System for reducing emission of chloride ions in waste incineration fly ash based on machine learning

Abstract

The invention discloses an emission reduction system for chloride ions in waste incineration fly ash based on machine learning, relates to the technical field of solid waste resourceful treatment, is used for data-driven fly ash multi-stage dosing collaborative removal, and is divided into fly ash pretreatment and data acquisition, online ion monitoring and dosing decision, multi-stage reaction and dynamic feedback, OTA upgrading and cross-plant migration. And removing the multi-dimensional pollutants synergistically. The dosing ratio and the reaction stage are optimized in real time through a gradient boosting tree-reinforcement learning model, and the concentrations of chloride ions, heavy metals and dioxin precursors are monitored on line. According to the system, monitoring and decision-making paths of all stages are stored in a cloud knowledge graph and can be upgraded through OTA, the chloride ion removal rate is increased by more than three percent, the agent dosage is reduced by 25%, the scheme can adapt to the difference of fly ash components in different factories, economy and environmental protection are both considered, and the system has remarkable popularization value in large-scale waste incineration and hazardous waste disposal.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of solid waste resource processing, in particular to a garbage incineration fly ash chlorine ion emission reduction system based on machine learning. BACKGROUND

[0002] In the process of dechlorination and multi-pollutant co-processing of garbage incineration fly ash, the key parameters such as reagent dosage, reaction order, stirring rate and residence time are significantly affected by the composition of incoming fly ash, grate load and environmental condition fluctuations, showing typical industrial process characteristics of multivariable, strong coupling, nonlinearity and time-varying. At present, most projects still regard fly ash emission reduction as a simple chemical disposal problem, and pay insufficient attention to dynamic disturbance and parameter uncertainty at the process control level, resulting in difficulty in maintaining long-term stability and economic optimization of the control strategy.

[0003] With the expansion of incineration scale, enterprises generally face the practical contradictions of "dechlorination efficiency fluctuation, reagent cost increase, and increasing pressure of secondary pollution control" in operation; although means such as staged dosing, online monitoring and data analysis are introduced, but for the rapid fluctuation of fly ash properties and the difference of cross-factory conditions, the self-adaptation and robustness of the existing control strategy are still insufficient.

[0004] The existing technology mainly uses multi-stage water / acid washing, dry / semi-dry method dosing and carbonation solidification processes, and is supplemented by fixed parameter PID or sequence control logic for adjustment. However, such methods generally have the following problems: online ion and key pollutant monitoring lags, feedback distortion caused by unmodeled data noise; in the face of periodic or random disturbance, fixed or empirical parameters are difficult to balance removal efficiency and reagent consumption economy; when centralized disposal or cross-factory deployment is used, the fly ash alkalinity, salt and heavy metal content differ significantly, and the static control scheme is difficult to migrate, resulting in unstable removal rate and co-processing reduction effect.

[0005] Therefore, there is an urgent need for a multi-stage dosing control system or method that takes process control as the core and can learn and self-adjust based on real-time multi-source data: it should have disturbance identification and compensation, rolling optimization and multi-objective trade-off capabilities, and can be quickly deployed in different factories to achieve comprehensive optimal control of chlorine ion removal rate, reagent consumption and multi-pollutant co-processing reduction. SUMMARY

[0006] (I) Technical problems solved In view of the deficiencies of the prior art, the present application provides a machine learning-based chloride ion emission reduction system in waste incineration fly ash, which comprises five steps of fly ash pretreatment and data acquisition, online ion monitoring and dosing decision, multi-stage reaction and dynamic feedback, OTA upgrade and cross-factory migration, and multi-dimensional pollutant synergistic removal. The gradient boosting tree-reinforcement learning model is used to optimize the dosing ratio and reaction order in real time, and the concentrations of chloride ions, heavy metals and dioxin precursors are monitored online. The system stores the monitoring and decision paths of each stage in the cloud knowledge graph and can be OTA upgraded, which realizes more than 300% improvement in chloride ion removal rate and 25% reduction in reagent consumption. The scheme can adapt to the differences in fly ash composition in different factories, taking into account the economy and environmental protection, thereby solving the technical problems described in the background art.

[0007] (Two) Technical solutions To achieve the above object, the present application is realized by the following technical solutions: a machine learning-based chloride ion emission reduction system in waste incineration fly ash, comprising: when the detection rate of large particles in fly ash is detected to be over standard, the pretreatment module performs multi-stage screening and grinding on fly ash raw materials, real-time detection and marking of fly ash particle homogeneity factor, pH and temperature, and the detection results are summarized as preliminary fly ash characteristic data set; The gradient boosting tree-reinforcement learning model is called by the online ion monitoring module, and the chloride ion concentration and removal rate target is obtained in real time based on the ion difference function, and the initial dosing amount and reaction order are output; If the removal rate does not reach the threshold after each reaction, the multi-stage dosing unit executes the candidate dosing scheme on the fly ash reaction system and maintains constant temperature stirring, records the multi-stage reaction synergy index and ion chromatography data, and dynamically modifies the dosing scheme according to the dynamic correction score; When the dosing cycle is completely finished, the cloud algorithm center archives the existing data set and generates an OTA update package, and upgrades the gradient boosting tree-reinforcement learning model to the latest version and sends it to each factory, synchronously migrating the best dosing scheme and fly ash spectrum adaptation parameters; When the concentrations of heavy metals or dioxin precursors are detected to be over limit, the multi-dimensional pollutant synergistic removal module adds special monitoring items and expands the multi-objective optimization function, dynamically introduces chelating agent or oxidizing agent dosing, and performs synchronous control of toxic components and chloride ions.

[0008] Further, in the fly ash pretreatment and data acquisition process, when the large particle impurity removal rate exceeds the preset threshold, the fly ash is subjected to multi-stage screening, grinding and auxiliary dispersion treatment to obtain pretreated fly ash with fly ash particle homogeneity factor, and the homogenization information and corresponding pH, temperature and stirring rate are recorded in real time.

[0009] Further, through the multi-point distributed sensor arrangement, a preliminary fly ash characteristic data set is established for the pretreated fly ash, and quality checking is performed during structured storage to realize adaptive correction when the fly ash homogeneity, sensor difference detection and backtracking grinding screening conditions are implemented.

[0010] Further, in the online ion monitoring and initial dosing decision stage, in order to continuously track the chloride ions and calcium, sodium ions in the fly ash system, the ion difference function is used to evaluate the ion concentration and process deviation degree in the system in real time, and the monitoring data is updated to the online ion monitoring data set at the same time; If the ion difference function is monitored to be higher than the preset difference threshold, according to the comparison of the fly ash characteristics in the preliminary fly ash characteristic data set, the comparison result is used to trigger an alarm or a correction measure.

[0011] Further, the gradient boosting tree-reinforcement learning model is called to calculate and output the initial dosing amount of CaO and NaAlO2 and the reaction order; and according to the real-time changes of water quality parameters, fly ash composition and operating conditions, the initial dosing amount and reaction order are corrected in real time; the selected initial dosing scheme includes: reagent dosing amount, preset reaction order and process time length suggestion.

[0012] Further, when the gradient boosting tree-reinforcement learning model determines that it has entered the first stage or subsequent stages of reaction, multi-stage circulating dosing and stirring operations are performed, and the multi-stage reaction synergy index is used to quantify the contribution degree of this stage of reaction, so as to realize the efficiency evaluation of fly ash chloride ion reaction in stages.

[0013] Further, after completing any stage of reaction, the chloride ion concentration is collected and real-time feedback is returned to the cloud algorithm center, if the chloride ion removal rate is lower than the set threshold, the next stage of reaction dosing formula is adjusted in real time according to the dynamic decision correction function or terminated in advance, and the correction result is stored in the multi-stage reaction feedback data set.

[0014] Further, the multi-stage reaction feedback data set, dosing decision path and treatment effect of the whole process of multi-stage reaction are uploaded to the cloud database and indexed to the knowledge graph; The preliminary fly ash characteristic data set and the online ion monitoring data set are linked in multiple dimensions by using semantic networks and association rules, which are used to support fast retrieval during subsequent OTA upgrade and cross-plant migration.

[0015] Further, the cloud algorithm center automatically generates an OTA update package after comparing the parameters of the gradient boosting tree-reinforcement learning model across batches or across plants, and distributes the updated model hyperparameters, kernel function parameters and decision strategies to each local plant, so that each plant can dynamically obtain the latest dosing scheme without stopping production.

[0016] Further, when the new plant area fly ash spectrum difference is greater than the expected, through knowledge graph retrieval and parameter mapping, based on the migration fitness, the applicability of the new fly ash treatment process is evaluated, and the existing model is adaptively optimized by means of the incremental updating function, so that the rapid adaptation of each heterogeneous fly ash batch is realized; wherein, if the migration fitness continuously decreases below the expected value within a specified time, the iteration intensity of the parameter mapping scheme is increased.

[0017] Further, when it is detected that the target pollutant concentration exceeds the set threshold, the real-time or timing sampling monitoring items of the multi-dimensional pollutants are increased, and the pollutant deviation function is established, the concentration of the toxic component in the pollutant deviation function is comprehensively evaluated with the indicators such as pH and chloride ions, for subsequent multi-target dosing decision calling.

[0018] Further, the gradient boosting tree-reinforcement learning model performs multi-objective optimization on the newly added heavy metals and dioxin precursors based on a multi-objective optimization function, outputs a comprehensive dosing vector containing CaO, NaAlO2 and chelating agents, oxidizing agents, and adaptively balances the removal efficiency of chloride ions and other highly toxic components. Wherein, based on the multi-objective optimization function target, the gradient boosting tree combines the reinforcement learning search scheme to automatically generate a multi-objective dosing scheme based on the prediction of multi-dimensional pollutant removal efficiency.

[0019] Further, when performing multi-stage reactions, if the online detection result shows that the pollutant deviation function continuously exceeds the expected value, the next stage dosing formula is dynamically corrected based on the correction gain function, and the correction process is recorded.

[0020] (Three) beneficial effects The application provides a machine learning-based chloride ion emission reduction system for waste incineration fly ash, which has the following beneficial effects: The fly ash particle homogeneity factor is used to quantitatively characterize the particle size distribution and large particle removal efficiency of fly ash, which not only highlights the innovation of the front-end treatment, but also lays a solid foundation for the contact efficiency of the medicament and fly ash during multi-stage dosing.

[0021] The ion difference function and the gradient boosting tree-reinforcement learning (GBT+RL) fusion model are used to form a data-driven adaptive dosing scheme, which can realize accurate prediction of the dosing amount of CaO and NaAlO2. This process interacts with the online ion monitoring data set in real time, which not only ensures efficient removal of chloride ions, but also dynamically balances the cost of medicaments and environmental pressure, showing strong inclusiveness for complex industrial conditions.

[0022] The multi-stage reaction synergy index is integrated into the reaction process, and the multi-stage reaction feedback data set is constructed by combining the online ion chromatography data, so that the actual contribution of each stage of drug administration to the removal of chloride ions can be measured in stages, and the next stage of drug administration formula can be corrected in real time under the dynamic decision correction function. This multi-stage cycle-feedback mode is closely coupled with the decision engine of the previous two steps, greatly improving the dechlorination rate, saving chemical consumption, and achieving Pareto optimality between technical and economic benefits.

[0023] The existing data set is integrated through the cloud to generate a knowledge graph and implement OTA iteration of the fusion model parameters, which greatly shortens the adaptation period under different factory areas or different batches of fly ash, and uses the cross-batch differential update and migration adaptability function to make the system more flexible.

[0024] The monitoring and optimization function of complex pollutants such as heavy metals and dioxin precursors is expanded on the chloride ion removal framework. The monitoring data is included in the multi-pollutant online monitoring data set by adding new sensors, and the synergistic governance effect is reflected in the multi-objective optimization function or pollutant deviation function and other indicators, so that the interaction between reagent addition and pollutant control is more holistic. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The figure is a structural diagram of the waste incineration fly ash chloride ion reduction system based on machine learning. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0027] Please refer to Figure 1 , the present application provides a waste incineration fly ash chloride ion reduction system based on machine learning, comprising, Step one, when the proportion of large particles in fly ash exceeds the preset range, the pretreatment module performs multi-stage screening and grinding on the fly ash, and real-time acquisition of fly ash particle homogenization factors , pH and temperature, integrate the processing results into the preliminary fly ash characteristic data set and mark the particle distribution state tends to be uniform, ensure the particle size uniformity and monitoring accuracy in the subsequent online ion monitoring stage; The step one includes the following contents: Step 101, homogenization and pretreatment of fly ash particles The fly ash is removed by a classification screening device to remove large size agglomerates and metal fragments and the like impurities, and the remaining particles are preliminarily refined. A specific stirring mechanism and controllable time grinding measure are used to further refine the larger particles, and a small amount of inert dispersant (such as low-activity ceramic microbeads) is used to improve the overall dispersion effect in the process. A propeller stirrer with an adjustable speed variable frequency motor is used to stir the fly ash and a small amount of inert dispersant (such as ceramic microbeads) in a closed grinding tank at a speed of 200-400 rpm. This method can quickly disperse large particles and ensure uniform distribution of the dispersant by shear force and collision force acting on the fly ash agglomerates. The stirring time can be freely set between 5 min and 20 min to adapt to the initial humidity and agglomeration degree of the incoming ash.

[0028] A planetary ball mill is used, which is equipped with zirconia or alumina grinding balls with a diameter of 2-4 mm. The planetary disc and the grinding tank rotate and revolve synchronously at different speeds. The mechanism can effectively refine the fly ash median particle size from the initial 100-120 μm to the target 50-60 μm by high-speed collision and grinding of the grinding balls at a comprehensive linear speed of 300-600 rpm. The grinding time can be accurately controlled in the range of 10 min to 30 min to obtain the best particle homogeneity.

[0029] In order to quantify the homogeneity of fly ash particles, a fly ash particle homogeneity factor is defined in this step The calculation formula is as follows:

[0030] In the formula: is the median particle size of the current batch of fly ash (usually in μm). ).

[0031] represents the particle size distribution vector obtained after classification test or particle size analysis of the current fly ash, where can be regarded as the ratio or frequency normalized value of the first particle size interval; represents an ideal or target distribution vector that can be used as a reference to measure the difference between the fly ash particle distribution and the target state (such as the process design expectation). represents the normalized proportion of the first particle size interval under the ideal or target distribution (i.e., the proportion of the mass or number of particles in this interval to the overall verification distribution).

[0032] Each component and is not negative and satisfies (if a probability distribution is used).

[0033] Euclidean norm (L2 norm) , which can measure the overall difference between the actual distribution and the reference distribution: the larger the value, the greater the gap between the current fly ash particle size distribution and the ideal / target state.

[0034] For quantifying the large particle removal rate or the reduction of coarse particle proportion, it is directly related to the fly ash classification process in the previous steps. It is directly related to the fly ash classification process in the previous steps, and the acceptable range is: ; is the distribution difference sensitivity coefficient, which controls the response speed when changes, is the particle removal contribution coefficient, which controls the attenuation or enhancement amplitude when changes, and the acceptable range is: , , and the specific size can be adjusted according to different fly ash characteristics and target process requirements.

[0035] The fly ash particle homogeneity factor is higher, indicating that the fly ash has a better particle homogenization degree at the end of the step.

[0036] The fly ash particle homogeneity factor is calculated by coupling the Euclidean distance between the actual particle size distribution vector of fly ash and the ideal distribution vector, the median particle size of fly ash and the large particle removal rate , which is a dimensionless index to quantify the particle homogeneity of fly ash after pretreatment; the purpose of constructing this function is to provide a repeatable and closed-loop physical and chemical evaluation benchmark in multiple batches and cross-plant processing, so that the system can automatically check and adjust the grinding and screening parameters before entering the online dosing and multi-stage reaction, thereby ensuring the consistency of the input of the subsequent decision-making model and the stability of the reaction efficiency.

[0037] By combining particle primary screening and auxiliary dispersion treatment, the fly ash particles are more concentrated in size distribution, which is beneficial to the effective contact between the subsequent reagents and the fly ash surface, thereby reducing the waste of chemical consumables from the source. By calculating the fly ash particle homogeneity factor, the pretreatment quality can be quantitatively fed back to guide the operator or automatic control system to adjust the grinding or screening parameters when needed.

[0038] Step 102, real-time monitoring data collection and structuring After step 101 is completed, the homogenized fly ash is put into a container, and multiple sensor nodes are arranged in the container to measure pH (denoted as pH), temperature (denoted as T), and fly ash particle size distribution (denoted as D). ), stirring rate (denoted as ) and major element content (such as aluminum, calcium, sodium, etc., all in the form of mass concentration) in fly ash, a multi-point distributed sensor arrangement is completed: The real-time monitoring data obtained by the above sensors, together with the fly ash particle homogeneity factor calculated in step 101 , are packaged as a preliminary fly ash characteristic data set , and are stored in a structured manner in the data management center: Particle size homogeneity information , pH, temperature , stirring rate , and major element content. All the above parameters will be recorded with the same timestamp to support subsequent linkage analysis of the processing state and fly ash characteristics.

[0039] After completing the structuring, the preliminary fly ash characteristic data set is subjected to quality verification, including abnormal screening of differences between sensors and judgment of whether there are fly ash agglomeration or local temperature gradient abnormalities. If abnormalities are found, the grinding and screening conditions of step 101 can be traced back to avoid invalidation of subsequent decision-making (the next step of the dosing model) due to inaccurate input data.

[0040] For example, in the quality verification of sub-step 102, the system compares the pH, temperature, stirring rate and element concentration data collected by the multi-point sensor array with the median and median absolute deviation (MAD) of each sensor reading, and removes abnormal values deviating from the median by more than 3xMAD. Secondly, gradient test is performed on the time series data of each sensor, such as reading change less than 0.01pH within 10 minutes or corresponding variable not within its physical range (pH0-14, temperature 0-100℃, stirring rate 50-400rpm), then determine the fault and automatically trigger maintenance or cleaning. If the number of valid readings is less than 70% of the total number of sensors, the system automatically traces back to the grinding or screening step to adjust the pre-processing parameters and re-collect data. All the removed and traced back operations are recorded in the quality audit log of the preliminary fly ash characteristic data set to ensure that the data basis for subsequent online dosing and multi-level feedback decision-making is reliable and consistent.

[0041] A multi-point distributed monitoring framework is constructed, allowing key characteristics of fly ash , pH, , All of them can be captured synchronously at the same moment, greatly improving the spatiotemporal resolution of data acquisition. By structurally integrating multiple key parameters, not only can the results of the previous step of pretreatment be quickly verified, but also high-quality, information-island-free initial data can be provided for the next step of drug decision-making based on the gradient boosting tree-reinforcement learning model. By using the verification mechanism and data backtracking method, abnormalities can be found and iteratively corrected in real time, preventing the situation where a large amount of reagent is used in the subsequent stage but the optimal dechlorination effect cannot be achieved.

[0042] Step two, when the initial fly ash characteristic data set the fly ash particle size and temperature information are completed, the online ion monitoring module is used for online ion monitoring data set The gradient boosting tree-reinforcement learning model is called, and the ion difference function The real-time evaluation of chloride ion concentration and removal rate target is carried out, and the initial dosing amount and reaction order suggestion are output, and the adaptive dosing benchmark is established; The step two includes the following contents: Step 201, online ion monitoring After the fly ash is pretreated and put into the reaction container, a real-time ion monitoring sensor array is configured in the container for detecting key indicators such as chloride ion concentration (denoted as ), calcium ion concentration (denoted as ), etc.; at the same time, the temperature probe records the temperature , the pH probe records , and the stirring rate sensor records . These data will be linked with the initial fly ash characteristic data set obtained in the previous step (step one) to form the online ion monitoring data set .

[0043] All monitoring parameters are sampled at fixed time intervals and stored in the cloud or local data management center for decision-making in the next step.

[0044] In order to measure the deviation of key ions in the current system, the ion difference function is constructed to quantify the gap between the ion concentration in the system and the ideal / target value online, and the environmental factors such as pH and temperature are also considered, and the example formula is as follows:

[0045] Wherein: : the measured chloride ion concentration at the current moment; : the chloride ion concentration from 0 to ​The kernel function that characterizes the coupling relationship between ion activity and environmental factors can be fitted from historical data or mechanism model when integrating the interval; : the current stirring rate, and The difference between the ideal design stirring rate will reflect the influence of stirring deviation through the exponential term; : stirring deviation amplification coefficient, , used to control the ideal design stirring rate deviation value to the attenuation of ion difference function; If the current stirring rate and the ideal design stirring rate is too large, tends to a smaller value, suggesting that the effectiveness of the monitoring data may be affected by uneven stirring, providing a risk warning for drug delivery decision-making; The ion difference function and , , , , etc. are recorded in the online ion monitoring data set , and are stored in association with parameters such as the fly ash particle homogeneity factor defined in the previous step.

[0046] If the ion difference function is higher than a certain difference threshold, according to its comparison with the fly ash characteristics in the preliminary fly ash characteristic data set , it is judged whether the fly ash itself is not in accordance with the expected, or caused by process factors such as real-time uneven stirring, and an alarm or correction measure is triggered.

[0047] By quantifying the deviation of online detection data through the ion difference function , the instantaneous changes in ion concentration level under different pH, temperature and stirring rate conditions can be more sensitively captured. By including the stirring rate deviation in the exponential decay factor, the effectiveness of the test data can be more accurately evaluated, providing high-fidelity input for subsequent drug delivery decision-making.

[0048] Step 202, initial drug delivery decision After obtaining the online ion monitoring data set in step 201, it is combined with the preliminary fly ash characteristic data set as input data to form a comprehensive state vector required for drug delivery decision-making.

[0049] Gradient Boosting Tree (GBT) part is used to predict the trend of initial chloride ion removal rate under different combinations of dosing amount; Reinforcement Learning (RL) part searches for the optimal or suboptimal dosing formula and reaction order for a given removal rate-drug consumption target on this basis.

[0050] In order to unify the results of GBT and RL, a multi-objective evaluation function is designed The candidate dosing scheme (including the dosing amount of CaO, NaAlO2 and the preset reaction order) is scored:

[0051] Among them: : the gradient boosting tree predicts the chloride ion removal rate that the current comprehensive state vector and the candidate dosing scheme can achieve (i.e. integral from 0 to the predicted removal rate); represents the reinforcement learning reward function of the current environment state based on the reinforcement learning scheme during the process of removing the removal rate from 0 to ; this function is determined by the system prior knowledge ( , ) and real-time monitoring results, and can reflect the contribution of fly ash properties, temperature, pH, etc. to the chemical reaction.

[0052] is the drug consumption penalty coefficient, used to balance the relationship between removal rate and drug consumption, and the value is greater than 0; is the norm of the candidate dosing scheme after power correction . If , then , represents the combination of reagents of the candidate scheme at time ; , drug consumption penalty index (dimensionless), controls the sensitivity of the penalty term to large-dose dosing; This formula combines the positive reward (integral term) brought by the removal rate and the negative penalty (subtraction term) brought by the drug consumption, the higher the value, the more advantageous the candidate dosing scheme is under the current state comprehensive state vector .

[0053] The model (Gradient Boosting Tree-Reinforcement Learning Model (GBT+RL)) will have several candidate dosing schemes at the same time one by one The final solution with the highest score is selected.

[0054] reinforcement learning reward function Can be defined as:

[0055] Wherein: is the integral variable, representing the intermediate value of the chloride ion removal rate from 0 to the target removal rate; is the gain coefficient calculated based on the current state vector (including , , pH, , , etc.) through linear combination or small feedforward network, used to amplify the reward sensitivity under different working conditions; is the reward power index, used to emphasize the marginal effect of the improvement of the middle and late stage removal rate; is the decay coefficient, which can be or its extension, to reflect the suppression of stirring deviation on the overall reward; The selected initial dosing scheme will output including: Dosing amount of reagent (CaO, NaAlO2, etc.); Preset reaction order (such as four-stage circulation or other hierarchical settings); Process time length suggestion (initial estimate of each stage or total time length).

[0056] The output result will be directly transmitted to step three as the starting operation instruction, entering the multi-stage reaction and real-time feedback control.

[0057] When used, the prediction value of GBT and the scheme search result of RL are organically integrated through the multi-objective evaluation function , which can achieve Pareto optimality or near optimality between chloride ion removal rate and reagent use. The introduction of penalty term can nonlinearly regulate reagent consumption and adaptively offset the decreasing benefit caused by excessive dosing, effectively reducing chemical waste and cost. The integral form of the chloride ion removal rate interval (0 to ) predicted by GBT is incorporated into the reinforcement learning reward function , which reflects the measurement of cumulative benefits throughout the reaction, rather than only focusing on a single endpoint or average.

[0058] Step three, if the chloride ion removal rate is not up to the threshold, the multi-stage dosing unit executes the candidate dosing scheme on the fly ash reaction system and maintains constant temperature stirring, and real-time monitors the multi-stage reaction synergy index Evaluate the contribution of each level of reaction, and dynamically correct the score at the end of each level Dynamically correct the next level scheme The third step includes the following: Step 301, multi-stage dosing-stirring reaction execution According to the initial dosing amount output by step two and the reaction order suggestion (for example, the reaction order And the dosing amount of CaO, NaAlO2, etc. , configure the corresponding dosing scheme in a constant temperature water bath and set the stirring rate .

[0059] In step 202, for multiple groups of candidate dosing formulations and corresponding reaction orders, relying on historical data and real-time monitoring values, combined with multi-objective evaluation functions , comprehensive score is given to each group of schemes; according to , the initial dosing vector and the reaction order are determined After scoring, the dosing amount with the highest score and the order are automatically selected as the initial scheme, and the information of the scheme is written into the online ion monitoring data set in structured text ; Subsequently, through the interface with the on-site PLC / SCADA control system, the initial dosing amount (including the dosing concentration of each medicament) and the reaction cycle order are immediately issued to the execution unit to trigger the next step of multi-stage dosing and stirring reaction This scheme designs four-stage cyclic reaction as an example, so is taken as the default order, which can also be adjusted according to actual working conditions. The length of each reaction is determined by the initial length suggestion given by step two At the beginning of the first stage reaction, the initial dosing vector (given by step two) is put in, the stirring rate , and the reaction temperature is constant at 40℃, or other more suitable temperature This operation continues until the end of this stage reaction (for example, 1.5 hours or other more appropriate time), and then according to the subsequent feedback, it is selected whether to start the next stage reaction. If the next stage reaction is carried out, the dosing vector or other level of medicament formulation is added again and stirred according to the immediately corrected dosing vector of step 302

[0060] In each stage of the reaction, chlorine ions in fly ash and dosing agents undergo multiple chemical and physical interactions; to measure the overall conversion efficiency in this process, a multi-stage reaction synergy index is introduced to describe the contribution of the stage reaction at time to the overall dechlorination progress. The form can be designed as: ; in: In the First-order reaction from time Run to time At the same time, an index is used to comprehensively measure the synergistic effect of chloride ion removal in the fly ash system and the consistency of process operation.

[0062] For a moment Below is a vector used to describe various characteristics of the reaction system. Typical examples of dimensions may include:

[0063] The ion difference function introduced in step two is used to measure the deviation of key indicators such as chloride ions from the target level.

[0064] The fly ash particle homogeneity factor reflects the potential impact of fly ash particle size distribution on reaction efficiency; A multi-objective evaluation function combining gradient boosting trees and reinforcement learning (defined in step two) is used to characterize the current drug delivery strategy. In state The overall score is as follows: Indicates the first The state vector at the end of the first reaction stage; , , These are the real-time monitoring values ​​for chloride ion concentration, pH, and temperature, respectively.

[0065] For the first The weight vector of the order reaction, the length of which is related to the multidimensional state features. Same, used to linearly combine different components (such as ion difference function, fly ash homogeneity, dosing scheme score, etc.) onto the same scale.

[0066] Acceptable range: ,in Equal to multidimensional state features Each component can be positive, negative, or zero, depending on the application requirements.

[0067] This is a term that exhibits exponential decay in the stirring rate deviation. For this moment and the ideal stirring rate The Euclidean distance between them; if only a single speed scalar is considered, it can be... Replace with absolute values.

[0068] Sensitivity coefficient set in the first stage reaction, used to control the tolerance to stirring instability; The larger the value, the more sensitive the current stage is to stirring deviation.

[0069] The larger the value, the more significant the contribution of the current stage reaction to the overall chloride removal between time length 0 and , which helps to determine whether to extend the current stage time length or switch to the next stage as soon as possible. In use, by combining multi-stage dosing with constant temperature water bath stirring, the staged optimization of chloride removal can be achieved under different dosing conditions; the multi-stage reaction synergy index

[0070] can be used to measure the contribution of each stage to the overall dechlorination in a numerical way, providing quantitative basis for step 302; the staged operation reduces the risk of excessive or insufficient reaction of fly ash and reagent in a single environment, greatly improving the efficiency of reagent consumption. The integral form is adopted, and the process parameters such as temperature, pH, and ion concentration are embedded in the kernel function

[0071] , which can nonlinearly capture the synergistic effect of multi-component reaction of fly ash, independent of traditional linear or mean assumptions; the hierarchical sensitivity is set in the exponential penalty term for stirring rate deviation, allowing the tolerance or sensitivity of each stage to operation deviation to be designed differently, improving the flexibility and accuracy of multi-stage reaction. Step 302, real-time feedback and dynamic decision correction

[0072] At the end of each stage of reaction, trigger the online ion chromatograph to obtain the current chloride ion concentration , and combine pH, temperature, stirring rate, etc. to generate a new round of monitoring record. Write all monitoring results (including dechlorination rate calculation,

[0073] numerical values, etc.) into the multi-stage reaction feedback dataset , which will be used together with the preliminary fly ash characteristic dataset , online ion monitoring dataset for real-time decision correction.

[0074] To evaluate the next dosing scheme immediately after the end of each stage of reaction, follow the multi-objective evaluation idea in step two, and combine the newly added multi-stage reaction synergy index in this step, as follows:

[0075] ​Wherein: is the dynamic decision-making correction function used to dynamically correct the next level of dosing scheme after the end of the first level reaction; : the multi-objective evaluation function used in step two is used to score the candidate dosing scheme under the current state ; : the value of the multi-level reaction synergy index defined in this step at the end of the first level reaction; Positive gain coefficient, value between 0.1 and 2.0, used to adjust the influence weight of the contribution of the previous level on the subsequent decision; the larger the value, the more positive the influence of the good synergy of the previous level on the next level dosing scheme; Dynamic correction score is higher, indicating that the next candidate dosing scheme has more overall advantages in the current multi-level reaction environment.

[0076] At the end of the first level reaction, the dynamic decision-making correction function and the multi-level reaction feedback data set are used to screen the candidate dosing amount combination; If the value of a dosing scheme is higher than the preset threshold, it can be used as the execution instruction of the next level candidate dosing scheme ; if all the scheme scores cannot meet the requirements, the subsequent reaction can be terminated in advance or rolled back to the second step fusion model for more substantial reconstruction of the dosing scheme; This process is repeated at the end of each level until all levels are completed or terminated in advance.

[0077] When used, the multi-level reaction synergy index and the latest ion monitoring results are introduced into the dynamic decision-making correction formula , so that the subsequent dosing scheme can be updated at any time according to the actual progress of the chemical reaction, ensuring that the reaction efficiency and drug consumption are always optimal or near-optimal. The higher the positive synergy of the previous level reaction, the more targeted optimization of the next level dosing scheme can be achieved under the promotion of , truly realizing the benign linkage between levels.

[0078] If the monitoring results show that the dechlorination efficiency or synergy index is stagnant, the subsequent reaction can be automatically terminated according to the downward trend of the dynamic correction score , avoiding the continued consumption of reagents; the combination of the multi-level reaction synergy index and the multi-objective evaluation function of step two forms a dynamic correction system of integral + multi-objective scoring.​

[0079] Step four, when the multi-stage dosing process is completed and the multi-stage reaction feedback dataset is generated, the existing dataset is analyzed by knowledge graph and OTA update package is prepared, the upgraded version of gradient boosting tree-reinforcement learning model is issued to each factory, the best dosing scheme and fly ash spectrum adaptation parameters are migrated synchronously, the dechlorination performance of different batches of fly ash is continuously strengthened and the chemical consumption is saved; The step four includes the following contents: Step 401, data collection and knowledge graph construction When step three is completed, the multi-stage reaction feedback dataset generated in the whole process of multi-stage reaction , the dosing scheme decision log (including , and other evaluation results), and the preliminary fly ash characteristic dataset and online ion monitoring dataset are packaged and uploaded to the cloud database to form the complete data record of this batch.

[0080] The above data are uniformly indexed in the cloud, and the batch ID, fly ash source, physical and chemical indicators, and dosing decision state are associated with multiple-dimensional labels, so as to facilitate subsequent quick retrieval and calling.

[0081] After data collection is completed, the fly ash treatment knowledge graph is constructed by further using semantic network and association rule mining technology, and the fly ash composition spectrum, chlorine ion removal mechanism, dosing scheme score and multi-stage reaction synergy index are represented in the form of node-relation. Through the graph model expression, the interaction between the components of fly ash and the mapping relationship with the dosing scheme can be better displayed, providing a more intuitive semantic basis for subsequent algorithm upgrade and migration.

[0082] In use, through the data collection and indexing mechanism, the massive discrete data of steps one, two and three are uniformly organized into structured and semi-structured databases, which can efficiently trace any one-time dosing decision or chlorine ion removal process. Relying on the knowledge graph building, the potential association between multiple process elements (fly ash components, reagent ratio, temperature, stirring rate, etc.) is explicit, providing more context information for the algorithm and improving the accuracy of subsequent analysis and optimization.

[0083] Step 402, OTA upgrade and algorithm optimization The gradient boosting tree-reinforcement learning (GBT+RL) fusion model, the kernel function of ion difference function , and the core algorithm components such as multi-stage reaction synergy index are iterated continuously in the cloud.

[0084] When a new model version is determined, an update package (containing parameter weight matrix, kernel function parameter, decision scheme, etc.) is issued to each factory or device in an OTA manner, so that the local control system can complete the algorithm upgrade without stopping.

[0085] By comparing the new batch of fly ash treatment data with the historical records in the knowledge graph, cross-batch model evolution training can be performed in the cloud server to correct the mechanism hypothesis or enhance the recognition ability of rare fly ash components.

[0086] Using the differential update method of cross-batch iteration, the experience of the new batch is fed back to the reinforcement learning model, so that it has stronger generalization performance when OTA updates are issued; To demonstrate the update of the decision module during OTA upgrade, define the following incremental update function , the existing decision parameters are corrected:

[0087] Among them: and represent the scheme or model parameter vector of the old and new versions respectively; represents the new batch of fly ash treatment data set; To update the kernel function, integrate the difference between historical data and new data to output the correction amount of ; can be regarded as the integral upper limit of the update process in the data dimension or time dimension, and is flexibly defined to accommodate data of different orders of magnitude.

[0088] In step 402, the so-called decision parameters refer to the weights and hyperparameters of the model, which are initially obtained by gradient boosting tree regression and reinforcement learning strategy training on the cloud historical data set (including the preliminary fly ash feature data set generated in steps 1 to 3 , online ion monitoring data set , multi-stage reaction feedback data set ); Then, upload the new batch of fly ash treatment data to the cloud, and use the incremental update function and the updated kernel function to integrate the original parameters to generate a new parameter set ; These optimized through cross-batch comparison and knowledge graph mining are packaged as OTA update packages and issued to the local control system in each factory for replacement or superposition of the original parameters, so that the local dosing decision always uses the latest and optimal model configuration.

[0089] Update the kernel function The parameter correction amount is generated by calculating the difference between the actual removal rate of the new batch and the predicted removal rate of the old model, and combining a set of weight functions reflecting the importance of the removal rate section. Specifically, it will traverse different removal rate levels, calculate the difference between the measured value and the predicted value at that level, multiply the residual by the corresponding weight coefficient (for example, based on batch similarity or Gaussian kernel of the attention interval), and sum all the weighted residuals in the entire removal rate range to form a smooth incremental update of each model parameter. The difference between the measured value and the predicted value at that level, multiply the residual by the corresponding weight coefficient (for example, based on batch similarity or Gaussian kernel of the attention interval), and sum all the weighted residuals in the entire removal rate range to form a smooth incremental update of each model parameter.

[0090] In use, with the help of OTA upgrade, any waste incineration plant can quickly obtain the latest algorithm after the model version iteration completed in the cloud, without manual deployment or repeated debugging; through cross-batch evolution training, the processing experience accumulated in a special fly ash scenario can be spread to the entire system network, thereby greatly improving the overall efficiency of the industry.

[0091] Step 403: Cross-plant migration deployment and continuous learning When a new plant or a new fly ash pedigree is connected, first search its similarity and difference with the existing fly ash pedigree by knowledge graph, then select the closest model version and appropriate differential update package from the cloud; after local deployment for a period of time, continuously return the actual fly ash processing data to the cloud, and retrain through the incremental update function or reinforcement learning scheme search, and finally form a new model adapted to the plant.

[0092] If the fly ash of the new plant has special components such as heavy metal content, salt structure, etc., and the difference with the fly ash feature set in the known model is too large, the system will start dynamic parameter mapping to perform nonlinear transformation on the hyperparameter space of the Gradient Boosting Tree-Reinforcement Learning (GBT+RL) fusion model, to expand the explainability and applicability range of the model. This process will also reference similar cases in the knowledge graph in step 401 to reduce blind attempts.

[0093] To measure the adaptability of the current model to the fly ash treatment of the new plant, the migration adaptability can be defined as follows:

[0094] Wherein: : adaptability kernel function, used to evaluate the applicability of the model to new fly ash according to online monitoring of chloride ion concentration, pH, current decision score, and multi-level reaction synergy index defined in step three .

[0095] Adaptability kernel function For the comprehensive evaluation of the matching degree of the current model to the new fly ash environment at each time, it will four key indicators: real-time chloride ion concentration deviation (difference from target concentration), pH deviation, initial dosing decision score And multi-stage reaction synergy index : normalized and weighted fusion according to preset weights, and mapped to interval using an exponential or Gaussian function, so as to output a continuous fitness value at different times; this function takes into account both chemical removal efficiency and process stability, and can also reflect the immediate adaptive ability of the model under new working conditions, for migration fitness integral calculation.

[0096] If the migration fitness significantly increases within a specified time (such as after multiple dosing cycles), it indicates that the migration learning effect is good; if it remains low, the system can automatically increase the iteration intensity of the parameter mapping scheme.

[0097] When used, migration learning not only enables different plants to quickly have a mature dosing scheme, but also avoids the waste of resources caused by repeated development and training of one model per plant; through dynamic parameter mapping, the adaptive ability to extreme or rare fly ash components can be significantly improved, and an interface is provided for the integration of more new technologies in the future.

[0098] Step five, when the operation end detects that the concentration of heavy metals or dioxin precursors in fly ash exceeds the limit, the multi-dimensional pollutant synergy removal module creates a new monitoring item in the multi-pollutant online monitoring data set and calls the multi-objective optimization function , while expanding the dosing vector Integrate chelating agents or oxidizing agents, update the multi-pollutant reduction effect through multi-stage reactions and dynamic correction links, and finally complete the real-time inhibition of heavy metals and dioxin precursors together with the OTA upgrade mechanism; The step five includes the following contents: Step 501, multi-dimensional pollutant monitoring expansion On the basis of the original online ion monitoring module for chloride ions, expand the online monitoring devices or timed sampling detection units for heavy metals (such as Pb, Cd, Hg) and dioxin precursors; The pollutant concentration values generated during the monitoring process, together with the element content labels obtained in the fly ash pretreatment stage (step one), form new items and are added to the multi-pollutant online monitoring data set ; To quantify the synergistic monitoring effect of the new pollutants and chloride ions, construct a pollutant deviation function , at time Aggregate the differences between the detection values of multiple components (heavy metals, dioxin precursors, chloride ions) and the target values:

[0099] wherein: may represent the maximum value of the concentration vector of heavy metals or dioxin precursors at time (or a certain term thereof) converted from vector to scalar, or its proportion; is a coupling kernel function that combines the interaction of toxic components with chloride ions and pH, used to evaluate the deviation of multi-dimensional pollutants from each other under different environmental conditions.

[0100] In sub-step 501, the coupling kernel function is used to simultaneously reflect the comprehensive influence of the deviation of toxic component concentration, chloride ion concentration and pH on the multi-pollutant removal strategy in the same integration framework. The calculation process is as follows: first, the concentration of heavy metals or dioxin precursors to be integrated is amplified in a power exponential manner to highlight the contribution of high concentration sections to the overall deviation; then the power exponential amplification is also used to the chloride ion concentration measured at the same time to reflect the coupling effect of chloride ion level on the behavior of toxic components; finally, a negative exponential function is used to attenuate and weight the difference between the system pH and the preset optimal pH (such as 8.0), so that when the pH is close to the optimal value, the coupling effect is fully reflected, and when the deviation is too large, the coupling strength is automatically suppressed. This coupling design not only faithfully captures the nonlinear interaction between high-concentration pollutants and chloride ions, but also dynamically reflects the influence of the acid-base environment on the synergistic removal efficiency, thereby providing a highly realistic and adjustable kernel function input for the pollutant deviation function .

[0101] By obtaining more comprehensive information of harmful components of fly ash and maintaining the same time base with the chloride ion index, a data foundation is laid for subsequent multi-target dosing schemes; the pollutant deviation function can handle the influence of multiple types of pollutants on the removal scheme in the same integration framework, so that the removal of chloride ions and other toxic components is no longer mutually exclusive.

[0102] Step 502, multi-target dosing scheme optimization Based on the expansion in step 501, the system needs to incorporate the newly detected heavy metal and dioxin precursor concentration data into the original gradient boosting tree-reinforcement learning (GBT+RL) fusion model; from which a new comprehensive state vector is formed, which adds at least one set of dynamic components of heavy metal concentration and precursor concentration to the original state vector , and also expands the type of medicament, such as adding chelating agents, oxidizing agents, etc.

[0103] In order to consider the removal efficiency of chloride ions and the synergistic removal effect of multiple pollutants, the original multi-objective function or the dynamic modified score extended with a comprehensive toxicity consideration, for example:

[0104] wherein: represents the new multi-drug dosing vector, including the original CaO, NaAlO2 and other needed chelating agents or oxidants; the pollutant deviation function defined from step 501; a synergistic removal weight coefficient, used to control the proportion of the influence of the removal of toxic components on the final dosing scheme, with a value greater than 0; with a multi-objective optimization function as the goal, the gradient boosting tree combines reinforcement learning to search for the scheme based on the prediction of multi-dimensional pollutant removal efficiency, and automatically generates a multi-objective dosing scheme.

[0105] The output result is also presented in a similar form, and can be executed in a hierarchical manner through the same multi-stage reaction (step three).

[0106] The multi-objective optimization function combines the removal of chloride ions with the reduction of toxic substances such as heavy metals and dioxin precursors, avoiding the situation of losing one or the other or ignoring the accumulation of other pollutants due to the pursuit of chloride ion removal; the enhanced state vector and the dosing vector provide a broader interface for subsequent expansion of other co-production or recycling processes. Incorporating toxic substance removal into the same reinforcement learning scheme framework not only expands the input and output dimensions of the model, but also brings new synergistic reaction mechanism exploration space, which has great application potential and practical significance.

[0107] Step 503, synergistic reaction and dynamic correction In conjunction with step three, during the execution of multi-stage dosing-stirring reaction, additional drugs such as chelating agents or oxidants need to be added according to the new multi-drug dosing vector vector; after each stage ends, online monitoring will simultaneously detect the concentrations of chloride ions and heavy metals, dioxin precursors and update the results to the multi-pollutant online monitoring dataset .

[0108] If at the end of a certain stage operation, it is found that the pollutant deviation function is significantly high, indicating that the removal of toxic components is insufficient; then the decision model extended in step two can be called for dynamic correction before the start of the next stage reaction (such as increasing the chelating agent or prolonging the constant temperature stirring time).

[0109] In the decision correction, a correction gain function is constructed with the priority of toxic reduction, for example:

[0110] wherein: is the final comprehensive score, which is used to select or adjust the dosing formula of the first stage at the end of the first stage to make the system capable of simultaneously considering the synergistic removal of chloride ions and other toxic components; is the dynamic decision correction function from step three; is the multi-pollutant deviation function measured at this moment; is the multi-pollutant deviation function measured at this moment; is the multi-agent dosing vector of the first stage, including the dosages (in g / L) of CaO, NaAlO2, and newly added chelating agents or oxidizing agents, representing the candidate comprehensive dosing strategy; is the extended state vector at the end of the first stage reaction, including the concentrations of heavy metals and dioxin precursors and their deviation functions in addition to the original chloride ions, pH, temperature, and other parameters ; is the toxicity priority coefficient, which is greater than 0, and the larger the value, the higher the requirement for multi-pollutant removal; in the form of presents the nonlinear gain of the toxicity correction requirement, exponential decay mapping of the deviation amount, limiting the gain to a larger value when the deviation is high, and the gain tends to zero when the deviation is low; In use, real-time detection is combined with the aforementioned multi-stage reaction synergistic index , dynamic correction score and other content to ensure a more three-dimensional balance and dynamic optimization between different pollutants. When toxic components are efficiently removed, the reaction may tend to save reagents; conversely, if the toxicity is high, the system will immediately make corrections to ensure that the content of heavy metals or dioxin precursors is also controlled within a safe range.

[0111] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0112] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device, and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.​​

[0113] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the embodiments of the device described above are merely schematic, and the division of the units is merely logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0114] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e. can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to the actual needs to achieve the purposes of the embodiments of the present application.

[0115] The above describes only the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A machine learning based system for reducing chloride ion in municipal solid waste incineration fly ash, characterized in that: Comprising, When the detection rate of large particles in fly ash exceeds the standard, the pretreatment module performs multi-stage screening and grinding on the fly ash raw material, and real-time detection and labeling of fly ash particle homogeneity factor, pH and temperature, and the detection results are summarized as preliminary fly ash characteristic data set; The gradient boosting tree-reinforcement learning model is called by the online ion monitoring module, and the chloride ion concentration and removal rate target is obtained in real time based on the ion difference function, and the initial dosage and reaction order are output; If the removal rate does not reach the threshold after each stage of reaction, the multi-stage dosing unit executes the candidate dosing scheme on the fly ash reaction system and maintains constant temperature stirring, records the multi-stage reaction synergy index and ion chromatography data, and dynamically corrects the dosing scheme according to the dynamic correction score; When the dosing cycle is completely finished, the cloud algorithm center archives the existing data set and generates an OTA update package, and upgrades the gradient boosting tree-reinforcement learning model to the latest version and distributes it to each factory, synchronously migrating the best dosing scheme and fly ash lineage adaptation parameters; When the concentration of heavy metals or dioxin precursors is detected to be out of limit, the multi-dimensional pollutant synergistic removal module adds special monitoring items and expands the multi-objective optimization function, dynamically introduces chelating agent or oxidizing agent dosing, and executes synchronous control of toxic components and chloride ions.

2. The waste incineration fly ash chloride ion emission reduction system according to claim 1, wherein: During the fly ash pretreatment and data collection process, when the removal rate of large particle impurities exceeds the preset threshold, the fly ash is subjected to multi-stage screening, grinding and auxiliary dispersion treatment to obtain pretreated fly ash with fly ash particle homogeneity factor, and the homogenization information and corresponding pH, temperature and stirring rate are recorded in real time.

3. The waste incineration fly ash chloride ion emission reduction system according to claim 2, wherein: A preliminary fly ash characteristic data set is established for the pretreated fly ash by arranging multiple point distributed sensors, and quality checking is performed during structured storage to realize adaptive correction when the fly ash homogeneity, sensor difference detection and backtracking grinding and screening conditions are met.

4. The waste incineration fly ash chloride ion emission reduction system according to claim 3, wherein: During the online ion monitoring and initial dosing decision-making stage, in order to continuously track the chloride ions and calcium and sodium ions in the fly ash system, the ion difference function is used to evaluate the ion concentration and process deviation in the system in real time, and the monitoring data is updated to the online ion monitoring data set in real time; If the ion difference function is monitored to be higher than the preset difference threshold, according to the comparison of the fly ash characteristics in the preliminary fly ash characteristic data set, the comparison result triggers an alarm or a correction measure.

5. The waste incineration fly ash chloride ion emission reduction system according to claim 4, wherein: The gradient boosting tree-reinforcement learning model is called to calculate and output the initial dosing amount of CaO and NaAlO2 and the reaction order; and according to the real-time changes of water quality parameters, fly ash composition and operating conditions, the initial dosing amount and reaction order are corrected in real time; The selected initial dosing scheme includes: recommended dosage, preset reaction order and process duration.

6. The waste incineration fly ash chloride ion emission reduction system according to claim 5, wherein: When the gradient boosting tree-reinforcement learning model determines that it has entered the first or subsequent levels of reaction, a multi-stage circulation dosing and stirring operation is performed, and the multi-stage reaction synergy index is used to quantify the contribution of the corresponding reaction, achieving phased performance evaluation of fly ash chloride ion reaction.

7. The waste incineration fly ash chloride ion emission reduction system according to claim 6, characterized in that: After completing any level of reaction, the chloride ion concentration is collected and real-time feedback is returned to the cloud algorithm center. If the chloride ion removal rate is lower than the set threshold, the next level of reaction is adjusted according to the dynamic decision correction function, or the process is terminated in advance, and the correction results are stored in the multi-stage reaction feedback dataset.

8. The waste incineration fly ash chloride ion emission reduction system according to claim 7, characterized in that: The multi-stage reaction feedback dataset, dosing decision path, and treatment effect of the whole multi-stage reaction process are uploaded to the cloud database and indexed to the knowledge graph; The preliminary fly ash characteristic dataset and online ion monitoring dataset are used to establish a multi-dimensional link using semantic networks and association rules, which supports fast retrieval during subsequent OTA upgrades and cross-plant migration.

9. The waste incineration fly ash chloride ion emission reduction system according to claim 8, characterized in that: The cloud algorithm center automatically generates OTA update packages after comparing the parameters of the gradient boosting tree-reinforcement learning model across batches or across plants, and distributes the updated model hyperparameters, kernel function parameters, and decision strategies to each plant's local, allowing each plant to dynamically obtain the latest dosing scheme without stopping production.

10. The waste incineration fly ash chloride ion emission reduction system according to claim 9, characterized in that: When the fly ash pedigree of a new plant differs significantly from the expected, the applicability of the new fly ash treatment process is evaluated based on the migration fitness, and the existing model is adaptively tuned using an incremental update function, achieving fast adaptation to heterogeneous fly ash batches; if the migration fitness continues to be lower than expected within a specified time, the iteration intensity of the parameter mapping scheme is increased.

11. The waste incineration fly ash chloride ion emission reduction system according to claim 10, characterized in that: When the target pollutant concentration exceeds the set threshold, real-time or timed sampling monitoring items for multi-dimensional pollutants are added, and a pollutant deviation function is established for comprehensive deviation evaluation of toxic component concentrations, pH, chloride ions, and other indicators for subsequent multi-objective dosing decision calls.

12. The waste incineration fly ash chloride ion emission reduction system according to claim 11, characterized in that: The gradient boosting tree-reinforcement learning model uses a multi-objective optimization function to optimize the newly added heavy metals and dioxin precursors, outputs a comprehensive dosing vector including CaO, NaAlO2, and chelating agents, and adaptively balances the removal efficiency of chloride ions and other highly toxic components; Where the multi-objective optimization function aims to balance the removal efficiency of chloride ions and other highly toxic components, the gradient boosting tree combines reinforcement learning search schemes to automatically generate multi-objective dosing schemes based on the prediction of multi-dimensional pollutant removal efficiency.

13. The system of claim 12, wherein the system further comprises a controller configured to: determine a correction gain function based on the online measurement of the contaminant in the waste incineration fly ash; and dynamically correct the dosing recipe for the next stage of the multi-stage reaction based on the correction gain function. In performing the multi-stage reaction, if the online detection result shows that the contaminant deviation function continuously exceeds the expectation, the next-stage dosing formula is dynamically corrected based on the correction gain function, and the correction process is recorded.

Images