Intelligent prediction and migration control method for key indicators of solid waste incineration system

Abstract

This invention discloses an intelligent prediction and migration control method for key indicators of solid waste incineration systems. The method first constructs a hybrid expert model based on historical operating data from a benchmark incinerator, and incorporates a physical information neural network structure, introducing soft physical constraints such as oxygen balance, total pollution load conservation, and energy balance into the model training to improve the physical consistency of the prediction results. Based on this, cross-plant migration is achieved through domain adaptation, parameter fine-tuning, or plant-specific embedding vectors, enabling the model to quickly adapt to different incineration plants' data distributions and operating characteristics, maintaining high-precision prediction capabilities for multi-pollutant emissions. This invention further constructs a carbon-pollution synergistic comprehensive risk index to achieve real-time quantitative assessment of environmental risks during the incineration process. This method can stably predict key emission indicators of solid waste incineration systems under limited data conditions and provides effective decision-making basis for incineration condition optimization and pollution control.

Description

Technical Field

[0001] This invention belongs to the field of solid waste energy disposal technology, and in particular relates to an intelligent prediction and migration control method for key indicators of a solid waste incineration system. Background Technology

[0002] Solid waste incineration, as a crucial component of modern urban waste management systems, directly impacts regional environmental quality, workplace safety, and public health risks through its operational stability and emission compliance. With accelerating urbanization and increasingly complex waste composition, the numerous heterogeneous and highly volatile operational characteristics of incineration systems result in highly time-varying and coupled pollutant emissions. Traditional pollutant monitoring and prediction methods largely rely on fixed-mechanism models or single statistical models, which often assume relatively stable pyrolysis and oxidation patterns under specific operating conditions. However, in actual industrial settings, factors such as feed composition, load variations, air distribution, and furnace temperature fluctuations often exhibit complex nonlinear relationships, making it difficult for traditional models to maintain long-term, cross-condition prediction accuracy.

[0003] With the widespread adoption of online monitoring systems, incineration plants have accumulated a wealth of multi-source, multi-scale operational data, providing a foundation for intelligent prediction. However, machine learning models trained at a single plant generally lack cross-plant generalization capabilities. Due to significant differences in furnace type, flue gas system, air-fuel ratio strategy, combustion organization, and waste composition between different incineration plants, the same model often performs poorly in new plants, requiring substantial retraining with new data and computing power, resulting in high costs for practical deployment. Furthermore, traditional deep learning methods rely solely on the data itself for pattern fitting, which easily falls into the "black box" problem when effective physical constraints are lacking. The model may produce predictions that violate physical laws under sparse data or extreme operating conditions, which is detrimental to the reliability and stability requirements of engineering applications.

[0004] On the other hand, there are generally complex interactions among multiple pollutant emissions, such as NO x SO2 generation is closely related to furnace temperature, SO2 emissions are influenced by the pyrolysis pathways of chlorides and sulfur-containing components, CO concentration is often a characterization of incomplete combustion, and PM emissions are affected by air volume, boiler load, and flue gas purification efficiency. Understanding the synergistic effects among these pollutants and developing measurable and quantifiable comprehensive evaluation indicators is crucial for the operation and control of incineration plants. However, existing methods mostly focus on single-pollutant prediction or modeling of single physical mechanisms, lacking a holistic description of the risks associated with multiple pollutants.

[0005] In recent years, Physically Informed Neural Networks (PINNs) have attracted widespread attention in modeling complex engineering systems. PINNs introduce physical conservation constraints during training, enabling models to maintain physical consistency even when undersampling or expanding across scenarios. However, the application of PINNs in real-world industrial systems still has limitations: the physical mechanisms involved in industrial processes are often incomplete and unavailable, and physical parameters have high uncertainty; using strict equations may actually reduce the model's flexibility. Therefore, how to transform the advantages of PINNs into "soft constraint" mechanisms suitable for industrial data scenarios is a pressing issue in this field. Furthermore, existing research on cross-plant prediction typically relies on transfer learning methods, but most transfer methods are still limited to parameter fine-tuning or feature alignment, failing to fully absorb the common structures implicit in multi-plant data, lacking integration of physical laws, and their generalizability needs improvement. Therefore, there is an urgent need to construct an intelligent prediction framework that can integrate data from different plants, take into account physical constraints, and possess cross-scenario transfer capabilities, giving it both the flexibility of machine learning models and the stability provided by physical constraints.

[0006] In summary, the main technical shortcomings in this field are as follows: (1) weak cross-plant prediction capability, and the model cannot be quickly adapted to new plants; (2) lack of a comprehensive risk assessment index that can express the synergistic relationship of multiple pollutants; (3) traditional data-driven methods are prone to generating results that lack physical consistency; and (4) existing PINN models are difficult to directly apply to complex data scenarios in industrial sites. In order to overcome the above shortcomings, it is necessary to propose a new intelligent prediction and transfer learning method for solid waste incineration systems. By combining multi-task hybrid expert models, multi-plant pre-training, PINN soft physical constraints and carbon-pollutants synergetic index (CPSI) for synergistic risk quantification, high-precision prediction of pollutant emission indicators and rapid cross-plant transfer capability can be achieved, thereby supporting the intelligent operation and environmental risk control of incineration plants. Summary of the Invention

[0007] This invention aims to overcome the shortcomings of existing technologies, such as insufficient cross-plant prediction capability of solid waste incineration process, lack of quantitative description of synergistic relationship between pollutants, insufficient physical consistency of traditional data-driven models, and lack of effective structured mechanism for transfer learning. It proposes an intelligent prediction and transfer control method for key indicators of solid waste incineration system.

[0008] The objective of this invention is achieved through the following technical solution: an intelligent prediction and migration control method for key indicators of a solid waste incineration system, applied to a solid waste incineration process (with heterogeneity across multiple plant areas) to monitor data, comprising the following steps:

[0009] 1) Collect historical operating data of the benchmark incinerator;

[0010] 2) The multi-task hybrid expert model is pre-trained using the historical operating data of the benchmark incinerator, and the data fitting loss and physical constraints of the physical information neural network are jointly optimized. The physical constraints include at least oxygen balance constraints, total pollution load conservation constraints and energy balance constraints.

[0011] 3) Transfer learning is performed on the target incinerator. The pre-trained model is adapted to the domain or fine-tuned by the operation data of the target furnace area so that the model can be adapted to the operation characteristics of different incineration plant areas.

[0012] 4) Real-time prediction of multiple pollutant emission indicators in the target furnace area based on the migrated model.

[0013] Furthermore, the operating data includes steam flow rate, furnace temperature, primary air volume, secondary air volume, flue gas temperature, flue gas O2 content, and emission concentrations of multiple pollutants.

[0014] Furthermore, the oxygen balance constraint in step 2) uses the residual between the equivalent excess air coefficient calculated based on the primary air volume, secondary air volume and steam flow rate and the excess air coefficient calculated based on the dry basis volume fraction of flue gas O2 as a physical constraint term.

[0015] Furthermore, the total pollution load conservation constraint constructs a comprehensive pollution load by weighting pollutant concentrations according to their weights and estimates the flue gas volume flow rate using flue gas velocity, thus ensuring that the pollution load on the input side is consistent with the pollution load on the output side.

[0016] Furthermore, the energy balance constraint establishes a soft constraint residual by using steam flow rate, primary air and secondary air sensible heat as input energy, and using furnace average temperature and flue gas temperature as output energy.

[0017] Furthermore, the transfer learning includes any or a combination of parameter freezing transfer, partial layer fine-tuning, or constructing plant area embedding vectors for plant area differences.

[0018] Furthermore, the multi-task hybrid model includes multiple expert networks and a gating network, wherein the gating network automatically selects the most suitable expert network for pollutant prediction based on different operating conditions.

[0019] Furthermore, after predicting the multi-pollutant emission indicators of the target furnace area in real time based on the migrated model, it also includes calculating the carbon-pollution synergistic comprehensive risk index to realize intelligent prediction and risk assessment of pollutant emissions across plant areas;

[0020] The carbon-pollution synergistic risk index is constructed based on the concentration of multiple pollutants, the toxicity weight of pollutants, the stringency of emission limits, and the regulatory priority, and is obtained by weighted aggregation.

[0021] Furthermore, the pollutant weights in the carbon-pollution synergistic risk index are obtained by combining pollutant toxicity, emission limit stringency, and regulatory priority using weighting coefficients.

[0022] Furthermore, the physical constraint loss and data fitting loss are weighted by adjustable hyperparameters to guide the model to achieve a balance between physical consistency and prediction accuracy.

[0023] Compared with the prior art, the beneficial effects of the present invention are: 1) Based on the long-term accumulated operation data of multiple plant areas, a pre-trained model with cross-scenario generalization ability is constructed, and the transfer learning mechanism enables the model to accurately adapt to the characteristics of specific incineration plant areas; in this way, the model can quickly learn the operation mode of new plant areas with the support of a very small amount of target plant area data, thereby greatly reducing the model deployment cost and improving the cross-plant area adaptation speed and prediction stability.

[0024] 2) By embedding physical conservation relationships into the neural network structure in a soft constraint manner, the model not only relies on data fitting, but also reflects the basic physical laws of the combustion system, thereby fundamentally improving prediction accuracy and stability.

[0025] 3) The proposed carbon-pollution synergistic risk index integrates pollutant toxicity, emission limit stringency, and regulatory priority through weighted aggregation, enabling the synergistic hazards among multiple pollutants to be presented in a unified index format. Based on model-based predictions of pollutant concentrations, this index provides incineration plant operators with more intuitive risk information, thereby assisting them in adjusting airflow ratios and controlling loads, thus forming a more scientific and refined emission control path. Attached Figure Description

[0026] To more clearly illustrate the technical solution of the present invention, this specification provides further explanation of the structure and method flow of the present invention in conjunction with the accompanying drawings. It should be noted that these drawings are only used to illustrate the technical concept of the present invention and are not intended to limit the scope of protection of the present invention. The functional modules or data flows shown in the drawings can be equivalently replaced or functionally combined without changing the core principles.

[0027] Figure 1 This is a schematic diagram of a method for intelligent prediction and migration control of key indicators of a solid waste incineration system provided in an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of a hybrid expert model structure for a solid waste incineration system provided in an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram illustrating the embedding of physical information constraints in a solid waste incineration system, provided as an embodiment of the present invention.

[0030] Figure 4 This is a schematic diagram of a cross-plant migration learning process provided in an embodiment of the present invention.

[0031] Figure 5 This is a schematic diagram illustrating the composition principle of a carbon-pollution synergistic comprehensive risk index provided in an embodiment of the present invention. Detailed Implementation

[0032] To make the technical solution of this invention clearer, the method flow, network structure, physical constraint construction, and migration process of this invention are described in detail below with reference to specific embodiments. It should be noted that the embodiments described herein are only for explaining this invention and are not intended to limit the scope of protection of this invention. For those skilled in the art, appropriate adjustments and modifications can be made based on the methods provided herein without departing from the core ideas of this invention.

[0033] In this embodiment, the data for the solid waste incineration system comes from a typical municipal solid waste incineration power plant, and includes steam flow rate, furnace temperature at nine o'clock, primary air volume, secondary air volume, flue gas temperature, flue gas O2 content, and emission concentrations of various pollutants (CO, CO2, HCl, NO). x The system collects multi-dimensional real-time operating parameters such as SO2 and PM. These data are accumulated over a long period using a sampling method of 1 data point per hour, providing a sufficient data foundation for constructing the intelligent prediction and transfer learning model described in this invention. Figure 1 This is a schematic diagram illustrating the intelligent prediction and migration control method for key indicators of a solid waste incineration system provided by an embodiment of the present invention. The diagram, from left to right, shows the overall process of collecting baseline plant data, building a pre-trained model, introducing physical constraints, adapting through transfer learning, and finally predicting and controlling the output. The diagram demonstrates the design concept of parallel integration of data-driven and physical constraints, and also reflects the hierarchical relationship of cross-plant migration paths, giving the entire method framework a coherent logical process from a macroscopic perspective. Specifically, it includes the following steps:

[0034] 1) First, the long-term collected data from each plant area needs to be cleaned, aligned, and normalized to remove obviously abnormal sensor drift points and data from shutdown periods. Analysis of data feature distribution reveals significant differences between different plant areas in terms of load variation, air distribution strategies, waste composition ratios, and flue gas treatment system efficiency. These differences constitute the main source of "cross-plant heterogeneity" in this invention. To construct an intelligent prediction model with universality and transferability, this invention first trains a hybrid expert model based on data from a benchmark plant area, enabling the model to grasp the core patterns of multi-pollutant emissions during solid waste incineration in the initial stage.

[0035] 2) In the construction of the baseline model, the input layer uses steam flow rate, primary air volume, secondary air volume, furnace multi-point temperature, flue gas temperature, flue gas O2 content, and emission concentrations of various pollutants as feature vectors. The structure is transformed into a hybrid expert structure through multi-layer nonlinear transformation, as shown in equation (1):

[0036] (1)

[0037] In the formula, M represents the number of prediction tasks / the number of pollutant types, f e (·) represents the mapping of the e-th expert subnetwork, θ e For the parameters of the e-th expert, Let be the prediction vector for each pollutant by the e-th expert.

[0038] Figure 2 This diagram illustrates a hybrid expert model (MoE) structure for a solid waste incineration system, as provided in an embodiment of the present invention. The three expert networks shown are merely illustrative; other numbers of expert networks can be designed according to specific scenarios. The diagram illustrates the relationship between the gating network and multiple expert sub-networks from a network hierarchy perspective. Different expert networks correspond to the response patterns of pollutant emissions under different operating conditions, while the gating network is responsible for automatically selecting the appropriate expert combination based on real-time input characteristics. Furthermore, this diagram demonstrates the importance of the MoE framework in handling the variable operating conditions of incineration systems and highlights its advantages in predicting complex nonlinear pollutants.

[0039] Different expert networks maintain a consistent structure, but their learned parameters focus on the pollutant generation characteristics corresponding to different operating conditions. For example, some experts capture NO under high-load operation during training. xDue to the rapid increase in the number of experts, another group of experts is more suitable as a predictive sub-model for CO concentration changes under low-oxygen combustion conditions. The outputs of these expert networks are weighted and aggregated by a gating network to form the final prediction result, enabling the model to flexibly adapt to complex and variable operating conditions. As shown in equations (2) and (3), the gating network judges the input features and automatically selects the expert combination that is closest to the current operating condition, thereby making the prediction result more stable and adaptable.

[0040] (2)

[0041] (3)

[0042] In the formula, g(·) is the gated network, ϕ is the gated network parameter, and a t For gating the scores (logits) of each expert, π t =[π t,1 ,…, π t,E ] T For expert weights ( The superscript T indicates the transpose operation, E is the number of expert networks, and a t,e Let π be the score given by the e-th expert at time t. t,e The weight of the e-th expert is selected for time t, which is the expert fit degree under the corresponding working condition.

[0043] Taking the benchmark model training conducted at a large-scale municipal solid waste incineration plant as an example, the hybrid expert model demonstrated high accuracy in predicting multiple pollutants. As shown in Table 1, PM, SO2, NO... x The prediction results for HCl, CO, and CO2 showed low mean absolute error (MAE) and root mean square error (RMSE), ranging from 0.0167 to 1.773 mg / Nm³. 3 and 0.0259-3.95 mg / Nm 3 and a high coefficient of determination (R²) 2 =0.74-0.97). Among them, NO x The R² values ​​for CO2 and CO2 are 0.97 and 0.96, respectively, indicating that the model performs particularly well in predicting the emissions of these two components. The overall model has an average R² of 0.8582, demonstrating strong predictive ability and the ability to accurately predict the emissions of different pollutants, providing effective support for the regulation and optimization of the incineration process.

[0044] Table 1: Prediction Results of the Benchmark Hybrid Expert Model

[0045] PM concentration 0.0167 0.0259 0.7435 <![CDATA[SO2 concentration]]> 0.2281 0.3145 0.8814 <![CDATA[NO x Concentration 0.1307 0.1984 0.9701 HCl concentration 0.1598 0.2370 0.8524 CO concentration 1.7733 3.9479 0.7376 <![CDATA[CO2 content]]> 0.0275 0.1112 0.9641 Overall average R² - - 0.8582

[0046] 3) Building upon this, the present invention introduces a physical information neural network structure, superimposing physical constraint terms on top of the data fitting loss. This ensures that the model does not deviate from the fundamental physical laws of the incineration process, maintaining physical rationality, especially under conditions of data scarcity, sudden changes in operating conditions, or inter-plant migration. See also Figure 3 The figure shows three parallel physical constraint paths representing three types of constraints: oxygen balance, total pollution load conservation, and energy balance. These, along with the data fitting loss, constitute the total loss function for model training. The attached figure visually illustrates the process of integrating physical constraints into the deep network in a soft constraint manner, demonstrating how these constraints influence the model's hidden layers and output results during training, ensuring that the prediction results conform to data patterns without deviating from basic physical logic.

[0047] Taking oxygen balance as an example, the equivalent excess air coefficient on the input side can be constructed by the primary air volume, secondary air volume, and steam flow rate. As shown in equation (4):

[0048] (4)

[0049] In the formula, Q 1,i Q 2,i Total primary / secondary air volume in the i-th hour; Q s,i k is the steam flow rate. AF As a trainable scalar parameter, it serves as a proportional constant for the theoretical air volume required per unit of steam volume. It can be shared across plants or embedded between different plants.

[0050] Meanwhile, as shown in equation (5), the excess air coefficient on the output side is calculated by using the O2 content of the flue gas. This makes the deviation between the two (which can be obtained from equations (6) and (7)) a soft constraint for network training, thereby avoiding the prediction result of "impossible burning state" in the model.

[0051] (5)

[0052] (6)

[0053] (7)

[0054] In the formula, It is the dry basis volume fraction of O2 in the flue gas at that moment (%). For oxygen balance residual, This is due to the loss of oxygen balance.

[0055] The total pollution load conservation constraint estimates the outlet pollutant mass flux by weighting the concentrations of each pollutant and combining this with the flue gas velocity, ensuring that the mass flux is close to the input pollutant flux determined by the load. Specifically, it assumes that the flue gas volumetric flow rate is proportional to the flue gas velocity, and the flue cross-sectional area A... duct If we consider it as a constant (trainable), then we can obtain equation (8):

[0056] (8)

[0057] In the formula, Flue gas volumetric flow rate (Nm³) 3 / s), The cross-sectional area of ​​the flue (m²) 2 ), The flue gas velocity (m / s) that can be measured within the system.

[0058] Construct a total pollution concentration index by weighting all pollutant concentrations according to their respective weights. As shown in equation (9):

[0059] (9)

[0060] In the formula, k iterates through PM, SO2, PCDD / Fs, NOx, HCl, CO, CO2; a k These are trainable weights, representing the contribution of different pollutants to the total elemental load. This represents the real-time concentration of each pollutant at that moment.

[0061] From this, the total pollution output load can be obtained. As shown in equation (10):

[0062] (10)

[0063] In the formula, b poll It is a trainable scalar based on the molecular weight and unit conversion of different pollutants.

[0064] Furthermore, assuming that the flux of pollutants entering the incinerator per unit steam flow rate is approximately constant, the total pollutant input load can be obtained from equation (11). :

[0065] (11)

[0066] In the formula, c poll It is a trainable scalar, that is, the comprehensive input of Cl+S+N elements corresponding to a unit of steam; Q s,i Main steam flow rate (kg / s).

[0067] Therefore, the residual of total pollution load conservation can be obtained. and losses As shown in equations (12) and (13) respectively:

[0068] (12)

[0069] (13)

[0070] The energy conservation constraint establishes a relative balance between input and output energy based on steam flow rate, primary / secondary air sensible heat, and furnace average temperature, preventing unreasonable temperature trends in the model under abnormal operating conditions. Specifically, as shown in equation (14), steam flow rate and primary / secondary air sensible heat are used as energy inputs. Approximation:

[0071] (14)

[0072] In the formula, T 1,i , T 2,i Primary / secondary air temperature (K); α s α1 and α2 are trainable scalars, constants such as specific heat absorption and efficiency, and can also be set to different values ​​by embedding according to the factory.

[0073] Let the weighted average of the temperatures at J points in the furnace be the average temperature of the furnace. As shown in equation (15):

[0074] (15)

[0075] In the formula, This represents the temperature (K) at various points in the furnace.

[0076] Simultaneously utilizing the flue gas velocity Temperature after the reaction tower Dust collector outlet temperature etc., construct simplified output energy As shown in equation (16):

[0077] (16)

[0078] In the formula, , , It is a trainable scalar.

[0079] Therefore, the residual of energy conservation can be obtained. and losses As shown in equations (17) and (18) respectively:

[0080] (17)

[0081] (18)

[0082] In summary, PINN's The total loss can be constructed using equation (19):

[0083] (19)

[0084] In the formula, , , These are the hyperparameters that are tuned during training.

[0085] The physical constraints described in this invention are all incorporated into the training process using soft constraints, avoiding common problems in rigorous mechanistic models such as difficulty in determining parameters and model convergence failure. This allows physical information and data-driven models to be naturally integrated in this invention. After PINN constraints, the baseline model prediction results are shown in Table 2, demonstrating a more stable improvement in performance.

[0086] Table 2: Prediction results of the baseline model after PINN constraints

[0087] PM concentration 0.0150 0.0225 0.7550 <![CDATA[SO2 concentration]]> 0.2150 0.3001 0.8932 <![CDATA[NO x Concentration 0.1205 0.1841 0.9752 HCl concentration 0.1452 0.2055 0.8750 CO concentration 1.7010 3.4211 0.7451 <![CDATA[CO2 content]]> 0.0255 0.0925 0.9700 Overall average R² - - 0.8689

[0088] 4) After completing the training of the benchmark plant area model, see [link / reference]. Figure 4 This invention enters the cross-plant transfer learning stage. Due to significant differences in data distribution, operational logic, and combustion characteristics across different plants, directly applying the baseline model to a new plant can easily lead to prediction bias. This invention addresses this problem through two types of transfer learning methods:

[0089] First, parameter fine-tuning: keeping the deep structure of the model unchanged, only fine-tuning the high-level layers or parts of the expert network, allowing the model to adapt more quickly to the feature distribution of the target plant area. Second, constructing plant area embedding vectors: by learning the latent features of different plants, the model automatically switches to the data space suitable for that plant area when receiving input from the target plant area, thus maintaining structural consistency and prediction stability during cross-plant area migration. The plant area embedding vector is a set of trainable low-dimensional vector parameters that correspond one-to-one with specific incineration plants, used to characterize the comprehensive differences between plants in terms of furnace structure, air distribution strategy, waste composition, flue gas treatment system, and operation management strategy. Each plant area corresponds to an independent embedding vector, which is updated synchronously with the network weights during model training and automatically learns the latent representations reflecting the differences between plants through backpropagation. In practical implementation, the plant area embedding vector can be used as an additional feature and input into the model along with the original operating parameters, or injected into the intermediate layer of the hybrid expert model to modulate the weight allocation of the expert network output or the gating network. This allows the model to automatically adjust the calling strategy of each expert sub-model according to the value of different plant area embedding vectors while keeping the overall structure unchanged.

[0090] Taking the aforementioned incineration plant as an example, the prediction results for the baseline and target plant areas were compared. After fine-tuning with a small sample dataset of the target plant area, the prediction results for the main pollutants are shown in Table 3, with an overall average R... 2 The value of 0.7728 indicates that the model still has good predictive ability after transfer learning in the target factory area.

[0091] Table 3: Relocation Prediction Results of the Target Plant Area

[0092] PM concentration 0.0535 0.0792 0.4906 <![CDATA[SO2 concentration]]> 0.2713 0.3726 0.8485 <![CDATA[NO x Concentration 0.1664 0.3143 0.8682 HCl concentration 0.1739 0.2406 0.7889 CO concentration 1.6836 3.3864 0.6770 <![CDATA[CO2 content]]> 0.0299 0.1027 0.9639 Overall average R² - - 0.7728

[0093] 5) After the model achieves stable predictive ability, see [link to relevant documentation]. Figure 5 This invention further incorporates a carbon-pollution synergistic risk index, enabling the complex synergistic effects among multiple pollutants to be represented by a unified indicator. This index aggregates real-time predicted values ​​of various pollutants and combines them with pollutant toxicity, the stringency of national emission limits, and regulatory priority to form a weighted structure. Specifically, it assumes six major pollutants (in this embodiment, these correspond to CO, CO2, HCl, NO, etc.) x The predicted values ​​C1, C2, ..., C6 for (SO2, PM2.5) are denoted by T1, T2, ..., T6, L1, L2, ..., L6, and R1, R2, ..., R6, respectively. CPSI can be calculated using equation (20):

[0094] (20)

[0095] In the formula, These are the weighting coefficients for each pollutant, defined according to equation (21):

[0096] (twenty one)

[0097] In the formula, , ,and It is an adjustable weighting coefficient used to control the contribution of toxicity, emission limits, and regulatory priorities to the risk index.

[0098] Therefore, this indicator structure can not only intuitively present the overall pollution risk level of a plant area at any time, but also reveal the direction of the impact of air volume adjustment, furnace temperature change, load change and other factors on the risk level through sensitivity analysis, providing a direct basis for on-site control strategies.

[0099] In engineering applications, this invention can be deployed in the real-time monitoring system of an incineration plant. By interfacing with the DCS system, the model can continuously receive the latest operational data and generate real-time pollutant prediction curves. Simultaneously, the CPSI index can be presented as a trend line or color-coded levels on the control room interface, enabling operators to take preventative measures before pollution trends worsen. These measures include timely adjustments to the primary and secondary air ratios, optimizing the feeding rhythm, or adjusting ammonia injection rates, ensuring the incineration system operates within a safe and environmentally friendly timeframe.

[0100] In a further implementation of the present invention, the model can be continuously learned according to the needs of the site, so that it can be continuously updated when new working conditions, new materials or new operating rules appear, so as to maintain the effectiveness of long-term prediction capability.

[0101] In summary, firstly, the model pre-trained based on multi-plant data significantly improves cross-scenario prediction capabilities, enabling rapid adaptation to new plants with minimal data, fundamentally solving the problem of traditional models' strong dependence on specific plants. Secondly, by embedding physical constraints into model training, the prediction framework constructed in this invention overcomes the logical deviation problem that may occur in deep learning models under extreme conditions, making the prediction results more reliable under different operating conditions. Thirdly, because the hybrid expert structure can capture the differentiated patterns of pollutant emissions under different operating conditions, this invention significantly outperforms traditional regression models and single-expert models in terms of accuracy in multi-pollutant collaborative prediction. Finally, the carbon-pollution synergistic comprehensive risk index proposed in this invention provides a scientific basis for optimizing the operation of the incineration process, enabling operators to obtain clearer risk assessments when facing complex emission characteristics, thereby effectively improving emission control efficiency and reducing environmental risks. This invention not only achieves systematic breakthroughs in cross-plant generalization ability, physical reliability, multi-pollutant collaborative prediction, and risk quantification, but also provides a new technical path for the construction of intelligent incineration plants, possessing significant application prospects and promotional value.

[0102] The above embodiments are not intended to limit the present invention. Based on the core technical ideas of the present invention, those skilled in the art can make various modifications, improvements or extensions, including more complex expert network structures, more refined physical constraint construction methods, different transfer learning strategies, etc. As long as they do not deviate from the spirit of the present invention, they should be considered within the scope of protection of the present invention.

Claims

1. An intelligent prediction and migration control method for key indicators of solid waste incineration system, applied to solid waste incineration process to monitor data, characterized in that, Includes the following steps: 1) Collect historical operating data of the benchmark incinerator; the operating data includes steam flow rate, furnace temperature, primary air volume, secondary air volume, flue gas temperature, flue gas O2 content, and emission concentration of multiple pollutants; 2) The multi-task hybrid expert model is pre-trained using the historical operating data of the benchmark incinerator. The model is jointly optimized using data fitting loss and physical constraints of the physical information neural network. The physical constraints include at least oxygen balance constraints, total pollution load conservation constraints, and energy balance constraints. The physical constraints are all incorporated into the training in a soft constraint manner. The oxygen balance constraint is constructed by the residual between the input-side equivalent excess air coefficient calculated based on the primary air volume, secondary air volume and steam flow rate and the output-side excess air coefficient calculated based on the dry basis volume fraction of flue gas O2. The total pollution load conservation constraint constructs a comprehensive pollution load by weighting pollutant concentrations and estimating flue gas volume flow rate using flue gas velocity, thereby constructing the residual between the input-side pollution load and the output-side pollution load. The energy balance constraint is formed by the input energy of steam flow rate, primary air and secondary air sensible heat, and soft constraint residuals are established by the output energy of the furnace average temperature, flue gas velocity, reaction tower temperature and dust collector outlet temperature. The multi-task hybrid expert model includes multiple expert networks and a gating network. The gating network automatically selects the most suitable expert network for pollutant prediction based on different operating conditions. 3) Transfer learning is performed on the target incinerator. The pre-trained model is adapted to the domain or fine-tuned by the operating data of the target furnace area, so that the model can be adapted to the operating characteristics of different incineration plant areas. The transfer learning includes any or a combination of parameter freezing transfer, partial layer fine-tuning, or constructing plant area embedding vectors for plant area differences. The plant area embedding vector is a set of trainable low-dimensional vector parameters that correspond one-to-one with specific incineration plant areas. It is used as an additional feature and is input into the model along with the original operating parameters, or injected into the intermediate layer of the hybrid expert model to modulate the output of the expert network or the weight allocation of the gating network. 4) Real-time prediction of multiple pollutant emission indicators in the target furnace area based on the migrated model.

2. The method of claim 1, wherein, The oxygen balance constraint in step 2) uses the residual between the equivalent excess air coefficient calculated based on the primary air volume, secondary air volume, and steam flow rate, and the excess air coefficient calculated based on the dry basis volume fraction of O2 in the flue gas, as a physical constraint term; specifically: In the formula, Q 1,i Q 2,i Q represents the total primary / secondary air volume in the i-th hour; s,i k is the steam flow rate. AF It is a trainable scalar parameter, serving as a proportional constant for the theoretical air volume required per unit of steam; To construct the equivalent excess air coefficient on the input side using primary air volume, secondary air volume, and steam flow rate, The excess air coefficient, It is the dry basis volume fraction of O2 in the flue gas during the i-th hour. For oxygen balance residual, This is due to the loss of oxygen balance.

3. The method of claim 1 wherein, The total pollution load conservation constraint constructs a comprehensive pollution load by weighting pollutant concentrations and estimating flue gas volumetric flow rate using flue gas velocity, ensuring consistency between the input and output pollution loads. Specifically: The flue gas volume flow is proportional to the flue gas flow rate, the flue cross-sectional area A duct Considered constant and trainable, then: wherein is the flue gas volumetric flow rate, is the flue gas cross-sectional area, is the flue gas flow rate available from measurements within the system; All pollutant concentrations are weighted by weight to construct the total pollution concentration index : Thus the total pollution output load : In the formula, b poll is a trainable scalar based on different pollutant absorption molecular weights and unit conversion; The flux of pollution elements entering the incinerator corresponding to the unit steam flow is constant, and the total pollution input load is obtained : where c poll is a trainable scalar, Q s,i is the steam flow rate; total pollution load conservation residual and losses : 。 4. The method of claim 1 wherein, The energy balance constraint is established by using steam flow rate, primary air and secondary air sensible heat as input energy, and using the output energy, which consists of furnace average temperature, flue gas velocity, reaction tower outlet temperature and dust collector outlet temperature, to create a soft constraint residual. Specifically: With steam flow and primary / secondary air sensible heat as energy input : In the formula, T 1,i , T 2,i Primary / secondary air temperature; α s α1, α2 are trainable scalars, and Q is a trainable scalar. 1,i Q 2,i Q represents the total primary / secondary air volume in the i-th hour; s,i Steam flow rate; Let the weighted average of the temperatures at J points in the furnace be the average temperature of the furnace. Simultaneously utilize the flue gas velocity Temperature after the reaction tower Dust collector outlet temperature Construct output energy : wherein , , is a trainable scalar; Thus, the energy-conserving residual and loss : 。 5. The method of claim 1 wherein, The multi-task hybrid expert model includes multiple expert networks and a gating network. The gating network automatically selects the most suitable expert network for pollutant prediction based on different operating conditions. In the formula, g(·) is a gated network. For gating network parameters, Let a be the feature vector of the running data. t To control the scoring of each expert, π t =[π t,1 ,…, π t,E ] T For expert weighting, The superscript T indicates the transpose operation, E is the number of expert networks, and a t,e Let π be the score given by the e-th expert at time t. t,e The weight of the e-th expert is selected for time t, which is the expert fit degree under the corresponding working condition.

6. The method of claim 1 wherein, After real-time prediction of multiple pollutant emission indicators of the target furnace area based on the migrated model, it also includes calculating the carbon-pollution synergistic comprehensive risk index to realize intelligent prediction and risk assessment of pollutant emissions across plant areas. The carbon-pollution synergistic risk index is constructed based on the concentration of multiple pollutants, the toxicity weight of pollutants, the stringency of emission limits, and the regulatory priority, and is obtained by weighted aggregation.

7. The method of claim 6 wherein, The pollutant weights in the carbon-pollution synergistic risk index are obtained by combining pollutant toxicity, emission limit stringency, and regulatory priority using weighting coefficients.

8. The method of claim 1 wherein, The physical constraint loss and data fitting loss are weighted by adjustable hyperparameters to guide the model to achieve a balance between physical consistency and prediction accuracy.

Images