A method and system for collaborative regulation of pollutants in a total discharge port of multiple furnaces

Abstract

The present application relates to the technical field of flue gas emission treatment, and discloses a total exhaust port pollutant multi-furnace collaborative regulation method and system, which comprises the following steps: collecting multi-source heterogeneous data to generate a multi-furnace collaborative time sequence feature matrix, constructing a multi-furnace coupling space-time prediction model to extract the space-time coupling contribution weight of each furnace position, solving the global optimal distribution value of the ammonia injection amount and the slurry circulation amount of each furnace with double-target constraints, outputting a real-time optimal regulation instruction set after iterative optimization, performing local feedback correction on the ammonia injection instruction as a feedforward ammonia injection amount after amplitude and rate limitation to obtain a final adjustment valve opening degree instruction and an ammonia water flow set value, scheduling the number of operating stages and the frequency of the slurry circulating pumps of each furnace in stages, calculating the concentration deviation rate in real time, triggering model rolling recalculation when the concentration deviation rate exceeds the standard, obtaining real-time operation feedback data, obtaining four-dimensional evaluation indexes to calculate time sequence concentration residual error, reversely updating the multi-furnace coupling space-time prediction model and generating a regulation effect evaluation report, and solving the problem of concentration fluctuation exceeding the standard of the multi-furnace total exhaust.

Description

Technical Field

[0001] This invention relates to the field of flue gas emission treatment technology, specifically to a method and system for coordinated control of pollutants from multiple furnaces at the total discharge outlet. Background Technology

[0002] Industrial flue gas ultra-low emission distributed coordinated control systems typically collect flue gas from various boilers at a central discharge outlet for unified desulfurization, denitrification, and other end-of-pipe treatment. However, the operating load, coal quality, and combustion conditions of each boiler frequently fluctuate, leading to drastic changes in the amount of flue gas and pollutant concentrations entering the central discharge outlet. Traditional control methods independently regulate each boiler but lack global coordination, or rely solely on reactive adjustments based on central discharge outlet monitoring data. This results in control lag and strong coupling between multiple boiler operating conditions, easily causing large fluctuations in emission concentrations and making it difficult to stably meet ultra-low emission requirements. It also leads to over- or under-adjustment of control equipment, increasing energy consumption and operating costs.

[0003] Chinese patent CN120502214B discloses a multi-pollutant collaborative control system for boiler exhaust gas based on PLC and frequency converter linkage, including a data acquisition module, a PLC control module, and a collaborative optimization module. This invention uses the data acquisition module to collect real-time concentration data of pollutants in boiler exhaust gas and uses an inlet induced draft fan to introduce the boiler exhaust gas into a desulfurization and denitrification tower. The PLC control module uses the PLC's logic controller to set the drive rules of the inlet induced draft fan and dynamically adjusts the operating frequency of the inlet induced draft fan according to the pollutant content in the boiler exhaust gas. The collaborative optimization module returns the non-compliant exhaust gas to the boiler for secondary desulfurization and denitrification, while the compliant exhaust gas is introduced into a dust removal tower through an outlet induced draft fan. The electrostatic precipitator voltage is adjusted according to the ammonia content in the exhaust gas, and a genetic algorithm is used to achieve coordinated control of desulfurization, denitrification, and dust removal with the goal of minimizing operating costs.

[0004] Although existing multi-pollutant collaborative control systems for boiler flue gas based on PLC and frequency converter linkage achieve optimal cost-effective collaborative control of desulfurization, denitrification, and dust removal units by combining PLC with fuzzy adaptive PID and frequency converter linkage to control various environmental protection equipment, and using genetic algorithms, they also achieve dynamic adjustment of flue gas pollutants in a single boiler, graded threshold early warning, and secondary treatment of excessive flue gas recirculation, stabilizing single boiler flue gas to meet standards and reducing operating energy consumption. However, they lack collaborative dynamic balance control capabilities under the condition of multiple boilers sharing a single exhaust outlet, leading to the total exhaust outlet... and The problem of fluctuating emission concentrations and the tendency for instantaneous values ​​to exceed standards. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for the coordinated control of pollutants from multiple furnaces at the total discharge outlet. This method generates a multi-furnace coordinated temporal feature matrix by acquiring multi-source heterogeneous data, constructs a multi-furnace coupled spatiotemporal prediction model, and extracts the spatiotemporal coupling contribution weights of each furnace position through a dual-branch approach. A dual objective function is constructed with the minimum total discharge concentration deviation and the lowest reducing agent consumption. A sequential quadratic programming rolling solver is used to iteratively solve for the globally optimal allocation values ​​of ammonia injection rate and slurry circulation rate for each furnace. Through iterative optimization, a real-time optimal control instruction set is output, performing amplitude rate limiting, local feedback correction, and hierarchical scheduling. The concentration deviation rate is calculated in real time, and if it exceeds the standard, the model is recalculated. Incremental model back-updates are completed using the temporal concentration residuals, and a control effect evaluation report is generated, stabilizing the pollutant emission concentration at the total discharge outlet and improving the compliance rate of the coordinated control of multiple furnaces.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for coordinated control of pollutants from multiple furnaces at the total discharge outlet includes:

[0008] After preprocessing multi-source heterogeneous data, derived features are extracted to generate a multi-furnace collaborative time-series feature matrix.

[0009] A multi-furnace coupled spatiotemporal prediction model is constructed, the spatiotemporal coupling contribution weight of each furnace position is extracted, and the global optimal allocation value of ammonia injection and slurry circulation of each furnace is solved with dual objective constraints. The real-time optimal control instruction set is output through iterative optimization.

[0010] The amplitude and rate of the real-time optimal control command set are limited, and the ammonia injection command is used as the feedforward ammonia injection quantity for local feedback correction to obtain the final control valve opening command and ammonia water flow set value; the number and frequency of operation of each boiler slurry circulation pump are scheduled in stages, the concentration change trend of the total discharge port is monitored in real time to calculate the concentration deviation rate, and when the concentration deviation rate is greater than the warning threshold, the model is triggered to recalculate and obtain real-time operation feedback data.

[0011] The time-series concentration residuals are calculated by obtaining four-dimensional evaluation indicators, and the multi-furnace coupled spatiotemporal prediction model is updated in reverse to generate a regulation effect evaluation report.

[0012] Specifically, the steps for constructing a multi-furnace coupled spatiotemporal prediction model include:

[0013] A multi-furnace coupled spatiotemporal prediction model is constructed by obtaining a target concentration value, including an input layer, a topology map construction layer, a spatiotemporal feature extraction layer, a rolling optimization decision layer, and an output layer.

[0014] The input layer receives the multi-furnace collaborative time-series feature matrix and the concentration target value.

[0015] The topology map construction layer extracts process features from the multi-furnace collaborative time sequence feature matrix and assigns values; calculates the transmission delay from each furnace to the main outlet and sets basic weights; introduces the inter-furnace load deviation correction factor to perform nonlinear weighted correction to generate a dynamic weighted adjacency matrix, forming a spatiotemporal topology map of the furnace group.

[0016] The spatiotemporal feature extraction layer extracts spatial coupling features from the dynamically weighted adjacency matrix through spatial convolution branches and temporal evolution features through temporal convolution branches. After concatenation, the spatiotemporal feature extraction layer is trained with the measured concentration at the total discharge outlet as supervision. The spatiotemporal coupling contribution weights of each denitrification-desulfurization purification unit are decoupled and output to construct a contribution weight matrix, and the initial predicted value of the total discharge outlet concentration is output.

[0017] Specifically, the steps for constructing a multi-furnace coupled spatiotemporal prediction model also include:

[0018] The rolling optimization decision layer receives the initial predicted value of the total discharge outlet concentration and the contribution weight matrix, constructs a dual objective function that minimizes the deviation of the total discharge outlet concentration and the consumption of reducing agent, and transforms it into a single objective function using the dynamic weighting coefficient method.

[0019] Four types of process constraints—actuator range, single-step adjustment rate, compliance red line, and inter-furnace balance—are mapped to the optimization variable space to form a nonlinear constraint optimization problem. A sequential quadratic programming rolling solver is used to synchronously iterate and optimize the ammonia injection rate and slurry circulation rate of each furnace to obtain the global optimal allocation value. The initial optimal control instruction set is then smoothly generated through a sliding window.

[0020] The output layer substitutes the globally optimal allocation value into the contribution weight matrix to calculate the predicted concentration of the total discharge outlet at the next time step for hierarchical correction, and verifies the actuator range and single-step adjustment rate limit. The verified initial optimal control command is then encapsulated into a real-time control command set.

[0021] Specifically, the steps for outputting the real-time optimal control instruction set include:

[0022] Based on the multi-furnace coordinated time series characteristic matrix of 30 historical time steps and the total discharge outlet at the corresponding time, Measured concentration data The measured concentration data were used as the training set. The mean square error between the predicted concentration and the measured concentration at the total discharge outlet was used as the loss function. The Adam optimizer was used to train the multi-furnace coupled spatiotemporal prediction model.

[0023] After training, the multi-furnace collaborative time-series feature matrix at the current moment is input in real time. A real-time spatiotemporal topology map of the furnace group is generated through the topology map construction layer. The contribution weight matrix and the initial predicted value of the total discharge concentration are obtained through forward calculation by the spatiotemporal feature extraction layer. The sequential quadratic programming rolling solver is called to perform iterative optimization and output the real-time optimal control instruction set.

[0024] Specifically, the steps for applying amplitude and rate limits to the real-time optimal control instruction set include:

[0025] The real-time optimal control instruction set is sent to the DCS system, and the amplitude and rate of the real-time optimal control instruction set are limited.

[0026] The difference between the ammonia injection quantity instruction value in the real-time optimal control instruction set and the current actual ammonia injection quantity is calculated to obtain the ammonia injection quantity instruction increment.

[0027] When the absolute value of the ammonia injection quantity command increment exceeds the preset first maximum step size, it is clamped to the positive and negative boundary values ​​of the first maximum step size and then added to the current actual ammonia injection quantity to obtain the limited ammonia injection quantity setting value.

[0028] The difference between the slurry circulation volume command value in the real-time optimal control command set and the current actual slurry circulation volume is calculated to obtain the slurry circulation volume command increment.

[0029] When the absolute value of the increment of the slurry circulation volume command exceeds the preset second maximum step size, it is clamped to the boundary value of the second maximum step size and then added to the current actual slurry circulation volume to obtain the limited slurry circulation volume setting value.

[0030] Specifically, the steps to obtain the final control valve opening command and the ammonia flow rate setpoint include:

[0031] Using the ammonia injection command from the real-time optimal control command set as the feedforward ammonia injection quantity, combined with the furnace outlet... Soft measurement values ​​are locally corrected using feedback.

[0032] furnace outlet Soft measurement values ​​are used as a basis for real-time feedback to obtain the preset furnace outlet. The concentration target value is calculated by using an incremental PI control algorithm to calculate the ammonia injection feedback correction increment. The feedforward ammonia injection amount is then superimposed with the ammonia injection feedback correction increment to obtain the final ammonia injection amount setpoint.

[0033] By combining the inherent flow characteristic curve of the ammonia injection regulating valve with the valve opening command, the corresponding ammonia flow setpoint is obtained through synchronous calculation. This yields the final regulating valve opening command and ammonia flow setpoint, which are then linked to the measured total discharge concentration and furnace outlet concentration. The soft measurement values ​​constitute a time series sample library.

[0034] Specifically, the steps to obtain real-time operational feedback data include:

[0035] The number and frequency of slurry circulation pumps operating in each furnace are scheduled in stages. The limited slurry circulation volume setpoint is decomposed into a discrete combination of the number of fixed-speed pumps and the operating frequency of variable-frequency pumps. The first adjustment is... The circulating pump group corresponding to the furnace position with the largest contribution weight of spatiotemporal coupling;

[0036] A pump set cumulative running time rotation mechanism is introduced. Under the same operating conditions, the circulating pump set with the shorter cumulative running time is started first. The final scheduling command is output after verifying the absorption tower liquid level safety constraint.

[0037] During the regulation process, the trend of total discharge outlet concentration changes is monitored in real time, and the concentration deviation rate is calculated based on the measured total discharge outlet concentration and the concentration target value.

[0038] When the concentration deviation rate exceeds the preset warning threshold, the model is triggered to recalculate and obtain real-time operation feedback data.

[0039] Specifically, the steps for calculating the time-series concentration residuals include:

[0040] Real-time data collection of total discharge outlets during the control cycle Measured concentration and The measured concentration was combined with real-time operational feedback data to calculate a four-dimensional evaluation index. The four-dimensional evaluation index was then normalized and weighted to obtain a comprehensive control score.

[0041] Retrieve the final control valve opening command, ammonia flow setpoint, measured total discharge concentration, and furnace outlet concentration corresponding to the control cycle from the time-series sample library. Soft measurement value;

[0042] The predicted concentration at the total discharge outlet from the multi-furnace coupled spatiotemporal prediction model at the same timestamp is subtracted point by point from the measured concentration value at the total discharge outlet to calculate the time-series concentration residual. The residual is then integrated into a time-series concentration residual sequence by arranging the residuals in ascending order.

[0043] Specifically, the steps for generating a regulation effectiveness evaluation report include:

[0044] Based on the comprehensive control score and the time-series concentration residual sequence, an incremental learning strategy is used to back-update the multi-furnace coupled spatiotemporal prediction model.

[0045] The updated multi-furnace coupled spatiotemporal prediction model is used to backtest the effective samples of the current control cycle, and the comprehensive control score is recalculated.

[0046] If the recalculated comprehensive control score is greater than the previous comprehensive control score, the incremental update is deemed valid, and the update record of the multi-furnace coupled spatiotemporal prediction model is saved.

[0047] An evaluation report on the control effect is generated by integrating four-dimensional evaluation indicators, time-series concentration residual sequences, and update records of multi-furnace coupled spatiotemporal prediction models.

[0048] A multi-furnace coordinated control system for pollutants at the total discharge outlet includes: a multi-source data preprocessing module, a coordinated prediction and optimization module, an instruction execution module, and an evaluation and update module;

[0049] The multi-source data preprocessing module is used to collect multi-source heterogeneous data, preprocess it, extract derived features, and generate a multi-furnace collaborative time-series feature matrix.

[0050] The collaborative prediction and optimization module is used to construct a multi-furnace coupled spatiotemporal prediction model, extract the spatiotemporal coupling contribution weight of each furnace position, solve the global optimal allocation value of ammonia injection and slurry circulation for each furnace with dual objective constraints, and output the real-time optimal control instruction set through iterative optimization.

[0051] The instruction execution module is used to limit the amplitude and rate of the real-time optimal control instruction set, use the ammonia injection instruction as the feedforward ammonia injection amount for local feedback correction, and obtain the final control valve opening instruction and ammonia water flow set value; it also performs hierarchical scheduling of the number and frequency of operation of each boiler slurry circulation pump, monitors the concentration change trend of the total discharge port in real time to calculate the concentration deviation rate, and triggers the model rolling recalculation when the concentration deviation rate is greater than the warning threshold to obtain real-time operation feedback data;

[0052] The evaluation and update module is used to obtain the time-series concentration residuals of the four-dimensional evaluation index calculation, update the multi-furnace coupled spatiotemporal prediction model in reverse, and generate a regulation effect evaluation report.

[0053] The beneficial effects of this invention are:

[0054] 1. This invention constructs a topology map layer, calculates the transmission delay from each furnace to the main discharge port based on the fluid pipeline resistance formula, and introduces an inter-furnace load deviation correction factor to generate a dynamic weighted adjacency matrix, overcoming the defect of fixed topology being unable to adapt to changes in operating conditions; the spatiotemporal feature extraction layer uses a dual-branch structure to extract spatial coupling features and temporal evolution features, and outputs the connection between each denitrification-desulfurization purification unit and the main discharge port via Grad-CAM decoupling. , The contribution weight matrix enables the constructed multi-furnace coupled spatiotemporal prediction model to have physical interpretability; the rolling optimization decision layer constructs a linear prediction relationship between the control increment and the total discharge concentration using the contribution weight matrix, with the minimum concentration deviation and the minimum reducing agent consumption as dual objectives, and solves the global optimal allocation value of ammonia injection and slurry circulation for each furnace through sequential quadratic programming iteration, thus realizing the synergistic optimization of multi-furnace load balance and emission compliance.

[0055] 2. This invention limits the amplitude and rate of the real-time optimal control command set to avoid oscillations in the denitrification and desulfurization system caused by large movements of the actuator; it uses ammonia injection commands as feedforwards to combine with the furnace outlet. Incremental PI feedback correction of soft measurement values ​​forms a feedforward-feedback composite control, improving the response speed and anti-interference capability of the denitrification system to fluctuations in inlet operating conditions; a tiered scheduling system is adopted for the slurry circulation pumps, using a combination of fixed-speed pumps and variable-frequency pumps, first adjusting... The circulating pump group corresponding to the furnace position with the largest contribution weight in spatiotemporal coupling is selected, and a cumulative runtime rotation mechanism is introduced. Combined with the absorber tower liquid level safety constraint, the stable operation of the desulfurization system is ensured. During the control process, the concentration change trend at the total discharge outlet is monitored in real time and the concentration deviation rate is calculated. When the concentration deviation rate exceeds the warning threshold, the model is triggered to recalculate and suppress the total discharge outlet. , The instantaneous concentration exceeded the standard. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of a multi-furnace coordinated control method for pollutants at the total discharge outlet;

[0057] Figure 2 This is a flowchart of the multi-furnace coupled spatiotemporal prediction model constructed in this invention;

[0058] Figure 3 This is a flowchart of the amplitude and rate limiting of the real-time optimal control instruction set in this invention;

[0059] Figure 4 This is a flowchart illustrating the process of obtaining real-time operational feedback data in this invention;

[0060] Figure 5 This is a structural diagram of a multi-furnace coordinated control system for pollutants at the total discharge outlet. Detailed Implementation

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0062] Example 1:

[0063] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a method for coordinated control of pollutants from multiple furnaces at the total discharge outlet, including the following steps:

[0064] Step 1: Collect multi-source heterogeneous data, including combustion operation parameters of multiple boilers, actuator parameters of denitrification and desulfurization systems, monitoring data from continuous emission monitoring systems at each measuring point, and data from the total chimney exhaust outlet. , Real-time concentration data; preprocessing of multi-source heterogeneous data based on the timestamp of the total discharge outlet data, including second-level time-series resampling and interpolation alignment; using the isolated forest algorithm to identify anomalies such as instrument failures and data jumps; soft measurement repair of missing or invalid monitoring data through weighted regression of the correlation features of adjacent furnace positions, reducing direct dependence on process instruments; and extracting load rate, ammonia injection rate, and inlet parameters from the preprocessed and soft measurement repaired full-series data. The ratio of slurry circulation volume and inlet The ratio and other derived features are normalized and standardized to generate a standardized multi-furnace collaborative time series feature matrix, with each sampling time step as the row and each furnace's operating parameters, process parameters and derived feature variables as the column.

[0065] Step 2: Based on the multi-furnace collaborative time-series feature matrix, obtain the target value of the total discharge pollutant concentration, construct a multi-furnace coupled spatiotemporal prediction model that integrates spatiotemporal graph convolution and rolling optimization control, construct a spatiotemporal topology map of the furnace group with each boiler as a graph node and the flue gas manifold as a topology edge, extract the spatiotemporal coupling contribution weights of load fluctuations and pollutant generation at different furnace positions to the total discharge concentration through the spatiotemporal feature extraction layer; in the rolling optimization decision layer, with the minimum total discharge concentration deviation and the minimum reducing agent consumption as dual objective constraints, simultaneously solve the global optimal allocation values ​​of ammonia injection and slurry circulation for each furnace within each control step, iteratively optimize through a constrained sequential quadratic programming rolling solver, and finally output the real-time optimal control instruction set of the denitrification and desulfurization actuators of each furnace;

[0066] Step 3: The real-time optimal control command set is sent to the DCS (Distributed Control System) via the OPC (Open Platform Communications) interface. Amplitude and rate limits are applied to the real-time optimal control command set to prevent system oscillations caused by large actuator movements. To address the poor stability of ammonia flow rate regulation, the ammonia injection command in the real-time optimal control command set is used as the feedforward ammonia injection amount, combined with the furnace outlet... The soft measurement values ​​are locally corrected to obtain the final control valve opening command and ammonia flow rate setpoint for each furnace actuator. This is combined with the measured total discharge concentration and furnace outlet concentration. The soft measurement values ​​are bound one by one to form a time-series sample library; the number and frequency of operation of the slurry circulation pumps of each furnace are simultaneously scheduled in a hierarchical manner; during the control process, the concentration change trend of the total discharge port is monitored in real time and the concentration deviation rate is calculated; when the concentration deviation rate is greater than the warning threshold, the model is triggered to recalculate in a rolling manner to obtain the real-time operation feedback data of the actuators of each furnace.

[0067] Step 4: Collect online monitoring data of the total discharge outlet in real time during the control cycle, and calculate four-dimensional evaluation indicators by combining real-time operation feedback data, including comprehensive concentration compliance rate, comprehensive fluctuation range, reducing agent consumption and energy cost. Retrieve the final regulating valve opening command, ammonia flow setpoint, and measured total discharge outlet concentration from the time series sample library, calculate the time series concentration residual between the predicted concentration of the total discharge outlet and the measured concentration value of the total discharge outlet, and use an incremental learning strategy to update the multi-furnace coupled spatiotemporal prediction model in reverse to generate a control effect evaluation report.

[0068] Specifically, the steps in step 2 for outputting the real-time optimal control command set for the denitrification and desulfurization actuators of each furnace include:

[0069] Based on the multi-furnace collaborative temporal feature matrix, the concentration target value of pollutants at the total discharge outlet is obtained through a real-time data interface. A multi-furnace coupled spatiotemporal prediction model integrating spatiotemporal graph convolution and rolling optimization control is constructed, including an input layer, a topology graph construction layer, a spatiotemporal feature extraction layer, a rolling optimization decision layer, and an output layer.

[0070] The input layer is used to receive the multi-furnace collaborative time-series feature matrix and the target value for the concentration of pollutants at the total discharge outlet;

[0071] The topology graph construction layer is used to extract load rate and inlet parameters from the multi-boiler collaborative time-series feature matrix, with each boiler's corresponding denitrification-desulfurization purification unit as an independent graph node. Concentration, inlet Concentration, ammonia injection rate, slurry circulation rate, outlet Soft measurement values ​​and exports The soft measurement values ​​consist of 7 process characteristics, which are then used to assign node values. The transmission delay from each furnace to the main discharge port is calculated based on the fluid pipeline resistance formula. The basic weight is set based on the transmission delay, and a nonlinear weighting correction factor is introduced between the furnaces to obtain the dynamic weight. A directional dynamic weighted adjacency matrix with a weight range limited to 0.2~1.0 is generated, forming a spatiotemporal topology map of the furnace group that strictly follows the physical confluence path of the flue gas.

[0072] The calculation process of the fluid pipeline resistance formula includes: the transmission delay is obtained by dividing the equivalent pipeline length from each boiler to the total discharge port by the average flue gas velocity; the average flue gas velocity is obtained by dividing the real-time flue gas flow rate of each boiler by the cross-sectional area of ​​the flue; the cross-sectional area of ​​the flue is the geometric cross-sectional area of ​​the denitrification-desulfurization outlet flue of each boiler, calculated according to the flue design drawings or measured inner diameter, and is a fixed structural parameter; the basic weight is the reciprocal of the transmission delay plus a small positive number to avoid division by zero; the inter-boiler load deviation correction factor is calculated by adding 1 to the adjustable coefficient, multiplying by the absolute value of the difference between the load rates of the two boilers, and then dividing by the rated load rate; the adjustable coefficient is a sensitivity parameter set according to process experience, with a typical value range of 0.1~0.5, used to adjust the correction strength of the load deviation on the weight; the rated load rate is the load percentage corresponding to the output of the boiler design nameplate, usually 100%; the final dynamic weight is the product of the basic weight and the inter-boiler load deviation correction factor, and then truncated and limited to the range of 0.2~1.0, forming a directional edge weight.

[0073] The spatiotemporal feature extraction layer extracts spatiotemporal coupling contribution weights through a dual-branch approach. The spatial convolution branch receives a dynamically weighted adjacency matrix from the spatiotemporal topology of the furnace group and aggregates the spatial coupling effects of multiple denitrification-desulfurization purification units along the confluence path using a second-order convolution kernel, outputting spatial coupling features. The temporal convolution branch employs causal dilation convolution with progressive dilation factors, covering all 30 historical time steps to capture the temporal evolution of pollutant generation and purification reactions, outputting temporal evolution features. The spatial coupling features and temporal evolution features are concatenated and then passed through a fully connected mapping layer. End-to-end training is performed using the measured concentration at the total discharge outlet as supervision. The Grad-CAM method is used to decouple and output the spatiotemporal coupling contribution weights of each denitrification-desulfurization purification unit to the pollutants at the total discharge outlet. Each spatiotemporal coupling contribution weight is then used to construct a contribution weight matrix according to the node and pollutant type. The number of rows in the contribution weight matrix equals the number of furnaces, and the number of columns corresponds to... and For two pollutants, each column stores the spatiotemporal coupling contribution weight of each furnace to the pollutant in that column, and the sum of all spatiotemporal coupling contribution weights in each column is 1; at the same time, it outputs the initial predicted value of the total discharge concentration for the next control step, which represents the trend prediction of the total discharge concentration under the condition of no control adjustment.

[0074] The rolling optimization decision layer receives the initial predicted value of the total discharge concentration and the contribution weight matrix. To achieve the dual objectives of compliance separation and economic operation, a linear prediction relationship between the control increment and the total discharge concentration is constructed based on the initial predicted value of the total discharge concentration and the contribution weight matrix: the predicted total discharge concentration equals the initial predicted value of the total discharge concentration plus the amount caused by the control increment of ammonia injection from each boiler. The change in concentration, plus the increase caused by the control of the slurry circulation volume in each furnace. The change in concentration; among which, The change in concentration is from The spatiotemporal coupling contribution weight is obtained by multiplying the product of the ammonia injection control increment of each furnace and the gain coefficient element by element and then summing the results. The change in concentration is determined by The spatiotemporal coupling contribution weight is obtained by multiplying the product of the control increment of the slurry circulation volume of each furnace and the gain coefficient element by element and then summing the results. The gain coefficient is determined based on the ratio of the corresponding outlet concentration change caused by the change in control quantity under historical stable operating conditions to the control increment. The comparison value is normalized and taken as 0~1 to reflect the degree of influence of unit control quantity change on outlet concentration. Based on this, a dual objective function is constructed to minimize the total discharge concentration deviation and the reducing agent consumption. The total discharge concentration deviation is the difference between the predicted concentration of the total discharge and the set target value of the total discharge pollutant concentration.

[0075] The dual-objective function is transformed into a single-objective function using a dynamic weighting coefficient method. The transformation process is as follows: An environmental-economic dynamic weighting factor, adjustable between 0 and 1, is introduced. A larger value indicates a stronger bias towards the environmental objective, while a smaller value indicates a stronger bias towards the economic optimization objective. When constructing the single-objective function, the environmental-economic dynamic weighting factor is multiplied by the square of the total emission concentration deviation to obtain the environmental objective term. The environmental-economic dynamic weighting factor is subtracted from 1 to obtain the economic optimization weight. The economic objective term is obtained by multiplying the economic optimization weight by the comprehensive economic cost term. The comprehensive economic cost term consists of the reducing agent consumption cost corresponding to the ammonia injection amount of each furnace and... The operating energy consumption cost corresponding to the circulation volume of each furnace slurry is calculated by weighted summation after being converted according to their respective unit price coefficients; the environmental protection-economic dynamic weight factor is adjusted in real time according to the initial predicted value of the total discharge concentration and the emission limit: when the initial predicted value of the total discharge concentration exceeds 90% of the emission limit, the environmental protection-economic dynamic weight factor is increased to strengthen the environmental protection target; when the initial predicted value of the total discharge concentration is lower than 80% of the emission limit, the environmental protection-economic dynamic weight factor is decreased to focus on the economic optimization target; among which, the emission limit is determined according to the "Emission Standard of Air Pollutants for Thermal Power Plants", which is read from the DCS environmental protection parameter interface or set manually;

[0076] Four types of process constraints—actuator range, single-step adjustment rate, compliance red line, and inter-furnace balancing—are mapped to the optimization variable space. The mapping process is as follows: the actuator range constraint requires that the ammonia injection rate and slurry circulation rate of each furnace must be within their respective allowable upper and lower limits; the single-step adjustment rate constraint requires that the amplitude of each control adjustment of each furnace be lower than the preset maximum step size, which is determined based on the actuator's physical adjustment capability and process stability requirements: the preset maximum step size for ammonia injection rate can be 5%~10% of the current ammonia injection rate, and the preset maximum step size for slurry circulation rate can be 3%~5% of the rated circulation rate; the compliance red line constraint requires that the total discharge concentration at the next moment be calculated through a linear prediction relationship. The degree must be less than or equal to the emission limit; the inter-furnace balance constraint requires that the standard deviation of the purification efficiency of each furnace be lower than the set tolerance threshold in order to maintain the load balance of each furnace. The tolerance threshold is set according to the environmental protection assessment requirements, and the typical value is 5%~10%, that is, the standard deviation of the purification efficiency of each furnace does not exceed 10%; a standard nonlinear constraint optimization problem is formed, and a constrained sequential quadratic programming rolling solver is adopted. The ammonia injection amount and slurry circulation amount of each furnace are synchronously iterated and optimized with a control step size of 5 seconds, a prediction time domain of 3 steps, and a control time domain of 1 step. The global optimal allocation value of ammonia injection amount and slurry circulation amount of each furnace is obtained by solving the problem, and the initial optimal control instruction set is generated by combining the 3-step sliding window smoothing mechanism.

[0077] The output layer is used to perform closed-loop correction of the initial optimal control command set for process compliance: The globally optimal allocation value obtained from the solution is substituted into the contribution weight matrix to calculate the predicted concentration of the total discharge outlet at the next time step, and then graded correction is performed. If the predicted concentration of the total discharge outlet exceeds 95% of the emission limit, the environmental-economic dynamic weight factor is increased to 1.0 and the sequential quadratic programming rolling solver is re-executed; if the predicted concentration of the total discharge outlet is lower than 80% of the emission limit, the value of the environmental-economic dynamic weight factor is decreased to increase the proportion of the economic optimization target. Simultaneously, the actuator range and single-step adjustment rate limits are checked one by one, and the initial optimal control commands that exceed the limits are limited and the deviation is recorded as the initial correction item for the next round of optimization. Finally, the verified initial optimal control commands are encapsulated into a real-time control command set that can be issued according to the DCS communication protocol.

[0078] Based on the multi-furnace coordinated time series characteristic matrix of 30 historical time steps and the total discharge outlet at the corresponding time, Measured concentration data The measured concentration data is used as the training set. The mean square error between the predicted concentration and the measured concentration at the total emission outlet is used as the loss function. The Adam optimizer is used to train the multi-furnace coupled spatiotemporal prediction model. During training, the dynamic edge weights are updated synchronously with the operating conditions to avoid generalization error caused by fixed topology. The measured concentration values ​​are obtained from the continuous emission monitoring system for flue gas.

[0079] After training, the standardized multi-furnace collaborative time-series feature matrix of the current moment is input in real time. A real-time spatiotemporal topology map of the furnace group is generated through the topology map construction layer. The contribution weight matrix and the initial predicted value of the total discharge concentration are obtained through forward calculation by the spatiotemporal feature extraction layer. Finally, the constrained sequential quadratic programming rolling solver is called to perform iterative optimization and output the real-time optimal control instruction set of the denitrification ammonia injection amount and desulfurization slurry circulation amount of each furnace.

[0080] For example, taking the coordinated denitrification and desulfurization of furnace 1 and furnace 2 as a scenario, the input layer is used to receive the multi-furnace coordinated time-series feature matrix and the set target value of the total discharge pollutant concentration; wherein the multi-furnace coordinated time-series feature matrix at the current moment contains 7-dimensional process features of the past 30 time steps. Taking the latest time step as an example, the load rate of furnace 1 is 78%, and the inlet... Concentration 320mg / m³, ingestion Concentration 1800 mg / m³, ammonia injection rate 45 kg / h, slurry circulation rate 380 m³ / h, outlet Soft measurement value 38mg / m³, outlet Soft measurement value: 25 mg / m³; Furnace 2 load rate: 92%; Inlet Concentration 405 mg / m³, ingestion Concentration 2150 mg / m³, ammonia injection rate 62 kg / h, slurry circulation rate 460 m³ / h, outlet Soft measurement value 52mg / m³, outlet Soft measurement value: 34 mg / m³; target value for pollutant concentration at total discharge outlet. 40mg / m³ It is 30 mg / m³;

[0081] The topology graph construction layer uses the denitrification-desulfurization purification units corresponding to the two boilers as independent graph nodes. Node values ​​are assigned using 7-dimensional process features, and the transmission delay from each boiler to the main exhaust port is calculated: Boiler 1 has an equivalent pipe length of 25m, a flue gas flow rate of 55m³ / s, and a flue cross-sectional area of ​​6.0m². Dividing 55 by 6.0 yields an average flue gas velocity of 9.17m / s, and dividing 25 by 9.17 yields a transmission delay of 2.73s. Boiler 2 has an equivalent pipe length of 35m, a flue gas flow rate of 65m³ / s, and a flue cross-sectional area of ​​7.5m². Dividing 65 by 7.5 yields an average flue gas velocity of 8.67m / s, and dividing 35 by 8.67 yields a transmission delay of 4.04s. The base weight is taken as the reciprocal of the transmission delay plus a small positive number 0.001, resulting in the base weight of Boiler 1. The basic weight of Furnace 1 is 0.366, and the basic weight of Furnace 2 is 0.247. The adjustable coefficient is 0.3, the rated load rate is 100%, and the load rate difference is the absolute difference of 78% and 95% (15%). The inter-furnace load deviation correction factor of 1.045 is obtained by calculating 1 + 0.3 × 15% ÷ 100%. The dynamic weight is truncated to 0.2~1.0 after multiplying the basic weight and the inter-furnace load deviation correction factor, forming a directional dynamic weighted adjacency matrix: the basic weight of Furnace 1 is 0.366, the basic weight of Furnace 2 is 0.247, the directional weight from Furnace 1 to Furnace 2 is the product of 0.366 and 1.045 (0.382), and the directional weight from Furnace 2 to Furnace 1 is the product of 0.247 and 1.045 (0.258). This generates a spatiotemporal topology map of the furnace group that strictly follows the physical convergence path of flue gas.

[0082] The spatial convolution branch in the spatiotemporal feature extraction layer receives the dynamically weighted adjacency matrix from the spatiotemporal topology map of the furnace group. It uses a second-order convolution kernel to aggregate the spatial coupling effect between the two denitrification-desulfurization purification units, outputting spatial coupling features, such as furnace 1 [0.42, -0.15, 0.78, ...] and furnace 2 [0.63, 0.21, -0.34, ...]. The temporal convolution branch uses dilated causal convolution to cover all 30 historical time steps, outputting temporal evolution features, such as furnace 1 [0.55, 0.12, -0.41, ...]. After concatenating the spatial coupling features and temporal evolution features, the result is passed through a fully connected mapping layer, using the measured concentration at the total discharge outlet. 47 mg / m³ End-to-end training was conducted with a target concentration of 33 mg / m³. Grad-CAM was used to decouple and output the spatiotemporal coupling contribution weights of each denitrification-desulfurization purification unit to the total discharge pollutants. A contribution weight matrix was constructed according to the node and pollutant type: the number of rows 2 corresponds to the number of furnaces, and the number of columns corresponds to... and Furnace 1 and Furnace 2 The spatiotemporal coupling contribution weights are 0.42 and 0.58, respectively, for furnace 1 and furnace 2. The spatiotemporal coupling contribution weights are 0.38 and 0.62, respectively; simultaneously, the initial predicted total discharge concentration for the next control step is output as... 46.5 mg / m³ It is 34.2 mg / m³;

[0083] The rolling optimization decision layer receives the initial predicted value of the total discharge concentration and the contribution weight matrix, and constructs a linear prediction relationship between the control increment and the total discharge concentration: the ammonia injection gain coefficients of furnace 1 and furnace 2 are -0.20 mg / m³ and -0.15 mg / m³, respectively, and the slurry circulation volume gain coefficients of furnace 1 and furnace 2 are -0.10 mg / m³ and -0.08 mg / m³, respectively. Assuming that the optimized solution yields ammonia injection control increments of +4.0 mg / m³ and +7.0 mg / m³ for furnace 1 and furnace 2, and slurry circulation volume increments of +10.0 mg / m³ and +12.0 mg / m³, respectively, then the result is calculated using 0.42 × 4.0 × (-0.20) + 0.58 × 7.0 × (-0.15). The concentration change of -0.945 mg / m³ was calculated using 0.38 × 10.0 × (-0.10) + 0.62 × 12.0 × (-0.08). If the concentration change is -0.975 mg / m³, then the predicted concentration at the total discharge outlet is... Adding -0.945 mg / m³ to 46.5 mg / m³ yields 45.555 mg / m³. Adding -0.975 mg / m³ to 34.2 mg / m³ yields 33.225 mg / m³, which is the target value for the total discharge concentration. 40mg / m³ If the concentration is 30 mg / m³, then the total discharge outlet concentration deviations are +5.555 mg / m³ and +3.225 mg / m³, respectively; introducing an environmental-economic dynamic weighting factor, because... The initial predicted total discharge concentration of 46.5 mg / m³ exceeds the emission limit of 50 mg / m³ by 90%. The environmental-economic dynamic weighting factor is increased from 0.7 to 0.85. The single objective function is composed of 0.85 multiplied by the square of the concentration deviation term plus 0.15 multiplied by the total cost of ammonia injection and slurry circulation for each furnace. The square of the concentration deviation term is 5.555² + 3.225². The unit price of liquid ammonia is 8 yuan / kg, and the unit price of electricity consumption is 0.15 yuan / m³. After adjustment... The total cost is approximately 1390 yuan / h. Simultaneously, the actuator range is constrained to allow ammonia injection of 15~100 kg / h and slurry circulation of 150~600 m³ / h. The single-step adjustment rate constrains the maximum ammonia injection step size to be 5% of the current value, i.e., 2.25 kg / h for furnace 1 and 3.1 kg / h for furnace 2. The maximum slurry circulation step size is set to 3% of the rated circulation size, i.e., 13.5 m³ / h for furnace 1 and 15.0 m³ / h for furnace 2. Compliance red lines constrain the predicted concentration. Less than or equal to 50 mg / m³ Four types of constraints—less than or equal to 35 mg / m³, standard deviation of purification efficiency less than 10% of tolerance threshold—are mapped to the optimization variable space. A sequential quadratic programming approach is used with a 5-second control step size, 3-step prediction time domain, 1-step control time domain rolling solution, and a 3-step sliding window smoothing method. This determines if the ammonia injection increment exceeds the limit. After limiting, the initial optimal control command set is obtained: ammonia injection increment for furnace 1 +2.25 kg / h, slurry circulation increment +10.0 m³ / h; ammonia injection increment for furnace 2 +3.1 kg / h, slurry circulation increment +12.0 m³ / h.

[0084] The output layer substitutes the limited control increment into the contribution weight matrix to recalculate the total emission outlet predicted concentration, obtaining 0.42×2.25×(-0.20)+0.58×3.1×(-0.15). The change was -0.458 mg / m³, which was obtained by calculating 0.38 × 10.0 × (-0.10) + 0.62 × 12.0 × (-0.08). The change is -0.975 mg / m³, and the predicted concentration at the total discharge outlet at the next time step is... 46.042 mg / m³ The concentration was 33.225 mg / m³; the predicted concentration at the total discharge outlet was determined to be... 95% of the emission limit of 50 mg / m³ and If the standard is met, the environmental protection-economic dynamic weighting factor remains at 0.85; the range and rate limits of the actuators are checked one by one, and all initial optimal control commands are within the constraints. The final verified initial optimal control commands are: ammonia injection rate of 47.25 kg / h and slurry circulation rate of 390 m³ / h for furnace 1, and ammonia injection rate of 65.1 kg / h and slurry circulation rate of 472 m³ / h for furnace 2. These are then encapsulated into a set of real-time control commands that can be issued according to the DCS communication protocol.

[0085] Based on the multi-furnace coordinated time series characteristic matrix of 30 historical time steps and the total discharge outlet at the corresponding time, Measured concentration data The measured concentration data is used as the training set, for example, the total discharge outlet data over 30 time steps within the past 150 seconds. The measured concentration sequence is 45, 46, 44, 48, 50, 49, 47, 43, 42, 41, 44, 46, 48, 47, 45, 43, 42, 44, 46, 48, 49, 47, 45, 44, 43, 46, 48, 47, 45, 47, with units of mg / m³. The measured concentration sequence is 32, 31, 33, 34, 35, 33, 32, 30, 29, 28, 30, 32, 33, 34, 32, 31, 30, 31, 33, 34, 35, 33, 32, 31, 30, 32, 34, 33, 32, 33, in mg / m³. The mean square error between the predicted concentration and the measured concentration at the total discharge outlet is used as the loss function. The Adam optimizer is used to train the multi-furnace coupled spatiotemporal prediction model with a learning rate of 0.001. The training is carried out for 200 epochs. Each epoch contains 16 time series samples with a step size of 30. During training, the dynamic edge weights are updated synchronously with the operating conditions to avoid generalization errors caused by fixed topology.

[0086] After training, the standardized multi-furnace collaborative time-series feature matrix of the current moment is input in real time. A real-time spatiotemporal topology map of the furnace group is generated through the topology map construction layer. Then, the contribution weight matrix and the initial predicted value of the total discharge concentration are obtained by forward calculation through the spatiotemporal feature extraction layer. The constrained sequential quadratic programming rolling solver is called to iteratively optimize and solve the global optimal allocation value of ammonia injection and slurry circulation for each furnace. The real-time optimal control command set of denitrification ammonia injection and desulfurization slurry circulation for each furnace is output: ammonia injection of 47.25 kg / h and slurry circulation of 390 m³ / h for furnace 1, and ammonia injection of 65.1 kg / h and slurry circulation of 472 m³ / h for furnace 2.

[0087] Specifically, the steps in step 3 for classifying and scheduling the number and frequency of operation of each furnace slurry circulation pump include:

[0088] The real-time optimal control command set is sent to the DCS system via the OPC communication interface. The system reads the current opening degree, frequency, and other actual status of each furnace actuator in the DCS system. The received real-time optimal control command set is then subjected to amplitude and rate limiting. The amplitude and rate limiting process includes: calculating the difference between the ammonia injection command value in the real-time optimal control command set and the current actual ammonia injection quantity to obtain the ammonia injection command increment; when the absolute value of the ammonia injection command increment exceeds the first maximum step size, it is clamped to the positive and negative boundary values ​​of the first maximum step size, and the clamped ammonia injection command increment is added to the current actual ammonia injection quantity to obtain the limited ammonia injection quantity set value; wherein, the first maximum step size is determined by the dynamic response characteristics of the actuator and the process safety limit, and can be taken as 5% to 8% of the rated ammonia injection flow rate, which is used to constrain the single-step ammonia injection adjustment range, avoid ammonia escape exceeding the standard and outlet concentration overshoot caused by ammonia injection sudden change, and ensure the stable operation of the denitrification reaction;

[0089] The difference between the slurry circulation volume command value in the real-time optimal control command set and the current actual slurry circulation volume is calculated to obtain the slurry circulation volume command increment. When the absolute value of the slurry circulation volume command increment exceeds the second maximum step size, it is clamped to the boundary value of the second maximum step size, and the clamped slurry circulation volume command increment is added to the current actual slurry circulation volume to obtain the limited slurry circulation volume set value. The second maximum step size is determined by the physical constraints of the equipment and the inertia of the slurry circulation system, and can be taken as 3% to 5% of the rated circulation flow rate. It is used to constrain the adjustment range of the single-step slurry circulation volume, avoid drastic fluctuations in the absorber level and large fluctuations in desulfurization efficiency, and ensure the stability of the desulfurization absorption process.

[0090] To address the issue of poor stability in ammonia flow rate regulation, the ammonia injection command from the real-time optimal control command set is used as the feedforward ammonia injection quantity, combined with the furnace outlet... The soft measurement value is locally corrected by feedback. The local feedback correction process is as follows: taking the furnace outlet... The soft measurement value is used as the basis for real-time feedback. The preset furnace outlet value is obtained by back-calculating the target value based on the total discharge concentration and the spatiotemporal coupling contribution weight of the corresponding furnace position. The concentration target value is calculated using an incremental PI control algorithm to determine the ammonia injection feedback correction increment. The calculation process is as follows: calculate the furnace outlet value at the current time. Soft measurement value and furnace outlet The deviation from the concentration target value is multiplied by a preset proportional coefficient to obtain a proportional correction term; the deviation value at the current moment is accumulated with the deviation values ​​at all historical moments and then multiplied by a preset integral coefficient to obtain an integral correction term; the proportional correction term and the integral correction term are added together to obtain the ammonia injection feedback correction increment for the current control step; wherein, the preset proportional coefficient is adjusted according to the change in ammonia injection rate at the furnace outlet. The sensitivity of the concentration effect and the system response speed requirements are determined, specifically by calculating the proportional relationship between the ammonia injection increment and the concentration change, which is used to reasonably map the deviation value to the ammonia injection adjustment range. The preset integral coefficient is determined based on the inertial time constant of the denitrification system and the requirement for anti-integral saturation. The specific determination process is as follows: the inertial time constant of the denitrification system is determined based on the time required for the furnace outlet concentration to reach a stable state after the ammonia injection adjustment during historical stable operation. The ratio of the sampling period corresponding to the control step size to the inertial time constant of the denitrification system is multiplied by the preset proportional coefficient to obtain the preset integral coefficient. The determination of whether to accumulate the deviation value is based on the integral separation threshold. When the absolute value of the deviation value is greater than the integral separation threshold, the integral calculation is stopped. When the absolute value of the deviation value is less than or equal to the integral separation threshold, the integral calculation is performed. The preset integral coefficient is usually 1 / 10 to 1 / 5 of the preset proportional coefficient to ensure steady-state accuracy and avoid integral overshoot. The integral separation threshold is determined based on the allowable range of furnace outlet concentration control deviation and is taken as 1% to 5% of the target value of the furnace outlet concentration.

[0091] The final ammonia injection rate setpoint is obtained by superimposing the feedforward ammonia injection rate and the ammonia injection feedback correction increment. This setpoint is then converted into a control valve opening command based on the inherent flow characteristic curve of the ammonia injection control valve, and the corresponding ammonia flow rate setpoint is simultaneously calculated. Finally, the final control valve opening command and ammonia flow rate setpoint for each boiler actuator are obtained. This is combined with the measured total discharge concentration and boiler outlet concentration. The soft measurement values ​​are bound one by one to form a time series sample library, which is used to periodically perform offline or online incremental training on the multi-furnace coupled spatiotemporal prediction model and update the contribution weight matrix and gain coefficient.

[0092] The number and frequency of operating slurry circulation pumps for each furnace are scheduled in a tiered manner. The tiered scheduling process is as follows: the slurry circulation volume setpoint after the limit is decomposed into a discrete combination of the number of fixed-speed pumps and the operating frequency of variable-frequency pumps. First, adjust... The circulating pump set corresponding to the furnace position with the largest contribution weight in spatiotemporal coupling ensures the flow of water to the main discharge port. The system features rapid response for the highly impactful denitrification-desulfurization purification unit. A pump set cumulative runtime rotation mechanism is introduced, prioritizing the activation of circulating pump sets with shorter cumulative runtimes under the same operating conditions. The final dispatch command is output after verification based on the absorber level safety constraints. During the control process, the system monitors the total discharge outlet concentration trend in real time, calculates the difference between the measured total discharge outlet concentration and the set concentration target value, and divides the absolute value of the difference by the set concentration target value to obtain the concentration deviation rate. When the concentration deviation rate exceeds the warning threshold, the model is triggered to recalculate, obtaining real-time operational feedback data for each furnace actuator. The final dispatch command is sent to the slurry circulating pump control loop of the DCS system via the OPC communication interface, driving the actuator to complete gear switching and frequency conversion adjustment. The actual operating status of the circulating pump set is synchronously collected and saved to the real-time operational feedback data for the next round of optimization and deviation verification. The warning threshold is determined based on the allowable instantaneous exceedance margin in environmental emission assessments and the process response delay, and can be set to 20%.

[0093] Specifically, the steps for generating the regulation effect evaluation report in step 4 include:

[0094] Real-time acquisition of online monitoring data from the total discharge outlet during the control cycle, including the total discharge outlet's... Measured concentration and The measured concentration, combined with the real-time operation feedback data of each furnace actuator, is used to calculate four-dimensional evaluation indicators, including the comprehensive concentration compliance rate, comprehensive fluctuation range, reducing agent consumption and energy consumption cost.

[0095] The calculation process is as follows: within the statistical control cycle Measured concentration The number of sampling points with measured concentrations below the emission limit, the total number of sampling points during the control cycle, and the ratio of the number of sampling points to the total number of sampling points are calculated to obtain the following results. Instantaneous pass rate and The instantaneous compliance rate is calculated by using a 1-hour sliding window to determine the hourly average concentration compliance rate for each pollutant. Overall compliance rate and The overall compliance rate is calculated by weighting the pollutants according to their respective assessment weights and then summing the results to obtain the overall concentration compliance rate, which is used to characterize the overall compliance level of flue gas pollutant emissions; among these, due to the environmental assessment... The accountability for exceeding emission standards and the stringency of emission limits are both higher than those in other countries. ,but The overall pass rate is weighted at 0.65; according to The assessment weight of the overall pass rate and The overall pass rate is weighted at 1, and the final pass rate is determined accordingly. The overall pass rate is weighted at 0.35.

[0096] Based on the continuously collected online monitoring data of the total discharge outlets during the control cycle, the data are integrated into the total discharge outlet data. Measured concentration sequence and Measured concentration sequence, calculation Measured concentration sequence and The standard deviation and mean of the measured concentration sequence are obtained by dividing the standard deviation by the mean. Concentration fluctuation range and The concentration fluctuation range is calculated by weighting the values ​​according to environmental assessment criteria to obtain the comprehensive fluctuation range, which reflects the severity of the oscillation in the total discharge concentration during the regulation process; among which, due to More sensitive to fluctuations in combustion conditions, instantaneous fluctuations are more likely to lead to excessive emissions. The environmental assessment weight for concentration fluctuation amplitude can be set to 0.6; according to Concentration fluctuation range and The environmental assessment weights for concentration fluctuation ranges are summed to 1, thus determining... The environmental assessment weight for concentration fluctuation range is set at 0.4;

[0097] The total ammonia consumption is obtained by summing up the actual ammonia flow rates of each furnace during the control period, and then converted into the reducing agent consumption per unit of pollutant removal by combining the cumulative amount of pollutants removed during the control period. At the same time, the total operating power consumption of slurry circulation is obtained by summing up the actual slurry circulation volume of each furnace, and then converted into the energy cost per unit of pollutant removal by combining the cumulative amount of pollutants removed during the control period.

[0098] The comprehensive control score is obtained by normalizing and weighting the four-dimensional evaluation indicators. The specific calculation process is as follows: normalize the comprehensive concentration compliance rate to obtain the normalized compliance rate; normalize the comprehensive fluctuation amplitude to obtain the normalized fluctuation amplitude; normalize the reducing agent consumption to obtain the normalized consumption; normalize the energy cost to obtain the normalized energy cost; based on the principle of environmental protection priority, the compliance rate is set with the largest weight, which can be taken as 0.5; based on the requirement of process stability, the fluctuation amplitude is set with the second largest weight, which can be taken as 0.3; according to... To reduce operational consumption demand, a consumption weight can be set, which can be 0.15. Since the weights of compliance rate, fluctuation range, consumption, and energy cost are all equal to 1, the energy cost weight is set to 0.05. Calculate the product of normalized compliance rate and compliance rate weight, normalized fluctuation range and fluctuation range weight, normalized consumption and consumption weight, and normalized energy cost and energy cost weight. Sum these four products to obtain the comprehensive control score. The value range is determined to be 0~1 based on the normalization result of the comprehensive evaluation index.

[0099] Retrieve the final control valve opening command, ammonia flow setpoint, measured total discharge concentration, and furnace outlet concentration corresponding to the control cycle from the time-series sample library. The soft measurement values ​​are obtained by subtracting the predicted concentration at the total discharge outlet from the multi-furnace coupled spatiotemporal prediction model at the same time stamp point by point, calculating the time-series concentration residual, and integrating them into a time-series concentration residual sequence by arranging them in ascending order.

[0100] Based on the comprehensive control score and the time-series concentration residual sequence, an incremental learning strategy is used to back-update the multi-furnace coupled spatiotemporal prediction model. The update process includes: calculating the absolute value of the mean time-series concentration residual within the current sliding window based on the time-series concentration residual sequence to obtain the average time-series concentration residual of the sliding window; if the average time-series concentration residual exceeds 3% of the emission limit or the comprehensive control score is lower than the qualified threshold, an incremental update is triggered; the qualified threshold is set according to the on-site operation and maintenance qualification judgment standard, and can be set to 0.85, corresponding to the score loss caused by normal operating condition disturbances within 15% allowed on-site; the most recent data is extracted from the time-series sample library. 500 valid samples are used, and exponential decay weights are assigned based on the time elapsed since the current time. Valid samples with shorter time elapsed have a higher exponential decay weight than those with longer time elapsed. Weighted samples are obtained by multiplying the feature vector of each valid sample by its corresponding exponential decay weight. These weighted samples are used as training data, and the mean square error between the predicted and measured concentrations at the total discharge outlet is used as the loss function. Thirty rounds of incremental training are performed on the parameters of the multi-furnace coupled spatiotemporal prediction model, with the learning rate set to 1 / 8 of the initial training rate. After the incremental training converges, Grad-CAM decoupling is re-executed, and the updated values ​​for each furnace are output. and The contribution weight matrix and corresponding gain coefficients are determined; the updated multi-furnace coupled spatiotemporal prediction model is used to backtest the effective samples of the current control cycle, and the comprehensive control score is recalculated. If the recalculated comprehensive control score is greater than the comprehensive control score before the update, the incremental update is determined to be effective, and the update record of the multi-furnace coupled spatiotemporal prediction model is saved.

[0101] An evaluation report on the control effect is generated by integrating four-dimensional evaluation indicators, time-series concentration residual sequences, and update records of multi-furnace coupled spatiotemporal prediction models.

[0102] Example 2

[0103] Please see Figure 5 Another embodiment of the present invention provides: a multi-furnace coordinated control system for pollutants at the total discharge outlet, comprising: a multi-source data preprocessing module, a coordinated prediction and optimization module, an instruction execution module, and an evaluation and update module;

[0104] The multi-source data preprocessing module is used to collect multi-source heterogeneous data, preprocess it based on the timestamp of the total discharge outlet data, extract derived features and normalize and standardize them to generate a multi-furnace collaborative time series feature matrix.

[0105] The collaborative prediction and optimization module, based on the multi-furnace collaborative time-series feature matrix, obtains the target value of the total discharge outlet pollutant concentration, constructs a multi-furnace coupled spatiotemporal prediction model, and builds a spatiotemporal topology map of the furnace group with each boiler as a graph node and the flue gas manifold as a topology edge. It extracts the spatiotemporal coupling contribution weights of load fluctuations and pollutant generation at different furnace positions to the total discharge outlet concentration. With the minimum total discharge outlet concentration deviation and the minimum reducing agent consumption as dual objective constraints, it simultaneously solves the global optimal allocation values ​​of ammonia injection and slurry circulation for each furnace within each control step. It iterative optimization is performed through a constrained sequential quadratic programming rolling solver, and the real-time optimal control instruction set is output.

[0106] The instruction execution module is used to send the real-time optimal control instruction set to the DCS system for amplitude and rate limiting. Addressing the issue of poor stability in ammonia flow rate regulation, it uses the ammonia injection instruction as a feedforward to determine the ammonia injection quantity, combined with the furnace outlet... The soft measurement values ​​are locally corrected to obtain the final control valve opening command and ammonia flow rate setpoint, which are then linked to the measured total discharge concentration and furnace outlet concentration. The soft measurement values ​​constitute a time series sample library; the number and frequency of operation of each furnace slurry circulation pump are scheduled in a hierarchical manner; the concentration change trend of the total discharge outlet is monitored in real time and the concentration deviation rate is calculated; when the concentration deviation rate is greater than the warning threshold, the model is triggered to recalculate and obtain real-time operation feedback data.

[0107] The evaluation and update module is used to collect online monitoring data of the total discharge outlet during the control cycle, calculate four-dimensional evaluation indicators by combining real-time operation feedback data, calculate the time-series concentration residual between the predicted concentration of the total discharge outlet and the measured concentration value of the total discharge outlet by combining the time-series sample library, and use an incremental learning strategy to back-update the multi-furnace coupled spatiotemporal prediction model to generate a control effect evaluation report.

[0108] Working principle and effects:

[0109] Multi-source heterogeneous data is acquired and preprocessed based on the total discharge outlet timestamp. The monitoring data of missing or failed processes are repaired by soft measurement through weighted regression of the correlation features of adjacent furnace positions. Derived features are extracted from the preprocessed and repaired full time series data and normalized and standardized to form a multi-furnace collaborative time series feature matrix. This ensures the initial consistency of the operating data of each furnace on the time axis in the subsequent prediction and optimization process, and provides reliable time series data for multi-furnace collaborative control.

[0110] Based on the multi-furnace coordinated temporal characteristic matrix and the target value of the total discharge outlet pollutant concentration, a multi-furnace coupled spatiotemporal prediction model is constructed. A spatiotemporal topology map of the furnace group is built, with each boiler as a graph node and the flue gas manifold as a topological edge. The spatiotemporal coupling contribution weights of load fluctuations and pollutant generation at different furnace positions on the total discharge outlet concentration are extracted. With the minimum total discharge outlet concentration deviation and the minimum reducing agent consumption as dual-objective constraints, the globally optimal allocation values ​​of ammonia injection and slurry circulation for each furnace are solved synchronously within each control step. Iterative optimization is achieved through a constrained sequential quadratic programming rolling solver, outputting a real-time optimal control command set. This process ensures that each control command is based on a quantitative evaluation of the spatiotemporal coupling contribution weights and the dynamic balance constraints of the multi-furnace system. Inter-furnace coordinated allocation suppresses instantaneous spikes in the total discharge outlet concentration, avoiding the global optimality loss and emission fluctuation problems caused by neglecting inter-furnace flue gas coupling in traditional independent control.

[0111] The real-time optimal control command set is sent to the DCS system after amplitude and rate limiting, with the ammonia injection command as the feedforward ammonia injection quantity, combined with the furnace outlet... The soft measurement value is locally corrected by feedback to generate the final control valve opening command and ammonia flow rate setpoint, which is then linked to the measured total discharge concentration and furnace outlet concentration. The soft measurement values ​​constitute a time-series sample library; the number and frequency of operation of each furnace slurry circulation pump are scheduled in a hierarchical manner; the concentration change trend of the total discharge outlet is monitored in real time and the concentration deviation rate is calculated; when the concentration deviation rate is greater than the warning threshold, the model is triggered to recalculate and output real-time operation feedback data; based on the concentration deviation rate as a quantitative basis, the prediction deviation caused by operating condition drift and model mismatch is dynamically compensated, and the fluctuation amplitude of the total discharge outlet concentration is continuously limited within the set range to ensure that the instantaneous value of the emission concentration stably meets the environmental protection limit requirements;

[0112] Online monitoring data of the total discharge outlet is collected during the control cycle. Combined with real-time operational feedback data, a four-dimensional evaluation index is calculated. The time-series concentration residual between the predicted and measured concentration values ​​of the total discharge outlet is calculated using a time-series sample library. An incremental learning strategy is employed to update the multi-furnace coupled spatiotemporal prediction model, generating a control effect evaluation report. The spatiotemporal coupling contribution weights of the furnace group are continuously adjusted to reduce the time-series concentration residual and lower the total discharge outlet concentration. , Concentration fluctuation range.

[0113] Overall, through a four-layer architecture of multi-source data preprocessing, collaborative prediction optimization, instruction execution, and evaluation updates, a closed-loop system is achieved, encompassing data acquisition, spatiotemporal coupling modeling, rolling optimization execution, and feedback iterative self-learning. This system specifically addresses the lack of collaborative dynamic balance control capabilities under the condition of multiple furnaces sharing a single outlet, which leads to problems with the total outlet... and The technical issues of large fluctuations in emission concentration and easy exceedance of instantaneous values ​​have been addressed by improving the operational adaptability and emission compliance reliability of the multi-furnace collaborative purification process, and reducing the risk of instantaneous exceedance at the total discharge outlet.

[0114] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for coordinated control of pollutants from multiple furnaces at the total discharge outlet, characterized in that, include: After preprocessing multi-source heterogeneous data, derived features are extracted to generate a multi-furnace collaborative time-series feature matrix. A multi-furnace coupled spatiotemporal prediction model is constructed, the spatiotemporal coupling contribution weight of each furnace position is extracted, and the global optimal allocation value of ammonia injection and slurry circulation of each furnace is solved with dual objective constraints. The real-time optimal control instruction set is output through iterative optimization. The amplitude and rate of the real-time optimal control command set are limited, and the ammonia injection command is used as the feedforward ammonia injection quantity for local feedback correction to obtain the final control valve opening command and ammonia water flow set value; the number and frequency of operation of each boiler slurry circulation pump are scheduled in stages, the concentration change trend of the total discharge port is monitored in real time to calculate the concentration deviation rate, and when the concentration deviation rate is greater than the warning threshold, the model is triggered to recalculate and obtain real-time operation feedback data. The time-series concentration residuals are calculated by obtaining four-dimensional evaluation indicators, and the multi-furnace coupled spatiotemporal prediction model is updated in reverse to generate a regulation effect evaluation report.

2. The method for coordinated control of pollutants from multiple furnaces at the total discharge outlet according to claim 1, characterized in that, The specific steps for constructing a multi-furnace coupled spatiotemporal prediction model include: A multi-furnace coupled spatiotemporal prediction model is constructed by obtaining a target concentration value, including an input layer, a topology map construction layer, a spatiotemporal feature extraction layer, a rolling optimization decision layer, and an output layer. The input layer receives the multi-furnace collaborative time-series feature matrix and the concentration target value. The topology map construction layer extracts process features from the multi-furnace collaborative time sequence feature matrix and assigns values; calculates the transmission delay from each furnace to the main outlet and sets basic weights; introduces the inter-furnace load deviation correction factor to perform nonlinear weighted correction to generate a dynamic weighted adjacency matrix, forming a spatiotemporal topology map of the furnace group. The spatiotemporal feature extraction layer extracts spatial coupling features from the dynamically weighted adjacency matrix through spatial convolution branches and temporal evolution features through temporal convolution branches. After concatenation, the spatiotemporal feature extraction layer is trained with the measured concentration at the total discharge outlet as supervision. The spatiotemporal coupling contribution weights of each denitrification-desulfurization purification unit are decoupled and output to construct a contribution weight matrix, and the initial predicted value of the total discharge outlet concentration is output.

3. The method for coordinated control of pollutants from multiple furnaces at the total discharge outlet according to claim 2, characterized in that, The specific steps for constructing a multi-furnace coupled spatiotemporal prediction model also include: The rolling optimization decision layer receives the initial predicted value of the total discharge outlet concentration and the contribution weight matrix, constructs a dual objective function that minimizes the deviation of the total discharge outlet concentration and the consumption of reducing agent, and transforms it into a single objective function using the dynamic weighting coefficient method. Four types of process constraints—actuator range, single-step adjustment rate, compliance red line, and inter-furnace balance—are mapped to the optimization variable space to form a nonlinear constraint optimization problem. A sequential quadratic programming rolling solver is used to synchronously iterate and optimize the ammonia injection rate and slurry circulation rate of each furnace to obtain the global optimal allocation value. The initial optimal control instruction set is then smoothly generated through a sliding window. The output layer substitutes the globally optimal allocation value into the contribution weight matrix to calculate the predicted concentration of the total discharge outlet at the next time step for hierarchical correction, and verifies the actuator range and single-step adjustment rate limit. The verified initial optimal control command is then encapsulated into a real-time control command set.

4. The method for coordinated control of pollutants from multiple furnaces at the total discharge outlet according to claim 3, characterized in that, The specific steps for outputting the real-time optimal control instruction set include: Based on the multi-furnace coordinated time series characteristic matrix of 30 historical time steps and the total discharge outlet at the corresponding time, Measured concentration data The measured concentration data were used as the training set. The mean square error between the predicted concentration and the measured concentration at the total discharge outlet was used as the loss function. The Adam optimizer was used to train the multi-furnace coupled spatiotemporal prediction model. After training, the multi-furnace collaborative time-series feature matrix at the current moment is input in real time. A real-time spatiotemporal topology map of the furnace group is generated through the topology map construction layer. The contribution weight matrix and the initial predicted value of the total discharge concentration are obtained through forward calculation by the spatiotemporal feature extraction layer. The sequential quadratic programming rolling solver is called to perform iterative optimization and output the real-time optimal control instruction set.

5. The method for coordinated control of pollutants from multiple furnaces at the total discharge outlet according to claim 4, characterized in that, The specific steps for applying amplitude and rate limiting to the real-time optimal control instruction set include: The real-time optimal control instruction set is sent to the DCS system, and the amplitude and rate of the real-time optimal control instruction set are limited. The difference between the ammonia injection quantity instruction value in the real-time optimal control instruction set and the current actual ammonia injection quantity is calculated to obtain the ammonia injection quantity instruction increment. When the absolute value of the ammonia injection quantity command increment exceeds the preset first maximum step size, it is clamped to the positive and negative boundary values ​​of the first maximum step size and then added to the current actual ammonia injection quantity to obtain the limited ammonia injection quantity setting value. The difference between the slurry circulation volume command value in the real-time optimal control command set and the current actual slurry circulation volume is calculated to obtain the slurry circulation volume command increment. When the absolute value of the increment of the slurry circulation volume command exceeds the preset second maximum step size, it is clamped to the boundary value of the second maximum step size and then added to the current actual slurry circulation volume to obtain the limited slurry circulation volume setting value.

6. The method for coordinated control of pollutants from multiple furnaces at the total discharge outlet according to claim 5, characterized in that, The specific steps to obtain the final control valve opening command and ammonia flow rate setpoint include: Using the ammonia injection command from the real-time optimal control command set as the feedforward ammonia injection quantity, combined with the furnace outlet... Soft measurement values ​​are locally corrected using feedback. furnace outlet Soft measurement values ​​are used as a basis for real-time feedback to obtain the preset furnace outlet. The concentration target value is calculated by using an incremental PI control algorithm to calculate the ammonia injection feedback correction increment. The feedforward ammonia injection amount is then superimposed with the ammonia injection feedback correction increment to obtain the final ammonia injection amount setpoint. By combining the inherent flow characteristic curve of the ammonia injection regulating valve with the valve opening command, the corresponding ammonia flow setpoint is obtained through synchronous calculation. This yields the final regulating valve opening command and ammonia flow setpoint, which are then linked to the measured total discharge concentration and furnace outlet concentration. The soft measurement values ​​constitute a time series sample library.

7. The method for coordinated control of pollutants from multiple furnaces at the total discharge outlet according to claim 6, characterized in that, The specific steps to obtain real-time operational feedback data include: The number and frequency of slurry circulation pumps operating in each furnace are scheduled in stages. The limited slurry circulation volume setpoint is decomposed into a discrete combination of the number of fixed-speed pumps and the operating frequency of variable-frequency pumps. The first adjustment is... The circulating pump group corresponding to the furnace position with the largest contribution weight of spatiotemporal coupling; A pump set cumulative running time rotation mechanism is introduced. Under the same operating conditions, the circulating pump set with the shorter cumulative running time is started first. The final scheduling command is output after verifying the absorption tower liquid level safety constraint. During the regulation process, the trend of total discharge outlet concentration changes is monitored in real time, and the concentration deviation rate is calculated based on the measured total discharge outlet concentration and the concentration target value. When the concentration deviation rate exceeds the preset warning threshold, the model is triggered to recalculate and obtain real-time operation feedback data.

8. The method for coordinated control of pollutants from multiple furnaces at the total discharge outlet according to claim 7, characterized in that, The specific steps for calculating time-series concentration residuals include: Real-time data collection of total discharge outlets during the control cycle Measured concentration and The measured concentration was combined with real-time operational feedback data to calculate a four-dimensional evaluation index. The four-dimensional evaluation index was then normalized and weighted to obtain a comprehensive control score. Retrieve the final control valve opening command, ammonia flow setpoint, measured total discharge concentration, and furnace outlet concentration corresponding to the control cycle from the time-series sample library. Soft measurement value; The predicted concentration at the total discharge outlet from the multi-furnace coupled spatiotemporal prediction model at the same timestamp is subtracted point by point from the measured concentration value at the total discharge outlet to calculate the time-series concentration residual. The residual is then integrated into a time-series concentration residual sequence by arranging the residuals in ascending order.

9. The method for coordinated control of pollutants from multiple furnaces at the total discharge outlet according to claim 8, characterized in that, The specific steps for generating a regulation effectiveness evaluation report include: Based on the comprehensive control score and the time-series concentration residual sequence, an incremental learning strategy is used to back-update the multi-furnace coupled spatiotemporal prediction model. The updated multi-furnace coupled spatiotemporal prediction model is used to backtest the effective samples of the current control cycle, and the comprehensive control score is recalculated. If the recalculated comprehensive control score is greater than the previous comprehensive control score, the incremental update is deemed valid, and the update record of the multi-furnace coupled spatiotemporal prediction model is saved. An evaluation report on the control effect is generated by integrating four-dimensional evaluation indicators, time-series concentration residual sequences, and update records of multi-furnace coupled spatiotemporal prediction models.

10. A multi-furnace coordinated control system for pollutants at a total discharge outlet, used to implement a multi-furnace coordinated control method for pollutants at a total discharge outlet as described in any one of claims 1-9, characterized in that, It includes a multi-source data preprocessing module, a collaborative prediction and optimization module, an instruction execution module, and an evaluation and update module; The multi-source data preprocessing module is used to collect multi-source heterogeneous data, preprocess it, extract derived features, and generate a multi-furnace collaborative time-series feature matrix. The collaborative prediction and optimization module is used to construct a multi-furnace coupled spatiotemporal prediction model, extract the spatiotemporal coupling contribution weight of each furnace position, solve the global optimal allocation value of ammonia injection and slurry circulation for each furnace with dual objective constraints, and output the real-time optimal control instruction set through iterative optimization. The instruction execution module is used to limit the amplitude and rate of the real-time optimal control instruction set, use the ammonia injection instruction as the feedforward ammonia injection amount for local feedback correction, and obtain the final control valve opening instruction and ammonia water flow set value; it also performs hierarchical scheduling of the number and frequency of operation of each boiler slurry circulation pump, monitors the concentration change trend of the total discharge port in real time to calculate the concentration deviation rate, and triggers the model rolling recalculation when the concentration deviation rate is greater than the warning threshold to obtain real-time operation feedback data; The evaluation and update module is used to obtain the time-series concentration residuals of the four-dimensional evaluation index calculation, update the multi-furnace coupled spatiotemporal prediction model in reverse, and generate a regulation effect evaluation report.

Images