An artificial intelligence-based waste power generation boiler combustion control method

Abstract

This invention relates to the field of waste-to-energy incineration technology and provides an artificial intelligence-based combustion control method for waste-to-energy boilers. The method includes: collecting data from a distributed control system, flue gas monitoring, and feed metering, aligning these data with a unified time frame, and generating combustion condition data; calculating drift residuals and missing rates, classifying and labeling them to obtain confidence weights; calculating equivalent lower heating value and moisture content from steam flow and feed flow based on energy balance, updating parameters only during the stable steam and carbon monoxide windows; identifying the equivalent time delay from control quantities to steam and emissions based on cross-correlation peak values, writing this information into the prediction model after stabilization, and performing shift compensation; constructing hard constraints for steam and emissions based on aligned predictions, writing confidence into the cost function, using MPC to calculate the feasible region, selecting incremental deployments within the policy network domain, and backing down and auditing for low confidence levels.

Description

Technical Field

[0001] This invention relates to the field of waste incineration power generation technology, specifically an artificial intelligence-based combustion control method for waste-to-energy boilers. Background Technology

[0002] In the application of grate-type waste-to-energy boilers with high proportion of wet waste and sludge co-firing, the incoming materials exhibit low calorific value, high moisture content, and fluctuating chlorine and sulfur content. The combustion in the furnace releases heat in segments along the grate, and the flue gas forms a definite transmission lag in the furnace, flue, and recirculation loop. This causes the primary air, secondary air, grate speed, and flue gas recirculation to form a strongly coupled dynamic response and time lag superposition with the steam flow, furnace temperature field, and emissions. As a result, the dual objective of ultra-low emission constraints and stable steam parameters is difficult to achieve in a closed-loop stable manner under uncertain incoming material conditions.

[0003] Current control schemes use steam pressure or steam flow rate as the main loop to construct PID cascade control, and use the volume fraction of oxygen in the flue gas as the feedback quantity of the excess air coefficient to adjust the total air volume or air ratio. At the same time, existing technologies use the carbon monoxide and nitrogen oxide concentrations output by CEMS as correction quantities to perform feedback correction on the secondary air nozzle opening, segmented air ratio, or flue gas recirculation volume. Some schemes introduce MPC with a simplified furnace model to jointly optimize control quantities such as air volume, air ratio, and grate speed in a constrained manner, so as to achieve coordinated control of steam and emissions under stable operating conditions.

[0004] However, the aforementioned existing technologies suffer from time-lapse adjustment defects caused by hysteresis coupling. When PID or rule control directly adjusts the air volume and air distribution in real time based on the flue gas oxygen volume fraction, carbon monoxide and nitrogen oxide concentration measured at the current moment, the actual object of the control quantity is the combustion state at the previous moment rather than the current combustion state, forming a compensation based on hysteresis measurement on the time axis, thereby generating erroneous compensation and closed-loop oscillation. Summary of the Invention

[0005] In view of the existing problems mentioned above, an artificial intelligence-based combustion control method for waste-to-energy boilers is proposed.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a combustion control method for waste-to-energy boilers based on artificial intelligence, comprising: Step S1: Collect data from the distributed control system, flue gas online monitoring, and feed metering, align them with a unified timestamp, and obtain the combustion operation data of the waste-to-energy boiler; calculate the drift residual and data missing rate for each sensor output, generate confidence weights according to preset grading rules, and write them into the data tags; Step S2: Based on the energy balance equation, the equivalent lower heating value and equivalent moisture content of the waste fed into the furnace are calculated from the steam flow rate and the feed mass flow rate; the steam flow rate fluctuation rate and carbon monoxide concentration fluctuation rate are set as dual criteria, and the soft measurement model parameters are updated within the stable window; Step S3: Within the sliding time window, the peak method of cross-correlation function is used to determine the equivalent time delay from each control quantity to the steam flow and emission index. The time delay parameter is updated only when consistent time delays are obtained in multiple consecutive windows and the effectiveness score is consistently higher than the threshold. The time delay parameter is written into the dynamic prediction model. A time shift consistent with the equivalent time delay is introduced into the control execution quantity at the model input to achieve feedforward compensation. Step S4: Construct hard constraints for steam stability and emission limits using the time-delay aligned predictions, and write the confidence weights into the cost function; use model predictive control to solve the feasible control domain, and then the policy network selects the control increment within the feasible domain and outputs it as the setpoint correction amount to be sent to the distributed control system; when the confidence weights are lower than the threshold, stop automatic adjustment and fall back to the historical conservative setpoint, and at the same time generate an audit record containing time delay, soft measurement and constraint trigger fields.

[0007] In a preferred embodiment, step S1 further includes: Using the sampling time of the distributed control system as a unified time reference, data with different sampling frequencies are time-aligned according to the rule of taking the nearest value backward and keeping it until the current reference time. Data exceeding the preset physical upper and lower limits are subjected to amplitude limiting processing. At the same time, the original acquisition time, the aligned time, the original value and the aligned value are written into the record to obtain the combustion operation data of the waste-to-energy boiler. For each sensor output, a quality assessment is performed within a preset sliding time window. Data points that exceed the range, experience communication timeouts, continuously maintain a constant value for more than a preset duration, or contradict the device's operating logic are marked as invalid, and the reasons for invalidity are written into the reason field. Robust reference values ​​and robust dispersion reference values ​​are calculated for the valid data within the window. The degree of deviation between the current data and the robust reference values ​​is used as the drift residual level, and the proportion of invalid data points in the window to the total number of data points in the window is used as the missing rate level. The drift residual level and missing rate level are mapped to confidence weights and written into the data labels. The highest confidence weight is assigned when both the drift residual level and the missing rate level do not exceed the first level; the medium confidence weight is assigned when either one is at the second level; and the lowest confidence weight is assigned when either one exceeds the second level, and the data for that measurement point is marked as unavailable. When a critical measurement point is marked as unavailable, the most recently available data is used as the alternative input for that measurement point, and an alarm record is output.

[0008] As a preferred implementation, the specific steps for adding operational phase identification and phased labeling after data collection and quality assessment are as follows: The system calculates stage discrimination features from the aligned combustion operation data, classifies the current operation into operation stage labels in the preset stage library according to preset stage discrimination rules, and calculates the confidence score of the discrimination result. When the confidence score is lower than a preset threshold, the operation stage label is not updated and the previous operation stage label is used, and the stage freeze reason field is written. The confidence score comprehensively considers feature deviation, duration stability, and historical switching frequency. After generating the runtime stage label, the runtime stage label is written into the stage field of the data label, and the data is segmented and archived using the stage field: separate sliding window statistics, separate confidence weight grading thresholds, and separate valid sample counts that can be used for model updates are maintained for different runtime stages; when a stage switch is detected, the stage switch flag and switch reason are written into the data label, and the statistical baseline and model update entry of the previous stage are frozen. The corresponding model parameters and strategy parameters for the new stage are only updated after the cumulative number of samples in the new stage reaches the preset minimum.

[0009] In a preferred embodiment, step S2 further includes: Based on the steam property table, the steam-side enthalpy difference is multiplied by the steam flow rate to convert it into the boiler's heat absorption within the specified window. The steam-side enthalpy difference is the difference between the enthalpy of the main steam and the enthalpy of the feedwater. Combined with the feed mass flow rate obtained from the feed metering, and using the boiler thermal efficiency correction coefficient, the heat absorption is divided by the thermal efficiency correction coefficient and converted into the furnace heat input. The equivalent lower heating value of the waste entering the furnace is then calculated by dividing the furnace heat input by the feed mass flow rate. The heat consumption for moisture evaporation is preset to a fixed value, and a dry-basis combustible component equivalent calorific value benchmark is introduced. This benchmark is taken from the back-calculation result of the most recent stable window. The difference between the equivalent lower heating value and the calorific value benchmark is converted into equivalent moisture content according to the fixed evaporation heat consumption. The equivalent lower heating value and equivalent moisture content are written into the data tag as soft measurement output. Within the same sliding window, a stable window is determined based on the fluctuation rate of steam flow, the fluctuation rate of carbon monoxide concentration, and the average value of carbon monoxide concentration within the window. When a stable window is established and the data label is available and the confidence level is not lower than the medium confidence level, the soft sensor model parameters are updated. The update adopts a weighted fusion of historical parameters and the back-calculated parameters of the current window, and when the update conditions are not met, the unupdated state and the triggering reason are written.

[0010] As a preferred implementation, the specific steps for improving the maintenance and updating of soft measurement model parameters from a global single parameter set to a staged parameter set, with independent maintenance of soft measurement parameters in different operating stages and the adoption of staged steady-state criteria, are as follows: An independent set of soft measurement parameters is maintained for each operating stage. The set of soft measurement parameters includes the boiler thermal efficiency correction coefficient, the equivalent calorific value benchmark of dry-based combustible components, and the correction coefficient for the heat consumption converted from water evaporation. When a change in the operating stage label is detected, the soft measurement parameter update of the previous operating stage is immediately frozen, and the soft measurement calculation is performed using the soft measurement parameter set corresponding to the new operating stage. If the new operating stage is the first occurrence or its effective sample size is less than the preset minimum, the initialization parameter set of the same type as the operating stage in the historical database is called, and the stage initialization status and its source field are written into the data label. When the steady-state criteria corresponding to the current operating stage are met, and the key measurement point data labels are available and the confidence level is not lower than medium confidence level, the soft measurement model parameters corresponding to the operating stage are updated. When the staged steady-state criteria or confidence level conditions are not met, the soft measurement model parameters corresponding to the operating stage are kept unchanged, and the unupdated state and its triggering reason are written into the data label.

[0011] In a preferred embodiment, step S3 further includes: Within a recently preset duration window, a control execution sequence and an output sequence are constructed. Time delay identification is only performed when at least a preset proportion of data points within the window meet medium or higher confidence levels. The time delay identification includes decentralizing the sequence using the window median and normalizing the window absolute deviation median; calculating the correlation between the input sequence and the output sequence after alignment within a preset candidate time shift range; and marking the time shift as invalid if the number of valid data points in the overlapping interval is less than a preset proportion. The time shift corresponding to the maximum correlation value is determined as the equivalent time delay, and the maximum correlation value is used as... The validity score is written into the audit log. When the maximum correlation is lower than the preset threshold or the difference between the maximum correlation and the second largest correlation is less than the preset difference, the identification result is determined to be unreliable and the previous reliable time delay is used. The equivalent time delay is written into the dynamic prediction model. A time shift consistent with the equivalent time delay is introduced into the control execution quantity at the model input. The first missing segment generated by the shift is filled in by the first valid value retention method and a filling flag is written. When the prediction validity flag is invalid, conservative control is switched and model updates are frozen. At the same time, the triggering reason and the corresponding time delay parameter are written into the audit log.

[0012] As a preferred embodiment, the specific steps for maintaining the time delay table in the time delay identification are as follows: Before performing time delay identification and prediction alignment, the running stage field is read, and the input and output sequences are filtered for the same stage using this running stage field. Only data points whose running stage is consistent with the current stage within the most recent 15-minute window are used to participate in the time delay calculation. It is required that the number of data points of this type in the same stage accounts for no less than a preset proportion of the total number of points in the window, and the proportion of data points with a confidence level of no less than medium confidence level is no less than 90%. If the above proportion conditions are not met, the time delay parameters are not updated, and the locked time delay of the previous window is used. The audit record is frozen due to insufficient samples and the reasons. Each operational phase maintains an independent time delay parameter table and locking status. That is, the same control execution quantity to the same output measurement corresponds to different equivalent time delay parameters in different operational phases. When the operational phase changes, the time delay parameters of the previous phase are immediately stopped, and the time delay parameter table of the new phase is called as the initial time delay to participate in the relevant peak search. The phase locking rule is set: the time delay parameters of the new phase are only updated and locked when the new phase obtains consistent time delay results in a number of consecutive sliding windows and the effectiveness score is not lower than the preset threshold. Otherwise, the initial time delay of the new phase remains unchanged.

[0013] In a preferred embodiment, step S4 further includes: Within a preset prediction time domain, a prediction sequence of steam flow rate, steam pressure, flue gas oxygen volume fraction, carbon monoxide concentration, and nitrogen oxide concentration is generated. Hard constraints are constructed using the time-delay-aligned prediction sequence. These hard constraints include: steam pressure prediction deviation not exceeding a preset pressure deviation upper limit; steam flow rate prediction deviation not exceeding a preset flow rate deviation upper limit; flue gas oxygen volume fraction within a preset range; and carbon monoxide concentration and nitrogen oxide concentration not exceeding their respective upper limits. The credibility weight of key measurement points is introduced into the cost function, and penalty terms are set for steam stability deviation, emission risk approaching the upper limit, and control increment change magnitude. When the credibility decreases, the penalty for steam and emission deviations is increased, and the safety margin of the emission upper limit is tightened. Model predictive control is used to solve for the feasible control domain that satisfies hard constraints. The feasible control domain is then input into the strategy network along with the current operating condition characteristics. The output control increment is then pruned according to the feasible control domain and distributed to the distributed control system. When the confidence weight of any key measurement point is the lowest or the window average of the confidence weight of the key measurement point is lower than the preset threshold, the system reverts to the historical conservative setting value and generates an action-level audit record.

[0014] As a preferred implementation, the specific steps for improving constraint optimization from a global single constraint set and weight set to a staged constraint set and weight set are as follows: Based on the operation phase field, maintain and call the phased constraint set, phased cost weight set, and phased strategy parameter set corresponding to each operation phase. The phased cost weight set includes steam pressure deviation penalty weight, steam flow deviation penalty weight, flue gas oxygen volume fraction deviation penalty weight, carbon monoxide exceedance risk penalty weight, and nitrogen oxide exceedance risk penalty weight, and stipulate that different weight sets are used for different operation phases. The phased strategy parameter set is a set of parameters stored by the strategy network according to the running stage. The strategy network only uses data samples that are consistent with the current running stage and have a confidence level of not less than medium confidence level for updating, and it is prohibited to use cross-stage samples to update parameters.

[0015] In a preferred embodiment, the policy network is trained and its parameters are invoked independently in stages. The specific steps for gating and transitional contraction during stage switching are as follows: When a phase transition occurs, a transition gating is executed, freezing the strategy parameter updates and cost weight adaptive updates of the previous phase. The constraint set, weight set, and strategy parameter set corresponding to the new phase are then invoked. Within a preset transition period after the transition, the upper limit of the allowable increment of the feasible control domain is contracted to a fixed proportion of the stable phase, allowing only control increments that cause changes in steam pressure, steam flow, and emission indicators in the direction of satisfying hard constraints. Simultaneously, the transition contraction is lifted and the normal feasible control domain and strategy output of the new phase are restored only when the new phase continuously meets a preset number of control cycles and the hard constraints are not triggered.

[0016] Beneficial effects Compared with the prior art, the present invention has the following advantages: 1. By identifying the equivalent time delay from the control action to the steam and emission measurements within the sliding time window, and writing the time delay parameters into the dynamic prediction model to perform time shift alignment of the control action, the prediction and optimization are made based on the same time axis, avoiding closed-loop oscillation, steam parameter overshoot and instantaneous emission exceedance caused by compensation based on lag measurements.

[0017] 2. Calculate the drift residual level and missing rate level for each sensor output, map them to generate confidence weights and write them into the data label. Use the confidence weights for model update gating and control backoff triggering to ensure that no wrong learning or wrong adjustment is driven when key measurement points are abnormal, thereby reducing control deviations caused by sensor drift, communication interruption, and constant value hysteresis. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0019] Fig. 1 This is a flowchart illustrating the present invention; Fig. 2 This is a comparison diagram of the effects of the present invention and the prior art, where gray bars represent the prior art and black bars represent the present invention. Detailed Implementation

[0020] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0021] Example 1: To achieve the above objectives, please refer to Figs. 1-2 This invention provides an artificial intelligence-based combustion control method for waste-to-energy boilers, the method comprising the following steps: Step S1: Collect data from the distributed control system, flue gas online monitoring, and feed metering, align them with a unified timestamp, and obtain the combustion operation data of the waste-to-energy boiler; calculate the drift residual and data missing rate for each sensor output, generate confidence weights according to preset grading rules, and write them into the data tags; Step S2: Based on the energy balance equation, the equivalent lower heating value and equivalent moisture content of the waste fed into the furnace are calculated from the steam flow rate and the feed mass flow rate; the steam flow rate fluctuation rate and carbon monoxide concentration fluctuation rate are set as dual criteria, and the soft measurement model parameters are updated within the stable window; Step S3: Within the sliding time window, the equivalent time delay from each control quantity to the steam flow rate and emission index is determined using the peak method of the cross-correlation function, and the time delay parameters are written into the dynamic prediction model to achieve feedforward compensation. Step S4: Construct hard constraints for steam stability and emission limits using the time-delay aligned predictions, and write the confidence weights into the cost function; use model predictive control to solve the feasible control domain, and then the policy network selects the control increment within the feasible domain and outputs it as the setpoint correction amount to be sent to the distributed control system; when the confidence weights are lower than the threshold, stop automatic adjustment and fall back to the historical conservative setpoint, and at the same time generate an audit record containing time delay, soft measurement and constraint trigger fields.

[0022] In this embodiment, step S1 specifically includes: The combustion control execution data and process status measurement data output by the distributed control system are collected. The combustion control execution data includes the primary and secondary air volumes, grate operating speed, and flue gas recirculation volume. The process status measurement data includes steam flow rate and steam pressure. Simultaneously, the flue gas oxygen volume fraction, carbon monoxide concentration, and nitrogen oxide concentration output by the flue gas online monitoring device are collected, and the feed mass flow rate output by the feed metering device is collected. Using the sampling time of the distributed control system as a unified time reference, the data of different sampling frequencies are time-aligned according to the rule of taking the nearest value backward and keeping it until the current reference time. Data exceeding the preset physical upper and lower limits are subjected to amplitude limiting processing. At the same time, the original acquisition time, the aligned time, the original value, and the aligned value are written into the record to obtain the combustion operation data of the waste-to-energy boiler. The preset physical upper and lower limits are the range boundaries given by the equipment nameplate as the basic upper and lower limits, and are combined with the 99th percentile and 1st percentile of the historical valid data of the boiler within no less than one month of operation for convergence correction. Each sensor output is evaluated for quality within a preset sliding time window of 5 minutes. Data points that exceed the range, cause communication timeouts, continuously constant values ​​exceeding the preset duration, or contradict the equipment's operating logic are marked as invalid, such as zero feed mass flow rate but steam flow rate continuously exceeding 20% ​​of the rated value. The reason for invalidity is written to the reason field, with a preset duration of 1 minute. Robust reference values ​​and robust dispersion reference values ​​are calculated for the valid data within the window. The degree of deviation between the current data and the robust reference value is used as the drift residual level, and the proportion of invalid data points in the window to the total data points in the window is used as the missing rate level. Specifically, the absolute difference between the current data and the robust reference value is used as the deviation. This deviation is compared with the robust dispersion reference value by a multiple. When the deviation is no more than twice the robust dispersion reference value, it is classified as Level 1. When the deviation exceeds twice but does not exceed five times the robust dispersion reference value, it is classified as Level 2. When the deviation exceeds five times the robust dispersion reference value, it is classified as Level 3 and treated as a drift anomaly. When the robust dispersion reference value is less than a preset minimum dispersion, the preset minimum dispersion is used as the comparison benchmark. For example, when the steam flow range is 0-100 t / h, the preset minimum dispersion is 0.5 t / h. The missing rate is quantified and graded according to the following rules: within the sliding time window, the number of invalid data points is counted against the total number of data points in the window, and the percentage of invalid data points is calculated. When the percentage does not exceed 5%, it is classified as Level 1. When the percentage exceeds 5% but does not exceed 20%, it is classified as Level 2. When the percentage exceeds 20%, it is classified as Level 3, and the measurement point is marked as unavailable in the current window. Specifically, the calculation of robust reference values ​​and robust dispersion reference values ​​for valid data within the window involves the following steps: First, the data points marked as valid within the sliding time window are sorted by numerical value. The data point in the middle of the sorted list is taken as the robust reference value. Then, based on this robust reference value, the absolute difference between each valid data point within the window and the robust reference value is calculated. These absolute differences are then sorted again, and the absolute difference in the middle of the sorted list is taken as the robust dispersion reference value. When the number of valid data points within the window is less than a preset minimum (e.g., a preset minimum of 3), the robust reference value and robust dispersion reference value from the previous window are used, and the reason for insufficient samples is written in the data label. According to the preset grading rules, the drift residual level and the missing rate level are mapped to the confidence weight. Specifically, when neither the drift residual level nor the missing rate level exceeds the first level, the highest confidence weight is assigned, for example, 1. When either of them is at the second level, a medium confidence weight is assigned, for example, 0.5. When either of them exceeds the second level, the lowest confidence weight is assigned, and the data of that measurement point is marked as unavailable, for example, 0. The confidence weight, drift residual level, missing rate level, validity status, and cause fields are written into the data label corresponding to the combustion operation condition data at that moment. When a key measurement point is marked as unavailable, such as steam flow, steam pressure, nitrogen oxide concentration, etc., the most recent available data is used as the alternative input for that measurement point in the subsequent control cycle, and an alarm record is output. The control cycle is uniformly set to 30 seconds, that is, a complete data acquisition, quality assessment, soft measurement, time delay identification, prediction, and optimization control process is executed every 30 seconds.

[0023] In this embodiment, step S2 specifically includes: Based on the steam flow rate, main steam pressure and temperature, and feedwater temperature, the corresponding enthalpy value is obtained by calling a preset steam property table. This steam property table adopts the IAPWS-IF97 international standard industrial water steam thermodynamic property table. The steam-side enthalpy difference is multiplied by the steam flow rate to convert it into the boiler's heat absorption within that window. The steam-side enthalpy difference is the difference between the main steam enthalpy and the feedwater enthalpy. Then, combined with the feed mass flow rate obtained from the feed metering, and using a boiler thermal efficiency correction coefficient, the heat absorption is converted into the furnace heat input, for example, 0.75. The equivalent lower heating value of the waste entering the furnace is calculated by dividing the furnace heat input by the feed mass flow rate. The heat consumption for moisture evaporation is preset to a fixed value, for example, 2.6 liters per kilogram of moisture. MJ introduces a dry-based equivalent calorific value benchmark for combustible components, which is taken from the back-calculation result of the most recent stability window; the difference between the equivalent lower heating value and the calorific value benchmark is converted into equivalent moisture content according to the fixed evaporation heat consumption, and the equivalent lower heating value and equivalent moisture content are written into the data tag as soft measurement output; To determine a steady state, the fluctuation rate of steam flow and carbon monoxide concentration are calculated within the same sliding window. The fluctuation rate is obtained by dividing the difference between the maximum and minimum values ​​within the window by the average value within the window. When the fluctuation rate of steam flow is no greater than 2%, the fluctuation rate of carbon monoxide concentration is no greater than 25%, and the average value of carbon monoxide concentration within the window is no greater than 100 mg / m³, the window is determined to be a stable window. When a stable window is established and the data labels corresponding to steam flow, steam pressure, feedwater temperature, feed mass flow, and carbon monoxide concentration are available and have a confidence level no lower than medium confidence level, the soft sensor model parameters are updated. The soft sensor model parameters include the boiler thermal efficiency correction coefficient and the dry basis combustible component equivalent calorific value benchmark. The new parameters are weighted and fused by taking 90% of the historical parameters and 10% of the back-calculated parameters of the current window. When the stable window or confidence level conditions are not met, the model parameters of the previous stable window remain unchanged, and the unupdated state and its triggering reason are written into the data label. The triggering reasons include excessive steam flow fluctuation, excessive carbon monoxide fluctuation, excessive average carbon monoxide concentration, and insufficient data confidence.

[0024] In this embodiment, step S3 specifically includes: The setpoints for primary and secondary air volume or damper opening, grate speed, flue gas recirculation valve opening or recirculation fan frequency corresponding to the recirculation amount, and ammonia injection flow rate are aligned with a unified timestamp and then spliced ​​in chronological order to form a control execution sequence. The primary air volume, secondary air volume, grate operating speed, flue gas recirculation amount, and ammonia injection amount when needed are selected as the input sequence to be identified. The steam flow rate, steam pressure, flue gas oxygen volume fraction, carbon monoxide concentration, and nitrogen oxide concentration are selected as the output sequence. The time window length of the input and output sequences is the most recent 15 minutes, and the sampling period is consistent with the control period. The calculation is only performed when the data labels of the input and output sequences within the window are available and the confidence level is not lower than medium confidence level. The confidence level determination requires that at least 90% of the data points within the window meet medium or higher confidence level. Otherwise, the time delay result of the previous window remains unchanged, and the time delay freeze and trigger reason are written in the audit record. For each sequence within the window, the window median is first used for decentering, and then the amplitude is normalized using the median of the absolute deviation of the window. Outliers exceeding the physical boundary are removed. For each pair of input and output sequences, the correlation is calculated one by one within a preset candidate time shift range, where the candidate time shift range is set to zero seconds to ten minutes, and the step size is equal to the control period. Specifically, for each time shift, the input sequence is shifted forward by that time shift and aligned with the output sequence in the overlapping interval. The forward shift means that the timestamp of the input sequence is advanced by that time shift, so that the input action time and the output response time are aligned. After alignment, the product of the two is calculated for each pair of synchronous sampling points and averaged in the overlapping interval. The average value is used as the correlation degree under that time shift. When the number of valid data points in the overlapping interval is less than 50% of the total window length, the correlation degree corresponding to that time shift is marked as invalid and does not participate in the subsequent peak filtering. The time shift corresponding to the maximum correlation is determined as the equivalent time delay from the control execution quantity to the output measurement, and the maximum correlation is written into the audit record as the effectiveness score. A peak filtering rule is set: when the maximum correlation is lower than a preset threshold (e.g., 0.25), or the difference between the maximum and second-highest correlation is less than a preset difference (e.g., 0.05), the identification result is deemed unreliable, and the previous reliable time delay is used. Only when multiple consecutive windows obtain consistent time delays and the effectiveness score remains above the threshold is a new time delay parameter allowed to be updated and locked. The initial time delays from each control execution quantity to each output measurement are set as follows: 60 seconds for primary and secondary air to flue gas oxygen volume fraction, 90 seconds to carbon monoxide concentration, 120 seconds to nitrogen oxide concentration, and 180 seconds to steam flow rate; the time delay from grate speed to each output measurement is increased by 60 seconds; 45 seconds for flue gas recirculation to nitrogen oxide concentration and 90 seconds to other outputs; and 30 seconds for ammonia injection to nitrogen oxide concentration. The obtained equivalent time delay parameters are written into the dynamic prediction model. A time shift consistent with the equivalent time delay is introduced into the control execution quantity at the model input, so that the model uses the control action quantity aligned with the same time axis as the output response when predicting steam stability and emission indicators. In the subsequent optimization solution, the currently issued control increment is regarded as having an effect on steam and emissions only after the equivalent time delay, realizing feedforward compensation at the prediction level and avoiding overshoot and oscillation caused by time-lapse adjustment based on lag measurement. At the same time, the time delay value, effectiveness score, freeze or update reason of each update is written into the action-level audit record for traceability. Specifically, the dynamic prediction model employs a time-delay explicit embedded sliding window regression model. Its input is the historical sequence of control execution quantities and state measurements from the time-aligned combustion operation data. The output is the predicted values ​​of steam flow rate, steam pressure, and flue gas oxygen volume fraction, carbon monoxide concentration, and nitrogen oxide concentration for the next prediction interval. The model input window uses the most recent ten minutes of historical data, and each control execution quantity is time-shifted according to the equivalent time delay obtained in step S3. For example, the primary air volume, secondary air volume, grate speed, and flue gas recirculation volume are shifted forward according to their corresponding equivalent time delays to align with the response times of steam flow rate and emission measurements. For any missing segments that cannot be aligned after shifting, the first effective value is retained to fill them, and a completion flag is written. A feature set is constructed from the aligned input window. The feature set includes the current value of each control execution variable, past multi-step difference values, window mean and window range, as well as the current value and difference value of steam flow and flue gas oxygen volume fraction. The feature set is input into a lightweight predictor to obtain predicted values ​​for the next few steps. In this embodiment, the lightweight predictor is a multi-output regression model with regularization constraints. The predictor output provides both the predicted value and the prediction validity flag, where the prediction validity flag is determined by the minimum and average values ​​of the confidence weights within the input window. When the prediction validity flag is invalid, the control step automatically switches to conservative control and freezes the model update. Conservative control means that the control increment is limited to 50% of the normal range, and only control actions that adjust steam and emissions in a safe direction are allowed. At the same time, the triggering reason for the model invalidation, the corresponding time delay parameters, and the input feature summary are written into the audit record.

[0025] In this embodiment, step S4 specifically includes: The input and output of the dynamic prediction model are time-series aligned based on time delay parameters, and prediction sequences of steam flow rate, steam pressure, flue gas oxygen volume fraction, carbon monoxide concentration, and nitrogen oxide concentration are generated within a preset prediction time domain, which is 10 minutes. Hard constraints are constructed based on the time-delay aligned prediction sequences. Specifically, the deviation of the predicted steam pressure value from the current steam pressure setpoint does not exceed 0.2 MPa, the deviation of the predicted steam flow rate value from the current steam flow rate setpoint does not exceed 2%, the predicted flue gas oxygen volume fraction is in the range of 6% to 9%, the predicted carbon monoxide concentration does not exceed 100 mg / m³, and the predicted nitrogen oxide concentration does not exceed 150 mg / m³. The reliability weights are incorporated into the cost function, with penalty terms set for steam stability deviations, the risk of emissions approaching the upper limit, and the magnitude of changes in control increments. The weighted result of the reliability weights of key measurement points is used as the adjustment amount for the penalty coefficient and safety margin. This automatically increases the penalty for emissions and steam deviations and tightens the safety margin of the emission upper limit when reliability decreases, suppressing misaligned adjustments and overshoot caused by measurement drift or missing data. The cost function is as follows: , in The total cost of model predictive control is represented by k, where k is the prediction step index and N is the number of prediction steps. This represents the confidence weight of key measurement points, with a value range of [0,1]. For example, the highest value is 1, the medium value is 0.5, and the lowest value is 0. This represents the penalty weighting coefficient for the steam pressure deviation term. This represents the predicted steam pressure value at the k-th prediction step. This indicates the target setpoint for steam pressure. This represents the penalty weighting coefficient for the steam flow deviation term. This represents the predicted steam flow rate at the k-th prediction step. This indicates the target setpoint for steam flow rate. This indicates the penalty weight for the oxygen volume fraction constraint in the flue gas. The constraint risk penalty function represents the oxygen volume fraction in the flue gas. This represents the predicted value of the flue gas oxygen volume fraction at the k-th prediction step. This indicates the penalty weight for carbon monoxide constraints. Indicates the penalty weight for nitrogen oxide constraints. This represents the predicted carbon monoxide concentration at the k-th prediction step. This represents the predicted nitrogen oxide concentration at the k-th prediction step. Let T represent the control increment vector at step k, and let T represent the transpose matrix. This represents the control increment penalty matrix; Specifically, the penalty weighting coefficient is set according to the following rules: Prioritize ensuring steam pressure stability; To ensure the stability of steam flow; The oxygen content in the flue gas deviates from the penalty; It is highly sensitive to carbon monoxide concentration; Nitrogen oxide concentration is highly sensitive; where w is the weighted average of the confidence weights of key measurement points, with a value range of [0,1]. Model predictive control is used to solve for the feasible control domain that satisfies the hard constraints. The feasible control domain is represented by the allowable increment range of each control execution variable, including the increment range of primary air volume, secondary air volume, grate speed, flue gas recirculation, and ammonia injection. The feasible control domain and the current operating condition characteristics are input into the strategy network. The strategy network outputs the control increment of each control execution variable, and the control increment is trimmed according to the feasible control domain to ensure that it falls within the feasible control domain. Then, it is converted into the corresponding setpoint correction value and sent to the distributed control system. When any... When the confidence weight of a key measuring point within the current sliding window is zero, or when the average confidence weight of the key measuring point within the window is less than 0.5, automatic correction stops and the system reverts to the historical conservative setting value. The historical conservative setting value is the average of the control setting values ​​that were determined to be in a stable window within the last 30 minutes and whose emissions did not trigger the upper limit. An action-level audit record is generated in the same control cycle. The audit record includes time delay parameters, soft measurement outputs of equivalent lower heating value and equivalent moisture content, a prediction sequence summary, a hard constraint trigger field, a feasible control domain boundary, the actual correction amount issued, and a revert trigger reason field. Specifically, the strategy network adopts a lightweight time-series decision architecture, which concatenates the combustion operation characteristics after time delay alignment, the equivalent lower heating value and equivalent moisture content obtained by soft measurement, the statistical summary of the steam and emission prediction sequence in the prediction time domain, and the confidence weights of key measurement points and the time delay frozen state into an input vector. First, the combined features are extracted through two layers of fully connected nonlinear mapping, and then the feature sequences of the most recent control cycles are fused through a gated cyclic unit. Finally, the control increment suggestions for primary air, secondary air, grate speed, flue gas recirculation and ammonia injection are output at once through a linear output layer. The output suggestions are pruned according to the feasible control domain, that is, each control increment is restricted within the allowable upper and lower bounds and rate of change constraints obtained by model predictive control solution. Only after pruning are the results converted into setpoint corrections and issued.

[0026] like Fig. 2A comparative chart of the effects of an AI-based waste-to-energy boiler combustion control method is presented. From the perspective of indicators: reduced steam pressure and steam flow fluctuations indicate that the invention suppresses parameter oscillations caused by mistimed adjustments under the effects of prediction alignment and hard constraints; reduced carbon monoxide and nitrogen oxide exceedance rates indicate that the invention proactively avoids exceedance risks under emission upper limit constraints and a credibility-driven safety margin tightening mechanism; reduced control oscillation frequency indicates that the invention first solves the feasible control domain using model predictive control, and then outputs and prunes the data within the feasible domain using a policy network, avoiding out-of-bounds errors and frequent reverse corrections; reduced prediction bias indicates that after incorporating equivalent time delay parameters into the dynamic prediction model, the control action and measurement response are aligned on the time axis, reducing systematic deviations caused by monitoring delays and in-furnace transmission lags; and reduced stage switching stability indicates that by adopting operational stage identification and staged model parameters, staged time delay tables, and staged constraint weights, control biases caused by using the same set of parameters across stages are avoided, making stage switching processes such as load ramping and denitrification switching smoother. In summary, the figure reflects that the present invention is superior to the prior art in four aspects: steam stability, emission compliance, control smoothness, and cross-stage robustness. This solves the problems of the prior art's time-shifting adjustment defects and the systematic deviation caused by the switching of operation stages.

[0027] Example 2: This embodiment 2 provides an improved solution based on embodiment 1. Embodiment 1 achieves time axis alignment between control action and measurement response by aligning multi-source operating data, generating confidence weights for sensor drift and missing data, and writing time delay parameters into the dynamic prediction model based on cross-correlation to determine equivalent time delays. This eliminates the problem of misaligned adjustment caused by the superposition of online flue gas monitoring delay and in-furnace transmission lag. However, under complex operating conditions such as rapid fluctuations in waste calorific value and moisture content, frequent load increases and decreases, and switching of the denitrification system, although the dynamic prediction model and strategy network have achieved time delay alignment, another type of problem still occurs: the combustion mechanism and dominant constraint factors switch at different operating stages, causing systematic prediction deviations and control biases when the same set of model parameters and the same set of strategies are used across stages. In this embodiment, after data acquisition and quality assessment in step S1, an operation phase identification and phase labeling sub-step is added, whereby the operation phase label is written into the phase field of the data label: Within each control cycle, phase discrimination features are calculated from the aligned combustion operation condition data. These features include the steam flow rate variation, steam pressure variation, feed mass flow rate variation, primary air to secondary air ratio variation, flue gas recirculation variation, ammonia injection variation, and the proximity of carbon monoxide and nitrogen oxide concentrations to the emission limits. The current operating condition is classified into operation phase labels in the preset phase library according to preset phase discrimination rules, and the confidence score of the discrimination result is calculated. The confidence score comprehensively considers feature deviation, duration stability, and historical switching frequency. Specifically, it is calculated by: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] The deviation from the typical characteristic range of this stage is normalized to [0,1], and the weighted average of its reciprocal is taken as the feature matching degree. The typical characteristic range is the 10th to 90th percentile interval of the historical valid samples of this stage. When there are less than 50 historical valid samples of this stage, a preset initial typical range is used. For example, the typical range of steam flow fluctuation rate in the stable combustion stage is [0.5%, 2%], and the typical range of steam flow change amplitude in the load rise stage is [3%, 8%]. The ratio of the current stage duration to the typical duration of this stage is limited to the range of [0,1] as the continuous stability. The reciprocal of the number of stage switching times in the last 30 minutes is normalized to [0,1] as the switching stability. Confidence score = 0.5 × feature matching degree + 0.3 × continuous stability + 0.2 × switching stability. When the confidence score is lower than 0.At 6 o'clock, the system is determined to be in a transitional operating condition. The stage label is not switched temporarily, and a transitional operating condition flag is written into the data label. The operating stage labels include the startup stage, stable combustion stage, load increase stage, load decrease stage, denitrification switching stage, rapid feed fluctuation stage, shutdown stage, and abnormal operating condition stage. The priority order is: abnormal operating condition stage, startup and shutdown stage, switching and fluctuation stage, load adjustment stage, and stable combustion stage. The stage discrimination rule is set using a multi-condition simultaneous satisfaction, duration determination, and hysteresis maintenance method. Specifically, when the steam flow rate changes by more than 3% of the rated steam flow rate within four consecutive control cycles and the direction of change is consistent, it is determined to be in a load increase stage or a load decrease stage. The process is divided into several stages: a decrease phase; a decrease in ammonia injection rate within a control cycle exceeding 10% of its range, or ammonia injection switching from zero to non-zero and vice versa, is considered the denitrification switching phase; a rapid feed fluctuation phase is defined as the feed mass flow rate fluctuating more than 5% of its window average over two consecutive control cycles and the carbon monoxide concentration fluctuating more than 25%; a stable combustion phase is defined as the phase where none of the above conditions are met, the steam flow rate fluctuating more than 2%, and the carbon monoxide concentration fluctuating more than 25%. Each stage determination result must be maintained continuously for at least two minutes before switching stage labels, and the switched stage label must be maintained for at least five minutes to eliminate frequent jumps. After generating the operational phase label, this embodiment writes the operational phase label into the phase field of the data label and archives the data in segments using the phase field: maintaining independent sliding window statistics, independent confidence weight grading thresholds, and independent valid sample counts that can be used for model updates for different operational phases; when a phase switch is detected, a phase switch flag and switch reason are written into the data label, and the statistical baseline and model update entry of the previous phase are frozen. The subsequent steps are only allowed to start updating the model parameters and strategy parameters corresponding to the new phase after the cumulative number of samples in the new phase reaches a preset minimum, such as 30 valid samples. This avoids directly using the parameters of the stable phase for the switching phase and directly using the parameters of the switching phase for the stable phase, thus eliminating the systematic prediction bias and control bias caused by using the same set of parameters across phases.

[0028] In this embodiment, for step S2 of embodiment 1, the maintenance and updating of soft sensor model parameters is improved from a global single parameter set to a staged parameter set. Soft sensor parameters are maintained independently for different operating stages, and staged steady-state criteria are adopted: the operating stage label generated in step S1 is used as the segment index for this step, and an independent soft sensor parameter set is maintained for each operating stage. The soft sensor parameter set includes the boiler thermal efficiency correction coefficient corresponding to the stage, the equivalent calorific value benchmark of dry-based combustible components corresponding to the stage, and the correction coefficient for the heat consumption converted from water evaporation corresponding to the stage. When a change in the operating stage label is detected, the soft sensor parameter update of the previous operating stage is immediately frozen, and the soft sensor calculation is performed using the soft sensor parameter set corresponding to the new operating stage. If the new operating stage is the first occurrence or its effective sample count is less than the preset minimum, the initialization parameter set of the same type as the operating stage in the historical database is called, and the stage initialization status and its source field are written into the data label. A phased steady-state criterion is adopted. When the operating phase label is stable combustion phase, the steady-state criterion is still used. When the operating phase label is load increase phase or load decrease phase, the steam flow fluctuation rate threshold is relaxed to no more than 5%, and an additional criterion is added that the steam flow change direction is consistent and the steam pressure deviation does not exceed the preset pressure bandwidth. When the operating phase label is denitrification switching phase, the steady-state criterion is switched to the ammonia injection rate remaining unchanged for two consecutive minutes and the average value of the nitrogen oxide concentration window not exceeding the threshold after deducting the safety margin from the emission limit. The average value threshold of carbon monoxide concentration is replaced with the threshold after deducting the safety margin from the emission limit for carbon monoxide concentration. When the operating phase label is rapid fluctuation phase of incoming materials, the parameter update entry is canceled, only the soft measurement results are allowed to be output without updating the soft measurement parameters, and the fluctuation phase update prohibition status is written in the data label. When the steady-state criteria corresponding to the current operating stage are met, and the key measurement point data labels are available and the confidence level is not lower than medium confidence level, the soft measurement model parameters corresponding to the operating stage are updated. The update adopts a weighted fusion method, which combines 90% of the historical parameters with 10% of the parameters back-calculated in the current window. When the staged steady-state criteria or confidence level conditions are not met, the soft measurement model parameters corresponding to the operating stage remain unchanged, and the unupdated state and its triggering reason are written in the data label. The triggering reason includes at least steam flow fluctuation exceeding the limit, carbon monoxide fluctuation exceeding the limit, nitrogen oxides approaching the upper limit, ammonia injection not stable, insufficient data confidence, and stage switching freeze. This avoids directly using the thermal efficiency correction coefficient and calorific value benchmark obtained in the stable stage for the switching stage or the rapid fluctuation stage of incoming materials, eliminates the systematic soft measurement deviation caused by using the same set of parameters across stages, and reduces the bias of subsequent prediction and control.

[0029] In this embodiment, for step S3 of embodiment 1, the time delay identification is improved from a global single time delay table to a staged time delay table. Only data from the same stage are used in the time delay calculation, and time delay parameters are maintained independently for different operating stages. Before performing time delay identification and prediction alignment, the operating stage field written in the data tag in step S1 is read first, and the input sequence and output sequence are filtered for the same stage using this operating stage field. Only data points whose operating stage is consistent with the current stage within the most recent 15-minute window are used in the time delay calculation. It is required that the number of data points of this type in the same stage accounts for no less than a preset proportion of the total number of points in the window, and the proportion of data points with a confidence level of no less than medium confidence level is no less than 90%. The preset proportion is as follows: in the stable combustion stage: data points in the same stage account for no less than 80% of the total number of points in the window; in the load adjustment stage: no less than 70%; in the denitrification switching stage: no less than 60%; and in the rapid fluctuation stage of incoming materials: no less than 50%. During startup and shutdown phases: no less than 50%; during abnormal operating conditions: time delay identification is not performed, and the locking time delay of the stable combustion phase is directly adopted; if the above percentage conditions are not met, the time delay parameters are not updated and the locking time delay of the previous window is used, and the insufficient sample freezing and its reasons are written in the audit record; furthermore, an independent time delay parameter table and locking status are maintained for each operating phase, that is, the same control execution quantity to the same output quantity corresponds to different equivalent time delay parameters in different operating phases; when the operating phase is switched, the time delay parameters of the previous phase are immediately stopped, and the time delay parameter table of the new phase is called as the initialization time delay to participate in the relevant peak search, and a phased locking rule is set: only when the new phase obtains consistent time delay results in four consecutive sliding windows and the effectiveness score is not lower than the preset threshold, the time delay parameters of the new phase are updated and locked, otherwise the initialization time delay of the new phase remains unchanged; By writing the stage label, stage time delay parameter, and freeze or lock reason into the action-level audit record, and by using the above-mentioned data from the same stage to participate in the calculation, the stage time delay parameter table, and the improvement of the stage switching freeze and lock conditions, the time delay of the stable stage is avoided from being directly used for the switching stage or load ramping stage. This eliminates the systematic prediction deviation and control bias caused by cross-stage misalignment, thereby ensuring that subsequent dynamic prediction and optimization control maintain time axis consistency and constraint consistency during stage switching.

[0030] In this embodiment, for step S4 of embodiment 1, the constraint optimization is improved from a global single constraint set and weight set to a staged constraint set and weight set. The policy network is trained and calls parameters independently according to the stage, and gating and transitional contraction are performed when switching stages. According to the running stage field in the data label written in step S1, the staged constraint set, staged cost weight set, and staged policy parameter set corresponding to each running stage are maintained and called respectively. Among them, the staged constraint set includes steam pressure deviation bandwidth, steam flow deviation bandwidth, allowable range of flue gas oxygen volume fraction, and carbon monoxide concentration. The emission limit is set with upper limits for nitrogen oxide concentration, and a phased safety margin is introduced for the upper limit. This safety margin is determined by the reliability weights of key measurement points; as reliability decreases, the safety margin increases by a fixed proportion, effectively tightening the upper limit. The phased cost weight set includes penalties for steam pressure deviation, steam flow deviation, flue gas oxygen volume fraction deviation, carbon monoxide exceedance risk, and nitrogen oxide exceedance risk. Different weight sets are specified for different operating phases, with a weighting system that balances steam stability and emission constraints during the stable combustion phase. The solution is reorganized, increasing the penalties for deviations in steam pressure and steam flow rate during load increase and decrease phases, and simultaneously increasing the penalty for changes in control increments to suppress ramp-up oscillations. During the denitrification switching phase, the penalty for exceeding nitrogen oxide limits is increased, and the penalty for changes in control increments related to ammonia injection is reduced to a fixed proportion of the stable phase to ensure controllability. The phased strategy parameter set is a set of parameters stored by the strategy network according to the operating phase. In each control cycle, the strategy network only uses data samples that are consistent with the current operating phase and have a confidence level not lower than medium confidence level for updating parameters, and it is prohibited to use cross-phase samples to update parameters. When the operating phase changes, a switching gating is executed: the strategy parameter update and cost weight adaptive update of the previous phase are frozen, the constraint set, weight set and strategy parameter set corresponding to the new phase are switched and called, and within the preset transition time after the switch, the upper limit of the allowable increment of the feasible control domain is contracted to a fixed proportion of the stable phase, and only control increments that allow changes in steam pressure, steam flow rate and emission indicators in the direction of satisfying hard constraints are allowed; at the same time, only when the new phase continuously meets the preset number of control cycles and the hard constraints are not triggered, the transition contraction is released and the normal feasible control domain and strategy output of the new phase are restored. Through the improvements in staged constraints and weights, staged strategy parameters, stage switching freeze and transition contraction, the model trade-offs and control biases of the previous stage are avoided when switching between stages, thus eliminating the systematic prediction bias and control bias caused by using the same set of strategies across stages.

[0031] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. A method for combustion control of a waste-to-energy boiler based on artificial intelligence, characterized in that, include: Step S1: Collect data from the distributed control system, flue gas online monitoring, and feed metering, and align them with a unified timestamp to obtain the combustion operation data of the waste-to-energy boiler; Calculate the drift residual and data missing rate for each sensor output, generate confidence weights according to preset grading rules, and write them into the data labels; Step S2: Based on the energy balance equation, the equivalent lower heating value and equivalent moisture content of the waste fed into the furnace are calculated from the steam flow rate and the feed mass flow rate; the steam flow rate fluctuation rate and carbon monoxide concentration fluctuation rate are set as dual criteria, and the soft measurement model parameters are updated within the stable window; Step S3: Within the sliding time window, the peak method of cross-correlation function is used to determine the equivalent time delay from each control quantity to the steam flow and emission index. The time delay parameter is updated only when consistent time delays are obtained in multiple consecutive windows and the effectiveness score is consistently higher than the threshold. The time delay parameter is written into the dynamic prediction model. A time shift consistent with the equivalent time delay is introduced into the control execution quantity at the model input to achieve feedforward compensation. Step S4: Construct hard constraints for steam stability and emission limits using the time-delay aligned predictions, and write the confidence weights into the cost function; use model predictive control to solve the feasible control domain, and then the policy network selects the control increment within the feasible domain and outputs it as the setpoint correction amount to be sent to the distributed control system; when the confidence weights are lower than the threshold, stop automatic adjustment and fall back to the historical conservative setpoint, and at the same time generate an audit record containing time delay, soft measurement and constraint trigger fields.

2. The method for combustion control of a waste-to-energy boiler based on artificial intelligence according to claim 1, characterized in that: Step S1 also includes: Using the sampling time of the distributed control system as a unified time reference, data with different sampling frequencies are time-aligned according to the rule of taking the nearest value backward and keeping it until the current reference time. Data exceeding the preset physical upper and lower limits are subjected to amplitude limiting processing. At the same time, the original acquisition time, the aligned time, the original value and the aligned value are written into the record to obtain the combustion operation data of the waste-to-energy boiler. For each sensor output, a quality assessment is performed within a preset sliding time window. Data points that exceed the range, experience communication timeouts, continuously maintain a constant value for more than a preset duration, or contradict the device's operating logic are marked as invalid, and the reasons for invalidity are written into the reason field. Robust reference values ​​and robust dispersion reference values ​​are calculated for the valid data within the window. The degree of deviation between the current data and the robust reference values ​​is used as the drift residual level, and the proportion of invalid data points in the window to the total number of data points in the window is used as the missing rate level. The drift residual level and missing rate level are mapped to confidence weights and written into the data labels. The highest confidence weight is assigned when both the drift residual level and the missing rate level do not exceed the first level; the medium confidence weight is assigned when either one is at the second level; and the lowest confidence weight is assigned when either one exceeds the second level, and the data for that measurement point is marked as unavailable. When a critical measurement point is marked as unavailable, the most recently available data is used as the alternative input for that measurement point, and an alarm record is output.

3. The method for combustion control of a waste-to-energy boiler based on artificial intelligence according to claim 2, characterized in that: After the execution data collection and quality assessment, the specific steps for adding operational phase identification and phased labeling are as follows: The system calculates stage discrimination features from the aligned combustion operation data, classifies the current operation into operation stage labels in the preset stage library according to preset stage discrimination rules, and calculates the confidence score of the discrimination result. When the confidence score is lower than a preset threshold, the operation stage label is not updated and the previous operation stage label is used, and the stage freeze reason field is written. The confidence score comprehensively considers feature deviation, duration stability, and historical switching frequency. After generating the runtime stage label, the runtime stage label is written into the stage field of the data label, and the data is segmented and archived using the stage field: separate sliding window statistics, separate confidence weight grading thresholds, and separate valid sample counts that can be used for model updates are maintained for different runtime stages; when a stage switch is detected, the stage switch flag and switch reason are written into the data label, and the statistical baseline and model update entry of the previous stage are frozen. The corresponding model parameters and strategy parameters for the new stage are only updated after the cumulative number of samples in the new stage reaches the preset minimum.

4. The method for combustion control of a waste-to-energy boiler based on artificial intelligence according to claim 1, characterized in that: Step S2 also includes: Based on the steam property table, the steam-side enthalpy difference is multiplied by the steam flow rate to convert it into the boiler's heat absorption within the specified window. The steam-side enthalpy difference is the difference between the enthalpy of the main steam and the enthalpy of the feedwater. Combined with the feed mass flow rate obtained from the feed metering, and using the boiler thermal efficiency correction coefficient, the heat absorption is divided by the thermal efficiency correction coefficient and converted into the furnace heat input. The equivalent lower heating value of the waste entering the furnace is then calculated by dividing the furnace heat input by the feed mass flow rate. The heat consumption for moisture evaporation is preset to a fixed value, and a dry-basis combustible component equivalent calorific value benchmark is introduced. This benchmark is taken from the back-calculation result of the most recent stable window. The difference between the equivalent lower heating value and the calorific value benchmark is converted into equivalent moisture content according to the fixed evaporation heat consumption. The equivalent lower heating value and equivalent moisture content are written into the data tag as soft measurement output. Within the same sliding window, a stable window is determined based on the fluctuation rate of steam flow, the fluctuation rate of carbon monoxide concentration, and the average value of carbon monoxide concentration within the window. When a stable window is established and the data label is available and the confidence level is not lower than the medium confidence level, the soft sensor model parameters are updated. The update adopts a weighted fusion of historical parameters and the back-calculated parameters of the current window, and when the update conditions are not met, the unupdated state and the triggering reason are written.

5. The method for combustion control of a waste-to-energy boiler based on artificial intelligence according to claim 4, characterized in that: The specific steps for improving the maintenance and updating of soft measurement model parameters from a global single parameter set to a staged parameter set, with independent maintenance of soft measurement parameters in different operating stages and the adoption of staged steady-state criteria, are as follows: An independent set of soft measurement parameters is maintained for each operating stage. The set of soft measurement parameters includes the boiler thermal efficiency correction factor for that stage, the equivalent calorific value of dry-based combustible components for that stage, and the correction factor for the heat consumption due to water evaporation for that stage. When a change in the running phase label is detected, the soft measurement parameter update of the previous running phase is immediately frozen, and the soft measurement calculation is performed on the soft measurement parameter set corresponding to the new running phase. If the new running phase is the first time it appears or its number of valid samples is less than the preset minimum number, the initialization parameter set of the same type as the running phase in the historical database is called, and the phase initialization status and its source field are written into the data label. When the steady-state criteria corresponding to the current operating stage are met, and the key measurement point data labels are available and the confidence level is not lower than medium confidence level, the soft measurement model parameters corresponding to the operating stage are updated. When the staged steady-state criteria or confidence level conditions are not met, the soft measurement model parameters corresponding to the operating stage are kept unchanged, and the unupdated state and its triggering reason are written into the data label.

6. The method for combustion control of a waste-to-energy boiler based on artificial intelligence according to claim 1, characterized in that: Step S3 also includes: Within a recently preset duration window, a control execution sequence and an output sequence are constructed. Time delay identification is only performed when at least a preset proportion of data points within the window meet medium or higher confidence levels. The time delay identification includes decentralizing the sequence using the window median and normalizing the window absolute deviation median; calculating the correlation between the input sequence and the output sequence after alignment within a preset candidate time shift range; and marking the time shift as invalid if the number of valid data points in the overlapping interval is less than a preset proportion. The time shift corresponding to the maximum correlation value is determined as the equivalent time delay, and the maximum correlation value is used as... The validity score is written into the audit log. When the maximum correlation is lower than the preset threshold or the difference between the maximum correlation and the second largest correlation is less than the preset difference, the identification result is determined to be unreliable and the previous reliable time delay is used. The equivalent time delay is written into the dynamic prediction model. A time shift consistent with the equivalent time delay is introduced into the control execution quantity at the model input. The first missing segment generated by the shift is filled in by the first valid value retention method and a filling flag is written. When the prediction validity flag is invalid, conservative control is switched and model updates are frozen. At the same time, the triggering reason and the corresponding time delay parameter are written into the audit log.

7. The method for combustion control of a waste-to-energy boiler based on artificial intelligence according to claim 6, characterized in that: The specific steps for time delay identification using a staged time delay table maintenance are as follows: Before performing time delay identification and prediction alignment, the running stage field is read, and the input and output sequences are filtered for the same stage using this running stage field. Only data points whose running stage is consistent with the current stage within the most recent 15-minute window are used to participate in the time delay calculation. It is required that the number of data points of this type in the same stage accounts for no less than a preset proportion of the total number of points in the window, and the proportion of data points with a confidence level of no less than medium confidence level is no less than 90%. If the above proportion conditions are not met, the time delay parameters are not updated, and the locked time delay of the previous window is used. The audit record is frozen due to insufficient samples and the reasons. Maintain an independent time delay parameter table and lockout status for each operating stage, that is, the same control execution quantity to the same output quantity corresponds to different equivalent time delay parameters in different operating stages; When the operation phase changes, the time delay parameters of the previous phase are immediately stopped. Instead, the time delay parameter table of the new phase is called as the initial time delay to participate in the relevant peak search. The phase locking rule is set: the time delay parameters of the new phase are only updated and locked when the new phase obtains consistent time delay results in a number of consecutive sliding windows and the validity score is not lower than the preset threshold. Otherwise, the initial time delay of the new phase remains unchanged.

8. The method for combustion control of a waste-to-energy boiler based on artificial intelligence according to claim 1, characterized in that: Step S4 also includes: Within a preset prediction time domain, a prediction sequence of steam flow rate, steam pressure, flue gas oxygen volume fraction, carbon monoxide concentration, and nitrogen oxide concentration is generated. Hard constraints are constructed using the time-delay-aligned prediction sequence. These hard constraints include: steam pressure prediction deviation not exceeding a preset pressure deviation upper limit; steam flow rate prediction deviation not exceeding a preset flow rate deviation upper limit; flue gas oxygen volume fraction within a preset range; and carbon monoxide concentration and nitrogen oxide concentration not exceeding their respective upper limits. The credibility weight of key measurement points is introduced into the cost function, and penalty terms are set for steam stability deviation, emission risk approaching the upper limit, and control increment change magnitude. When the credibility decreases, the penalty for steam and emission deviations is increased, and the safety margin of the emission upper limit is tightened. Model predictive control is used to solve for the feasible control domain that satisfies hard constraints. The feasible control domain is then input into the strategy network along with the current operating condition characteristics. The output control increment is then pruned according to the feasible control domain and distributed to the distributed control system. When the confidence weight of any key measurement point is the lowest or the window average of the confidence weight of the key measurement point is lower than the preset threshold, the system reverts to the historical conservative setting value and generates an action-level audit record.

9. The method for combustion control of a waste-to-energy boiler based on artificial intelligence according to claim 8, characterized in that: The specific steps for improving constraint optimization from a global single constraint set and weight set to a staged constraint set and weight set are as follows: Based on the operation phase field, maintain and call the phased constraint set, phased cost weight set, and phased strategy parameter set corresponding to each operation phase. The phased cost weight set includes steam pressure deviation penalty weight, steam flow deviation penalty weight, flue gas oxygen volume fraction deviation penalty weight, carbon monoxide exceedance risk penalty weight, and nitrogen oxide exceedance risk penalty weight, and stipulate that different weight sets are used for different operation phases. The phased strategy parameter set is a set of parameters stored by the strategy network according to the running stage. The strategy network only uses data samples that are consistent with the current running stage and have a confidence level of not less than medium confidence level for updating, and it is prohibited to use cross-stage samples to update parameters.

10. The method for combustion control of a waste-to-energy boiler based on artificial intelligence according to claim 9, characterized in that: The policy network is trained and its parameters are invoked independently in stages. The specific steps for gating and transitional shrinkage during stage switching are as follows: When a phase transition occurs, a transition gating is executed, freezing the strategy parameter updates and cost weight adaptive updates of the previous phase. The constraint set, weight set, and strategy parameter set corresponding to the new phase are then invoked. Within a preset transition period after the transition, the upper limit of the allowable increment of the feasible control domain is contracted to a fixed proportion of the stable phase, allowing only control increments that cause changes in steam pressure, steam flow, and emission indicators in the direction of satisfying hard constraints. Simultaneously, the transition contraction is lifted and the normal feasible control domain and strategy output of the new phase are restored only when the new phase continuously meets a preset number of control cycles and the hard constraints are not triggered.

Images