Method for predicting tail pollutant emission of solid waste incineration process facing raw material strong fluctuation

Abstract

The application discloses a kind of raw material strong fluctuation-oriented solid waste incineration process tail pollutant emission prediction method, method includes: from solid waste incineration system real-time acquisition including solid waste characteristic parameters and operating parameter multi-source data;Based on the static characteristic parameters of multiple solid waste into furnace and the real-time flow information corresponding thereto, the dynamic feature vector is generated by weighted fusion calculation;The dynamic feature vector and operating parameter are input to the cascade prediction model constructed in advance;The cascade prediction model is according to the physical and chemical causal chain of incineration process, and the emission concentration prediction value of tail pollutant is finally output by recursive state prediction of multistage series connection submodel;Based on the prediction result, execute early warning, operation guidance or feedforward control.The application effectively solves the emission prediction and control problem caused by the composition of solid waste raw material violent fluctuation and the intermediate state of process cannot be measured, realizes the change from passive end feedback to active intelligent feedforward, improves process stability and economy.

Description

Technical Field

[0001] This invention relates to the field of solid waste treatment and environmental monitoring technology, specifically a method for predicting pollutant emissions at the tail end of solid waste incineration processes that are subject to strong fluctuations in raw material emissions. Background Technology

[0002] With the rapid development of urbanization and industrialization, the safe disposal of solid waste has become a serious challenge. Incineration technology has been widely used because it can achieve volume reduction, harmlessness and energy recovery.

[0003] However, solid waste incineration is a complex system involving multiphase flow, complex heat and mass transfer, and deep coupling of chemical reactions. Compared to traditional fuels, solid waste itself has the unique characteristics of drastic fluctuations and high uncertainty in composition. Its key properties, such as moisture content, calorific value, chlorine and sulfur content, exhibit strong spatiotemporal heterogeneity depending on factors such as source, season, and pretreatment process. This strong fluctuation of raw materials fundamentally undermines the stable thermal and mass balance of the incineration process, leading to unstable combustion and unpredictable pollutant generation potential. At the same time, due to the harsh operating conditions of high temperature and high dust in the furnace, reliable online monitoring of key intermediate states of the combustion process (such as local temperature and initial pollutant concentration) is technically extremely difficult and costly.

[0004] This results in a serious flaw in the currently widely adopted feedback control model based on monitoring pollutant concentrations at the chimney end: the dosage of purification agents can only be adjusted in reverse after pollutants have been generated and reached the chimney. This end-monitoring and reverse adjustment method is severely lagging in response to drastically fluctuating solid waste raw materials. It not only relies on human experience for frequent trial and error and cannot guarantee the accuracy of control, but is also prone to excessive consumption of agents and secondary pollution risks due to inaccurate forecasts and conservative strategies with lagging adjustments. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting pollutant emissions at the tail end of solid waste incineration processes that are subject to strong fluctuations in raw material emissions, in order to solve the above-mentioned problems.

[0006] The technical solution of this invention is: A method for predicting tail-end pollutant emissions from solid waste incineration processes with strong fluctuations in raw material prices, comprising: Multi-source data is collected in real time from the solid waste incineration system. The multi-source data includes solid waste characteristic parameters corresponding to various types of solid waste entering the furnace and operation parameters characterizing the operating status. Based on the static characteristic parameters of the various solid wastes fed into the furnace, and the real-time flow information of the various solid wastes corresponding to the operation parameters, a dynamic feature vector that can reflect the characteristics of the mixed fuels fed into the furnace in real time is generated through weighted fusion calculation. The dynamic feature vector and the operating parameters are input into a pre-constructed cascaded prediction model. The cascaded prediction model is based on the physicochemical causal chain of the incineration process, which proceeds sequentially as follows: fuel, combustion, heat transfer, purification, and emission. The cascaded prediction model performs recursive state predictions through multiple interconnected sub-models, ultimately outputting predicted emission concentrations of tail-end pollutants.

[0007] Based on the pollutant concentration prediction results output by the model, perform at least one of the following actions: early warning, operational guidance, or feedforward control.

[0008] Furthermore, the cascaded prediction model is a three-stage prediction model structure connected in series, with at least three stages. The number of stages can be increased if there are too many intermediate model parameters or the combustion system is more complex. The optimal stage can be determined by the effective dimensionality ratio (output / input < 0.5). The three-stage prediction model structure includes: The first-level prediction model is used to predict the combustion state parameters in the furnace based on the input dynamic feature vector and operating parameters. The second-level prediction model is used to predict the flue gas state parameters entering the flue gas purification system based on the combustion state parameters output by the first-level prediction model. The third-level prediction model is used to predict the emission concentration of pollutants at the tail end based on the flue gas state parameters output by the second-level prediction model and in combination with the operating parameters of the flue gas purification system. The three-level prediction models are executed sequentially, with the output of the preceding model serving as part of the input for the subsequent model.

[0009] Furthermore, the combustion state parameters output by the first-level prediction model include the temperatures of at least three key areas within the furnace; the flue gas state parameters output by the second-level prediction model include at least one of the temperature, pressure, moisture content, oxygen content, and carbon dioxide content at the inlet of the flue gas purification system; and the tail pollutants output by the third-level emission concentration prediction model include at least one of sulfur dioxide, nitrogen oxides, hydrogen chloride, carbon monoxide, and ammonia.

[0010] Furthermore, the dynamic feature vector includes: the combined values ​​of the received basis moisture content, dry basis ash content, received basis lower heating value, and dry basis chlorine content of the mixed fuel fed into the furnace; the multi-source data also includes monitoring parameters characterizing the emission results, which are used to periodically retrain or adaptively update the parameters of the cascade prediction model online.

[0011] Furthermore, the feedforward control steps include: inputting the predicted emission concentration value to the advanced process controller, which calculates and outputs adjustment commands in advance based on the deviation between the predicted emission concentration value and the target set value, and directly controls the dosage of denitrification reducing agent or deacidification absorbent.

[0012] Furthermore, the method for establishing a cascaded prediction model includes the following steps: Historical multi-source operation data of the solid waste incineration system are collected and preprocessed to obtain a regular historical dataset, which includes solid waste characteristic parameters, operation parameters and corresponding pollutant emission monitoring parameters. A dynamic feature fusion module is constructed, which is configured to: perform weighted fusion calculation on the static characteristic parameters of various solid wastes based on their real-time flow information, and output a dynamic feature vector; Based on the process causal chain of fuel → combustion → heat transfer → purification → emission in the incineration process, a cascade prediction model framework is constructed, and the inputs and outputs of each sub-model of the cascade prediction model are defined. The cascaded prediction model framework is trained using the historical dynamic feature vector and the operation parameters in the historical dataset. The hyperparameters of each sub-model are jointly tuned with the goal of minimizing the overall prediction error of the cascaded system, resulting in the trained and optimized cascaded prediction model.

[0013] Furthermore, the inputs and outputs of each sub-model of the cascaded prediction model are as follows: The input to the first-level prediction model is the dynamic feature vector and operating parameters, and the output is the combustion state parameters in the furnace. The input to the second-level prediction model includes at least the combustion state parameters output by the first-level prediction model, and the output is the flue gas state parameters entering the flue gas purification system. The input to the third-level prediction model includes at least the output of the second-level prediction model and the operating parameters of the flue gas purification system, with the output being the concentration of pollutants emitted at the tail end.

[0014] Furthermore, random forest, gradient boosting decision tree, or deep neural network algorithms are used to construct sub-models at each level, and hyperparameters are jointly tuned using Bayesian optimization or grid search methods.

[0015] Furthermore, the model training and co-optimization steps also include data downsampling frequency optimization: determining the lag time between key variables based on cross-correlation analysis, and evaluating the data smoothness and trend preservation ability under different downsampling frequencies, with the frequency with the highest comprehensive score being used as the unified sampling frequency for model training.

[0016] Furthermore, the trained and optimized cascaded prediction model is integrated with the dynamic feature fusion module into an executable prediction system, and the integrated prediction system is encapsulated as a microservice with standard data interfaces, including OPC-UA and / or RESTful-API.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention generates a dynamic feature vector that accurately characterizes the current mixed fuel properties by real-time acquisition and fusion of static characteristic parameters and dynamic flow information of various solid wastes fed into the furnace. This vector, along with operating parameters, is input into a cascade prediction model constructed based on the physicochemical causal chain of the incineration process, achieving forward, recursive, and high-precision prediction of tail-end pollutant emission concentrations. It effectively overcomes the challenges of process instability and prediction caused by drastic fluctuations in raw material composition, enabling the system to anticipate emission trends and allowing operators to understand emission changes in advance. This facilitates proactive early warning of emission exceedance risks, precise optimization of operating parameters, and intelligent feedforward adjustment of the purification system. Compared to existing monitoring and control modes that rely on end-point monitoring and delayed feedback, this invention fundamentally achieves a paradigm shift from passive response to proactive prediction and optimization, significantly improving the system's stable operation under conditions of drastic changes in raw material properties. While ensuring stable and compliant emissions, it also enables refined management of purification agent dosing, effectively controlling operating costs and suppressing secondary pollution problems that may be caused by excessive agent use. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for predicting pollutant emissions at the tail end of a solid waste incineration process that addresses strong fluctuations in raw material emissions, according to the present invention.

[0019] Figure 2 This is a flowchart of the input-output framework of a multi-level prediction model in a method for predicting pollutant emissions from the tail end of a solid waste incineration process that addresses strong fluctuations in raw material emissions, as described in this invention.

[0020] Figure 3 This is a simplified process flow diagram of the bubbling fluidized bed solid waste incineration system in an embodiment of the present invention.

[0021] Figure 4 This is a matrix diagram illustrating variable selection based on Spearman correlation coefficient analysis in an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram illustrating the acquisition of lag steps and correlation coefficients of multiple variables using mixed sludge flow rate as a baseline sequence in an embodiment of the present invention.

[0023] Figure 6 This diagram illustrates the smoothness index, trend preservation capability, and comprehensive evaluation of the main characteristic indicators after downsampling in this embodiment of the invention. (a) shows the number of remaining samples at different downsampling frequencies, (b) shows the changes in the smoothness index at different downsampling frequencies, (c) shows the trend preservation capability index at different downsampling frequencies, and (d) shows the comprehensive evaluation diagram at different downsampling frequencies.

[0024] Figure 7This is a comparison of pollutant NH3 before and after hyperparameter optimization of the prediction model in this embodiment of the invention.

[0025] Figure 8 This is a comparison of pollutant CO before and after hyperparameter optimization of the prediction model in this embodiment of the invention.

[0026] Figure 9 This is a comparison of SO2 pollutant before and after hyperparameter optimization of the prediction model in this embodiment of the invention.

[0027] Figure 10 The example of this invention shows the prediction model before and after hyperparameter optimization for pollutant NO. x Comparison results.

[0028] Figure 11 This is a comparison of pollutant HCl before and after hyperparameter optimization of the prediction model in this embodiment of the invention.

[0029] Figure 12 This shows the changes in the relative errors of various pollutant results before and after hyperparameter optimization of the prediction model in this embodiment of the invention.

[0030] Figure 13 This is the main interface for the industrial application of the prediction model in this embodiment of the invention.

[0031] Figure 14 This is a flowchart of the prediction method of the present invention. Detailed Implementation

[0032] The following is combined Figures 1 to 14 The specific embodiments of the present invention will be described in detail below.

[0033] It should be noted that the circuit connections involved in this invention all adopt conventional circuit connection methods and do not involve any innovation.

[0034] Example like Figure 14 As shown, a method for predicting tail-end pollutant emissions from solid waste incineration processes with strong fluctuations in raw material emissions includes: Based on the static characteristic parameters of various solid wastes fed into the furnace, and the real-time flow information of various solid wastes corresponding to the operating parameters, a dynamic feature vector that can reflect the characteristics of the mixed fuels fed into the furnace in real time is generated through weighted fusion calculation. The dynamic feature vector and operating parameters are input into a pre-constructed cascaded prediction model. Based on the physicochemical causal chain of the incineration process, which proceeds sequentially as: fuel, combustion, heat transfer, purification, and emission. The cascaded prediction model performs recursive state predictions through multiple interconnected sub-models, ultimately outputting predicted emission concentrations of tail-end pollutants.

[0035] Based on the pollutant concentration prediction results output by the model, perform at least one of the following actions: early warning, operational guidance, or feedforward control.

[0036] Multi-source data is collected in real time from the solid waste incineration system. The multi-source data includes solid waste characteristic parameters corresponding to various types of solid waste entering the furnace and operation parameters characterizing the operating status. Low-frequency solid waste characteristic data: industrial analysis parameters (moisture content, ash content, volatile matter, fixed carbon), elemental analysis parameters (C, H, O, N, S, Cl, etc.), and fuel characteristic parameters including calorific value of various solid wastes fed into the furnace; High-frequency operating condition data: including adjustable fuel system parameters, steam-water system parameters, auxiliary system parameters, etc., including solid waste feed flow rate, primary / secondary air volume and temperature, auxiliary fuel flow rate, temperature and pressure data of key parts in the incineration facility, water flow rate and temperature, crusher frequency, etc. Purification system data: including flue gas treatment system parameters such as the dosage of reducing agent (e.g., ammonia, urea) in the denitrification system, the dosage of absorbent (e.g., lime, alkali solution) in the deacidification system, and the amount of activated carbon injected; Final emission monitoring data includes flue gas parameters from major equipment, such as temperature, pressure, moisture content, flow rate, oxygen concentration, and carbon dioxide concentration, as well as SO2 and NO emissions at the chimney. x The target pollutant concentrations include HCl, CO, and NH3.

[0037] The dynamic feature vector includes: the received basis moisture content, dry basis ash content, received basis lower heating value of the mixed fuel fed into the furnace, and the fusion value of at least one dry basis element content corresponding to the target predicted pollutant type. If the pollutant is HCl, the dry basis element is dry basis chlorine; if it is NOx, the dry basis nitrogen is required; and if it is SO2, the dry basis sulfur is required. The multi-source data also includes monitoring parameters characterizing the emission results, which are used for periodic online retraining or adaptive parameter updates of the cascade prediction model.

[0038] The data fusion preprocessing shown in the figure includes: Dimensionality reduction of solid waste characteristic data: When multiple solid wastes with different characteristics are fed into the furnace at the same time, unlike systems using a single stable fuel, the solid waste characteristic parameters are numerous and have low frequency. Dynamic fusion is required to reduce the dimensionality of fuel characteristic parameters. For example, the characteristic parameters of each solid waste can be weighted and fused according to the flow value corresponding to each solid waste to generate a dynamic parameter vector representing the characteristics of the mixed solid waste currently fed into the furnace. Based on the real-time flow rates of x types of solid waste entering the furnace, a weighted average algorithm is used to convert the monthly updated static characteristic parameters into dynamic parameters that vary with the flow rate ratio, generating y dynamic characteristics of the "virtual mixed solid waste," including moisture content, ash content, calorific value, and chlorine content. Specifically, for the parameter category of the received basis (such as moisture content, calorific value, etc.), its fusion value at time t is calculated as follows:

[0039] ; in This represents the fusion value of the j-th type of parameter at time t. Let be the mass flow rate of the i-th group of solid waste. Let be the j-th type of characteristic parameter of the i-th group of solid waste.

[0040] For the parameter category (air-dried moisture content) of the air-dried basis, its fusion value at time t is calculated as follows: ; in Let be the received basis moisture characteristic parameters of the i-th group of solid waste.

[0041] For the parameter categories of the empty basis (dry basis ash, carbon, hydrogen, oxygen, nitrogen, sulfur, chlorine), the fusion value at time t is calculated as follows: ; in Let be the air-dried moisture characteristic parameters of the solid waste corresponding to the i-th group.

[0042] By integrating solid waste parameters, on the one hand, static monthly monitoring data is transformed into transient parameters that change dynamically with flow ratio; on the other hand, by constructing a "virtual mixed solid waste" approach, the number of input dimensions involving x types of solid waste characteristic parameters is optimized from x*y to y without losing key process information.

[0043] Data quality improvement: Outlier detection and processing, interpolation to fill in missing values, and removal of abnormal operating conditions are performed on operational data, purification system data, and emission monitoring data to ensure data authenticity and usability. Outlier detection is based on the properties of Gaussian distribution and uses standard deviation to estimate the dispersion of data. Specifically, the mean μ and standard deviation σ of each parameter are calculated. According to the 3σ criterion, 68% of the data lies within ±1 standard deviation of the mean, 95% lies within ±2 standard deviations of the mean, and 99.7% lies within ±3 standard deviations of the mean. If a data point deviates from the mean of the dataset by more than 3 standard deviations, it is considered an outlier. Data points that are about to exceed the range of [μ-3σ, μ+3σ] are marked as potential outliers. Considering the volatility of solid waste incineration conditions, after outliers are detected, the following manual review methods are adopted according to the specific circumstances to avoid misjudgment: For obviously erroneous records (such as negative values ​​or out-of-range values), they are deleted directly; for extreme values ​​that may reflect the true operating conditions, local weighted regression (LOESS) is used for smoothing; and for outliers of important parameters, manual correction is performed in combination with process knowledge.

[0044] Missing values ​​are handled using different strategies depending on the proportion of missing data. When there are few missing data, linear interpolation is used to fill in the missing values. However, when there is a high proportion of missing data or 0 values, the production operation records are reviewed to determine if it is due to a furnace shutdown, instrument failure, or other reasons. Once confirmed, the variable or time period data is directly removed.

[0045] Data frequency unification: Using the data with the highest sampling frequency as the baseline sequence, the data collected by different devices or manually are first aligned according to the target frequency. For gaps caused by different frequencies, forward padding and other methods are used to fill in the gaps to form a complete dataset, ensuring the integrity and synchronization of the data set. Feature selection and importance evaluation: Based on data categories and sampling frequencies, combined with knowledge of process mechanisms and statistical correlation analysis, the analysis methods and importance evaluation indicators are determined, and comparative analysis methods, including correlation coefficient analysis such as Pearson and Spearman, are used.

[0046] Firstly, based on correlation analysis, different methods for determining the correlation coefficient can be selected depending on whether there is a linear relationship.

[0047] Pearson correlation coefficient (applicable to linear relationships): ; Where n is the number of samples, and are the means of the feature and the target variable, respectively. The closer |r| is to 1, the stronger the linear correlation.

[0048] Spearman rank correlation coefficient (applicable to nonlinear relationships): ; In practical applications, the connection between variables is irrelevant, so the difference in ranks (ρ) between the two observed variables can be calculated using simple steps. Therefore, ρ is: ; Remove features that are irrelevant to the target variable (such as NOx emissions) (|ρ|<0.1).

[0049] Next, a secondary screening based on process knowledge is required. While conducting correlation analysis, key variables should be retained in conjunction with the incineration mechanism, such as combustion control variables: primary air volume, auxiliary combustion, and bed temperature; solid waste characteristics variables: moisture content, volatile matter, and chlorine content; and purification system variables: ammonia injection rate and alkali solution rate.

[0050] Finally, the optimal feature subset is determined through recursive feature elimination (RFE) to minimize the RMSE of the model on the validation set.

[0051] Through the preprocessing steps described above, the raw data is transformed into a well-structured dataset suitable for modeling, laying the foundation for the subsequent construction of the prediction model. Key parameters at each stage of preprocessing (such as outlier thresholds and interpolation models) are determined through cross-validation to ensure the robustness of the processing method.

[0052] In the fields of machine learning and deep learning, evaluating and validating model prediction performance is a crucial step in ensuring the reliability and practicality of model performance. This embodiment uses mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 MSE is used as the primary evaluation metric to comprehensively measure the predictive performance of the model. MSE amplifies larger errors through squaring and is highly sensitive to outliers. Its calculation formula is shown below.

[0053] ; in, This is the actual value. It is a predicted value. It refers to the number of samples.

[0054] RMSE, as the square root of MSE, maintains the same dimensions as the original data, making it easier to interpret errors in practical applications. Its calculation formula is shown below.

[0055] ; MAE directly calculates the absolute difference between the predicted value and the actual value, which has better robustness and intuitiveness. Its calculation formula is shown in the following formula.

[0056] ; Coefficient of determination (R) 2 This measures how well a regression model fits the data. It represents the proportion of variation in the dependent variable (target variable) that can be explained by the independent variable, and its calculation formula is shown below.

[0057] ; The smaller the values ​​of the first three indicators, the better the model's predictive performance. 2 The closer the value is to 1, the more accurate the prediction model. Through systematic evaluation and verification, not only can the model's prediction accuracy be objectively reflected, but potential overfitting problems can also be effectively identified, providing a scientific basis for model optimization. Experimental results show that these indicators can comprehensively evaluate model performance from different perspectives. Among them, MAE exhibits better stability in the presence of data noise, while MSE and RMSE are more suitable for scenarios requiring the highlighting of severe prediction errors. This multi-dimensional evaluation system provides reliable theoretical support for model selection and tuning.

[0058] For highly correlated feature groups (such as multiple adjacent temperature measurement points along the flue gas flow path), only the most representative or process-significant feature is retained to avoid multicollinearity. The cross-correlation matrix between features is calculated; for any two features, if their absolute correlation coefficient is greater than a preset threshold (e.g., 0.85), they are considered highly redundant. Redundant features are then removed based on their importance at the process location.

[0059] Determining the optimal sampling frequency range for model construction: Based on methods such as cross-correlation analysis or transfer entropy, the original operating parameters such as solid waste feeding and air volume are selected as targets. The time lag relationship between the operating parameters and key state parameters (temperature, pollutant concentration) is obtained to determine the maximum sampling frequency. Based on this, high-frequency data is downsampled or aligned to eliminate causal time misalignment. At the same time, a comprehensive evaluation of smoothness and trend preservation ability is used to perform secondary processing on the data to determine the minimum data sampling frequency. Using data within the interval for model construction helps to improve generalization ability and retain important information. The core of the downsampling strategy based on time lag analysis lies in first identifying the lag times between key physical quantities, and then determining a reasonable sampling frequency. First, the time lag relationships between multiple variables are identified, and the cross-correlation function is used to analyze the lag relationship between two time series. For the baseline series... X and target sequence Y The lagged cross-correlation coefficient is defined as:

[0060] ; in and They are respectively X andY The mean, σX and σY They are respectively X and Y standard deviation E It represents the mathematical expectation.

[0061] In practical calculations, sample estimators are used: ; For negative hysteresis τ < 0, then: ; To improve the robustness of the estimation, a grouped cross-validation method is adopted. First, the original data is divided into K consecutive blocks, and the optimal lag is calculated independently within each block. ; Finally, the weighted mode is selected as the final lag estimate based on the weighting function. ; in For weighted samples, As a replication factor, For standardized weights, For the weight function, Let be the cross-correlation number of the i-th data block.

[0062] Next, the optimal downsampling frequency is determined. Although downsampling can improve the generalization ability of the model, continuously increasing the sampling frequency will cause the data to be overly smoothed and lose important information. At the same time, the reduction of available data samples will lead to problems such as overfitting, high variance, and model evaluation distortion. Therefore, the optimal sampling frequency is determined around the two core dimensions of smoothness and trend preservation ability.

[0063] Smoothness metrics quantify the smoothness of a data sequence and reflect the balance between eliminating random fluctuations and preserving useful information. Ideal smoothing should suppress high-frequency noise while maintaining the essential characteristics of the data. First, grouped averages are obtained. Let the original data sequence be... The sequence after group averaging is ,in N and M These represent the lengths of the original sequence and the processed sequence, respectively. The smoothness evaluation index based on the second-order difference of the sequence is defined as:

[0064] ; in , where is the absolute average of the second-order differences of the processed sequence, .

[0065] Trend preservation capability refers to the degree to which the data processing faithfully reproduces the macroscopic change pattern of the original sequence. To avoid the influence of amplitude differences, a normalized accumulation and sequence needs to be constructed.

[0066] ; Similarly, the normalized cumulative sum of the processed sequence is: ; The trend preservation capability assessment index based on cumulative sum series is defined as follows: ; in This represents the Pearson correlation coefficient between the original sequence and the processed sequence.

[0067] Given the special emphasis on trend information in industrial data analysis, and combining expert experience with engineering practice, a comprehensive evaluation function is established: ; in α + β =1, recommended value α =0.4, β =0.6.

[0068] Data standardization and normalization: Data standardization and normalization involves scaling the data to a standard range, such as 0 to 1 or -1 to 1. This step helps the model converge faster and improves its performance. Z-score normalization can be used. The main process involves obtaining the mean μ and standard deviation σ of historical samples x, and then transforming them to conform to the form of a standard normal distribution, i.e., x... stand ~ N(0,1), the specific transformation expression is shown in the following formula.

[0069] ; In the formula: ; .

[0070] After normalization, it can meet the form of a standard normal distribution, forming a regular modeling dataset to facilitate the sub-model division and construction of subsequent models.

[0071] like Figure 2 As shown, the cascaded prediction model is a three-level prediction model structure connected in series, including: The first-level prediction model is used to predict the combustion state parameters in the furnace based on the input dynamic feature vector and operating parameters. The second-level prediction model is used to predict the flue gas state parameters entering the flue gas purification system based on the combustion state parameters output by the first-level prediction model. The third-level prediction model is used to predict the emission concentration of pollutants at the tail end based on the flue gas state parameters output by the second-level prediction model and in combination with the operating parameters of the flue gas purification system. The three-level prediction models are executed sequentially, with the output of the preceding model serving as part of the input for the subsequent model.

[0072] The combustion state parameters output by the first-level prediction model include the temperature of at least three key areas within the furnace; the flue gas state parameters output by the second-level prediction model include at least one of the following: temperature, pressure, moisture content, oxygen content, and carbon dioxide content at the inlet of the flue gas purification system; and the tail pollutants output by the third-level emission concentration prediction model include at least one of sulfur dioxide, nitrogen oxides, hydrogen chloride, carbon monoxide, and ammonia.

[0073] The method for establishing a cascaded prediction model includes the following steps: Collect and preprocess historical multi-source operation data of the solid waste incineration system to obtain a regular historical dataset, which includes solid waste characteristic parameters, operation parameters and corresponding pollutant emission monitoring parameters; A dynamic feature fusion module is constructed, which is configured to perform weighted fusion calculations on the static characteristic parameters of various solid wastes based on their real-time flow information, and output dynamic feature vectors. Based on the process causal chain of incineration (fuel → combustion → heat transfer → purification → emission), a cascade prediction model framework is constructed, and the inputs and outputs of each sub-model of the cascade prediction model are defined. Using historical datasets, each sub-model of the cascaded prediction model is trained separately, and the hyperparameters of each sub-model are jointly tuned with the goal of minimizing the overall prediction error of the cascaded system. The trained and optimized cascaded prediction model is integrated with the dynamic feature fusion module into an executable prediction system and then encapsulated in an engineering manner.

[0074] The inputs and outputs of each sub-model in the cascaded prediction model are as follows: The input to the first-level prediction model is a dynamic feature vector and operating parameters, and the output is the combustion state parameters in the furnace. The input to the second-level prediction model includes at least the combustion state parameters output by the first-level prediction model, and the output is the flue gas state parameters entering the flue gas purification system. The input to the third-level prediction model includes at least the output of the second-level prediction model and the operating parameters of the flue gas purification system. The output is the concentration of pollutants emitted at the tail end.

[0075] Specifically: The first-level model needs to establish the relationship between the input parameters to be predicted and the combustion state in the furnace. The input parameters should include at least fuel-side characteristic parameters and operational characteristic parameters. The fuel-side characteristic parameters include the characteristic parameter vector of the "virtual mixed solid waste" obtained by dynamic fusion calculation in S1, including core parameters that can fully characterize the fuel combustion characteristics, such as moisture content, ash content, volatile matter, fixed carbon, elemental analysis parameters (C, H, O, N, S, Cl, etc.), and calorific value. The operational characteristic parameters are variables that can control the combustion process, including fuel system parameters such as solid waste feed rate and auxiliary fuel feed rate, air volume system parameters such as primary, secondary, tertiary, total air volume, induced draft air volume, and air temperature, steam-water system parameters such as feedwater temperature and feedwater flow rate, and auxiliary system parameters such as crusher frequency, auger mixer frequency, and soot blowing frequency that affect the particle size of solid waste.

[0076] The output parameters represent key combustion state parameters within the incinerator, primarily including temperature and pressure distribution in critical areas. Temperature distribution should be selected based on the location and path of fuel entering the furnace for combustion, and should at least include temperatures in different combustion zones within the furnace, such as the sand bed temperature, dense phase zone temperature, and dilute phase zone temperature in a fluidized bed; the ash hopper zone temperature, core combustion zone temperature, burnout zone temperature, and screen zone inlet temperature in a pulverized coal furnace; the preheating zone temperature, decomposition zone temperature, exothermic reaction zone temperature, sintering zone temperature, and secondary combustion chamber temperature in a rotary kiln; and the drying section temperature, combustion zone temperature, furnace center temperature, and burnout zone temperature in a grate furnace. Specific temperature measurement points should be selected based on the furnace type, but at least three temperatures should be included. Pressure distribution should be selected synchronously based on the temperature measurement point distribution. If no corresponding location is available, the inlet, center, or outlet locations can be optimally selected based on the distribution of measurement points within the furnace.

[0077] Second-level prediction sub-model M2 (flue gas state prediction model): The second-level model needs to establish a mapping relationship between the input parameters to be predicted and the preceding combustion state to the inlet physical conditions of each flue gas purification system. The input parameters mainly consist of three parts: the inherited root source parameter set, the upstream state parameter set, and the current stage operation / disturbance parameter set (optional). The inherited root source parameter set is all the input parameters of the previous-level prediction model M1. The state of the generated flue gas after leaving the furnace essentially still depends on the fuel characteristics and combustion conditions. The parameters inherited from the previous level ensure the causal continuity of the results of the second-level prediction model M2. The upstream state parameter set is all the output parameters of the previous-level prediction model M1. The temperature, pressure, and other parameters in the output of M1 define the initial energy state and physical state of the flue gas when it enters the subsequent process. The current stage operation / disturbance parameter set consists of a small number of operational variables that can significantly affect the flue gas state after the combustion stage and before the flue gas purification stage. These mainly include parameters that are specific to this stage and have a significant impact on parameters such as the flow rate of the heat exchange equipment, the air leakage coefficient adjustment, and the flue gas reheat equipment, which have a significant impact on parameters such as the flow rate, temperature, and oxygen content of the flue gas.

[0078] The output parameters represent the key physical state parameters of the flue gas before it enters the core purification unit, which determine the chemical reaction rate and removal efficiency of subsequent pollutants. These parameters include at least the inlet flue gas temperature of the core flue gas purification units such as desulfurization and denitrification, and the flue gas pressure, flow rate, moisture content, oxygen content, and CO2 content after passing through all flue gas treatment equipment. Additionally, parameters such as the inlet pressure, flow rate, moisture content, oxygen content, CO2 content, and dust content of each core flue gas purification unit can be added as outputs based on system configuration and predicted requirements.

[0079] The third-level prediction sub-model M3 (pollutant emission prediction model): The third-level model needs to establish a comprehensive mapping relationship between the overall process state information, purification system parameters, and the final pollutant emission concentration. The input parameter categories are consistent with the previous-level prediction model, mainly consisting of three parts: the inherited root source parameter group, the upstream state parameter group, and the current-stage operation / disturbance parameter group. The inherited root source parameter group consists of all the input parameters of the previous-level prediction model M2. M3 receives complete process chain information on pollutant generation, including fuel characteristics, combustion operation, and combustion state, ensuring causal continuity. The upstream state parameter group consists of all the output parameters of the previous-level prediction model M2. Parameters such as flue gas pressure, temperature, moisture content, oxygen content, and CO2 in the M2 output define the boundary conditions affecting pollutant removal when flue gas enters different flue gas treatment devices. The current-stage operation / disturbance parameter group consists of operational variables that can significantly affect the pollutant removal level during flue gas purification. These mainly include flue gas treatment system parameters such as the dosage, concentration, and particle size of reagents such as urea water, ammonia water, alkaline solution, lime water, and activated carbon.

[0080] The output parameters correspond to the emission concentrations of the target pollutants at the chimney outlet or online monitoring point, including at least the concentrations of carbon monoxide, sulfur dioxide, nitrogen oxides, and hydrogen sulfide. In addition, depending on the object being treated, the predicted concentrations of particulate matter, heavy metals, and dioxin toxicity equivalents can be extended.

[0081] More-level model cascaded prediction framework: For systems that do not fully conform to the three-level prediction structure of "in-furnace combustion state - flue gas state - emission concentration" due to differences in furnace type, heat treatment method, treatment object and system configuration requirements, they are divided into multiple levels according to the direction of material and energy flow. At the same time, sub-models with additional requirements such as more output parameters and higher prediction accuracy in the three-level prediction structure can also be divided into multiple levels. For example, sub-models can be built separately for parameters inside the waste heat boiler, and sub-models can be built separately for parameters before and after flue gas reheating, etc., to form a multi-level prediction model.

[0082] Random forest, gradient boosting decision tree, or deep neural network algorithms are used to construct sub-models at each level, and hyperparameters are jointly tuned using Bayesian optimization or grid search methods. Specifically:

[0083] Data flow organization and model training: First, the training data needs to be split. The preprocessed dataset in S1 is divided into three or more logically independent but progressively structured subsets according to the model framework defined in S2 and the input-output mapping relationship. The data in each subset is further divided into training, validation, and test subsets in a 7:2:1 ratio, ensuring that validation and testing represent future, unpredictable scenarios. Next, model algorithms are selected and initialized. Each sub-model is trained as an independent module. Appropriate machine learning algorithms are selected for initialization based on the characteristics of each sub-task. Random forests, gradient boosting decision trees, and deep neural networks, which can handle complex nonlinear relationships, can be chosen as the foundation for each sub-model. During implementation, multiple models need to be compared for each sub-task, and reasonable initial hyperparameters need to be set.

[0084] The main parameters of the random forest model are shown in the table below.

[0085] The main parameters of the gradient boosting decision tree model are shown in the table below.

[0086] The main parameters of the deep neural network model are shown in the table below.

[0087] Model training and co-optimization also include data downsampling frequency optimization. Data downsampling frequency optimization includes: determining the lag time between key variables based on cross-correlation analysis, evaluating the data smoothness and trend preservation ability under different downsampling frequencies, and using the frequency with the highest comprehensive score as the unified sampling frequency for model training.

[0088] Cooperative optimization strategy for multi-stage systems: Using independent models may amplify the overall error after multiple stages are cascaded, necessitating a cooperative optimization strategy guided by overall prediction accuracy. The ultimate goal of optimization is to minimize the pollutant emission prediction error of the entire cascaded system using independent validation sets. This follows the "weakest link effect," meaning the overall performance of the system is determined by the accuracy of the weakest link; therefore, the performance and error propagation of each stage model need to be considered.

[0089] The optimization strategy consists of two parts. One part is for the S2 framework system. Based on the dimensions of the input and output data, the dimensionality ratio (output data dimension: input data dimension) should be controlled to be at least below 0.5. When the dimensionality ratio exceeds 0.5, a deeper correlation analysis can be performed to increase the input or decrease the output. Alternatively, the sub-model can be divided into two sub-models. However, it should be noted that the more levels the framework system is divided into, the greater the error propagation, which involves more subsequent model training and optimization work.

[0090] One aspect involves model optimization. For the entire system, higher-order systems require greater robustness in prediction. As the order increases, the robustness requirement shifts to accuracy and performance metrics. It's necessary to determine the optimal model under different frequency reduction conditions and select the optimal model for each prediction sub-model based on the robustness and performance metrics at the optimal frequency reduction point. After fixing the model structure, using the overall validation set error as the loss function, automated hyperparameter optimization methods such as Bayesian optimization and grid search are employed to jointly search and fine-tune key hyperparameters (such as learning rate, tree depth, and subsampling rate) of the selected models for each sub-model. The model is expanded left and right by 3-5 variables based on the initial hyperparameter settings, as shown in the table below. The model is then adjusted based on the optimized parameters to further reduce the accumulation of error during propagation and ensure optimal overall performance.

[0091] Model integration and system solidification: After training and optimization, the models at each level are integrated into a robust prediction system. The optimized model parameters at each level are serialized and saved separately. At the same time, the corresponding feature scalers, feature selectors, and parameter mapping relationships required for dynamic fusion calculation are saved to construct a prediction pipeline.

[0092] The ensemble model receives the input parameters from the first-level model and the current-stage operation / perturbation parameter set from subsequent models as inputs. According to the defined model framework, it selects the inputs from the ensemble model as the inputs to the first-level sub-model for output. Subsequent models select the outputs from the preceding models and the inputs from the ensemble model as inputs for output, ultimately obtaining the predicted output of the target pollutant.

[0093] Model engineering deployment and application: In the model integration and deployment steps, the integrated prediction system is encapsulated as a microservice with standard data interfaces, including OPC-UA and / or RESTful-API.

[0094] Specifically, the integrated predictive model is encapsulated as a service and integrated into the existing distributed control system (DCS) or digital twin platform of the solid waste incineration plant through standard data interfaces (such as OPCUA and RESTAPI). This system supports two core application modes:

[0095] Offline simulation and optimization mode: Operators can input different material ratio schemes and preset operating parameters, and the model can quickly simulate and predict the corresponding emission results, which can be used for scheme evaluation before receiving new materials and daily operating parameter optimization.

[0096] Online real-time prediction and early warning mode: The model periodically reads real-time operating data and makes online rolling predictions of pollutant emission trends in the near future. When the predicted value approaches the emission limit, the system issues an early warning and can provide operational suggestions (such as adjusting air volume and fine-tuning reagent dosage) in conjunction with optimization algorithms, realizing a leap from passive feedback to active feedforward.

[0097] The steps of feedforward control include: inputting the predicted emission concentration value into the advanced process controller; the advanced process controller calculates and outputs adjustment commands in advance based on the deviation between the predicted emission concentration value and the target set value; and directly controls the dosage of denitrification reducing agent or deacidification absorbent.

[0098] This embodiment focuses on a bubble-fluidized bed sludge incineration line with a processing capacity of 42.7 tons of dry solids per day (tDS / d). The system has a complete process flow, as follows: Figure 3 As shown, the system mainly includes: a wet sludge receiving and storage system, a paddle dryer (to dry part of the wet sludge to a moisture content of 30-40%), a bubbling fluidized bed incinerator, a waste heat boiler system, and a complete flue gas purification system. The flue gas purification system employs a combined process of "selective non-catalytic reduction (SNCR) + semi-dry spray deacidification absorption tower + activated carbon spray adsorption + bag filter + wet alkaline scrubbing tower (NaOH solution)" to ensure that pollutants are discharged in compliance with standards. The system is equipped with a complete distributed control system (DCS) and a continuous emission monitoring system (CEMS).

[0099] like Figure 1 As shown, in order to build a predictive model, historical operational data of the system over a calendar year (365 days) was systematically collected, mainly divided into three categories: High-frequency process operation data: Sourced from the DCS system, automatically collected and stored at fixed intervals of 30 seconds. This mainly includes: real-time flow rates (t / h) of the four sludge feeders (wet sludge left / right, semi-dry sludge left / right); primary air fan flow rate (Nm³ / h) and preheated temperature (°C); auxiliary burner natural gas flow rate (Nm³ / h); flue gas temperature (°C) at different heights inside the furnace (sand bed, dense phase zone, dilute phase zone) and at the outlet; temperature (°C) and pressure (Pa) at each key node along the flue gas purification process (waste heat boiler outlet, electrostatic precipitator outlet, semi-dry tower outlet, bag filter outlet, wet tower inlet / outlet, reheater outlet, chimney); urea solution injection flow rate (L / h), alkali solution consumption flow rate (m³ / h), quicklime feed rate (kg / d); and total flue gas flow rate (Nm³ / h), humidity (%), oxygen content (O2, %), and carbon dioxide content (CO2, %).

[0100] Low-frequency solid waste characteristic data: derived from laboratory sampling and testing reports taken periodically (usually weekly or per batch). This includes industrial analysis (received basis moisture, air-dried basis moisture, dried basis ash), elemental analysis (dried basis carbon, hydrogen, oxygen, nitrogen, sulfur, chlorine), and received basis lower heating value for each batch of sludge. These data represent the essential characteristics of the fuel fed into the furnace, but are updated much less frequently than process data.

[0101] Final emission monitoring data: sourced from the CEMS system and synchronized in real time with DCS data. Monitoring parameters include O2, CO2, NH3, CO, SO2, and NO at the chimney outlet. x The measured concentrations (mg / Nm³, dry basis, standard state) of pollutants such as HCl will be used as the final target variables for model training and validation.

[0102] A total of 1,051,200 sets of sample data can be obtained at a sampling interval of 30 seconds.

[0103] The sludge fed into the furnace consists of wet sludge (80% moisture content) and semi-dry sludge (30-40% moisture content), added from the left and right sides respectively. Based on the real-time flow rates of the four types of sludge, a weighted average algorithm is used according to a dynamic fusion calculation method to generate 11 dynamic characteristics of the "virtual mixed sludge," including moisture content, ash content, calorific value, and chlorine content. Through sludge parameter integration, the input dimensions are optimized from 44 to 11 without losing key process information.

[0104] Outlier detection and missing value handling were performed, resulting in 748,333 valid sample data sets.

[0105] Feature filtering: The data collection details are shown in the table below: like Figure 4 As shown in the table, Spearman correlation coefficient analysis and process knowledge were used to analyze the parameters in the table. Finally, 37 key characteristics strongly correlated with pollutant emissions were selected from the original variables, including dynamic sludge characteristics (11), main operating variables (such as primary air volume, auxiliary fuel volume, 7), key temperatures (7), purification agent dosage (3), and some flue gas state parameters (9), as shown in the table below: Determining the optimal frequency reduction range: Downsampling analysis based on time lag analysis was performed, using mixed sludge flow rate as the baseline sequence to obtain the number of lag steps and correlation coefficients for multiple variables, as follows: Figure 5 As shown. The main characteristic indicators, including the smoothness index after downsampling, trend preservation ability, and overall evaluation, are as follows: Figure 6 As shown, (a) represents the number of remaining samples at different downsampling frequencies, (b) shows the changes in the smoothness index at different downsampling frequencies, (c) shows the trend preservation capability index at different downsampling frequencies, and (d) is a schematic diagram of the comprehensive evaluation at different downsampling frequencies. Where d = 0.4b + 0.6c.

[0106] The construction and training of cascaded prediction models includes the following steps: Based on the aforementioned data processing, a three-level cascaded prediction framework is constructed according to the process flow.

[0107] Task 1 (Temperature Prediction): The input is 18 dimensions (10 sludge characteristics + total sludge volume + primary air volume and temperature + 2 auxiliary combustion quantities), and the output is 3 dimensions (sand bed temperature, dense phase zone temperature, and dilute phase zone temperature).

[0108] Task 2 (Flue Gas State Prediction): The input is 21-dimensional (18 inputs from Task 1 + 3 outputs from Task 1), and the output is 11-dimensional (7 along-the-path flue gas temperatures + flue gas flow rate, humidity, O2, CO2).

[0109] Task 3 (Emission Prediction): The input is 32-dimensional (18 inputs from Task 1 + 3 outputs from Task 1 + 11 outputs from Task 2), and the output is 5-dimensional (emission concentrations of NOx, SO2, HCl, CO, and NH3).

[0110] The effective dimensionality ratio after the cascaded model is constructed is compared with that of the traditional end-to-end model in the table below. The cascaded model transforms the unfavorable dimensionality ratio into three favorable sub-tasks, reducing the risk of model overfitting.

[0111] Random forest (RF), gradient boosting decision tree (GBDT), and deep neural network (DNN) were selected as candidate models, and the initial hyperparameters of the sub-models were set according to S3.1.

[0112] Independent training: Three tasks were trained using corresponding subsets of data to obtain the main metrics after down-concentration using different models and data.

[0113] The main parameters for Task 1 prediction are shown in the table below.

[0114] The main prediction parameters for Task 2 are shown in the table below.

[0115] The main prediction parameters for Task 3 are shown in the table below.

[0116] Among the three tasks, the optimal model is not distributed at the same frequency reduction point. Considering the cumulative error of the three tasks in cascaded prediction, we need to pursue the best overall performance in terms of accuracy and avoid performance bottlenecks. We use the bottleneck effect to evaluate the frequency reduction point. The method is to find the best R-value that each task can achieve at each candidate frequency reduction point. 2 Then observe the lowest R among them. 2 The optimal unified frequency reduction point should ensure that the accuracy of the "worst-case" task remains high. The analysis results are shown in the table below.

[0117] In this task, selecting 60 groups (i.e., sampling interval of 30 minutes) as the unified frequency reduction point based on the results is appropriate.

[0118] When 60 groups were selected as the uniform reduction point, the GBDT model demonstrated excellent performance in all three tasks. However, the model's potential was not fully realized under the default parameter configuration, and optimization of hyperparameters can further improve its performance. With the regularization parameter and computation settings unchanged, five parameters—ensemble size, learning size, and tree structure control—were optimized. The optimization range of the hyperparameters is shown in the table below.

[0119] With RMSE, MAE, R 2 Using the hyperparameters as the benchmark, we optimized them within the selected range and obtained the optimal hyperparameter settings, as shown in the table below.

[0120] The results after hyperparameter optimization are shown in the table below. After hyperparameter optimization, the metrics of the three cascaded tasks improved to varying degrees, which helps to enhance the generalization performance of the industrial process data-driven model.

[0121] One hundred independent test set samples were used for full-process prediction validation. The model before and after hyperparameter optimization was used to predict pollutants, and the results were compared between the actual and predicted values. Figures 7-11 As shown, the model with optimized hyperparameters exhibits a higher degree of fit to the true values ​​on the vast majority of sample points, and its tracking response capability is significantly enhanced. Figure 12 As shown, the relative error distribution of the unoptimized model on the left exhibits a clear right skewness and broad peaks, indicating a systematic tendency to overestimate the model under the original parameter configuration, and a large dispersion in the prediction error. In contrast, the error distribution of the optimized model on the right shows an overall shift towards zero and a significantly narrowed distribution width. Hyperparameter optimization effectively corrects the model's systematic bias, reducing the interquartile range (IQR) by 35%-50%, resulting in a substantial improvement in sample prediction accuracy. The tail characteristics of the optimized error distribution also change significantly; the probability of extreme error samples with relative errors > ±15% decreases from 37% before optimization to 18% after optimization, indicating an enhanced generalization ability of the model for abnormal operating conditions. The degree of improvement in the symmetry of the prediction error distribution varies, and the skewness coefficient is close to the ideal normal distribution, effectively balancing the model's fitting ability and generalization performance.

[0122] This embodiment integrates with industrial applications of models.

[0123] The next step is to realize the actual deployment and application of the cascaded prediction model in the solid waste incineration production environment, complete the engineering encapsulation and system integration of the model, and form a complete "data-model-application" closed loop.

[0124] The trained and optimized three-level cascaded prediction model is standardized and lightweightly packaged. Specifically, the model inference logic, preprocessing code, and dependent environment are packaged into a Docker container to form an independent prediction microservice. This service provides prediction functionality through a predefined RESTful API interface, decoupling the computational logic from the business system and ensuring the model's portability, scalability, and ease of maintenance. The service is deployed on servers in the factory data center or a private cloud platform to ensure the stability and security of computing resources. The encapsulated system interface of this embodiment is shown below. Figure 13 As shown.

[0125] To achieve seamless integration between the predictive service and the factory's existing automation system, a standardized data exchange channel was established. The service establishes a secure connection with the factory's distributed control system (DCS) real-time database via industrial communication protocols (such as OPCUA), periodically reading the real-time operating parameters required by the model (such as feed rate, air volume, temperature, and reagent dosage). All input and output data are encapsulated and transmitted using lightweight JSON format. The input JSON object contains structured key-value pairs corresponding to the model's input features; the output JSON object contains the prediction results at various levels, such as temperature, flue gas conditions, and pollutant concentrations. This standardized JSON-based interface greatly simplifies the complexity of system integration and improves the flexibility and readability of data interaction.

[0126] The packaged predictive services are integrated into the factory's digital twin platform or advanced operational guidance system to form an offline simulation module. In this module, operators can input or adjust preset operating parameters, such as:

[0127] Change the characteristic parameters (moisture content, calorific value, chlorine and sulfur content) of various solid wastes planned to be fed into the furnace; set different dry and wet material co-firing ratios and total treatment loads; adjust the primary air volume, set the auxiliary fuel quantity, and reagent usage, etc.

[0128] The system completes cascade calculations within seconds by calling the prediction service's API, and then provides the user with the predicted furnace temperature distribution, flue gas conditions along the process, and final pollutant emission concentrations in a visual manner (such as curves and numerical values). This function provides a quantitative tool for the pre-assessment and optimization of operational plans, and can be used for feasibility studies of new material co-firing, optimization of daily operating parameters, and the development of contingency plans to cope with raw material fluctuations, realizing a refined management model of "simulation verification first, then actual operation".

[0129] Although this embodiment uses sludge incineration as an example, the method is universal: For municipal solid waste incineration, solid waste characteristic parameters can be replaced by the composition of the waste (proportion of kitchen waste, plastics, paper, etc.), calorific value, chlorine content, etc.

[0130] For industrial hazardous waste incineration, the content of specific toxic and harmful elements (such as heavy metals, F, Br, etc.) should be the key input characteristic.

[0131] The cascade framework can adjust the specific input and output variables of the model according to different incineration processes (grate furnace, rotary kiln, etc.) and purification processes, but its core cascade logic of "combustion-heat transfer-purification" remains unchanged.

[0132] The above-disclosed embodiments are merely preferred embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for predicting pollutant emissions from the tail end of a solid waste incineration process accommodating strong fluctuations in raw material emissions, characterized in that, include: Multi-source data is collected in real time from the solid waste incineration system. The multi-source data includes solid waste characteristic parameters corresponding to various types of solid waste entering the furnace and operation parameters characterizing the operating status. Based on the static characteristic parameters of the various solid wastes fed into the furnace, and the real-time flow information of the various solid wastes corresponding to the operation parameters, a dynamic feature vector that can reflect the characteristics of the mixed fuels fed into the furnace in real time is generated through weighted fusion calculation. The dynamic feature vector and the operation parameters are input into a pre-constructed cascaded prediction model. The cascaded prediction model is based on the physicochemical causal chain of the incineration process. The cascaded prediction model performs recursive state prediction through multi-level series of sub-models and finally outputs the predicted emission concentration of tail pollutants. Based on the pollutant concentration prediction results output by the model, perform at least one of the following actions: early warning, operational guidance, or feedforward control.

2. The method for predicting tail-end pollutant emissions from solid waste incineration processes with strong raw material fluctuations, as described in claim 1, is characterized in that... The cascaded prediction model is a three-level prediction model structure connected in series, including: The first-level prediction model is used to predict the combustion state parameters in the furnace based on the input dynamic feature vector and operating parameters. The second-level prediction model is used to predict the flue gas state parameters entering the flue gas purification system based on the combustion state parameters output by the first-level prediction model. The third-level prediction model is used to predict the emission concentration of pollutants at the tail end based on the flue gas state parameters output by the second-level prediction model and in combination with the operating parameters of the flue gas purification system. The three-level prediction models are executed sequentially, with the output of the preceding model serving as part of the input for the subsequent model.

3. The method for predicting tail-end pollutant emissions from solid waste incineration processes with strong raw material fluctuations, as described in claim 2, is characterized in that... The combustion state parameters output by the first-level prediction model include the temperature of at least three key areas within the furnace; the flue gas state parameters output by the second-level prediction model include at least one of the temperature, pressure, moisture content, oxygen content, and carbon dioxide content at the inlet of the flue gas purification system; and the tail pollutants output by the third-level emission concentration prediction model include at least one of sulfur dioxide, nitrogen oxides, hydrogen chloride, carbon monoxide, and ammonia.

4. The method for predicting tail-end pollutant emissions from solid waste incineration processes with strong raw material fluctuations, as described in claim 1, is characterized in that... The dynamic feature vector includes: the fusion value of the received basis moisture content, dry basis ash content, received basis lower heating value of the mixed fuel fed into the furnace, and the content characteristics of at least one dry basis element corresponding to the target predicted pollutant type; the multi-source data also includes monitoring parameters characterizing the emission results, which are used to periodically retrain or adaptively update the parameters of the cascaded prediction model online.

5. The method for predicting tail-end pollutant emissions from solid waste incineration processes with strong raw material fluctuations, as described in claim 1, is characterized in that... The feedforward control steps include: inputting the predicted emission concentration value into the advanced process controller, which calculates and outputs adjustment commands in advance based on the deviation between the predicted emission concentration value and the target set value, and directly controls the dosage of denitrification reducing agent or deacidification absorbent.

6. The method for predicting tail-end pollutant emissions from solid waste incineration processes with strong raw material fluctuations, as described in claim 1, is characterized in that... The method for establishing a cascaded prediction model includes the following steps: Historical multi-source operation data of the solid waste incineration system are collected and preprocessed to obtain a regular historical dataset, which includes solid waste characteristic parameters, operation parameters and corresponding pollutant emission monitoring parameters. A dynamic feature fusion module is constructed, which is configured to: perform weighted fusion calculation on the static characteristic parameters of various solid wastes based on their real-time flow information, and output a dynamic feature vector; Based on the process causal chain of incineration, a cascade prediction model framework is constructed, and the inputs and outputs of each sub-model of the cascade prediction model are defined. The cascaded prediction model framework is trained using the historical dynamic feature vector and the operation parameters in the historical dataset. The hyperparameters of each sub-model are jointly tuned with the goal of minimizing the overall prediction error of the cascaded system, resulting in the trained and optimized cascaded prediction model.

7. The method for predicting tail-end pollutant emissions from solid waste incineration processes with strong raw material fluctuations, as described in claim 6, is characterized in that... The inputs and outputs of each sub-model of the cascaded prediction model are as follows: The input to the first-level prediction model is the dynamic feature vector and operating parameters, and the output is the combustion state parameters in the furnace. The input to the second-level prediction model includes at least the combustion state parameters output by the first-level prediction model, and the output is the flue gas state parameters entering the flue gas purification system. The input to the third-level prediction model includes at least the output of the second-level prediction model and the operating parameters of the flue gas purification system, with the output being the concentration of pollutants emitted at the tail end.

8. The method for predicting tail-end pollutant emissions from solid waste incineration processes with strong raw material fluctuations, as described in claim 6, is characterized in that... Each sub-model is constructed using random forest, gradient boosting decision tree, or deep neural network algorithms, and its hyperparameters are jointly tuned using Bayesian optimization or grid search methods.

9. The method for predicting tail-end pollutant emissions from solid waste incineration processes with strong raw material fluctuations, as described in claim 6, is characterized in that... It also includes data downsampling frequency optimization, which includes: determining the lag time between key variables based on cross-correlation analysis, evaluating the data smoothness and trend preservation ability under different downsampling frequencies, and using the frequency with the highest comprehensive score as the unified sampling frequency for model training.

10. The method for predicting tail-end pollutant emissions from solid waste incineration processes with strong raw material fluctuations as described in claim 6, wherein the trained and optimized cascaded prediction model is integrated with the dynamic feature fusion module into an executable prediction system, and the integrated prediction system is encapsulated as a microservice with a standard data interface.

Images