Safety intelligent control method for thermal system of semi-coke tail gas boiler

Abstract

The present application relates to the technical field of industrial process control, in particular to a safety intelligent control method for a semi-coke tail gas boiler thermodynamic system. The method collects combustible components and flow at the tail gas inlet through a chromatographic analyzer, inputs a tail gas calorific value dynamic observer combining a long short-term memory network and a mechanism model to calculate and predict a calorific value sequence; according to the sequence, a target air-fuel ratio is obtained by table lookup in a three-dimensional mapping table, which is converted into a frequency and opening degree instruction to perform feedforward pre-regulation with time delay compensation; the main steam pressure is collected to correct the residual error through proportional integral derivative feedback with conditional integration, forming a closed-loop control. The present application moves the control trigger node to the fuel input end, overcomes the control lag caused by thermal inertia delay, avoids the mismatch between fuel quantity and air quantity, and suppresses the furnace combustion oscillation.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, specifically to a method for intelligent safety control of a semi-coke tail gas boiler thermal system. Background Technology

[0002] Conventional control schemes for semi-coke tail gas boiler thermal systems primarily rely on macroscopic thermal parameters such as main steam pressure or furnace temperature to construct a single-loop feedback control circuit. In practice, the boiler's fuel input and air supply are calculated and output commands via a proportional-integral-derivative (PID) controller based on the deviation between the main steam pressure and the set pressure, adjusting the opening of the tail gas inlet regulating valve and the frequency of the blower motor, respectively. When fluctuations in semi-coke production conditions cause changes in the proportions of combustible components such as carbon monoxide, hydrogen, and methane in the tail gas, the conventional scheme maintains a fixed air-fuel ratio setpoint or only performs slow, delayed corrections based on oxygen levels. The entire control logic is triggered after combustion heat release and heat transfer to the main steam side; the system lacks a direct detection and pre-regulation mechanism for changes in the tail gas's combustion characteristics.

[0003] In the actual operation of the aforementioned conventional control scheme, there is an inherent thermal inertia delay from the heat release of the exhaust gas during combustion in the furnace to the generation of a response in the main steam pressure. When the proportion of combustible components in the exhaust gas undergoes a drastic change, due to the lack of a real-time sensing and pre-adjustment mechanism for changes in the calorific value at the fuel input end, the single-loop feedback control cannot intervene during the thermal inertia delay period. This leads to a mismatch between the fuel input and the actual required air volume during the delay period, which in turn causes oscillations in the furnace combustion conditions and may even lead to safety accidents such as partial flameout or deflagration. Summary of the Invention

[0004] The purpose of this invention is to provide a safe and intelligent control method for the thermal system of semi-coke tail gas boiler, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for safe and intelligent control of the thermal system of a semi-coke tail gas boiler includes: collecting carbon monoxide volume fraction, hydrogen gas integral number, methane volume fraction, and total tail gas flow rate signals at the inlet of the tail gas conveying pipeline using an online gas chromatograph; inputting the carbon monoxide volume fraction, hydrogen gas integral number, methane volume fraction, and total tail gas flow rate signals into a tail gas calorific value dynamic observer, calculating a predicted calorific value sequence for the current moment and a preset future step size based on the combustion heat release time sequence characteristics of each combustible component; determining the target air-fuel ratio by looking up the predicted calorific value sequence in a preset three-dimensional mapping table of excess safety air coefficient; converting the target air-fuel ratio into a frequency adjustment command for the blower inverter and an opening command for the tail gas regulating valve, and outputting them to the corresponding actuators to perform feedforward pre-regulation; collecting the boiler main steam pressure signal, correcting the residual error generated by the feedforward pre-regulation based on the deviation between the main steam pressure signal and the set pressure value using a proportional-integral-derivative feedback controller, generating a correction command, and superimposing it with the feedforward pre-regulation command to form a closed-loop control command.

[0006] Preferably, the calculation process of the exhaust gas calorific value dynamic observer includes: establishing a mechanistic calculation model containing the standard calorific value constants of each combustible component; acquiring the carbon monoxide volume fraction, hydrogen gas integral, methane volume fraction, and the measured calorific value data sequence at the corresponding time point within a historical time period; constructing a residual sequence using the measured calorific value data sequence and the output value sequence of the mechanistic calculation model; inputting the residual sequence into a long short-term memory network for training to obtain a residual prediction model; and, during the real-time operation phase, adding the baseline calorific value output by the mechanistic calculation model to the residual prediction value output by the residual prediction model to generate the predicted calorific value sequence for the current time and the future preset step size.

[0007] Preferably, the construction and lookup process of the three-dimensional mapping table of the safety excess air coefficient includes: establishing an initial excess air coefficient space with the predicted calorific value sequence, the boiler real-time load command, and the flue gas temperature as three-dimensional coordinate axes; dividing the initial excess air coefficient space into grid nodes according to a preset step size, and storing the corresponding basic excess air coefficient in each grid node; during the lookup process, using the predicted calorific value at the current moment, the current boiler real-time load command, and the current flue gas temperature as index coordinates, and using a trilinear interpolation algorithm to perform interpolation calculations between 8 adjacent grid nodes, and using the interpolation calculation results as the target air-fuel ratio.

[0008] Preferably, the process of converting the target air-fuel ratio into the frequency adjustment command of the blower inverter and the opening command of the exhaust gas regulating valve includes: obtaining the current actual operating frequency of the blower and the current opening of the exhaust gas regulating valve; calculating the change in total air volume demand and the change in total fuel demand based on the difference between the target air-fuel ratio and the current actual air-fuel ratio; introducing the response time constant of the damper actuator and the response time constant of the valve actuator, and performing time misalignment allocation on the change in total air volume demand and the change in total fuel demand based on the response time constant, thereby generating the frequency adjustment command of the blower inverter and the opening command of the exhaust gas regulating valve with time delay compensation.

[0009] Preferably, the correction process via the proportional-integral-derivative (PID) feedback controller includes: setting a conditional integral enable switch in the integral stage of the PID feedback controller; closing the conditional integral enable switch to accumulate integral values ​​when the absolute value of the deviation between the main steam pressure signal and the set pressure value exceeds a preset dead zone threshold; opening the conditional integral enable switch and keeping the accumulated integral value unchanged when the absolute value of the deviation is less than or equal to the preset dead zone threshold; and adding the proportional stage output value, the derivative stage output value, and the kept accumulated integral value to generate the correction command.

[0010] Preferably, the process of acquiring signals through the online gas chromatograph further includes preprocessing the acquired signals: obtaining the raw chromatographic peak data of the online gas chromatograph in the current sampling period; performing baseline drift subtraction on the raw chromatographic peak data to obtain the initial volume fraction; using the sliding window method to extract the initial volume fraction of the current sampling period and several consecutive previous sampling periods to form a time window sequence; calculating the mean and standard deviation of the time window sequence; when the initial volume fraction of the current sampling period deviates from the mean by more than a preset multiple of the standard deviation, the initial volume fraction of the current sampling period is marked as an outlier and removed.

[0011] Preferably, the training process of the residual prediction model includes: introducing a forgetting gating variable in the hidden layer state update calculation of the long short-term memory network; iteratively updating the forgetting factor value of the forgetting gating variable according to the number of training batches, wherein the forgetting factor value decays exponentially with the increase of the number of training batches; during the forward propagation of the long short-term memory network, attenuating and weighting the historical hidden layer states according to the forgetting factor value of the current iteration, calculating the current hidden layer state based on the attenuated and weighted historical hidden layer states and the currently input residual sequence; and outputting the residual prediction value based on the current hidden layer state.

[0012] Preferably, the interpolation calculation process of the trilinear interpolation algorithm further includes boundary constraint correction: determining whether the index coordinates exceed the boundary of the initial excess air coefficient space in each dimension of the three-dimensional coordinate axis; when the predicted calorific value dimension exceeds the maximum boundary value, locking the predicted calorific value dimension to the maximum boundary value, and simultaneously obtaining the furnace negative pressure measurement value at the current moment; calculating the negative pressure bias coefficient based on the deviation between the furnace negative pressure measurement value and the preset negative pressure reference value; multiplying the negative pressure bias coefficient with the interpolation calculation result locked to the maximum boundary value to generate the corrected target air-fuel ratio.

[0013] Preferably, the process of allocating the changes in total air volume demand and total fuel demand in a time-staggered manner includes: constructing a second-order transfer function model that incorporates the dynamic characteristics of the damper actuator and the valve actuator; inputting a step signal into the second-order transfer function model and recording the output response curve; calculating the time difference between the response time constant of the damper actuator and the response time constant of the valve actuator based on the rise time and peak time of the output response curve; and allocating the changes in total fuel demand in real time and the changes in total air volume demand in lag behind the time difference, using the response time constant of the valve actuator as a reference.

[0014] Preferably, after marking the initial volume fraction of the current sampling period as an outlier and removing it, a missing data compensation process is also included: deploying multiple redundant gas sensors at different spatial locations upstream and downstream of the inlet of the exhaust gas pipeline; obtaining the spatial volume fraction matrix collected by the multiple redundant gas sensors at the removal time point marked as an outlier; extracting a historical spatial volume fraction matrix sequence by extending a preset time window forward and backward from the removal time point; performing spatial interpolation calculation on the spatial volume fraction matrix at the removal time point according to the correlation weight of each spatial location in the historical spatial volume fraction matrix sequence, and replacing the removed initial volume fraction with the spatial interpolation calculation result.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention collects the volume fraction and flow rate of combustible components by installing an online gas chromatograph at the inlet of the exhaust gas pipeline, and calculates the predicted calorific value sequence using a dynamic observer of exhaust gas calorific value combining a long short-term memory network and a mechanistic model. This moves the control trigger node from the steam parameter end after thermal inertia propagation to the fuel input end. Based on the predicted calorific value sequence, the target air-fuel ratio is determined in a three-dimensional mapping table and converted into a blower frequency adjustment command with time delay compensation and an exhaust gas regulating valve opening command for feedforward pre-adjustment. Combined with proportional-integral-derivative feedback of the main steam pressure to correct the residual error, this eliminates the control response lag caused by boiler thermal inertia delay, avoids fuel input and air quantity mismatch problems caused by sudden changes in exhaust gas combustible components, and suppresses furnace combustion condition oscillations.

[0016] 2. In the calorific value prediction stage, a forgetting factor that decays exponentially with the training batch is introduced to attenuate and weight the historical hidden layer states of the Long Short-Term Memory network, reducing the interference of early historical data on the current prediction. In the signal acquisition stage, baseline drift subtraction, sliding window mean standard difference anomaly removal, and multi-redundant sensor spatial interpolation compensation are performed on the original chromatographic peaks to eliminate the misleading influence of abnormal detection data on the control logic. In the command execution stage, the response time difference between the damper and valve actuator is calculated using a second-order transfer function model, and the changes in air volume and fuel quantity demand are allocated in a time-displaced manner accordingly to compensate for the timing differences of the physical actions of the actuators. Under boundary conditions, the predicted calorific value exceeding the boundary is locked, and the furnace negative pressure deviation is introduced to calculate the bias coefficient to correct the target air-fuel ratio, avoiding control command divergence under extreme calorific value fluctuations. In the feedback stage, the integral saturation phenomenon under small deviations is eliminated by setting a conditional integral enable switch based on the dead zone threshold in the integral stage. Attached Figure Description

[0017] Figure 1 This invention provides the overall process for the safe and intelligent control of the thermal system of a semi-coke tail gas boiler. Figure 2 The calculation process of the exhaust gas calorific value dynamic observer of the present invention; Figure 3 This is the lookup process for the three-dimensional mapping table of the safety air excess coefficient of the present invention; Figure 4 This is the time misalignment allocation process for the actuator of the present invention; Figure 5 This is the signal preprocessing flow for the online gas chromatograph of the present invention; Figure 6 The proportional-integral-derivative feedback control process with conditional integral is described in this invention. Detailed Implementation

[0018] refer to Figure 1In one embodiment, an online gas chromatograph collects carbon monoxide volume fraction, hydrogen volume fraction, methane volume fraction, and total exhaust gas flow rate signals at the inlet of the exhaust gas delivery pipeline. The sampling point of the online gas chromatograph is located on the straight section of the exhaust gas delivery pipeline inlet, at a distance of no less than 10 times the pipeline diameter from upstream bends, valves, reducers, or other flow obstructions, and at a distance of no less than 20 times the pipeline diameter from the downstream burner inlet. This ensures that the collected exhaust gas samples have sufficient flow field representativeness and avoids sampling concentration deviations caused by flow field disturbances within the pipeline. The sampling period of the online gas chromatograph is set to 1 second, consistent with the basic control cycle of the boiler thermal control system. The collected signals include carbon monoxide volume fraction, hydrogen volume fraction, and methane volume fraction, all dimensionless volume percentage values, as well as the total exhaust gas flow rate signal, in standard cubic meters per hour. The collected analog signals are transmitted to the analog input module of the boiler distributed control system via a 4-20mA hardwired connection. After 16-bit analog-to-digital conversion, digital signals are generated and enter the subsequent data processing stage.

[0019] The carbon monoxide volume fraction, hydrogen volume fraction, methane volume fraction, and total tail gas flow rate are input into the tail gas calorific value dynamic observer. Based on the combustion heat release time sequence characteristics of each combustible component, a predicted calorific value sequence for the current moment and a preset future step size is calculated. The preset step size is set to 5 sampling periods, corresponding to a time length of 5 seconds, matching the thermal inertia delay time from combustion in the semi-coke tail gas boiler furnace to the main steam pressure response. The combustion heat release time sequence characteristics of each combustible component correspond to the differences in combustion reaction rates of different combustible components under the rated operating conditions of the furnace. Among them, hydrogen has the fastest combustion reaction rate and the smallest combustion heat release time constant, followed by carbon monoxide, and methane has the slowest combustion reaction rate and the largest combustion heat release time constant. During the calorific value calculation, the time sequence changes of the volume fraction are weighted according to the combustion reaction time constants of each component, matching the actual heat release time sequence in the furnace, so that the calculated calorific value sequence can accurately reflect the heat release pattern after tail gas combustion.

[0020] The target air-fuel ratio is determined by looking up a table in a pre-defined three-dimensional mapping table of excess safety air coefficients based on the predicted calorific value sequence. The three-dimensional mapping table of excess safety air coefficients is pre-constructed and stored in the read-only storage area of ​​the boiler distributed control system, and is a fixed-length three-dimensional array data structure. The three dimensions of the mapping table correspond to the predicted calorific value, the real-time boiler load command, and the flue gas temperature, respectively. Each dimension has a corresponding value range and grid division step size, and each grid node stores a basic excess air coefficient. The target air-fuel ratio is obtained by multiplying the interpolated target excess air coefficient by the theoretical air volume required for complete combustion of a unit volume of exhaust gas. The theoretical air volume is pre-calculated based on the combustion chemical reaction equations of the combustible components in the exhaust gas. The basic grid division rules for the three-dimensional mapping table of excess safety air coefficients are shown in Table 1.

[0021] Table 1. Three-dimensional mapping table of excess safety air coefficient and basic mesh parameter table. Dimension Name Range of values Grid step size Number of nodes Predicted calorific value 4000-12000kJ / Nm³ 200kJ / Nm³ 41 Boiler real-time load command 20%-100% of rated load 2% of rated load 41 Smoke temperature 120-180℃ 2℃ 31 Table 1 provides a standardized coordinate space reference for table lookup interpolation calculations. The value ranges for each dimension are predetermined based on the rated operating conditions and extreme fluctuation conditions of the semi-coke tail gas boiler, and the grid step size is set according to the accuracy requirements of boiler combustion control to ensure that the resolution of the interpolation calculation meets the control requirements.

[0022] The target air-fuel ratio is converted into frequency adjustment commands for the blower inverter and opening commands for the exhaust gas regulating valve, and output to the corresponding actuators to perform feedforward pre-adjustment. The target air-fuel ratio is the ratio of the standard air volume to the fuel volume required for complete combustion of a unit volume of fuel. Based on the target air-fuel ratio and the total exhaust gas flow signal, the theoretical total air volume requirement is calculated. Combined with the pre-calibrated blower air volume-frequency characteristic curve, the theoretical total air volume requirement is converted into a frequency adjustment command for the blower inverter. The value range of the frequency adjustment command corresponds to the 0-50Hz rated operating range of the blower inverter. Simultaneously, based on the matching requirements between the target air-fuel ratio and the current furnace combustion conditions, the required adjustment value for the exhaust gas fuel quantity is calculated. Combined with the pre-calibrated exhaust gas regulating valve flow-opening characteristic curve, the required adjustment value for the exhaust gas fuel quantity is converted into an opening command for the exhaust gas regulating valve. The value range of the opening command corresponds to the 0-100% full stroke range of the valve. The feedforward pre-regulation has a higher execution priority than the feedback correction stage. When the predicted calorific value sequence changes, the feedforward regulation command is output to the actuator first, and the pre-matching of fuel quantity and air supply is completed during the boiler thermal inertia delay period.

[0023] The main steam pressure signal of the boiler is acquired. A proportional-integral-derivative (PID) feedback controller corrects the residual error generated by the feedforward pre-regulation based on the deviation between the main steam pressure signal and the set pressure value. This correction command is then superimposed on the feedforward pre-regulation command to form a closed-loop control command. The main steam pressure signal is acquired by a high-precision pressure transmitter installed on the boiler's main steam outlet pipeline. The sampling period is consistent with that of the online gas chromatograph. The set pressure value is the rated main steam pressure value for boiler operation, preset by the operator or issued by the upper-level unit coordination control system. The deviation between the main steam pressure signal and the set pressure value is input to the PID feedback controller. The correction command output by the controller compensates for operating condition fluctuations not covered by the feedforward pre-regulation, including changes in furnace heat loss, changes in heat transfer efficiency due to coking on the boiler heating surfaces, calorific value deviations caused by fluctuations in non-combustible components in the exhaust gas, and heat loss fluctuations caused by changes in ambient temperature. After being superimposed on the feedforward pre-regulation command, the correction command is output to the blower frequency converter and the exhaust gas regulating valve actuators, respectively, forming a complete closed-loop control circuit.

[0024] In this embodiment, by collecting the volume fraction of combustible components and the total flow rate of exhaust gas at the inlet of the exhaust gas delivery pipeline, direct detection of combustion characteristics on the fuel input side is achieved, moving the control trigger node from the main steam pressure end after thermal inertia transmission to the fuel input end. A predicted calorific value sequence is generated by the exhaust gas calorific value dynamic observer, enabling early prediction of fuel calorific value changes. The target air-fuel ratio is determined based on a three-dimensional mapping table, and feedforward pre-adjustment is performed, completing the pre-matching of fuel quantity and air supply during the thermal inertia delay period. Combined with proportional-integral-derivative feedback of the main steam pressure to correct the feedforward residual error, a complete feedforward-feedback closed-loop control system is formed, eliminating control response lag caused by boiler thermal inertia delay and avoiding fuel-air mismatch.

[0025] refer to Figure 2 In a preferred embodiment, the calculation process of the exhaust gas calorific value dynamic observer includes establishing a mechanistic calculation model that includes the standard calorific value constants of each combustible component. The mechanistic calculation model is based on the fundamental thermochemical principles of combustion reactions, using the volume fraction of each major combustible component in the exhaust gas as input and the lower heating value per unit volume of exhaust gas as output, covering the heat contribution of combustion of the three major combustible components: carbon monoxide, hydrogen, and methane. The basic physicochemical constants required for the mechanistic calculation model are shown in Table 2.

[0026] Table 2 Standard Combustion Thermal Parameters of Main Combustible Components in Semi-coke Tail Gas

[0027] The standard lower heating value is the lower heating value released by the complete combustion of a unit volume of combustible components under standard conditions, after deducting the latent heat of vaporization of water vapor. The combustion reaction time constant is used to characterize the combustion heat release rate characteristics of each component under the rated operating conditions of the furnace. The parameters are pre-calibrated through boiler hot combustion tests.

[0028] The baseline calorific value calculation process of the mechanistic calculation model is achieved through the following formula:

[0029] in, The baseline calorific value output by the mechanism calculation model at time t; , , The volume fractions of carbon monoxide, hydrogen, and methane collected at time t are respectively. , , The standard lower heating constants for carbon monoxide, hydrogen, and methane are respectively determined by the parameters in Table 2.

[0030] Acquire the data sequence of carbon monoxide volume fraction, hydrogen gas integral, methane volume fraction, and corresponding measured calorific value for a historical time period. The length of the historical time period is set to 720 hours, corresponding to the data accumulation period of the boiler's continuous operation under rated conditions. The sampling interval of the historical data is consistent with the sampling period of the online gas chromatograph, forming corresponding time-series data pairs. The measured calorific value data is obtained through boiler heat balance back-calculation. Specifically, the actual combustion efficiency of the boiler is calculated based on the flue gas composition data collected by the flue gas analyzer installed at the furnace outlet. Combined with the main steam flow rate, feedwater temperature, and main steam temperature, the effective heat output of the boiler is calculated, and the measured calorific value of the tail gas at the corresponding time is obtained by back-calculation, ensuring that the measured calorific value data completely matches the actual combustion heat release process of the boiler.

[0031] A residual sequence is constructed using the measured calorific value data sequence and the output value sequence of the mechanistic calculation model. The calculation process of the residual sequence is achieved through the following formula:

[0032] in, Let be the residual value at time t; Let be the measured calorific value at time t; The reference calorific value is the output of the mechanism calculation model at time t. The residual sequence consists of residual values ​​from multiple consecutive sampling times arranged in chronological order. The residual sequence includes nonlinear errors not covered by the mechanism calculation model, such as operating condition fluctuations in the combustion efficiency of combustible components, the calorific value contribution of trace combustible components in the exhaust gas, systematic biases in the measurement signal, and dynamic changes in boiler heat transfer efficiency.

[0033] The residual sequence is input into a Long Short-Term Memory (LSTM) network for training to obtain a residual prediction model. The LTM network has an input layer dimension of 3, corresponding to the residual values ​​at three consecutive sampling times; two hidden layers with 64 neurons per layer; and an output layer dimension of 6, corresponding to the predicted residual values ​​at the current time and five preset step sizes in the future. The network uses the hyperbolic tangent function as its activation function and the mean squared error function as its loss function. An adaptive moment estimation optimizer is used for iterative weight updates during training. The training set to test set ratio is 8:2. An early stopping mechanism is implemented during training: training stops when the loss function on the validation set does not decrease for 10 consecutive training batches to prevent overfitting.

[0034] A forgetting gating variable is introduced into the hidden layer state update calculation of the Long Short-Term Memory (LSTM) network. The forgetting factor value of the forgetting gating variable is iteratively updated according to the number of training batches, wherein the forgetting factor value decays exponentially with the increase of the number of training batches. The iterative update process of the forgetting factor is implemented by the following formula:

[0035] in, The forgetting factor value is assigned to the k-th training batch. The initial value for the forgetting factor is set to 0.99; α is the decay coefficient, set to... k represents the number of training batches, which increases from 1.

[0036] During the forward propagation of the Long Short-Term Memory (LSTM) network, the historical hidden layer states are decay-weighted according to the forgetting factor value of the current iteration. The current hidden layer state is calculated based on the decay-weighted historical hidden layer states and the current input residual sequence. The predicted residual value is then output based on the current hidden layer state. The update process of the forgetting gate variable with forgetting factor is implemented using the following formula:

[0037] in, Let be the output value of the forgetting gating variable at time t; It is the sigmoid activation function; Here is the weight matrix for the forget gate; The hidden layer state at time t-1; Let be the input residual sequence at time t; This is the bias term for the forget gate. The forget gate variable is used to control the proportion of historical cell states retained. The decay-weighted historical hidden layer states reduce the impact of historical data in early training batches on the current hidden layer state update, improve the network's ability to fit recent heat value fluctuations, and avoid the accumulation of prediction errors caused by operating condition deviations in early historical data.

[0038] During the real-time operation phase, the baseline calorific value output by the mechanism calculation model is added to the residual prediction value output by the residual prediction model to generate the predicted calorific value sequence for the current moment and the future preset step size. The calculation process of the predicted calorific value sequence is implemented through the following formula:

[0039] in, for Predicted calorific value at any given time; The prediction step size is set to 0, 1, 2, 3, 4, 5, where... =0 corresponds to the current time. =1 to 5 correspond to the preset step size for the next 5 sampling periods; This is obtained by extrapolation based on the combustion heat release time sequence characteristics of each combustible component. Time-based reference calorific value; For the output of the residual prediction model Predicted time-residual values.

[0040] In this embodiment, a benchmark calculation of the calorific value of exhaust gas was achieved through a mechanistic calculation model, covering the combustion heat contribution of the main combustible components. A residual sequence was constructed by comparing the measured calorific value with the benchmark calorific value, and nonlinear error features not covered by the mechanistic model were extracted. A long short-term memory network was used to train and predict the residual sequence, achieving fitting of the nonlinear fluctuations in calorific value. By introducing a forgetting factor that decays exponentially with the training batch to attenuate and weight the historical hidden layer states, the interference of early historical data on the current prediction was reduced, improving the adaptability of the calorific value prediction model to dynamic changes in operating conditions and providing a predicted calorific value sequence for feedforward pre-adjustment.

[0041] refer to Figure 3 In another preferred embodiment, the construction and lookup process of the three-dimensional mapping table for the safety excess air coefficient includes establishing an initial excess air coefficient space using the predicted calorific value sequence, the boiler real-time load command, and the flue gas temperature as three-dimensional coordinate axes. The value range of the predicted calorific value dimension corresponds to the extreme calorific value fluctuation range of the semi-coke tail gas, the value range of the boiler real-time load command dimension corresponds to the boiler's operating range of 20%-100% rated load, and the value range of the flue gas temperature dimension corresponds to the ±30% fluctuation range of the flue gas temperature under the boiler's rated operating conditions. These three dimensions together constitute a three-dimensional initial excess air coefficient space, covering all normal operating conditions and extreme fluctuation conditions of the boiler.

[0042] The initial excess air coefficient space is divided into grid nodes according to a preset step size, and the corresponding basic excess air coefficient is stored in each grid node. The basic excess air coefficient of the grid node is predetermined through boiler hot-state combustion test. For each grid node, based on the predicted calorific value, boiler load, and flue gas temperature combination, the excess air coefficient is adjusted to minimize the sum of flue gas heat loss and incomplete combustion heat loss, while ensuring that the carbon monoxide volume fraction in the flue gas at the furnace outlet is below a preset safety threshold. The excess air coefficient under this condition is stored as the basic excess air coefficient of the corresponding grid node, ensuring that the parameters of each grid node correspond to the safe and economical operating conditions of the boiler.

[0043] During the table lookup process, the predicted calorific value at the current moment, the current real-time boiler load command, and the current flue gas temperature are used as index coordinates. A trilinear interpolation algorithm is employed to perform interpolation calculations between eight adjacent grid nodes, and the interpolation result is used as the target air-fuel ratio. The calculation process of the trilinear interpolation algorithm is implemented using the following formula:

[0044] in, The target excess air coefficient is calculated by interpolation; , , , , , , , The index coordinate represents the basic excess air coefficient corresponding to the 8 adjacent grid nodes; x, y, and z are the normalized offsets of the index coordinates relative to adjacent grid nodes in the three dimensions of predicted calorific value, real-time boiler load command, and flue gas temperature, respectively, with values ​​ranging from [0,1]. The target air-fuel ratio is obtained by multiplying the target excess air coefficient by the theoretical air volume required for complete combustion of a unit volume of fuel. The theoretical air volume is pre-calculated based on the combustion chemical reaction equations of the combustible components.

[0045] The interpolation calculation process of the trilinear interpolation algorithm also includes boundary constraint correction, determining whether the index coordinates exceed the boundaries of the initial excess air coefficient space in each dimension of the three-dimensional coordinate axis. When a certain dimension of the index coordinates is lower than the minimum boundary value, that dimension is locked as the minimum boundary value; when the predicted calorific value dimension exceeds the maximum boundary value, the predicted calorific value dimension is locked as the maximum boundary value, and the furnace negative pressure measurement value at the current moment is obtained. The furnace negative pressure measurement value is acquired by a micro differential pressure transmitter installed at the furnace outlet, and the sampling period is consistent with the basic control period of the control system.

[0046] A negative pressure bias coefficient is calculated based on the deviation between the measured furnace negative pressure value and the preset negative pressure reference value. This negative pressure bias coefficient is then multiplied by the interpolation result locked at the maximum boundary value to generate the corrected target air-fuel ratio. The calculation process for the negative pressure bias coefficient is achieved using the following formula:

[0047] in, This is the negative voltage bias coefficient; The proportional gain coefficient for negative pressure deviation is set to 0.01 / Pa; This is the preset furnace negative pressure reference value; This represents the current furnace negative pressure measurement. When the furnace negative pressure measurement is lower than the reference value, it indicates that the combustion conditions in the furnace are weak, and the negative pressure offset coefficient is greater than 1. When the furnace negative pressure measurement is higher than the reference value, it indicates that the combustion conditions in the furnace are strong, and the negative pressure offset coefficient is less than 1. The target air-fuel ratio is reduced to decrease the air supply volume and avoid the divergence of control commands under extreme calorific value fluctuations.

[0048] The process of converting the target air-fuel ratio into the frequency adjustment command of the blower inverter and the opening command of the exhaust gas regulating valve includes obtaining the current actual operating frequency of the blower and the current opening of the exhaust gas regulating valve, and calculating the change in total air volume demand and the change in total fuel demand based on the difference between the target air-fuel ratio and the current actual air-fuel ratio. The calculation of the change in total air volume demand and the change in total fuel demand is achieved through the following formulas:

[0049] in, This represents the change in total air volume demand. This represents the change in total fuel demand. This represents the total exhaust gas flow rate at the current moment. This is the theoretical amount of air required for the complete combustion of a unit volume of exhaust gas. The target excess air coefficient; The current actual excess air coefficient is calculated from the current supply air volume and the total exhaust gas flow rate.

[0050] refer to Figure 4 The response time constants of the damper actuator and the valve actuator are introduced. Based on these response time constants, the time-displacement allocation of the changes in total air volume demand and total fuel demand is performed to generate the frequency adjustment command of the blower inverter and the opening command of the exhaust gas regulating valve with timing delay compensation. The dynamic characteristic parameters of the damper actuator and the valve actuator are shown in Table 3. Table 3 Dynamic Response Characteristic Parameters of Actuators Name of implementing agency Second-order transfer function damping ratio Natural oscillation frequency (rad / s) Response time constant (s) Rise time (s) Peak time (s) blower damper actuator 0.707 8.5 0.6 0.35 0.52 Exhaust gas regulating valve actuator 0.707 12.6 0.4 0.22 0.35 Table 3 provides benchmark data for the calculation of the actuator response time constant and the time misalignment allocation. The parameters are determined in advance through the step response test of the actuator. The response time constant is defined as the time required for the step response curve to rise from 0 to 95% of the steady state value.

[0051] A second-order transfer function model incorporating the dynamic characteristics of the damper actuator and the valve actuator is constructed. The expression of the second-order transfer function model is as follows:

[0052] in, For the second-order transfer function of the actuator; For the Laplace operator; The damping ratio of the actuator; The damping ratio and natural oscillation frequency are determined by the parameters in Table 3, corresponding to the dynamic characteristics of the damper actuator and valve actuator, respectively.

[0053] A step signal is input into the second-order transfer function model, and the output response curve is recorded. Based on the rise time and peak time of the output response curve, the time difference between the response time constant of the damper actuator and the response time constant of the valve actuator is calculated. The calculation of the time difference is achieved through the following formula:

[0054] in, This represents the difference in response time between the damper actuator and the valve actuator. The response time constant of the damper actuator; This is the response time constant of the valve actuator.

[0055] Based on the response time constant of the valve actuator, the change in total fuel demand is allocated in real time, while the change in total air volume demand is allocated with a time difference lag. When the response speed of the valve actuator is faster than that of the damper actuator, the time difference is positive, and the output command for the change in total air volume demand is sent to the blower frequency converter with a time difference lag. This ensures that the change in air volume and the change in exhaust gas fuel volume are synchronized in the furnace, compensating for the timing differences in the physical actions of the actuators. This avoids oxygen-deficient combustion caused by fuel entering the furnace before air volume, or furnace negative pressure fluctuations caused by air volume entering the furnace before fuel volume.

[0056] In this embodiment, a standardized lookup table is provided for determining the target air-fuel ratio by constructing an initial excess air coefficient space and grid nodes in a three-dimensional coordinate system. A trilinear interpolation algorithm is used to continuously calculate the target excess air coefficient under non-grid node conditions, improving the smoothness of air-fuel ratio control. By introducing boundary constraint correction and a furnace negative pressure bias coefficient, control command divergence under extreme calorific value fluctuations is avoided, ensuring combustion safety under boundary conditions. The dynamic characteristics of the actuator are constructed using a second-order transfer function model, the response time difference is calculated, and time-displaced allocation is performed on changes in airflow and fuel demand, compensating for the temporal differences in the physical actions of the actuator and achieving synchronous matching of fuel and air within the furnace.

[0057] refer to Figure 5 In another preferred embodiment, the process of acquiring signals through the online gas chromatograph further includes preprocessing the acquired signals to obtain the raw chromatographic peak data of the online gas chromatograph in the current sampling period. The raw chromatographic peak data is output by the detector of the online gas chromatograph, containing response signal values ​​corresponding to different retention times. The data length matches the scan cycle of the chromatographic analysis, covering the peak time range of all combustible components.

[0058] The original chromatographic peak data were subjected to baseline drift subtraction to obtain the initial volume fraction. The baseline drift subtraction process employed a polynomial fitting method, performing a third-order polynomial fit on the peakless regions of the original chromatographic peak data to obtain the baseline drift curve. The baseline drift curve was then subtracted from the original chromatographic peak data to obtain the corrected chromatographic peak data. Based on the corrected peak area and a pre-calibrated component concentration-peak area calibration curve, the initial volume fraction of each combustible component was calculated, thus eliminating the systematic drift error of the chromatographic detection signal.

[0059] A time window sequence is constructed by extracting the initial volume fractions from the current sampling period and several consecutive previous sampling periods using the sliding window method. The mean and standard deviation of the time window sequence are then calculated. The length of the sliding window is set to 20 sampling periods, corresponding to a time length of 20 seconds. The time window is updated as the sampling period progresses to ensure that the statistical characteristics can reflect recent signal changes.

[0060] When the initial volume fraction of the current sampling period deviates from the mean by more than a preset multiple of the standard deviation, the initial volume fraction of the current sampling period is marked as an outlier and removed. The outlier determination process is implemented using the following formula:

[0061] in, This represents the initial volume fraction for the current sampling period; The mean of the time window series; The standard deviation of the time window series; The standard deviation is set to 3. When the above inequality holds, the initial volume fraction of the current sampling period is marked as an outlier and removed, eliminating the interference of sudden abnormal detection data on subsequent control logic.

[0062] After marking and removing the initial volume fraction of the current sampling period as an outlier, a missing data compensation process is also included. Multiple redundant gas sensors are deployed at different spatial locations upstream and downstream of the exhaust gas delivery pipe inlet. The redundant gas sensors include a non-dispersive infrared gas sensor and a thermal conductivity gas sensor. The non-dispersive infrared gas sensor is used to detect the volume fraction of carbon monoxide and methane, while the thermal conductivity gas sensor is used to detect the volume fraction of hydrogen, forming a redundant backup with the main online gas chromatograph.

[0063] The spatial volume fraction matrix collected by the multiple redundant gas sensors at the removal time point marked as an outlier is obtained. Then, a historical spatial volume fraction matrix sequence is extracted from each preset time window extending forward and backward from the removal time point. The preset time window length is set to 10 sampling periods, corresponding to a time length of 10 seconds. The historical spatial volume fraction matrix sequence contains continuously collected data from each redundant sensor within the time window, used to calculate the temporal correlation characteristics between the sensors.

[0064] Based on the correlation weights of each spatial location in the historical spatial volume fraction matrix sequence, spatial interpolation is performed on the spatial volume fraction matrix at the removal time point, and the spatial interpolation result replaces the removed initial volume fraction. The spatial deployment locations and initial correlation weights of redundant gas sensors are shown in Table 4.

[0065] Table 4 Spatial Deployment and Correlation Weights of Redundant Gas Sensors Sensor number Deployment location Spatial distance from the main chromatograph (m) Initial association weight Sampling period (s) 1 upstream of the same cross section of the main chromatograph 0.5 0.3 1 2 Downstream of the main chromatograph 0.5 0.3 1 3 Pipe radial left side section 0 0.2 1 4 Pipe radial right side section 0 0.2 1 Table 4 provides baseline parameters for spatial interpolation calculation of missing data. The sensors are deployed at different cross-sectional locations at the inlet of the exhaust gas delivery pipeline, covering different spatial points in the radial and axial directions of the pipeline, to ensure that the collected signals can reflect the spatial distribution characteristics of the exhaust gas concentration in the pipeline.

[0066] Spatial interpolation is performed using the following formula:

[0067] in, The compensation volume fraction is denoted by t; N is the number of redundant gas sensors. Let be the association weight of the i-th redundant gas sensor at time t; Let t represent the volume fraction collected by the i-th redundant gas sensor at time t. The correlation weights are calculated based on the historical spatial volume fraction matrix sequence, specifically by calculating the temporal correlation between each redundant sensor and the main chromatograph using the Pearson correlation coefficient. The higher the correlation, the greater the corresponding correlation weight, with a total weight of 1. Spatial interpolation effectively compensates for missing data after outlier removal, ensuring the continuity and reliability of the detection signal.

[0068] refer to Figure 6 The correction process via the proportional-integral-derivative (PID) feedback controller includes setting a conditional integral enable switch in the integral stage of the controller. When the absolute value of the deviation between the main steam pressure signal and the set pressure value exceeds a preset dead zone threshold, the conditional integral enable switch is closed for integral accumulation. When the absolute value of the deviation is less than or equal to the preset dead zone threshold, the conditional integral enable switch is opened and the accumulated integral value remains unchanged. The dead zone threshold is set based on the rated operating fluctuation range of the boiler's main steam pressure, avoiding integral accumulation under small pressure fluctuations.

[0069] The correction command is generated by adding the output values ​​of the proportional element, the output values ​​of the derivative element, and the held integral cumulative value. The output calculation process of the proportional-integral-derivative controller is implemented through the following formula:

[0070] This is the correction command output by the controller at time t; This is the gain coefficient of the proportional element; This is the gain coefficient of the integrator; is the gain coefficient of the differentiating element; The deviation between the main steam pressure signal and the set pressure value at time t; for The deviation between the main steam pressure signal and the set pressure value at any given time; For the conditional integral enabling switch state variable, when hour, =1, the conditional integration enable switch is closed, and integration is performed; when hour, =0, the conditional integral enable switch is off, and the accumulated integral value remains unchanged; The preset dead zone threshold is used. The proportional element is used to quickly respond to pressure deviations, the derivative element is used to predict the changing trend of pressure deviations, the integral element is used to eliminate static deviations of the system, and the conditional integral enable switch balances the needs of static deviation elimination and integral saturation suppression.

[0071] In this embodiment, baseline drift subtraction is performed on the raw chromatographic peak data to eliminate system drift error in the chromatographic detection signal. A time window sequence is constructed using the sliding window method, and outlier detection and removal are performed based on the mean and standard deviation, eliminating interference from sudden abnormal detection data on the control logic. Spatial interpolation is used to compensate for missing data through the spatial deployment and correlation weight calculation of multiple redundant sensors, ensuring the continuity and reliability of the detection signal. By setting a conditional integral enable switch based on a dead-zone threshold in the integral stage of the proportional-integral-derivative controller, integral saturation under small deviations is eliminated, improving the stability and control accuracy of the feedback control loop.

Claims

1. A safe and intelligent control method for the thermal system of a semi-coke tail gas boiler, characterized in that, include: At the inlet of the exhaust gas delivery pipeline, the volume fraction of carbon monoxide, the integral number of hydrogen gas, the volume fraction of methane, and the total exhaust gas flow rate are collected using an online gas chromatograph. The carbon monoxide volume fraction, hydrogen gas integral, methane volume fraction, and total exhaust gas flow rate signals are input into the exhaust gas calorific value dynamic observer. Based on the combustion heat release time sequence characteristics of each combustible component, the predicted calorific value sequence for the current moment and the future preset step size is calculated. The target air-fuel ratio is determined by looking up the predicted calorific value sequence in a preset three-dimensional mapping table of excess safety air coefficient. The target air-fuel ratio is converted into frequency adjustment commands for the blower inverter and opening commands for the exhaust gas regulating valve, and then output to the corresponding actuators to perform feedforward pre-adjustment. The main steam pressure signal of the boiler is acquired, and the residual error generated by the feedforward pre-adjustment is corrected by the proportional-integral-derivative feedback controller based on the deviation between the main steam pressure signal and the set pressure value. The correction command is generated and superimposed with the feedforward pre-adjustment command to form a closed-loop control command.

2. The intelligent control method for the thermal system of a semi-coke tail gas boiler according to claim 1, characterized in that, The calculation process of the exhaust gas calorific value dynamic observer includes: establishing a mechanism calculation model that includes the standard combustion calorific value constants of each combustible component; Obtain the data sequence of carbon monoxide volume fraction, hydrogen gas integral, methane volume fraction, and corresponding measured calorific value at a historical time period; A residual sequence is constructed using the measured calorific value data sequence and the output value sequence of the mechanism calculation model; The residual sequence is input into a long short-term memory network for training to obtain a residual prediction model; During the real-time operation phase, the baseline calorific value output by the mechanism calculation model is added to the residual prediction value output by the residual prediction model to generate the predicted calorific value sequence for the current moment and the future preset step size.

3. The intelligent control method for the thermal system of a semi-coke tail gas boiler according to claim 1, characterized in that, The process of constructing and looking up the three-dimensional mapping table of the safety excess air coefficient includes: establishing an initial excess air coefficient space with the predicted calorific value sequence, real-time boiler load command, and flue gas temperature as three-dimensional coordinate axes; The initial excess air coefficient space is divided into grid nodes according to a preset step size, and the corresponding basic excess air coefficient is stored in each grid node. During the table lookup process, the predicted calorific value at the current moment, the current real-time load command of the boiler, and the current flue gas temperature are used as index coordinates. The trilinear interpolation algorithm is used to perform interpolation calculations between 8 adjacent grid nodes, and the interpolation calculation results are used as the target air-fuel ratio.

4. The intelligent control method for the thermal system of a semi-coke tail gas boiler according to claim 1, characterized in that, The process of converting the target air-fuel ratio into the frequency adjustment command of the blower inverter and the opening command of the exhaust gas regulating valve includes: obtaining the current actual operating frequency of the blower and the current opening of the exhaust gas regulating valve; The changes in total air volume demand and total fuel demand are calculated based on the difference between the target air-fuel ratio and the current actual air-fuel ratio. By introducing the response time constants of damper actuator and valve actuator, and by allocating the time misalignment of the total air volume demand change and the total fuel demand change based on the response time constants, the frequency adjustment command of the blower inverter and the opening command of the exhaust gas regulating valve with time delay compensation are generated.

5. The intelligent control method for the thermal system of a semi-coke tail gas boiler according to claim 1, characterized in that, The correction process via the proportional-integral-derivative feedback controller includes setting a conditional integration enable switch in the integral stage of the proportional-integral-derivative feedback controller. When the absolute value of the deviation between the main steam pressure signal and the set pressure value exceeds the preset dead zone threshold, the conditional integration enable switch is closed to perform integration accumulation. When the absolute value of the deviation is less than or equal to the preset dead zone threshold, the conditional integration enable switch is turned off and the cumulative integration value remains unchanged. The correction instruction is generated by adding the output values ​​of the proportional element, the output values ​​of the derivative element, and the accumulated integral value.

6. The intelligent control method for the thermal system of a semi-coke tail gas boiler according to claim 1, characterized in that, The process of acquiring signals through the online gas chromatograph also includes preprocessing the acquired signals: obtaining the raw chromatographic peak data of the online gas chromatograph in the current sampling period; The initial volume fraction was obtained by performing baseline drift subtraction on the original chromatographic peak data; The sliding window method is used to extract the initial volume fraction of the current sampling period and the previous multiple consecutive sampling periods to form a time window sequence; Calculate the mean and standard deviation of the time window series; When the initial volume fraction of the current sampling period deviates from the mean by more than a preset multiple of the standard deviation, the initial volume fraction of the current sampling period is marked as an outlier and removed.

7. The intelligent control method for the thermal system of a semi-coke tail gas boiler according to claim 2, characterized in that, The training process of the residual prediction model includes: introducing a forgetting gating variable into the hidden layer state update calculation of the long short-term memory network; The forgetting factor value of the forgetting gate variable is updated iteratively based on the number of training batches, wherein the forgetting factor value decays exponentially with the increase of the number of training batches; During the forward propagation of the Long Short-Term Memory network, the historical hidden layer state is decayed and weighted according to the forgetting factor value of the current iteration, and the current hidden layer state is calculated based on the decayed and weighted historical hidden layer state and the current input residual sequence. The residual prediction value is output based on the current hidden layer state.

8. The intelligent control method for the thermal system of a semi-coke tail gas boiler according to claim 3, characterized in that, The interpolation calculation process of the trilinear interpolation algorithm also includes boundary constraint correction: determining whether the index coordinates exceed the boundary of the initial air excess coefficient space in each dimension of the three-dimensional coordinate axis; When the predicted calorific value dimension exceeds the maximum boundary value, the predicted calorific value dimension is locked to the maximum boundary value, and the furnace negative pressure measurement value at the current moment is obtained. The negative pressure offset coefficient is calculated based on the deviation between the measured negative pressure value in the furnace and the preset negative pressure reference value; The negative pressure bias coefficient is multiplied by the interpolation result locked to the maximum boundary value to generate the corrected target air-fuel ratio.

9. The intelligent control method for the thermal system of a semi-coke tail gas boiler according to claim 4, characterized in that, The process of allocating the changes in total air volume demand and total fuel demand in a time-staggered manner includes: constructing a second-order transfer function model that includes the dynamic characteristics of the damper actuator and the valve actuator; Input a step signal into the second-order transfer function model and record the output response curve; The time difference between the response time constant of the damper actuator and the response time constant of the valve actuator is calculated based on the rise time and peak time of the output response curve. Based on the response time constant of the valve actuator, the change in total fuel demand is allocated in real time, and the change in total air volume demand is allocated with a time difference.

10. The intelligent control method for the thermal system of a semi-coke tail gas boiler according to claim 6, characterized in that, After marking the initial volume fraction of the current sampling period as an outlier and removing it, a missing data compensation process is also included: deploying multiple redundant gas sensors at different spatial locations upstream and downstream of the inlet of the exhaust gas delivery pipeline; Obtain the spatial volume fraction matrix collected by the multiple redundant gas sensors at the removal time points marked as outliers; Extract the historical spatial volume fraction matrix sequence from each preset time window before and after the removal time point; Based on the correlation weights of each spatial location in the historical spatial volume fraction matrix sequence, spatial interpolation is performed on the spatial volume fraction matrix at the removal time point, and the spatial interpolation result replaces the initial volume fraction that was removed.

Images