction model with a hybrid convolution-Transformer-temporal derivative structure extracts local temporal features through one-dimensional convolution, captures long-range dependencies through Transformer, and optimizes dynamic trends by combining temporal derivative constraints. This effectively overcomes the lag and noise sensitivity problems of traditional BP and LSTM models under complex flue gas pollutant conditions. The prediction accuracy, stability, and convergence speed of nitrogen oxide emission concentration and ammonia escape are significantly better than existing methods. The hybrid convolution-Transformer-temporal derivative temporal modeling architecture is specifically designed for multivariate dynamic prediction of flue gas pollutants, fusing one-dimensional convolution, Transformer multi-head attention, and temporal derivative constraint loss. It takes into account local feature extraction, global temporal dependencies, and consistency of dynamic trends, effectively overcoming the lag and insufficient coupled modeling of traditional models under complex conditions.

[0012] Furthermore, in the one-dimensional convolutional layer, the first... Each convolutional kernel at time step The output at this location is:

[0013] in, For the first Each convolutional kernel at time step Output at; It is a non-linear activation function; For the first In the nth convolutional kernel, the th The time offset, the first Weights of dimensional features; For the input sequence at time step + , No. The value of the dimensional feature; For the first The bias of each convolutional kernel; The time window size of the convolution kernel; for; Indexed by feature dimensions; The input feature dimension.

[0014] The beneficial effects of the above-mentioned further scheme are as follows: by combining multiple convolutional kernels with time windows to extract local features, it can accurately capture key temporal features in short time scales such as sudden changes in operating conditions, strengthen the local correlation information directly related to pollutant concentration in the input data, lay a high-quality feature foundation for subsequent long-range dependency learning, and improve the model's real-time response sensitivity to complex operating conditions.

[0015] Furthermore, the loss function of the prediction model of the convolution-Transformer-temporal derivative hybrid structure is:

[0016]

[0017]

[0018] in, This is the loss function for the prediction model; for The weights; For MSE loss; This refers to the time derivative error; To predict the total step size; for Predicted concentration at any given time; for Predicted concentration at time -1; For predicting the step size index; for The actual concentration at any given moment; for The true concentration at time -1.

[0019] The beneficial effects of the above-mentioned further scheme are as follows: by introducing a time derivative error term into the loss function, the model not only fits the concentration value, but also learns its rate of change simultaneously, which significantly improves the smoothness and time alignment of the prediction curve and avoids control misjudgment caused by response delay.

[0020] Furthermore, the expression for the reward function is:

[0021] in, In the state Take action in the following circumstances The reward function; This is the current state; To take action now; The weight of the nitrogen oxides item; It is the maximum value; This refers to the concentration of nitrogen oxides. For nitrogen oxide emissions limits; Weights for the ammonia escape term; This refers to the ammonia escape concentration. The emission limit for ammonia escape; The weight of the ammonia flow rate term; This refers to the ammonia water flow rate.

[0022] The beneficial effect of the above further scheme is that it balances NO through weight allocation. x The three objectives of achieving compliance, suppressing ammonia escape, and reducing ammonia costs avoid environmental violations or economic waste caused by optimizing a single objective. They also allow for flexible adjustment of priorities based on actual working conditions, ensuring that the decision-making logic of reinforcement learning aligns with the actual needs of the project and achieves a dynamic balance between environmental protection and economic efficiency.

[0023] Furthermore, the expression for the state space of the reinforcement learning algorithm is:

[0024] in, For state space; This is the current state; This refers to the concentration of nitrogen oxides. This refers to the ammonia escape concentration. This is the value of the first operating parameter in the feature set of the ammonia escape prediction model. This is the value of the second operating parameter in the feature set of the ammonia escape prediction model; The first feature set of ammonia escape prediction models The values ​​of the operating parameters; and Determined based on the predicted sequences of nitrogen oxide emission concentrations and ammonia escape concentrations.

[0025] The beneficial effects of the above-mentioned further scheme are: to fully incorporate pollutant concentration and core operating parameters, to completely depict the overall state of the current operating conditions, to ensure that the reinforcement learning agent can fully perceive the comprehensive information of "denitrification demand-operating conditions" when making decisions, to avoid the one-sidedness of the control strategy due to the lack of state information, and to improve the adaptability of the strategy to operating condition fluctuations.

[0026] Furthermore, the expression for the action space of the reinforcement learning algorithm is:

[0027] in, For action space; To take action now; This refers to the ammonia water flow rate. This is the minimum allowable value corresponding to the lower limit of the process safety for ammonia water flow rate; This represents the maximum allowable value corresponding to the upper limit of the ammonia water flow rate for process safety.

[0028] The beneficial effects of the above-mentioned further scheme are: clarifying the safe adjustment range of ammonia water flow rate, ensuring the engineering feasibility of the control strategy, providing a clear optimization boundary for reinforcement learning, reducing ineffective exploration, improving the strategy convergence speed, and ensuring the safety and stability of ammonia injection rate adjustment.

[0029] Furthermore, the reinforcement learning algorithm iteratively updates the Q function by minimizing the Bellman residual to obtain the ammonia flow control strategy.

[0030] The beneficial effects of the above-mentioned further scheme are as follows: by minimizing the Bellman residual to achieve precise iteration of the Q function, the reinforcement learning agent can continuously optimize the value assessment of "state-action", gradually approach the optimal control strategy, and ensure that the strategy can adapt to the dynamic changes of the working conditions. It has a greater ability to learn and improve online than traditional fixed rules or offline optimization algorithms.

[0031] Furthermore, the expression for the Q function is:

[0032] in, It is the Q function; For expectation operators; In the state Take action in the following circumstances The reward function; This is the current state; To take action now; Discount factor; It is the maximum value; The state for the next time step; This is the action for the next step.

[0033] The beneficial effects of the above-mentioned further scheme are as follows: by introducing the expectation operator and the discount factor, the immediate reward and long-term return are taken into account, and the subsequent overshooting or cost accumulation caused by the agent's short-sighted decision-making is avoided. At the same time, by predicting the value of the "next step state-action", the forward-looking nature of the strategy is enhanced, and the ammonia flow rate adjustment is made to better match the future trend of pollutant concentration, thereby improving the control accuracy. Attached Figure Description

[0034] Figure 1 This invention proposes a flue gas NO model based on a convolution-Transformer-time derivative model. x The overall flowchart of the co-prediction and closed-loop control method for ammonia slip.

[0035] Figure 2 This is a schematic diagram of the network structure of the convolution-Transformer-time derivative hybrid prediction model used in this invention.

[0036] Figure 3 This is a schematic diagram of the reinforcement learning-based process of the present invention.

[0037] Figure 4 NO in the embodiments of the present invention x A line graph comparing predicted and measured concentrations.

[0038] Figure 5 This is a line graph comparing the predicted and measured values ​​of ammonia escape concentration in an embodiment of the present invention.

[0039] Figure 6 NO in the embodiments of the present invention x A comparison chart of concentrations before and after optimization.

[0040] Figure 7 This is a comparison chart of ammonia escape concentration before and after optimization in an embodiment of the present invention.

[0041] Figure 8 This is a comparison chart showing the ammonia dosage before and after optimization in an embodiment of the present invention. Detailed Implementation

[0042] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0043] Example 1 like Figure 1 As shown, in one embodiment of the present invention, a method for synergistic predictive control of flue gas nitrogen oxides and ammonia slip based on time derivatives includes: Operating parameters, nitrogen oxide emission concentration, and ammonia slip concentration are collected synchronously from the distributed control system and continuous emission monitoring system of the flue gas boiler, and a feature set of the ammonia slip prediction model is constructed based on the operating parameters, nitrogen oxide emission concentration, and ammonia slip concentration. Historical feature data was collected based on the feature set of the ammonia escape prediction model. A prediction model with a hybrid convolution-transformer-temporal derivative structure is constructed, and the prediction model is trained based on historical feature data. The trained prediction model is then used to obtain the predicted sequences of nitrogen oxide emission concentration and ammonia escape concentration for future time steps. A reward function is designed based on the predicted sequences of nitrogen oxide emission concentrations and ammonia escape concentrations at future time steps; Based on the reward function, with ammonia flow rate as the decision variable, a reinforcement learning algorithm is used to solve the ammonia flow rate control strategy.

[0044] In this embodiment, a Markov decision process is constructed using ammonia flow rate as the decision variable, wherein the state space is defined by the NO output of the prediction model. x Based on ammonia escape concentration and historical operating parameters, and with the action space defined as the allowable ammonia flow rate adjustment range, the ammonia injection control strategy solution process is modeled as a learning task for a reinforcement learning agent. This agent, based on a defined reward function, employs a temporal difference reinforcement learning algorithm based on the Q-function (action value function). It iteratively updates the strategy by minimizing the Bellman residual and uses the ammonia flow rate output by the strategy in the current state as the control command. The ammonia flow rate control command is converted into an analog control signal and sent to the DCS system to achieve real-time regulation of the ammonia injection quantity, ensuring emission compliance while reducing ammonia consumption. The above steps are repeated in the next sampling cycle, forming a "prediction-decision-control" closed loop.

[0045] In this embodiment, the historical feature data collection time interval is 1 second, and the 1-second interval data collected by DCS is downsampled by a 15-second mean.

[0046] The feature set for constructing the ammonia escape prediction model is specifically as follows: By using Gini importance, operating parameters with importance higher than the threshold are selected for nitrogen oxide emission concentration and ammonia slip concentration respectively, to obtain a high-importance input feature set; Features with absolute Pearson correlation coefficients greater than the correlation threshold are removed from the high-importance input feature set. At the same time, nitrogen oxide emission concentration and ammonia escape concentration are introduced based on the stoichiometric relationship to obtain the ammonia escape prediction model feature set.

[0047] In this embodiment, operating parameters and NO are synchronously collected from the distributed control system (DCS) and continuous emission monitoring system (CEMS) of the flue gas furnace. x Emission concentration and ammonia slip concentration data were processed by imputing missing values, removing outliers, and downsampling the mean. Then, based on a random forest model (using Gini importance calculations to determine feature contribution), the data were analyzed for NO... x A highly important input feature set was constructed using ammonia escape. Subsequently, highly coupled variables with an absolute Pearson correlation coefficient greater than 0.7 were removed from the feature set, and historical pollutant concentrations were introduced to form the final training dataset, which was divided into training set, validation set, and test set at a ratio of 70%, 10%, and 20%, respectively.

[0048] In this embodiment, the ammonia escape prediction model explicitly incorporates NO. x Using emission concentration as an input characteristic, NO x Co-modeling and joint optimization control with ammonia slip. The input features of the ammonia slip prediction model include NO values ​​at the same time step as ammonia slip and several time steps prior. x Emission concentrations to reflect the emissions of NH3 and NO in the denitrification reaction. x The dynamic correlation characteristics.

[0049] In this embodiment, during the waste incineration process, NO x The generation of NH3 and its escape are influenced by multiple factors, including furnace temperature, oxygen content, load, flue gas flow rate, ammonia injection rate, and catalyst activity. Furthermore, significant nonlinear coupling and time lag relationships exist among these variables, making it difficult for traditional mechanistic models to accurately describe this complex dynamic system. This invention extracts multidimensional time-series data from DCS and CEMS systems and employs a data-driven approach to capture the implicit temporal dependencies and variable correlations within the system, providing a high-quality sample foundation for subsequent deep learning modeling.

[0050] like Figure 2As shown, the prediction model of the convolution-Transformer-temporal derivative hybrid structure includes: One-dimensional convolutional layers are used to extract local temporal features from historical feature data; Pooling layers are used to downsample and aggregate key features in the time dimension based on local temporal features, resulting in a simplified temporal feature sequence. The first Transformer encoder is used to capture short- to medium-range temporal dependencies between features based on the simplified temporal feature sequence through a multi-head attention mechanism, and combines layer normalization and residual connections to optimize feature propagation, outputting enhanced temporal features that fuse local correlations. The second Transformer encoder is used to mine long-range dependencies across time steps based on enhanced temporal features and through a multi-head attention mechanism, and outputs a final feature sequence that contains both local key features and long-range dependency information. A fully connected layer is used to map the final feature sequence to predicted sequences of nitrogen oxide emission concentration and ammonia escape concentration for future time steps.

[0051] In this embodiment, the prediction model includes: a one-dimensional convolutional layer and a pooling layer, a two-layer Transformer encoder, an input time step of 240, an output prediction time step of 20, an optimizer of Adam, a loss function of weighted sum of MSE and time derivative loss, and a maximum number of iterations of 100.

[0052] The first one-dimensional convolutional layer Each convolutional kernel at time step The output at this location is:

[0053] in, For the first Each convolutional kernel at time step Output at; It is a non-linear activation function; For the first In the nth convolutional kernel, the th The time offset, the first Weights of dimensional features; For the input sequence at time step + , No. The value of the dimensional feature; For the first The bias of each convolutional kernel; The time window size of the convolution kernel; for; Indexed by feature dimensions; The input feature dimension.

[0054] The loss function of the prediction model of the convolution-Transformer-temporal derivative hybrid structure is:

[0055]

[0056]

[0057] in, This is the loss function for the prediction model; for The weights; For MSE loss; This refers to the time derivative error; To predict the total step size; for Predicted concentration at any given time; for Predicted concentration at time -1; For predicting the step size index; for The actual concentration at any given moment; for The true concentration at time -1.

[0058] In this embodiment, a prediction model with a convolution-Transformer-temporal derivative hybrid structure is constructed, using the running parameters of the past 240 time steps and historical NO. x Using ammonia escape concentration as input, predict NO for the next 20 time steps. x Emission concentration and ammonia slip concentration are considered. The model loss function is a weighted sum of mean square error and time derivative error, and is trained to convergence through backpropagation. The time derivative error in the loss function is used to constrain the rate of change of pollutant concentration, improving the dynamic consistency of the predicted trend.

[0059] The expression for the reward function is:

[0060] in, In the state Take action in the following circumstances The reward function; This is the current state; To take action now; The weight of the nitrogen oxides item; It is the maximum value; This refers to the concentration of nitrogen oxides. For nitrogen oxide emissions limits; Weights for the ammonia escape term; This refers to the ammonia escape concentration. The emission limit for ammonia escape; The weight of the ammonia flow rate term; This refers to the ammonia water flow rate.

[0061] In this embodiment, the reward function comprehensively considers both the cost of ammonia consumption and the compliance rate of pollutant emissions. This reward function comprehensively considers NO... x By considering whether ammonia escape meets standards and the cost of ammonia water consumption, penalties are imposed when pollutants exceed standards, and cost constraints are applied when the amount of ammonia injected is too large, thus achieving a balance between emission compliance and economic efficiency.

[0062] The expression for the state space of the reinforcement learning algorithm is:

[0063] in, For state space; This is the current state; This refers to the concentration of nitrogen oxides. This refers to the ammonia escape concentration. This is the value of the first operating parameter in the feature set of the ammonia escape prediction model. This is the value of the second operating parameter in the feature set of the ammonia escape prediction model; The first feature set of ammonia escape prediction models The values ​​of the operating parameters; and Determined based on the predicted sequences of nitrogen oxide emission concentrations and ammonia escape concentrations.

[0064] The expression for the action space of the reinforcement learning algorithm is:

[0065] in, For action space; To take action now; This refers to the ammonia water flow rate. This is the minimum allowable value corresponding to the lower limit of the process safety for ammonia water flow rate; This represents the maximum allowable value corresponding to the upper limit of the ammonia water flow rate for process safety.

[0066] like Figure 3 As shown, the reinforcement learning algorithm iteratively updates the Q function by minimizing the Bellman residual to obtain the ammonia flow control strategy.

[0067] The expression for the Q function is:

[0068] in, It is the Q function; For expectation operators; In the state Take action in the following circumstances The reward function; This is the current state; To take action now; Discount factor; It is the maximum value; The state for the next time step; This is the action for the next step.

[0069] In this embodiment, the present invention uses a convolution-transformer-time derivative structure as the model skeleton to capture short-term local temporal features and long-term dependencies between different operating parameters, thereby achieving accurate modeling of emission dynamic response under complex operating conditions.

[0070] Based on the prediction results, a Markov Decision Process (MDP) is constructed. The ammonia injection control system is modeled as a reinforcement learning agent, using the predicted future pollutant concentration sequence and current operating conditions as state inputs and ammonia flow rate as action outputs. The optimal control strategy is learned through a reinforcement learning algorithm. The reward function is designed to impose a penalty when NOx or ammonia escape exceeds the standard, while simultaneously imposing a cost term on the ammonia usage, thereby minimizing operating costs while meeting emission standards. This method overcomes the shortcomings of traditional swarm intelligence algorithms in handling high-dimensional continuous action spaces and long-term reward optimization, and is more suitable for the dynamic characteristics of flue gas pollutant systems with strong fluctuations and large time lags.

[0071] Example 2 Historical operational data from a municipal solid waste incineration plant over 15 consecutive days was selected as the experimental data. The plant is equipped with a complete DCS control system and CEMS monitoring equipment. Data sampling frequency was 1 second, and after synchronization and preprocessing, a total of 1,296,000 valid samples were obtained. Operating parameters included 60 items such as furnace temperature, main steam flow rate, oxygen content, and air volume ratio. The pollutant targets were the SCR outlet NOx concentration (mg / m³) and ammonia slip concentration (mg / m³).

[0072] First, the data is cleaned, downsampled, and feature-filtered according to the process described in this invention, ultimately determining 44-dimensional input features. The input to the ammonia escape model explicitly includes the NOx concentration value at each time step. The data is then divided into a training set (70%), a validation set (10%), and a test set (20%) in chronological order.

[0073] To reduce data dimensionality and computational burden, the raw 1-second sampling data from the DCS is averaged over a 15-second window to align with the CEMS monitoring frequency. Each sample uses 44-dimensional features from the 60 minutes preceding the target time (i.e., 240 downsampling points) as input to construct a time-series window. This process compresses the hourly data from 3600×44 to 240×44. This embodiment yields 86,400 valid samples, which are then normalized using z-scores to ensure that the mean of each feature is 0 and the standard deviation is 1.

[0074] In this embodiment, the convolution-Transformer-temporal derivative collaborative prediction model is built using Python 3.10 and PyTorch 2.8 (CUDA 12.8). The model structure includes a one-dimensional convolutional layer, a max-pooling layer, and two Transformer encoder layers. The output layer is a fully connected layer, with an input time step of 240 and an output prediction step of 20. The optimizer used is Adam, and the loss function is the weighted sum of the mean squared error and the temporal derivative error. The maximum number of training epochs is set to 100. The hyperparameters to be tuned include the number of convolutional channels, the number of Transformer heads, the number of encoder layers, and the temporal derivative weights. The optimal combination determined by grid search is shown in Table 1. Table 1. Hyperparameter settings for the Convolution-Transformer-Time Derivative Model

[0075] After the model training is complete, the mean squared error (MSE) and the coefficient of determination (R²) are used. 2 Evaluate the predictive performance of the model.

[0076] Table 2. MSE and R² of the Convolution-Transformer-Time Derivative Model 2 result

[0077] In this embodiment, in order to more clearly understand the model's effect on NO... x The synergistic predictive effect with ammonia escape is illustrated in the following figures: Figure 4 and Figure 5 The graph shows a comparison between the predicted and actual measured values. As can be seen from the graph, the model's prediction for NO... x The predicted curves for ammonia slip concentration show a high degree of agreement with the measured values, accurately capturing the dynamic trends of both pollutants under fluctuating operating conditions, especially maintaining good tracking capabilities in regions of abrupt concentration changes. Therefore, the accuracy of these collaborative prediction results can provide a reliable state input guarantee for subsequent reinforcement learning-based optimized ammonia injection control.

[0078] Based on this, a closed-loop control strategy based on reinforcement learning is constructed. The future pollutant concentration sequence output by the collaborative prediction model is used as the state input of the agent to construct a Markov decision process (MDP).

[0079]

[0080] The state space includes the predicted future 20 steps NO x The concentration, ammonia slip concentration, and ammonia flow rate at the previous moment; the action space is the allowable adjustable ammonia injection flow rate range [0, 100] L / h, with a control resolution of 1 L / h to ensure adjustment accuracy; the reward function comprehensively considers emission compliance and operational economy, and is defined as follows:

[0081] A value function-based temporal difference reinforcement learning algorithm is used for policy optimization. The agent continuously collects transfer samples by interacting with the historical operating environment and iteratively updates the action value function Q(s,a) using the Bellman optimal equation, eventually converging to the optimal ammonia injection policy.

[0082] A rolling control experiment was conducted with a 15-second cycle based on the test data from the last 72 hours. The results show that, after adopting the method of this invention, as... Figure 6 , Figure 7 and Figure 8 shown, NO x The average emission concentration decreased by 18.5%, the average ammonia escape decreased by 25.4%, and the ammonia water consumption decreased by 7.7%.

[0083] In summary, this invention achieves the prediction of NO in flue gas processes by constructing a convolution-Transformer-time derivative hybrid prediction model. x This method achieves high-precision co-prediction of ammonia slip concentration. Furthermore, by incorporating reinforcement learning, the predicted results are used as state input to construct a closed-loop optimization control strategy aimed at achieving emission standards and minimizing ammonia consumption. Practical application verification shows that this method significantly reduces operating costs while ensuring environmental compliance, demonstrating good engineering practical value and promising prospects for widespread application.

Claims

1. A method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on time derivatives, characterized in that, include: Operating parameters, nitrogen oxide emission concentration, and ammonia slip concentration are collected synchronously from the distributed control system and continuous emission monitoring system of the flue gas boiler, and a feature set of the ammonia slip prediction model is constructed based on the operating parameters, nitrogen oxide emission concentration, and ammonia slip concentration. Historical feature data was collected based on the feature set of the ammonia escape prediction model. A prediction model with a hybrid convolution-transformer-temporal derivative structure is constructed and trained based on historical feature data. The trained prediction model is then used to obtain the predicted sequences of nitrogen oxide emission concentration and ammonia escape concentration for future time steps. A reward function is designed based on the predicted sequences of nitrogen oxide emission concentrations and ammonia escape concentrations at future time steps; Based on the reward function, with ammonia flow rate as the decision variable, a reinforcement learning algorithm is used to solve the ammonia flow rate control strategy.

2. The method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on the fusion time derivative as described in claim 1, characterized in that, The feature set for constructing the ammonia escape prediction model is specifically as follows: By using Gini importance, operating parameters with importance higher than the threshold are selected for nitrogen oxide emission concentration and ammonia slip concentration respectively, to obtain a high-importance input feature set; Features with absolute Pearson correlation coefficients greater than the correlation threshold are removed from the high-importance input feature set. At the same time, nitrogen oxide emission concentration and ammonia escape concentration are introduced based on the stoichiometric relationship to obtain the ammonia escape prediction model feature set.

3. The method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on the fusion time derivative as described in claim 1, characterized in that, The prediction model of the convolution-Transformer-temporal derivative hybrid structure includes: One-dimensional convolutional layers are used to extract local temporal features from historical feature data; Pooling layers are used to downsample and aggregate key features in the time dimension based on local temporal features, resulting in a simplified temporal feature sequence. The first Transformer encoder is used to capture short- to medium-range temporal dependencies between features based on the simplified temporal feature sequence through a multi-head attention mechanism, and combines layer normalization and residual connections to optimize feature propagation, outputting enhanced temporal features that fuse local correlations. The second Transformer encoder is used to mine long-range dependencies across time steps based on enhanced temporal features through a multi-head attention mechanism, and outputs a final feature sequence that simultaneously contains local key features and long-range dependency information. A fully connected layer is used to map the final feature sequence to predicted sequences of nitrogen oxide emission concentration and ammonia escape concentration for future time steps.

4. The method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on the fusion time derivative as described in claim 3, characterized in that, The first one-dimensional convolutional layer Each convolutional kernel at time step The output at this location is: in, For the first Each convolutional kernel at time step Output at; It is a non-linear activation function; For the first In the nth convolutional kernel, the th The time offset, the first Weights of dimensional features; For the input sequence at time step + , No. The value of the dimensional feature; For the first The bias of each convolutional kernel; The time window size of the convolution kernel; for; Indexed by feature dimension; The input feature dimension.

5. The method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on the fusion time derivative as described in claim 1, characterized in that, The loss function of the prediction model of the convolution-Transformer-temporal derivative hybrid structure is: in, This is the loss function for the prediction model; for The weights; For MSE loss; This refers to the time derivative error; To predict the total step size; for Predicted concentration at any given time; for Predicted concentration at time -1; For predicting the step size index; for The actual concentration at any given moment; for The true concentration at time -1.

6. The method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on the fusion time derivative as described in claim 1, characterized in that, The expression for the reward function is: in, In the state Take action in the following circumstances The reward function; This is the current state; To take action now; The weight of the nitrogen oxides item; It is the maximum value; This refers to the concentration of nitrogen oxides. For nitrogen oxide emissions limits; Weights for the ammonia escape term; This refers to the ammonia escape concentration. The emission limit for ammonia escape; The weight of the ammonia flow rate term; This refers to the ammonia water flow rate.

7. The method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on the fusion time derivative as described in claim 1, characterized in that, The expression for the state space of the reinforcement learning algorithm is: in, For state space; This is the current state; This refers to the concentration of nitrogen oxides. This refers to the ammonia escape concentration. This is the value of the first operating parameter in the feature set of the ammonia escape prediction model. This is the value of the second operating parameter in the feature set of the ammonia escape prediction model; The first feature set of ammonia escape prediction models The values ​​of the operating parameters; and Determined based on the predicted sequences of nitrogen oxide emission concentrations and ammonia escape concentrations.

8. The method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on the fusion time derivative as described in claim 1, characterized in that, The expression for the action space of the reinforcement learning algorithm is: in, For action space; To take action now; This refers to the ammonia water flow rate; This is the minimum allowable value corresponding to the lower limit of the process safety for ammonia water flow rate; This represents the maximum allowable value corresponding to the upper limit of the ammonia water flow rate for process safety.

9. The method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on the fusion time derivative according to claim 1, characterized in that, The reinforcement learning algorithm iteratively updates the Q-function by minimizing the Bellman residual to obtain the ammonia flow control strategy.

10. The method for synergistic prediction and control of flue gas nitrogen oxides and ammonia slip based on the fusion time derivative as described in claim 9, characterized in that, The expression for the Q function is: in, It is the Q function; For expectation operators; In the state Take action in the following circumstances The reward function; This is the current state; To take action now; Discount factor; It is the maximum value; The state for the next time step; This is the action for the next step.

Images