A Dynamic Prediction Method for Air-Fuel Ratio Based on Furnace Combustion Control

Abstract

This invention belongs to the field of combustion control technology for heating furnaces, specifically a dynamic prediction method for air-fuel ratio based on combustion control in heating furnaces. It includes preprocessing raw monitoring data from the heating furnace operation process based on combustion mechanisms, constructing a multi-dimensional input vector, outputting the optimal air-fuel ratio prediction value for future moments based on an improved LSTM neural network model, and adjusting the combustion air supply to the heating furnace according to the optimal air-fuel ratio prediction value. This invention effectively solves the problem of slow response to gas fluctuations and changes in operating conditions in traditional methods through the synergistic driving of mechanisms and data, significantly improving the accuracy and real-time performance of combustion control. Furthermore, by deeply integrating physical mechanisms, it enhances the model's generalization ability and interpretability, significantly improving the thermal efficiency of the heating furnace and reducing unit product energy consumption and harmful gas emissions.

Description

Technical Field

[0001] This invention relates to the field of furnace combustion control technology, specifically a dynamic air-fuel ratio prediction method based on furnace combustion control. Background Technology

[0002] In the combustion control process of industrial heating furnaces and walking beam heating lines, the air-fuel ratio (AFR, or air-fuel ratio by mass or volume) is a key operating parameter to ensure complete combustion of fuel, improve thermal efficiency, and reduce equipment damage and product defects.

[0003] Traditional industrial furnaces often employ control engineering-based strategies for combustion control: the upper combustion section typically uses a cascade control structure combined with limiting and cross-regulation to take into account the dynamic coupling relationship between temperature, fuel flow rate and gas supply; the lower combustion section often switches between cascade limiting and pulse control to meet heating requirements under different operating conditions.

[0004] These classic control architectures can achieve basic temperature and combustion stability requirements under deterministic and relatively stable operating conditions, but they still face the following unavoidable limitations under actual production conditions: First, changes in the calorific value of the gas have a significant impact on the performance of the combustion system. Most hot rolling or heat treatment processes use a mixture of multi-source mixed gas (such as coke oven gas, blast furnace gas, converter gas, etc.). The proportion of gas output from different sources and at different times will change with production plans, raw materials and process fluctuations, resulting in frequent fluctuations in the lower heating value (LHV) of the mixed gas within a certain range. Changes in calorific value directly affect the combustion chemical reaction and heat release rate under the gas-air ratio. If the calorific value decreases but the air volume does not decrease accordingly, the lean air-fuel ratio will carry away too much heat from the furnace. If the calorific value increases but the air volume is insufficient, incomplete combustion may occur. It is difficult to quickly and accurately compensate for and adaptively adjust the above-mentioned calorific value fluctuations by manually setting the air-fuel ratio on the HMI (human-machine interface), which will lead to increased fuel consumption and increased risk of burnout. Secondly, the air-fuel ratio set manually based on experience is difficult to guarantee a full reaction between air and fuel gas in actual operation. Operators set the air-fuel ratio statically based on experience and historical parameters. However, when the fuel composition, billet type, furnace temperature and operating conditions change frequently, manual adjustment is often lagging and lacks precision, making it impossible to achieve online real-time compensation. This results in the system being in a state of slight imbalance for a long time, which increases the unit fuel consumption, reduces the heat utilization rate, and may also lead to an increase in unburned gas emissions in the exhaust gas and damage to flue equipment. Furthermore, an improper setting of the excess air coefficient can lead to an oxidizing atmosphere or insufficient reducing conditions in the furnace. Excessive air not only carries away valuable heat and lowers the combustion temperature, but also creates an oxidizing atmosphere during metal heating, which significantly increases the oxidation and burn-off of the furnace lining and workpiece. Conversely, insufficient air can cause problems such as incomplete combustion of fuel, coking in the flue, and the emission of harmful gases. Since the steel grade, grade, and heating temperature often change, it is often difficult to manually optimize the excess air coefficient for each section in real time on the production site, resulting in poor combustion atmosphere control. In addition, the combustion system has obvious time-varying, strong nonlinear and operating condition coupling characteristics. The combustion process is affected by a variety of factors such as fuel composition, instantaneous heat load, dynamic characteristics of fans and valves, furnace heat capacity and measurement noise, resulting in significant delay and hysteresis effects. Traditional linear controllers or control strategies based on fixed models are prone to steady-state errors or oscillations when faced with these time-varying and nonlinear disturbances, making it difficult to simultaneously take into account response speed and combustion efficiency.

[0005] Given the above problems, data-driven prediction and control methods have gradually become a research hotspot for combustion system optimization. Compared with classical models, machine learning methods can learn complex nonlinear mapping relationships and time-dependent characteristics from a large amount of historical operating data, thereby achieving online prediction and adaptive decision-making. For time series-related problems, Long Short-Term Memory (LSTM) networks are widely used in the fields of prediction of industrial process variables and fault diagnosis due to their advantages in handling long-term dependencies, capturing sequence dynamic features and noise resistance. However, standard LSTM still has shortcomings in industrial combustion control scenarios. On the one hand, the original LSTM has limited robustness to abnormal disturbances, sensor distortion, and sudden changes in operating conditions. On the other hand, the combustion process involves multi-variable coupling (such as gas flow rate, air supply, calorific value, furnace temperature, product movement speed, etc.). When standard LSTM is dealing with multi-source heterogeneous inputs, time delays, and variable correlation modeling, it is prone to overfitting or prediction bias if it is not combined with appropriate feature engineering or network structure improvement. Furthermore, real-time online deployment also requires the model to have engineering feasibility in terms of inference latency, model compression, and incremental updates. Therefore, a solution is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a dynamic air-fuel ratio prediction method based on furnace combustion control, in order to solve the problem that in the prior art, it is difficult to combine the improved LSTM with online feature updates, robust loss functions, multi-channel input fusion and traditional control strategies (cascade / pulse control), and thus cannot provide a feasible path to achieve complete combustion, reduce gas consumption, and control the oxidizing atmosphere.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a dynamic prediction method for air-fuel ratio based on combustion control of a heating furnace, comprising the following steps: Step 1: Preprocess the raw monitoring data during the operation of the heating furnace based on the combustion mechanism to extract characteristic parameters including the total flow rate of the mixed gas, the real-time calorific value of the mixed gas, the theoretical air-fuel ratio, and the actual excess air coefficient. Step 2: Combine the preprocessed feature parameters with key real-time measurement parameters such as flue gas residual oxygen concentration, furnace pressure, furnace temperature and air valve opening to construct a multi-dimensional input vector, and combine the multi-dimensional input vectors of multiple consecutive time steps into a model input sequence; Step 3: Input the model input sequence into the pre-trained improved LSTM neural network model, which will output the optimal air-fuel ratio prediction for future time moments. Step 4: Adjust the combustion air supply to the heating furnace based on the obtained optimal air-fuel ratio prediction value, and periodically optimize the improved LSTM neural network model based on the actual combustion effect feedback.

[0008] Furthermore, the specific process of step one is as follows: The volume fraction and flow rate of each individual gas are detected in real time, the calorific value of each individual gas and the real-time calorific value of the mixed gas are calculated, and the real-time flow rate of each individual gas is obtained. Then, the volume fraction of each component of the mixed gas is calculated. Based on the calculated composition of the mixed gas, the theoretical air consumption required for complete combustion of a unit volume of mixed gas is calculated, and the theoretical air-fuel ratio is calculated in combination with the set theoretical excess air coefficient. The residual oxygen and carbon monoxide content in the flue gas are obtained through flue gas composition analysis, and the actual excess air coefficient is calculated in combination with the theoretical oxygen consumption, oxidation loss oxygen consumption and theoretical flue gas volume.

[0009] Furthermore, in step two, the multidimensional input vector is a six-dimensional vector, which includes the total flow rate of the mixed gas, the real-time calorific value of the mixed gas, the residual oxygen concentration in the flue gas, the furnace pressure, the furnace temperature, and the air valve opening; the model input sequence is a fixed-length time series matrix composed of the multidimensional input vectors from the current time and several consecutive previous times.

[0010] Furthermore, in step three, the improved LSTM neural network model structure sequentially includes an input layer, an LSTM layer, a Dropout layer, and a fully connected output layer; wherein, the LSTM layer is used to learn the temporal features of the input sequence, the Dropout layer is used to randomly discard some neuron outputs during the training phase to prevent overfitting, and the fully connected output layer is used to map the hidden state of the final time step to the optimal air-fuel ratio prediction value.

[0011] Furthermore, this includes improvements to the training process of the LSTM neural network model, as detailed below: A dataset was constructed by collecting historical operating data of the heating furnace and corresponding optimal air-fuel ratio label data. After preprocessing the data, it was divided into training set, validation set and test set. The training set was used to train the model, the validation set was used to monitor the training process and prevent overfitting, and finally the test set was used to evaluate the predictive performance of the model.

[0012] Furthermore, in step four, the strategy for periodically optimizing the improved LSTM neural network model based on actual combustion effect feedback is as follows: When the deviation between the actual excess air coefficient and the set value continues to exceed the preset threshold, the model update mechanism is triggered, and the model parameters are incrementally learned or fine-tuned using the latest operating data.

[0013] Furthermore, it also includes establishing a coal gas composition-calorific value-air-fuel ratio mapping database to associate and store mixed coal gas characteristics, operating conditions, and optimal air-fuel ratio data at different times, in order to support model training, optimization, and data analysis.

[0014] Compared with the prior art, the beneficial effects of the present invention are: In this invention, by changing the air-fuel ratio from traditional "lagging control" to "looking-ahead prediction", the accuracy and real-time performance of combustion control are greatly improved, ensuring complete combustion of fuel, significantly improving the thermal efficiency of the heating furnace, reducing gas consumption per unit product, reducing oxidation and burn-off of the furnace lining and workpiece, extending equipment service life and improving heating quality.

[0015] In this invention, by reasonably controlling the excess air coefficient, the problem of excessive oxidizing atmosphere or incomplete combustion of fuel is avoided, which effectively reduces the emission of harmful gases such as nitrogen oxides and unburned coal gas, meets environmental protection requirements, and at the same time reduces production risks such as flue gas coking, thus achieving a dual improvement in economic and environmental benefits. Attached Figure Description

[0016] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a flowchart of the improved LSTM model construction process in this invention; Figure 3 This is a diagram illustrating the LSTM model training process in this invention. Detailed Implementation

[0017] 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.

[0018] Example 1: As Figure 1 As shown, the dynamic prediction method for air-fuel ratio based on furnace combustion control proposed in this invention includes the following steps: Step 1: Preprocess the raw monitoring data during the operation of the heating furnace based on the combustion mechanism to extract characteristic parameters including the total flow rate of the mixed gas, the real-time calorific value of the mixed gas, the theoretical air-fuel ratio, and the actual excess air coefficient. Step 2: Combine the preprocessed feature parameters with key real-time measurement parameters such as flue gas residual oxygen concentration, furnace pressure, furnace temperature and air valve opening to construct a multidimensional input vector, and combine the multidimensional input vectors of multiple consecutive time steps into a model input sequence. Step 3: Input the model input sequence into the pre-trained improved LSTM neural network model, which will output the optimal air-fuel ratio prediction for future time moments. Step 4: Adjust the combustion air supply to the heating furnace based on the obtained optimal air-fuel ratio prediction value, and periodically optimize the improved LSTM neural network model based on the actual combustion effect feedback.

[0019] In other words, based on the combustion and thermodynamic mechanism formulas, the original data is first preprocessed and features are extracted to obtain a set of feature parameters that can directly and profoundly reflect the core state of the combustion process (such as the calorific value of the mixed gas and the theoretical air-fuel ratio). Then, this series of feature parameters is used as the input of the improved LSTM model. This method significantly reduces the learning difficulty of the model and improves the prediction accuracy, generalization ability and engineering practicality.

[0020] Furthermore, Long Short-Term Memory (LSTM) is a special type of recurrent neural network that uses cell states and threshold structures to transmit and add or subtract information. When modeling boiler combustion systems, the input variables are not independent of each other at different time points and are closely related to historical data. Using LSTM can achieve better training results.

[0021] Moreover, the core idea of ​​this method is that instead of directly inputting raw, unprocessed instrument data into the prediction model, it first constructs a rigorous mechanism preprocessing layer. This preprocessing layer is based on the principles of combustion and thermodynamics to perform real-time analysis and feature extraction on various basic data from the field.

[0022] Specifically, this method obtains the components of each individual gas (such as the volume fraction of carbon monoxide, hydrogen, hydrocarbons and oxygen) through an online component detection instrument, and calculates the precise component ratio and lower heating value of the mixed gas in real time based on its flow rate.

[0023] Furthermore, based on the stoichiometric relationship of the chemical reaction of complete combustion, the theoretical air volume and theoretical air-fuel ratio required for the current gas are calculated. At the same time, in order to characterize the deviation between the actual operating conditions and the ideal state, the method also introduces the analysis of residual oxygen concentration in flue gas and the calculation of oxidation loss, which are used to infer the excess air coefficient in actual operation, providing important feedback information and optimization basis for the model.

[0024] After completing the aforementioned high-physical-meaning feature engineering, these key state parameters are constructed into a well-defined six-dimensional input vector, which specifically includes: total flow rate of mixed gas, real-time calorific value of mixed gas, residual oxygen concentration in flue gas, furnace pressure, furnace temperature, and air valve opening. This input vector covers the main factors driving the combustion process (fuel and air volume), real-time feedback of combustion effect (residual oxygen and temperature), and dynamic stability of the system (pressure).

[0025] Furthermore, the sequence of this vector over continuous time steps is used as the input to an improved LSTM network. This improved LSTM model, through its internal sophisticated gating mechanism (including forget gate, input gate, and output gate), can effectively learn and memorize the complex temporal dynamic characteristics and long-term dependencies of the combustion process, thereby capturing the nonlinear law of the influence of operating condition changes on the optimal air-fuel ratio. Finally, the network outputs an accurate prediction of the optimal air-fuel ratio for future moments through a fully connected layer, which is used to guide the control system to perform feedforward adjustment.

[0026] In summary, by synergistically driving mechanisms and data, this approach effectively addresses the problem of slow response to gas fluctuations and changes in operating conditions in traditional methods, enabling a shift from "lagging control" to "proactive prediction" of the air-fuel ratio, significantly improving the accuracy and real-time performance of combustion control. Furthermore, by deeply integrating physical mechanisms, it enhances the model's generalization ability and interpretability, reducing the reliance of purely data-driven models on massive amounts of historical data and their "black box" risks. Ultimately, this significantly improves the thermal efficiency of the heating furnace, reduces energy consumption per unit of product, and decreases emissions of harmful gases such as nitrogen oxides, bringing significant economic and environmental benefits to enterprises.

[0027] Furthermore, such as Figure 2 As shown, the following mechanistic model is used to preprocess the data and extract features: (a) Determining the gas ratio: The physicochemical composition of each individual gas (such as coke oven gas, blast furnace gas, converter gas, etc.) before mixing is analyzed. The real-time calorific value of the mixed gas is obtained using a real-time gas calorific value analyzer. The calorific value of each individual gas is calculated using the detected components and their corresponding lower heating values. Based on the relationship between the calorific value of each individual gas and the total calorific value of the mixed gas, the flow rate ratio of the mixed gas is calculated, and the composition ratio of the mixed gas is further derived based on the flow rate ratio.

[0028] The physicochemical composition of each individual gas (such as coke oven gas, blast furnace gas, converter gas, etc.) before mixing is tested. The real-time calorific value of the mixed gas is obtained by combining the results with a real-time gas calorific value analyzer. The main combustible components of the converter gas in the heating furnace are CO and a small amount of H2, while the combustible components of the blast furnace gas are mainly CO, H2, and a small amount of hydrocarbons. The gas flow ratio formula is as follows: ; ; ; ; in, , and The calorific values ​​of blast furnace gas, converter gas, and mixed gas are respectively determined based on physicochemical experiments. and These are the flow rates of blast furnace gas and converter gas, respectively. Read the meter value to obtain the mixed gas flow rate.

[0029] and The calculation can be performed as follows: ; in, , and These are the volume fractions of carbon monoxide, hydrogen, and hydrocarbon wet components in each individual gas, obtained through component analysis.

[0030] The theoretical air volume is calculated using the following formula: ; in, This is the theoretical air consumption required for the complete combustion of a unit volume of gas. , and These represent the percentages of carbon monoxide, hydrogen, and oxygen in the wet composition of the mixed gas. This represents the percentage of various hydrocarbons in the fuel gas. This represents the percentage of oxygen in the air.

[0031] The proportions of the mixed gas components are calculated using the following formula: ; ; ; ; in, , and These are the flow rates of blast furnace gas, converter gas, and mixed gas, respectively. and These are the CO contents in blast furnace gas and converter gas, respectively. and These are the hydrogen contents in blast furnace gas and converter gas, respectively. and These are the oxygen contents in blast furnace gas and converter gas, respectively. and These represent the hydrocarbon content in blast furnace gas and converter gas, respectively.

[0032] (ii) Determining the air-fuel ratio: The air-fuel ratio is the ratio of the amount of air to the amount of fuel gas, and can be calculated using the following formula: ; in, Air-fuel ratio, This is the excess air coefficient.

[0033] Based on the composition of the mixed gas and the chemical reaction equations for complete combustion, the theoretical air volume required for the complete reaction of the mixed gas is calculated. The theoretical air-fuel ratio is obtained by calculating the ratio of the theoretical air volume to the actual gas supply. This theoretical air-fuel ratio, based on gas stoichiometry, can serve as an important input parameter for the optimal air-fuel ratio prediction model.

[0034] Calculate the calorific value of the gas: Based on the composition content and lower heating value of each individual gas, the theoretical calorific value of each individual gas is obtained using thermochemical calculation formulas. By recording multi-source data over different time periods, the measurement deviation of the real-time calorific value analyzer can be further verified, the reliability of calorific value data can be improved, and high-precision label data can be provided for model training.

[0035] The formula for calculating the calorific value of a single type of coal gas is as follows: ; ; in, and These are the calculated calorific values ​​of converter gas and blast furnace gas, respectively. , and These are the lower heating values ​​of CO, H2, and hydrocarbons, respectively. and These represent the percentages of CO and H2 in the converter gas, respectively. , and These represent the percentages of CO, H2, and hydrocarbons in the blast furnace gas.

[0036] Calculate the actual excess air coefficient: In actual operation, there are errors in the detection of the calorific value of the gas, the gas composition changes with the adjustment of production organization, and phenomena such as air suction and leakage exist in the furnace, resulting in a difference between the theoretical excess air coefficient and the actual operating conditions. This invention calculates the actual excess air coefficient by combining flue gas composition analysis (such as O2, CO, and CO2 concentrations) and oxidation loss measurement results, providing a basis for air-fuel ratio deviation analysis and model optimization.

[0037] The calculation process of the actual excess air coefficient.

[0038] The residual oxygen content is calculated using the following formula: ; in, The residual oxygen content in the flue gas (unit: %). The residual CO content in the flue gas (unit: %). The theoretical residual oxygen content after complete combustion (unit: %). The theoretical oxygen consumption is calculated using the following formula: ; in, This represents the theoretical unit oxygen consumption for coal gas.

[0039] Oxygen consumption due to oxidation is calculated using the following formula: ; in, Oxygen consumption per ton of steel billet due to oxidation and burning. , and These represent the amounts of ferrous oxide, iron(III) oxide, and iron oxide generated per ton of steel billet, respectively.

[0040] Actual excess air coefficient calculation: ; in, This is the actual excess air coefficient. For theoretical combustion flue gas volume, To increase airflow, This refers to the leakage airflow. For gas flow rate, This represents the percentage of oxygen in the air. This refers to the gas consumption per unit mass of steel billet.

[0041] ; in, This refers to the gas flow rate; One cycle time; This refers to the billet output within one cycle.

[0042] The theoretical flue gas volume is calculated using the following formula: ; in, This is the theoretical flue gas volume generated by the complete combustion of a unit volume of coal gas. This represents the percentage of non-combustible gases such as CO2 and N2 in the total volume of the gas.

[0043] Adjusting the composition of the gas: Based on the deviation between the actual excess air coefficient obtained in step four and the set theoretical excess air coefficient, when the deviation exceeds the preset threshold, the composition ratio of the mixed gas is corrected. The adjustment amount is determined comprehensively based on the gas composition, instantaneous gas flow rate, and calorific value fluctuation. Through the above adjustment, the composition stability of the mixed gas is maintained, the disturbance of the combustion process is reduced, and the system is more likely to achieve the optimal air-fuel ratio.

[0044] When the error between the actual excess air coefficient and the given excess air coefficient exceeds a certain value, the composition of the mixed gas is adjusted as follows: ; Wherein, represents the difference between the actual excess air coefficient and the theoretical excess air coefficient. At that time, the air quantity is calculated based on the excess air coefficient α0 and the current gas composition, where, The maximum error is set; when At that time, adjust the current gas composition ratio until... until.

[0045] The adjustment amount is obtained by solving a simultaneous heat balance equation based on the gas composition, gas flow rate, and gas calorific value. Since the hydrocarbon content in the gas is very small, the adjustment amount only considers CO and H2; the specific calculation process is shown in the following formula: like If the oxygen consumption is high, the proportion of high-oxygen-consumption gas should be reduced, and the proportion of low-oxygen-consumption gas should be increased; conversely, if the oxygen consumption is high, the proportion of low-oxygen-consumption gas should be reduced, and the proportion of high-oxygen-consumption gas should be increased. Specifically: ; ; in, and These represent the proportions of CO and H2 in the mixed gas; and These are the flow rates of the mixed gas and oxygen, respectively. and The adjustments are for CO and H2, respectively.

[0046] Establish a database of coal gas composition and calorific value: The calorific value of the mixed gas, the corresponding gas composition ratio, and the combustion condition parameters at each time point are recorded to form a multi-dimensional database of gas composition, calorific value, and air-fuel ratio. This database serves as a data source for training and updating the improved LSTM model and can be used to explore the complex mapping relationship between gas composition, calorific value fluctuations, theoretical air volume, and actual air-fuel ratio, providing a data foundation for intelligent prediction of air-fuel ratio.

[0047] The above steps establish a systematic analysis system for the physicochemical properties of coal gas, clarifying the relationship between the theoretical air-fuel ratio, the actual excess air coefficient, and the composition of the mixed gas. Furthermore, by combining the improved LSTM model, real-time prediction and dynamic adjustment of the optimal air-fuel ratio can be achieved, solving problems such as the lag in traditional manual setting, large fluctuations in the calorific value of the gas, and the difficulty in real-time control of the excess air coefficient, thus significantly improving the combustion efficiency and operational economy of the heating furnace.

[0048] (vii) Establishing a network model: The preprocessed furnace operating parameters are input into the trained air-fuel ratio prediction model to obtain the predicted air-fuel ratio value at time n. The air-fuel ratio prediction model is obtained by iteratively training on an improved LSTM neural network using a dataset constructed from the furnace operating parameters.

[0049] like Figure 3 As shown, the model predicts the optimal air-fuel ratio by learning the nonlinear relationship between the furnace operating state and the actual air-fuel ratio, so as to achieve full combustion and efficient energy utilization.

[0050] Data cleaning and preprocessing: After obtaining the operating parameters of the heating furnace, the data is further cleaned and preprocessed, including: Missing values ​​are filled using linear interpolation or moving average methods; Noise is smoothed using channel filtering; The air-fuel ratio prediction model is obtained by iteratively training a dataset constructed using furnace operating parameters on an improved LSTM neural network, specifically including: Collect continuous time series of furnace operating parameters to obtain a furnace operating parameter dataset; The dataset of heating furnace operating parameters was divided into a training set, a validation set, and a test set. Set the number of LSTM layer units in the air-fuel ratio prediction model; To prevent overfitting, a Dropout layer is set up at the level below the LSTM layer. A fully connected layer is set up at the level below the Dropout layer, and the output of this fully connected layer is the air-fuel ratio prediction value; The training set is input into the air-fuel ratio prediction model for training, the air-fuel ratio prediction model is evaluated using the test set, and the model is optimized to improve prediction accuracy.

[0051] The air-fuel ratio prediction model is trained by inputting the operating parameters into iteratively training the LSTM neural network on the dataset constructed by the operating parameters. The network structure of the air-fuel ratio prediction model is shown in the figure.

[0052] Furthermore, the air-fuel ratio prediction model in step S2 is obtained by iteratively training the dataset constructed from the running parameters on an LSTM neural network. The network structure of the air-fuel ratio prediction model is shown in the figure, and specifically includes: Collect the running parameters of a continuous time series to obtain the running parameter dataset; The engine operating parameters at each time step are encoded as a 6-dimensional vector. These operating parameters for 100 time steps are used as input to the air-fuel ratio prediction model. For example, the input to the air-fuel ratio prediction model for each time step is: ; Among them, the six-dimensional input parameters Including total flow of mixed gas calorific value of mixed gas residual oxygen concentration in flue gas Furnace pressure Furnace temperature air valve opening .

[0053] The overall input window is as follows: ; in, This represents the input vector at time step t, which contains the input data from t−99 to t.

[0054] The Gate of Oblivion is: ; in, For the sigmoid function, This is the weight matrix. This is the hidden state from the previous time step. For the current time step input, This is a bias term.

[0055] The input gate is: ; in, Here is the weight matrix of the input gate. This is the bias term for the input gate.

[0056] Candidate memories are: ; Among them, candidate memory units It is a new candidate value used to update the cell state. This is the activation function used to ensure that the output value is in the range [−1, 1]. This is the weight matrix for candidate memory units. The bias term for candidate memory units.

[0057] The memory unit is updated to: ; in, The cell state is updated by combining the outputs of the forget gate and the input gate. ⊙ represents the Hadamard product. It represents the cell state at the previous time step.

[0058] The input gate is: ; in, The activation value of the output gate determines the cell state. Which part will be output? This is the weight matrix of the output gate, used to connect the hidden state and the input. This is the hidden state of the previous time step. This is the input for the current time step. This is the bias term for the output gate.

[0059] The output in the hidden state is: ; in, The hidden state at the current time step is the final output of the network. Indicates cell state Nonlinear transformation is performed using the hyperbolic tangent function tanh.

[0060] The fully connected layer network is: ; in, The predicted output at time step t, This is the transpose of the weight vector of the fully connected layer. This is the bias term for the fully connected layer.

[0061] The working principle of this invention is as follows: When in use, through the synergistic driving of mechanism and data, it effectively solves the problem of slow response of traditional methods to gas fluctuations and changes in operating conditions, realizes the transformation of air-fuel ratio from "lagging control" to "forward prediction", significantly improves the accuracy and real-time performance of combustion control, and enhances the generalization ability and interpretability of the model by deeply integrating physical mechanisms, reduces the dependence of pure data-driven models on massive historical data and its "black box" risk, significantly improves the thermal efficiency of the heating furnace, reduces energy consumption per unit product, and reduces emissions of harmful gases such as nitrogen oxides.

[0062] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, enabling those skilled in the art to better understand and utilize it. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for dynamic prediction of air-fuel ratio based on combustion control of a heating furnace, characterized in that, Includes the following steps: Step 1: Preprocess the raw monitoring data during the operation of the heating furnace based on the combustion mechanism; Step 2: Combine the preprocessed feature parameters with key real-time measurement parameters such as flue gas residual oxygen concentration, furnace pressure, furnace temperature and air valve opening to construct a multi-dimensional input vector, and combine the multi-dimensional input vectors of multiple consecutive time steps into a model input sequence; Step 3: Input the model input sequence into the pre-trained improved LSTM neural network model, and the improved LSTM neural network model outputs the best air-fuel ratio prediction value for the future time. Step 4: Adjust the combustion air supply to the heating furnace based on the obtained optimal air-fuel ratio prediction value, and periodically optimize the improved LSTM neural network model based on the actual combustion effect feedback.

2. The method for dynamic prediction of air-fuel ratio based on combustion control of a heating furnace according to claim 1, characterized in that, The specific process of step one is as follows: The volume fraction and flow rate of each individual gas are detected in real time, the calorific value of each individual gas and the real-time calorific value of the mixed gas are calculated, and the real-time flow rate of each individual gas is obtained. Then, the volume fraction of each component of the mixed gas is calculated. Based on the calculated composition of the mixed gas, the theoretical air consumption required for complete combustion of a unit volume of mixed gas is calculated, and the theoretical air-fuel ratio is calculated in combination with the set theoretical excess air coefficient. The residual oxygen and carbon monoxide content in the flue gas are obtained through flue gas composition analysis, and the actual excess air coefficient is calculated in combination with the theoretical oxygen consumption, oxidation loss oxygen consumption and theoretical flue gas volume.

3. The method for dynamic prediction of air-fuel ratio based on combustion control of a heating furnace according to claim 1, characterized in that, In step two, the multidimensional input vector is a six-dimensional vector, and the model input sequence is a fixed-length time series matrix composed of the multidimensional input vectors from the current time and several consecutive previous time points.

4. The method for dynamic prediction of air-fuel ratio based on combustion control of a heating furnace according to claim 1, characterized in that, In step three, the improved LSTM neural network model structure includes, in sequence, an input layer, an LSTM layer, a Dropout layer, and a fully connected output layer.

5. The method for dynamic prediction of air-fuel ratio based on furnace combustion control according to claim 1, characterized in that, It also includes improvements to the training process of the LSTM neural network model, as detailed below: A dataset was constructed by collecting historical operating data of the heating furnace and corresponding optimal air-fuel ratio label data. After preprocessing the data, it was divided into training set, validation set and test set. The training set was used to train the model, the validation set was used to monitor the training process and prevent overfitting, and finally the test set was used to evaluate the predictive performance of the model.

6. The method for dynamic prediction of air-fuel ratio based on combustion control of a heating furnace according to claim 1, characterized in that, In step four, the strategy for periodically optimizing the improved LSTM neural network model based on feedback from actual combustion effects is as follows: When the deviation between the actual excess air coefficient and the set value continues to exceed the preset threshold, the model update mechanism is triggered, and the model parameters are incrementally learned or fine-tuned using the latest operating data.

7. The method for dynamic prediction of air-fuel ratio based on furnace combustion control according to claim 1, characterized in that, It also includes establishing a coal gas composition-calorific value-air-fuel ratio mapping database, which is used to associate and store the characteristics of mixed coal gas, operating conditions and optimal air-fuel ratio data at different times.

Images