Waste incineration combustion condition prediction method integrating rationalization mechanism and data driving

Abstract

The invention relates to the technical field of urban solid waste energy treatment, in particular to a waste incineration combustion condition prediction method and device based on a combined rationalization mechanism and data driving. The method comprises the steps that on the basis of a physical and chemical mechanism involved in the waste incineration process, characteristic values are collected, and a chain type model of a multi-level chain type characteristic system is constructed; constructing a new characteristic value set and a new target value set which reflect dynamic changes; and for each mechanism reaction unit in the multi-level chain type characteristic system, independently training a hybrid model fused with physical constraints, training the hybrid model under physical mechanism constraints based on the pre-processed new characteristic value set and new target value set, and constructing a chain type prediction model for predicting the garbage incineration combustion condition. According to the method, the physical and chemical laws of the incineration process are embedded into the data driving model, accurate prediction of the key parameters of the waste incineration working condition is achieved, and the method is suitable for optimization control of the waste incineration power generation process.

Description

Technical Field

[0001] The present invention relates to the technical field of urban solid waste energy treatment, and more specifically, to a method and device for predicting the combustion conditions of garbage incineration that integrates physical and chemical mechanisms with data-driven methods. Background Art

[0002] Waste-to-energy technology, a comprehensive waste treatment method that combines waste reduction, harmlessness, and resource utilization, has become the mainstream process for the harmless disposal of domestic waste in my country's cities. This technology decomposes toxic and hazardous substances through high-temperature incineration and recovers the resulting heat for power generation, offering significant environmental and energy benefits.

[0003] However, the waste incineration power generation process is a complex, continuous process involving the coexistence of gas, solid, and liquid phases, with significant nonlinear characteristics and large hysteresis. This poses a serious challenge to the optimal control of the incineration process. In this context, accurately predicting the combustion conditions of the incinerator has become a key prerequisite for improving the control level of the incineration process.

[0004] Combustion condition prediction is essentially a time series prediction problem. Its core lies in predicting the changing trends of key parameters in the future period through analysis and modeling of historical operating data. At present, the prediction method based on artificial neural network has become the mainstream technical solution in this field due to its powerful nonlinear mapping ability and adaptive learning characteristics. Traditional neural network derivative models, such as back propagation (BP) neural network and long short-term memory (LSTM) neural network, have been widely used in the time series prediction of key combustion parameters. However, this type of model has inherent limitations in capturing long-term time series dependencies and is difficult to effectively deal with the large lag characteristics of the incineration process.

[0005] In recent years, with the rapid development of artificial intelligence (AI), the Informer and Autoformer prediction models based on attention mechanisms, as well as the TimesNet model based on frequency domain analysis, have become research hotspots in the field of time series forecasting. The first two models, by modeling global data trends, have demonstrated outstanding performance in long-term time series forecasting tasks, particularly in areas such as power load forecasting and meteorological data forecasting. The TimesNet model, with its ability to extract multi-scale periodic features, demonstrates a strong advantage in scenarios with pronounced periodicity, such as power / heat load forecasting and sales forecasting.

[0006] It should be pointed out that both the classic LSTM model and emerging advanced models such as Informer, Autoformer, TimesNet, etc. are pure data-driven sequence-to-sequence (Seq2Seq) prediction methods, and their performance is highly dependent on the adequacy and quality of training data.

[0007] This type of method has the following two defects: on the one hand, model training requires a large amount of high-quality training data, while data collection in actual engineering often faces technical difficulties such as high cost, long cycle, and data missing; on the other hand, the model training process lacks the constraints of physical and chemical laws, which may lead to prediction results violating basic physical and chemical laws and insufficient generalization ability for abnormal working conditions.

[0008] Therefore, there is an urgent need for a new waste incineration combustion condition prediction technology to overcome the shortcomings of pure data-driven methods that rely heavily on data and lack physical mechanism constraints. Summary of the Invention

[0009] The purpose of the present invention is to provide a method and device for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms with data-driven methods, so as to solve the problems of strong data dependence and lack of physical constraints in the existing waste incineration combustion condition prediction schemes, and effectively improve the accuracy of the prediction results and engineering applicability.

[0010] To achieve the above objectives, the present invention provides a method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms with data-driven methods, comprising the following steps:

[0011] Step S1: Based on the physical and chemical mechanisms involved in the waste incineration process, characteristic values ​​are collected and a chain model of a multi-level chain characteristic system is constructed. The multi-level chain characteristic system is a modular architecture formed by disassembling the waste incineration process into a plurality of mechanism reaction units at multiple levels. The prediction results of the previous level are used as input features of the subsequent level. Each mechanism reaction unit includes a characteristic value set and a target value set, and the relationship between the characteristic value set and the target value set satisfies a mapping function driven by the physical and chemical mechanism.

[0012] Step S2: Preprocessing the eigenvalue data and target value data based on the dynamic and non-steady-state characteristics of the waste incineration combustion condition, constructing a new eigenvalue set and a new target value set that reflect the dynamic changes, wherein the relationship between the new eigenvalue set and the new target value set satisfies a mapping function driven by a physical and chemical mechanism, and performing data normalization processing on the new eigenvalue set and the new target value set;

[0013] In step S3, for each mechanism reaction unit in the multi-level chain feature system, a hybrid model integrating physical constraints is independently trained. Based on the preprocessed new feature value set and the new target value set, the hybrid model is trained under the physical mechanism constraints to construct a chain prediction model, which is used to predict the combustion conditions of waste incineration.

[0014] In some embodiments, step S1 further comprises:

[0015] Based on the transformation laws of matter and energy during waste incineration, the complex heat and mass transfer process of waste incineration is decomposed into multiple levels. The characteristic value data and target value data corresponding to each process are classified into corresponding levels. The multiple levels are progressive functional modules that reflect the incineration process from the actuator to the system level output. Each level corresponds to the heat and mass transfer behavior of a specific process.

[0016] For each level, several mechanism reaction units are constructed according to the different physical and chemical mechanisms involved in the waste incineration process;

[0017] The mechanism reaction units in different levels are connected through a modular architecture, and the output of the previous level drives the input of the subsequent level to build a chain model.

[0018] In some embodiments, the multi-level chain feature system constructed in step S1 includes at least three levels, including a first level, a second level, and a third level:

[0019] At the first level, the operating parameters of the actuator are used as characteristic values, and the generated material flow rate is used as the target value;

[0020] The second level uses part or all of the target values ​​predicted by the first level and other independent combustion process parameters as characteristic values, and uses the thermal parameters of the combustion process as target values;

[0021] The third level uses part or all of the target values ​​predicted by the first and second levels, combined with other independent system operating parameters as characteristic values, and uses the final system-level thermal parameters and emission parameters as target values.

[0022] In some embodiments, the first level is an air volume prediction model, which uses the frequency and wind pressure of each wind turbine as characteristic values ​​and the corresponding air volumes as target values;

[0023] The second level is the wind temperature prediction model, which uses the air volumes predicted by the first level, as well as the heat exchanger inlet and outlet flue gas temperatures and the air inlet temperatures as characteristic values, and the corresponding air outlet temperatures as target values;

[0024] The third level is the furnace temperature and flue gas volume prediction model, which uses the air volumes predicted by the first level, the air temperatures at each air outlet predicted by the second level, the amount of garbage entering the furnace and the amount of fuel for the auxiliary burner as characteristic values, and the temperatures in different areas of the incinerator and the flue gas parameters at the furnace outlet as target values.

[0025] In some embodiments, the step of preprocessing the eigenvalue data in step S2 further includes:

[0026] Calculate the first derivative of each eigenvalue with respect to time to obtain the rate of change of the eigenvalue with time f t , the rate of change f t Used to characterize the dynamic evolution trend of characteristic value parameters during the unsteady-state combustion process of waste incineration;

[0027] The original eigenvalue F t The rate of change of the characteristic value with time f t Combined with the above, a new eigenvalue set is constructed to reflect the static value and dynamic change trend of the eigenvalue {F t ,f t}.

[0028] In some embodiments, the step of preprocessing the target value data in step S2 further includes:

[0029] Calculate the difference between the value of each target value at the next time node and the value at the current time node, and use the difference to characterize the dynamic evolution trend of the target value parameter during the unsteady-state combustion process of waste incineration;

[0030] The difference is constructed as a target value difference set and used as a new target value set ΔT t , which is used to reflect the dynamic change characteristics of the target value in time series.

[0031] In some embodiments, the hybrid model is a neural network model that integrates physical prior knowledge, and the model training is performed based on the new feature value set and the new target value set preprocessed in step S2 as training data;

[0032] The neural network model integrating physical prior knowledge also includes a physical information long short-term memory network or a physical information informer model to achieve long time series prediction tasks.

[0033] In some embodiments, step S1 further comprises:

[0034] Collecting original eigenvalue data from the distributed control system of the waste incineration system;

[0035] Outlier detection is performed on the original eigenvalue data, and data points that are out of the set range are replaced to complete data cleaning.

[0036] In some embodiments, the step of performing outlier detection on the feature value data and replacing data points that are outside a set range further includes:

[0037] The interquartile range method was used to detect and clean outliers, and the cubic spline interpolation method was used to smooth and replace data points that exceeded the set range [Q1-1.5IQR, Q3+1.5IQR].

[0038] Among them, Q1 is the first quartile of the eigenvalue set, Q3 is the third quartile of the eigenvalue set, and IQR is the interquartile range.

[0039] In some embodiments, the data normalization step further comprises:

[0040] The new feature value set and the new target value set reflecting the dynamic changes are normalized by using a minimum-maximum normalization, a Z-Score normalization or a robust normalization method.

[0041] In some embodiments, the step of training the hybrid model under physical mechanism constraints further includes:

[0042] Construct a dual-objective loss function for model training;

[0043] The dual-objective loss function Loss is expressed as the weighted sum of the data loss term and the physical loss term, and the corresponding expression is:

[0044] Loss=w D L D +w P L P

[0045] Among them, L D is the data loss term, which is used to measure the deviation between the model prediction value and the true value;

[0046] L P are data loss terms, which are used to measure the deviation between the calculated value and the true value under the constraints of physical characteristics;

[0047] w D With w P are the weights of data loss item and physical loss item respectively.

[0048] In some embodiments, the step S3 further includes: optimizing the weight coefficients in the dual-objective loss function using an adaptive moment estimation optimizer.

[0049] To achieve the above objectives, the present invention provides a device for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms with data-driven methods, comprising:

[0050] a memory for storing instructions executable by the processor;

[0051] A processor is configured to execute the instructions to implement the above method.

[0052] In order to achieve the above object, the present invention provides a computer-readable medium having computer instructions stored thereon, wherein when the computer instructions are executed by a processor, the above method is executed.

[0053] The present invention provides a method and device for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms with data-driven methods. By embedding the physical and chemical laws of the incineration process into a data-driven model, accurate prediction of key parameters of the waste incineration combustion conditions is achieved, which is suitable for the optimized control of the waste incineration power generation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The above and other features, properties and advantages of the present invention will become more apparent from the following description taken in conjunction with the accompanying drawings and embodiments, in which like reference numerals denote like features throughout, wherein:

[0055] Figure 1 A step diagram of a method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to an embodiment of the present invention is disclosed;

[0056] Figure 2 A flow chart of a method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to an embodiment of the present invention is disclosed;

[0057] Figure 3 A schematic diagram of a simplified model of a combustion condition of a waste incinerator according to an embodiment of the present invention is disclosed;

[0058] Figure 4 A schematic diagram of a multi-level chain characteristic system of a waste incinerator combustion condition according to an embodiment of the present invention is disclosed. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the invention and are not intended to limit the invention.

[0060] In response to the shortcomings of the existing technology, the present invention proposes a method and device for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms with data-driven methods. By embedding the physical and chemical laws of the incineration process into a data-driven model, the method and device can be accurately predicted for the key parameters of the waste incineration combustion conditions. The method is suitable for the optimized control of the waste incineration power generation process and can be widely used in complex combustion systems such as waste incineration power generation and industrial boilers, and has significant engineering application value.

[0061] Figure 1 The following discloses a method for predicting the combustion conditions of waste incineration by integrating physical and chemical mechanisms with data-driven methods according to an embodiment of the present invention. Figure 1 As shown, the present invention proposes a method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven, including the following steps:

[0062] Step S1: Based on the physical and chemical mechanisms involved in the waste incineration process, characteristic values ​​are collected and a chain model of a multi-level chain characteristic system is constructed. The multi-level chain characteristic system is a modular architecture formed by disassembling the waste incineration process into a plurality of mechanism reaction units at multiple levels. The prediction results of the previous level are used as input features of the subsequent level. Each mechanism reaction unit includes a characteristic value set and a target value set, and the relationship between the characteristic value set and the target value set satisfies a mapping function driven by the physical and chemical mechanism.

[0063] Step S2: Preprocessing the eigenvalue data and target value data based on the dynamic and non-steady-state characteristics of the waste incineration combustion condition, constructing a new eigenvalue set and a new target value set that reflect the dynamic changes, wherein the relationship between the new eigenvalue set and the new target value set satisfies a mapping function driven by a physical and chemical mechanism, and performing data normalization processing on the new eigenvalue set and the new target value set;

[0064] In step S3, for each mechanism reaction unit in the multi-level chain feature system, a hybrid model integrating physical constraints is independently trained. Based on the preprocessed new feature value set and the new target value set, the hybrid model is trained under the physical mechanism constraints to construct a chain prediction model, which is used to predict the combustion conditions of waste incineration.

[0065] The present invention embeds physical prior knowledge such as heat and mass transfer into a data-driven model to construct a chain prediction model enhanced by physical and chemical information, so as to achieve high-precision prediction of the future long-term combustion conditions of the waste incinerator.

[0066] First, based on the physical and chemical mechanisms involved in the waste incineration process, this method breaks down the complex heat and mass transfer process in waste incineration into multiple levels and constructs multiple mechanism reaction units. With the help of the chain model construction method, the data dimension of each mechanism reaction unit model is effectively reduced, significantly improving the training efficiency.

[0067] Secondly, according to the dynamic non-steady-state characteristics of the waste incineration combustion condition, the data in each mechanism reaction unit is pre-processed to calculate the current eigenvalue F t The derivative f with respect to time t t and the original eigenvalue F t , construct a new eigenvalue set {F t ,f t}; At the same time, the target value is calculated at the next time node T t+1 With the current time node T t The change in the target value is constructed by the difference set ΔT t , as the new target value set; then, the newly created feature value set and target value set are normalized.

[0068] Finally, hybrid model training is carried out under the constraints of physical mechanisms. The model in each mechanism reaction unit can be trained, optimized and upgraded independently, and an adaptive weight mechanism is adopted to achieve prediction of higher-level target values ​​based on the chain model; physical property constraints are introduced into the loss function to further enhance the role of physical mechanisms in model training.

[0069] Figure 2 A flow chart of a method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to an embodiment of the present invention is disclosed. The above steps will be described in detail below with reference to the accompanying drawings. Within the scope of the present invention, the above-mentioned technical features of the present invention and the technical features specifically described below (such as the embodiments) can be combined and correlated with each other to form a preferred technical solution.

[0070] Step S1: Based on the physical and chemical mechanisms involved in the waste incineration process, characteristic values ​​are collected and a chain model of a multi-level chain characteristic system is constructed. The multi-level chain characteristic system is a modular architecture formed by disassembling the waste incineration process into a plurality of mechanism reaction units at multiple levels. The prediction results of the previous level are used as input features of the subsequent level. Each mechanism reaction unit includes a characteristic value set and a target value set, and the relationship between the characteristic value set and the target value set satisfies a mapping function driven by the physical and chemical mechanism.

[0071] In this embodiment, raw characteristic value data is collected from the distributed control system (DCS) of the waste incineration system;

[0072] Outlier detection is performed on the original eigenvalue data, and data points that are out of the set range are replaced to complete data cleaning.

[0073] More specifically, the outlier detection and cleaning step further includes:

[0074] The interquartile range (IQR) method was used to detect and clean outliers, and the data points that exceeded the set range [Q1-1.5IQR, Q3+1.5IQR] were smoothed and replaced using the cubic spline interpolation method;

[0075] Wherein, Q1 is the first quartile of the eigenvalue set, Q3 is the third quartile of the eigenvalue set, and IQR is the interquartile range, that is, IQR=Q3-Q1.

[0076] The interquartile range method is a nonparametric outlier identification method based on the quartiles of data distribution. It does not require the assumption that the data follows a specific distribution and has strong resistance to interference from extreme values: first, sort the eigenvalue set from small to large, and calculate the value of the 25% position in the first quartile Q1 data) and the value of the 75% position in the third quartile Q3 data). The difference between the two is the interquartile range.

[0077] IQR (IQR = Q3-Q1); then [Q1-1.5IQR, Q3+1.5IQR] is used as the normal data range, and the data points outside this range are smoothly replaced using the cubic spline interpolation method.

[0078] In addition to the interquartile range method, outlier detection can also use the standard deviation method (based on the data mean and standard deviation, data outside the range of "mean ± k times the standard deviation" is determined as an outlier, k = 3 is commonly used), the box plot method (consistent with the principle of the interquartile range method, visually identifying outliers through graphical methods), the DBSCAN clustering method (based on the density distribution of data points, isolated points in low-density areas are determined as outliers), or the isolation forest algorithm (by constructing multiple isolated trees, samples that are easily isolated are determined as outliers, suitable for large-scale data sets), etc.

[0079] Through the above outlier cleaning, the outliers that appear during the collection process can be effectively eliminated, and finally a high-quality feature data set at different time points t can be obtained. Provide a reliable data foundation for subsequent modeling.

[0080] The construction of the chain model is based on the collected characteristic values ​​and combined with the law of material-energy conversion in the waste incineration process to establish a multi-level chain characteristic system.

[0081] In this embodiment, the construction of the multi-level chain feature system includes:

[0082] Step S11, classifying characteristic values ​​and target values: Based on the transformation law of matter and energy in the waste incineration process, the complex heat and mass transfer process of waste incineration is decomposed into multiple levels, and the characteristic value data and target value data corresponding to each process are classified into corresponding levels. The multiple levels are progressive functional modules reflecting the incineration process from the actuator to the system level output, and each level corresponds to the heat and mass transfer behavior of a specific process;

[0083] Step S12, establishment of mechanism reaction units: for each level, construct several mechanism reaction units according to the different physical and chemical mechanisms involved in the waste incineration process;

[0084] Step S13, chain model construction: connect the mechanism reaction units in different levels through a modular architecture, use the output of the previous level to drive the input of the subsequent level, and build a chain model.

[0085] This multi-level chain feature system is constructed based on the transmission and transformation paths of material flow and energy flow in the waste incineration process, and is abstracted step by step from the bottom-level equipment control parameters to the top-level system performance parameters.

[0086] More specifically, in step S11, the complex heat and mass transfer process of waste incineration is decomposed into at least three levels, including the first level, the second level, and the third level:

[0087] At the first level, the operating parameters of the actuator are used as characteristic values, and the material flow rate generated by the actuator is used as the target value;

[0088] The second level uses part or all of the target values ​​predicted by the first level, combined with other independent combustion process parameters as characteristic values, and uses the thermal parameters of the combustion process as target values;

[0089] The third level uses part or all of the target values ​​predicted by the first and second levels, combined with other independent system operating parameters as characteristic values, and uses the final system-level thermal parameters and emission parameters as target values.

[0090] By breaking down the waste incineration process into different levels according to the heat transfer and mass transfer mechanism, and classifying the sets of characteristic values ​​and target values ​​of each process into different levels, the model data dimension of each mechanism reaction unit is effectively reduced, and the efficiency of model training is greatly improved.

[0091] In the waste incineration process, the key parameters involved can be divided into several types according to their physical meaning and scope of action:

[0092] The operating parameters of the actuator refer to the operating parameters of the equipment that drives the incineration system, such as feeder speed, fan speed, etc.

[0093] Material flow refers to the amount of material or medium transported by the operation of the actuator, such as feed volume, air volume, etc.

[0094] Combustion process parameters refer to independent state parameters that directly participate in the combustion reaction, such as the inlet and outlet flue gas temperatures of the heat exchanger;

[0095] The thermal parameters of the combustion process refer to the parameters that reflect the heat transfer and temperature changes in the combustion process, such as wind temperature;

[0096] Other independent system operating parameters refer to basic system operating parameters that do not depend on the prediction results of the previous level, such as the amount of waste entering the furnace;

[0097] The final system-level thermal parameters and emission parameters refer to the core parameters that reflect the overall operating status and environmental protection indicators of the incineration system, such as main steam temperature, furnace outlet flue gas volume, etc.

[0098] It should be noted that the characteristic values, target value divisions, and corresponding physical and chemical mechanism mappings of each level in the multi-level chain characteristic system can be flexibly configured according to the specific configuration and forecast requirements of the waste incineration system. The following is one implementation example:

[0099] The feeder speed, fan speed, and water pump frequency are the first-level characteristic values, and their corresponding feed rate, air volume, and water supply rate are the first-level target values. The first level focuses on the mapping of actuator operating parameters to material flow (e.g., fan speed to air volume), corresponding to the direct material transportation process.

[0100] Feed rate and air volume can be used as the characteristic values ​​of the second level. According to the heat and mass transfer principles of the combustion process, the furnace temperature and flue gas volume, which are affected by the characteristic values ​​of the second level, can be used as the target values ​​of the second level. The second level predicts the thermal parameters of the combustion state (such as air volume, fuel characteristics, and furnace temperature) based on the material flow rate and independent combustion parameters, corresponding to the combustion reaction and heat release process.

[0101] The target values ​​of furnace temperature and flue gas volume at the second level, and the target value of water supply volume at the first level can be used as the characteristic values ​​of the third level, while the main steam temperature and main steam flow rate, which are affected by parameters such as furnace temperature, flue gas volume and water supply volume, are used as the target values ​​of the third level. The third level integrates the outputs of the previous levels to predict the final system-level thermal and emission parameters (such as steam parameters), corresponding to the energy recovery and overall system performance evaluation process.

[0102] Through this progressive hierarchical division, not only the modular decoupling of complex processes is achieved, but also the input-output relationship of each level is ensured to strictly follow the physical and chemical laws of its link.

[0103] More specifically, in step S12, each mechanism reaction unit includes a characteristic value set and a target value set, and the relationship between the characteristic value set and the target value set satisfies a mapping function driven by a physical and chemical mechanism.

[0104] The establishment of the mechanism reaction unit is based on the data at each level. According to the different physical and chemical mechanisms involved in the waste incineration process, multiple corresponding mechanism reaction units are constructed.

[0105] For example, if Figure 2 As shown in the figure, the complex heat transfer and mass transfer process of waste incineration is decomposed into several levels, including level 1, level 2, ..., level n; within each level, several mechanism reaction units can be constructed according to the different physical and chemical mechanisms involved in the waste incineration process, including mechanism 1, mechanism 2, ..., mechanism n.

[0106] Each mechanism reaction unit contains a set of eigenvalues With 1 target value set And the characteristic value set and the target value set satisfy the physical and chemical mechanism, which can be expressed as formula (1):

[0107] T t =H(F t ) (1)

[0108] Among them, H: is the mechanism-driven mapping function, F t is the set of eigenvalues, T t is the target value set.

[0109] More specifically, in step S13, the chain model is constructed using a modular architecture, with the prediction results of the previous level being used as input features of the subsequent level to achieve information transfer and integration between levels.

[0110] This chain model breaks down the waste incineration process into multiple reaction levels that conform to physical and chemical laws, connects different mechanism reaction units in different levels, and trains the mechanism reaction units at each level independently. While maintaining the independence of their physical mechanisms, it achieves coupling and information interaction between modules through standardized data interfaces.

[0111] The multi-level chain feature system constructed in step S1 breaks down the complex combustion process system into multiple reaction levels that conform to the laws of physical and chemical reactions. This not only achieves data dimensionality reduction and modular independent training, but also improves the efficiency of model development and optimization. It also ensures that the prediction results always conform to the actual laws of matter-energy conversion through the mechanism-embedded chain architecture, significantly improving the engineering applicability and interpretability of the waste incineration process prediction model.

[0112] Step S2, based on the dynamic non-steady-state characteristics of the waste incineration combustion condition, pre-process the eigenvalue data and target value data, construct a new eigenvalue set and a new target value set that reflect the dynamic changes, wherein the relationship between the new eigenvalue set and the new target value set satisfies the mapping function driven by the physical and chemical mechanism, and perform data normalization on the new eigenvalue set and the new target value set.

[0113] The dynamic, non-steady-state nature of waste incineration combustion conditions refers to the constant dynamic change of key parameters during the incineration process. This is primarily manifested in fluctuations in the composition and calorific value of the incoming waste, shifts in combustion state due to load changes, and significant hysteresis in heat and mass transfer processes. These factors collectively lead to significant non-stationary and nonlinear variations in key process parameters over time. Traditional data processing methods based on steady-state assumptions struggle to accurately capture these transient evolution patterns, leading to significant deviations between predicted results and actual operating conditions.

[0114] Therefore, a specially designed preprocessing method is needed to extract features that can characterize the dynamic behavior of the system from the original data to effectively make up for the defect that traditional static data cannot reflect the trend of parameter changes, so as to adapt to the model's needs for mining time-series dependencies and provide a structured data foundation for the subsequent integration of physical mechanism constraints, thereby significantly improving the model's adaptability and prediction accuracy to non-steady-state conditions.

[0115] In this embodiment, the step of preprocessing the eigenvalue data in step S2 further includes:

[0116] Calculate the first derivative of each eigenvalue with respect to time to obtain the rate of change of the eigenvalue with time f t , the rate of change f t Used to characterize the dynamic evolution trend of characteristic value parameters during the unsteady-state combustion process of waste incineration, avoiding the loss of transient information caused by relying solely on static values;

[0117] The original eigenvalue F t The rate of change of the characteristic value with time f t Combined with the above, a new eigenvalue set is constructed to reflect the static value and dynamic change trend of the eigenvalue {F t ,f t}.

[0118] The new eigenvalue set {F t ,f t It can simultaneously cover the current state of the parameters and the direction of their change over time, effectively solving the problem of slow response of traditional methods to dynamic changes in parameters. It enables the model to more comprehensively perceive the dynamic evolution process of the combustion system, providing richer input dimensions for the subsequent embedding of physical mechanism constraints, and significantly improving the model's sensitivity to capturing transient behaviors of operating conditions.

[0119] In this embodiment, the step of preprocessing the target value data in step S2 further includes:

[0120] Calculate the difference between each target value at the next time node and the current time node. The difference is used to characterize the dynamic evolution trend of the target value parameters during the unsteady-state combustion process of waste incineration, which is more in line with the physical nature of the evolution of combustion conditions over time.

[0121] The difference is constructed as a target value difference set and used as a new target value set ΔT t , which is used to reflect the dynamic change characteristics of the target value in time series.

[0122] New target value set ΔT t It can effectively reduce the interference of static numerical fluctuations in non-steady-state processes on model training, enable the model to focus more on learning the laws of operating condition changes, and improve the predictive sensitivity to sudden changes in combustion states.

[0123] More specifically, with respect to the preprocessing and normalization of the characteristic data in step S2, the preprocessing of the characteristic data takes into account that the combustion conditions in the waste incineration process are in a constantly changing process.

[0124] Therefore, the value of the predicted target value at the next time node in the future depends on the value of the target at the current time node and its change at the current time node, as shown in formula (2):

[0125]

[0126] Where, is the target value at the next time node, is the value of the target value at the current time node, is the change in the target value at the current time node.

[0127] The change of target value is affected by the change of eigenvalue and its eigenvalue, so Based on the physicochemical mechanism of formula (1), formula (3) can be obtained:

[0128]

[0129] Where, h: A new mapping driven by the mechanism.

[0130] Therefore, by calculating the eigenvalue F t Solve its derivative at time t to obtain the derivative result f t , and combined with the original eigenvalue F t , construct a new eigenvalue set {F t ,f t}.

[0131] For each target value in the target value set, solve its t+1 With the current time node T t The change in the target value is constructed The difference value set is used as the new target value set to match the dynamic change characteristics of the waste incineration combustion conditions.

[0132] In this embodiment, the data normalization processing step further includes:

[0133] The minimum-maximum normalization, Z-Score normalization or robust normalization method is used to normalize the characteristic value set and the target value set reflecting the dynamic changes.

[0134] The data normalization step is to normalize the new feature value set {F t ,f t} and the target value set ΔT t Normalization processing is performed to achieve standardized conversion of feature values ​​and target values ​​at different times t.

[0135] In this embodiment, the minimum-maximum (Min-Max) normalization is calculated according to the following formula (4):

[0136]

[0137] Where y i is the normalized eigenvalue, x i is the eigenvalue before normalization, is the minimum value of the eigenvalue set, is the maximum value of the eigenvalue set.

[0138] The Z-Score standardization is calculated according to the following formula (5):

[0139]

[0140] Where y i is the normalized eigenvalue, x i is the eigenvalue before normalization, μ is the mean of the eigenvalue set, and σ is the standard deviation of the eigenvalue set.

[0141] The robust normalization is calculated according to the following formula (6):

[0142]

[0143] Where y i is the normalized eigenvalue, x i is the eigenvalue before normalization, Q2 is the median of the eigenvalue set, and IQR = Q3-Q1, which is the difference between the 75th percentile and the 25th percentile (i.e., the interquartile range).

[0144] The normalization method of the target value is the same as that of the eigenvalue. With the help of the above normalization method, the eigenvalue set {F t ,f t} and target value set ΔT t Normalization processing.

[0145] Step S2 constructs a dynamic feature set by introducing the time derivative of the eigenvalue and the time-series difference between the target value. This effectively captures the dynamic changes in the unsteady-state combustion process of waste incineration and significantly enhances the model's ability to characterize transient operating conditions. This preprocessing method, combined with multiple normalization strategies, significantly improves the convergence speed and generalization performance of model training while preserving the robustness of the data distribution. This provides a high-quality, dimensionally consistent data foundation for subsequent hybrid modeling under the constraints of physical mechanisms.

[0146] In step S3, for each mechanism reaction unit in the multi-level chain feature system, a hybrid model integrating physical constraints is independently trained. Based on the preprocessed new feature value set and the new target value set, the hybrid model is trained under the physical mechanism constraints to construct a chain prediction model, which is used to predict the combustion conditions of waste incineration.

[0147] In this embodiment, the hybrid model is a physics-informed neural network (PINN) model that integrates physical prior knowledge, and the model training is performed based on the new feature value set and the new target value set preprocessed in step S2 as training data;

[0148] Among them, PINN is a model that embeds physical laws into the neural network loss function in the form of differential equations. It can ensure the physical rationality of the prediction results when data is limited and is suitable for short time series or steady-state process prediction.

[0149] Furthermore, the neural network model integrating physical prior knowledge also includes a physical information long short-term memory network or a physical information informer model to achieve long time series prediction tasks.

[0150] Specifically, if long-term time series prediction tasks are to be implemented, LSTM or Informer models that integrate physical prior knowledge can be constructed, namely, physics-informed long-term and short-term memory networks (Physics-informed-LSTM, PI-LSTM) and physics-informer models (Physics-informer, PI-Informer).

[0151] PI-LSTM integrates physical constraints based on the time series modeling capabilities of traditional LSTM, captures the dynamic dependencies of the combustion process through memory units, and is suitable for medium-length time series prediction.

[0152] PI-Informer is based on the long-sequence modeling advantages of Informer and combines physical mechanism constraints to solve the error accumulation problem in long-series prediction. It is suitable for prediction scenarios of future long-cycle operating conditions of waste incinerators.

[0153] Since step S1 has broken down the waste incineration process into multiple multi-level mechanism reaction units, the models used in each unit remain independent of each other and can be optimized and upgraded individually. Each mechanism reaction unit uses an adaptive weight mechanism and relies on a chain model to achieve prediction of higher-level target values.

[0154] In this embodiment, the step of performing hybrid model training under physical mechanism constraints further includes:

[0155] Construct a dual-objective loss function for model training;

[0156] The dual-objective loss function Loss is expressed as the weighted sum of the data loss term and the physical loss term, and the corresponding expression is:

[0157] Loss=w D L D +w P L P (7) Among them, L D is the data loss term, which is used to measure the deviation between the model prediction value and the true value;

[0158] L P They are physical loss terms, which are used to measure the deviation between the calculated value and the true value under the constraints of physical characteristics;

[0159] w D With w P are the weights of data loss item and physical loss item respectively.

[0160] The above-mentioned dual-objective optimization function is constructed by introducing physical mechanism constraints into the loss function. When the total loss value Loss is greater than the expected loss ε, the model continues training; when the total loss value Loss is less than the expected loss ε, the training is terminated, thereby strengthening the constraint role of physical mechanism in model training.

[0161] Furthermore, an adaptive moment estimation optimizer (Adam) is used to optimize the weight coefficients in the dual-objective loss function.

[0162] The optimizer dynamically adjusts the learning rate of each weight coefficient by calculating the first-order moment (mean) and second-order moment (uncentered variance) of the loss function on the weight coefficient. During the training process, the Adam optimizer adaptively updates the weight coefficient w according to the gradient information. D With w P , so that it gradually converges to the optimal value during the iteration process, thereby effectively balancing the contribution ratio of data-driven accuracy and physical law constraints in the total loss value Loss, and further improving the stability and generalization ability of model training.

[0163] Step S3 achieves a deep integration of physical mechanisms and deep learning at the structural level by independently training a hybrid model that integrates physical constraints in each mechanism reaction unit. It innovatively adopts architectures such as the PINN physical information neural network. While maintaining the flexibility and upgradability of the model, it embeds physical conservation laws into the optimization process through a dual-objective loss function, significantly improving the physical consistency of the prediction results and the generalization ability for abnormal working conditions.

[0164] The method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms with data-driven methods proposed in this paper has the following technical advantages:

[0165] By building a chain model based on physical and chemical laws, the complex heat and mass transfer process of waste incineration is disassembled, effectively reducing the data dimension of each mechanism reaction unit model and significantly improving training efficiency;

[0166] Each mechanism reaction unit can adopt a differentiated hybrid model and support independent optimization and upgrading. Combined with the adaptive weight mechanism, it ensures that the system can maintain optimal prediction performance under different working conditions.

[0167] The pre-processing method of deriving the eigenvalues ​​within each mechanism reaction unit and calculating the difference between adjacent time nodes of the target value is consistent with the physical characteristics of unsteady-state combustion conditions;

[0168] Introducing physical property constraints in the loss function design and adopting a dual-objective optimization function further strengthens the core role of physical mechanisms in model training.

[0169] Figure 3 A schematic diagram of a simplified model of a waste incinerator combustion condition according to an embodiment of the present invention is disclosed. This embodiment uses the prediction of the temperature inside the waste incinerator furnace, the flue gas temperature at the furnace outlet, and the flue gas flow rate at the furnace outlet as an example to illustrate the specific process of the method. The specific implementation steps are as follows:

[0170] Step S1, collecting characteristic values ​​based on physicochemical analysis and constructing a chain model;

[0171] Through the data acquisition module, the characteristic value data of the waste incineration system DCS at time t is obtained, including:

[0172] Amount of waste entering the furnace Auxiliary burner fuel quantity

[0173] Primary fan air pressure Primary fan frequency Primary air volume Primary air temperature

[0174] Secondary fan air pressure Secondary fan frequency Secondary air volume Secondary air temperature

[0175] Return fan air pressure Return fan frequency Return air volume Return air temperature

[0176] Loose fan air pressure Loose fan frequency Loose air volume Loose air temperature

[0177] Heat exchanger inlet flue gas temperature Heat exchanger outlet flue gas temperature

[0178] Target values ​​include: Incinerator bottom temperature Temperature in the middle of the incinerator Incinerator upper temperature Furnace outlet temperature Furnace outlet flue gas volume

[0179] The collected data are cleaned of outliers: the interquartile range (IQR) method is used for outlier detection, and the data points outside the range of [Q1-1.5IQR, Q3+1.5IQR] are smoothed and replaced using cubic spline interpolation to eliminate outliers and obtain the characteristic data sets at different time t.

[0180] Chain model construction combines the laws of physical and chemical mechanisms to classify the collected characteristic data and construct a multi-level chain characteristic system. This invention deconstructs the complex heat and mass transfer process of waste incineration. Rather than modeling the entire process in a general way, it deconstructs the multi-coupled, nonlinear complex system into a modular architecture of multi-level, multi-mechanism reaction units based on the inherent laws of material flow and heat transfer during the incineration process.

[0181] The decomposition logic of the complex heat transfer and mass transfer process of waste incineration first divides the overall process into several levels according to the core physical processes of waste incineration (such as actuator drive, material transport, heat exchange, and system output). Each level corresponds to a clear type of physical and chemical mechanism (such as fluid mechanics, heat transfer equation, and energy balance). Each level is then further decomposed into independent mechanism reaction units. Each mechanism reaction unit only focuses on a single or a few related heat transfer and mass transfer mechanisms to ensure that the relationship between the characteristic value and the target value within the unit strictly conforms to the physical laws (such as the fluid mechanics relationship between fan frequency / wind pressure and air volume, the heat transfer relationship between air volume / flue gas temperature and wind temperature, and the energy balance relationship between multiple parameters and furnace temperature).

[0182] Figure 4 A schematic diagram of a multi-level chain characteristic system for a waste incinerator combustion condition according to an embodiment of the present invention is disclosed. In this embodiment, the chain model of the physical and chemical mechanism of the waste incinerator combustion condition is divided into three levels. The model construction, characteristic value and target value setting, and corresponding physical mechanism of each level are as follows:

[0183] The first level (Level 1) is the air volume prediction model. It uses the frequency and air pressure of each fan (primary fan, secondary fan, return fan, and loosening fan) as characteristic values, and the corresponding air volume (primary air volume, secondary air volume, return air volume, and loosening air volume) as target values. Its construction is based on the following physical mechanism:

[0184]

[0185] Where Q 1 is the air volume of the fan at time 1, Q 2 is the wind volume of the fan at time 2, n 1 is the frequency of the fan at time 1, n 2 is the frequency of the fan at time 2, P 1 is the wind pressure of the fan at time 1, P 2 is the wind pressure of the fan at time 2.

[0186] Based on the above mechanism, the air volume value is taken as the target value, and the fan pressure and fan frequency are taken as characteristic values.

[0187] like Figure 4 As shown, the first level of characteristic values ​​includes the primary fan frequency Primary fan air pressure Secondary fan frequency Secondary fan air pressure Return fan frequency Return fan air pressure Loose fan frequency And loose fan pressure The target value includes the primary air volume Secondary air volume Return air volume and loose air volume

[0188] Among them, the primary air volume The air pressure of the primary fan can be and primary fan frequency Make predictions;

[0189] Secondary air volume The secondary fan air pressure and secondary fan frequency Make predictions;

[0190] Return air volume The return fan pressure can be adjusted and return fan frequency Make predictions;

[0191] Loose air volume By loosening the fan pressure and loose fan frequency Make predictions.

[0192] In this embodiment, the first level focuses on the conversion of "actuator-material flow", corresponds the operating parameters (frequency) of equipment such as fans with the material delivery volume (air volume), and constructs a mechanism reaction unit based on the fluid mechanics mechanism.

[0193] The second level (level 2) is the air temperature prediction model. It uses the air volumes (primary air volume, secondary air volume, return air volume, and loosening air volume) predicted by the first level, as well as the flue gas temperatures at the heat exchanger inlet and outlet and the air inlet temperatures (primary air inlet temperature, secondary air inlet temperature, return air inlet temperature, and loosening air inlet temperature) as characteristic values. The corresponding air outlet temperatures (primary air outlet temperature, secondary air outlet temperature, return air outlet temperature, and loosening air outlet temperature) are used as target values. Its construction is based on the following physical mechanism:

[0194]

[0195] Where k is the heat transfer coefficient, A is the heat transfer area, T in and T out are the inlet and outlet air temperatures of the heat exchanger, t in and t out is the inlet and outlet flue gas temperatures of the heat exchanger, Q gas is the gas flow rate, c gas is the specific heat capacity of the flue gas.

[0196] According to Figure 3 The heat exchanger process flow is designed with the outlet air temperature as the target value, and the heat exchanger inlet temperature, heat exchanger outlet temperature, inlet air temperature and inlet air volume as the characteristic values.

[0197] like Figure 4 As shown, the characteristic value of the second level includes the primary air volume from the target value of the first level Secondary air volume Return air volume and loose air volume Heat exchanger inlet flue gas temperature Heat exchanger outlet flue gas temperature Inlet air temperature (primary air inlet temperature, secondary air inlet temperature, return air inlet temperature, loosening air inlet temperature), target value includes primary air outlet temperature Secondary air outlet temperature Return air outlet temperature Loose air outlet temperature

[0198] Among them, the primary air outlet temperature The smoke temperature can be inlet through the heat exchanger Heat exchanger outlet flue gas temperature Inlet air temperature and primary air volume Make predictions;

[0199] Secondary air outlet temperature The smoke temperature can be inlet through the heat exchanger Heat exchanger outlet flue gas temperature Inlet air temperature and primary air volume Make predictions;

[0200] Return air outlet temperature The smoke temperature can be inlet through the heat exchanger Heat exchanger outlet flue gas temperature Inlet air temperature and return air volume Make predictions;

[0201] Loose air outlet temperature The smoke temperature can be inlet through the heat exchanger Heat exchanger outlet flue gas temperature Inlet air temperature and loose air volume Make predictions.

[0202] The temperatures of primary air inlet temperature, secondary air inlet temperature, return air inlet temperature and loosening air inlet temperature are the same, which are all inlet air temperatures. The primary air, secondary air, return air and loosening air originate from the same wind and are formed by diverting them through only four branches.

[0203] The second level revolves around the transfer of "material flow-thermal parameters". The material flow (such as air volume) of the previous level is combined with independent parameters such as the flue gas temperature of the heat exchanger, and is decomposed into mechanism reaction units such as wind temperature prediction based on the heat transfer equation.

[0204] The third level (Level 3) is the furnace temperature and flue gas volume prediction model. It uses the air volumes predicted by the first level, the air outlet temperatures predicted by the second level, the amount of garbage entering the furnace, and the amount of fuel for the auxiliary burner as characteristic values. It uses the temperatures of different areas of the incinerator (the bottom temperature, the middle temperature, the upper temperature, and the furnace outlet temperature) and the flue gas volume parameters at the furnace outlet as target values. Its construction is based on the following physical mechanism:

[0205]

[0206] Where m i is the flow rate of the i-th gas entering the furnace;

[0207] h(T i) is the enthalpy of the i-th gas entering the furnace, which is determined by the temperature T i calculated;

[0208] m fuel is the amount of fuel fed into the furnace, q fuel is the calorific value of the fuel entering the furnace, m j is the incinerator outlet flow rate, h(T j ) is the enthalpy of the flue gas at the incinerator outlet, Q diss To dissipate heat from the incinerator.

[0209] According to Figure 3 In the incinerator process flow, the characteristic values ​​of the third level are introduced into the target values ​​in the first and second levels respectively.

[0210] The third level characteristic values ​​include the primary air volume from the first level target values Secondary air volume Return air volume and loose air volume Primary air outlet temperature from the second level target value Secondary air outlet temperature Return air outlet temperature Loose air outlet temperature and the amount of waste entering the furnace Auxiliary burner fuel quantity The target values ​​include the incinerator furnace bottom temperature Temperature in the middle of the incinerator furnace Temperature of upper part of incinerator furnace Incinerator furnace outlet temperature Furnace outlet flue gas volume

[0211] The third level is oriented towards "system-level output", integrating the thermal parameters (such as wind temperature) and material flow of the first two levels, and breaking it down into mechanism reaction units such as furnace temperature and flue gas volume prediction based on the energy conservation mechanism.

[0212] Each level realizes information transmission through the chain connection of "previous target value as subsequent characteristic value", which not only maintains the physical independence of each mechanism reaction unit (can be optimized independently), but also ensures the logical coherence of the overall heat transfer and mass transfer process through standardized data interfaces. Ultimately, while reducing the dimension and improving the efficiency, it ensures that the prediction results always conform to the basic laws of matter-energy conversion.

[0213] It should be noted that the above three-level chain structure with air volume prediction, wind temperature prediction and furnace temperature and flue gas volume prediction as the core is only an illustrative description and does not limit the scope of protection of the present invention. The waste incineration process can be disassembled into a variety of different forms of chain prediction structures according to the actual process complexity and prediction requirements. The core of the present invention lies in the idea of ​​constructing a modular chain feature system based on physical and chemical mechanisms, rather than being limited to a specific hierarchical division method or parameter combination. Any progressive prediction framework based on a material-energy transfer sequence that drives subsequent inputs through preceding outputs falls within the scope of protection of the present invention.

[0214] Step S2, preprocessing and normalization of feature data;

[0215] This embodiment takes the third level as an example to illustrate the preprocessing process:

[0216] Based on formula (3), the primary air volume Primary air outlet temperature Secondary air volume Secondary air outlet temperature Return air volume Return air outlet temperature Loose air volume Loose air outlet temperature and the amount of waste entering the furnace Auxiliary burner fuel quantity Construct the eigenvalue set F t , solve the derivative of each eigenvalue with respect to time t, and convert the derivative result f t and the original eigenvalue set F t Merge to form a new eigenvalue set {F t ,f t};

[0217] Based on formula (2), the target value data (the temperature at the bottom of the incinerator Temperature in the middle of the incinerator furnace Temperature of upper part of incinerator furnace Incinerator furnace outlet temperature Furnace outlet flue gas volume ), solve its next time node T t+1 With the current time node T t The difference ΔT t , and use the difference value set as the new target value set, thus completing the target data preprocessing.

[0218] Subsequently, the minimum-maximum normalization, Z-Score standardization or robust normalization method is used to normalize the above-mentioned characteristic value set reflecting dynamic changes and the newly constructed target value set to eliminate the dimensional differences between different parameters, laying the foundation for the stability of subsequent model training and improvement of prediction accuracy.

[0219] Step S3: training the hybrid model under physical mechanism constraints.

[0220] For the preprocessed feature set {F t ,f t} and the target value set ΔT t , build a neural network that integrates physical prior knowledge and conduct hybrid model training.

[0221] The hybrid model is a physics-informed neural network (PINN) model that incorporates physics prior knowledge. For long-term time series prediction, LSTM or informer models that incorporate physics prior knowledge can be constructed, namely PI-LSTM (Physics-informed-LSTM, PI-LSTM) and PI-Informer (Physics-informer, PI-Informer).

[0222] During the training process, based on the preprocessed feature value set {F t ,f t} and the target value set ΔT t A neural network integrating physical prior knowledge is constructed to perform hybrid model training. By introducing physical mechanism constraints into the loss function, a dual-objective optimization function is formed (as shown in Equation (7)).

[0223] In this embodiment, the data loss L D Calculate according to the following formula (12):

[0224]

[0225] Where, and are the true value and predicted value at time t+1 respectively.

[0226] Subtract the true value at time t from both The data loss L D It is equivalent to the actual deviation of the target value change and prediction bias The difference.

[0227] Taking the third level as an example, the physical loss L PBased on the energy balance mechanism of formula (8), the eigenvalue set {F t+1} is calculated as shown in formulas (13) to (15):

[0228]

[0229]

[0230] Where, the enthalpy of gas or flue gas h(T i ) and h(T j ) can be calculated through the flue gas enthalpy table, and the corresponding target value This can be checked through the flue gas enthalpy table.

[0231] Heat dissipation value Q at the bottom of the incinerator diss,bo , heat dissipation value Q in the middle of the incinerator diss,ce , heat dissipation value Q of the upper part of the incinerator diss,to Affected by the heat dissipation area of ​​the furnace, the heat dissipation area changes with different temperature zones, and the heat dissipation value of each temperature zone is a fixed constant.

[0232] The target value calculated by equations (13) to (15) is the calculated value under the physical mechanism constraint condition, that is, Then the physical loss is calculated according to formula (16):

[0233]

[0234] The loss value Loss based on physical feature constraints can be obtained by introducing the weight w of data loss and physical loss D With w P , calculated using the method of formula (7).

[0235] The dual-objective loss function Loss is the weighted sum of data loss and physical loss (see Equation (7), where the weight coefficient can be optimized by the adaptive moment estimation (Adam) optimizer. During the training process, when the total loss Loss value is greater than the expected loss ε, the model continues to iterate the training momentum and adaptive learning rate; when the total loss Loss value is less than or equal to ε, the training ends.

[0236] The waste incineration combustion condition prediction method proposed in this invention integrates physical and chemical mechanisms with data-driven, innovatively embeds physical laws such as mass conservation and energy balance in the form of constraints into the neural network training process, ensuring the prediction accuracy while ensuring the physical rationality of the results. On the one hand, it can screen characteristic values ​​through physical and chemical laws to achieve feature dimensionality reduction and strengthen the training data set; on the other hand, by introducing physical and chemical condition constraints in model training, it significantly improves the accuracy, robustness and physical interpretability of the prediction results.

[0237] Based on the above-mentioned method for predicting the combustion condition of garbage incineration that integrates physical and chemical mechanisms and data-driven, the present invention also proposes a device for predicting the combustion condition of garbage incineration that integrates physical and chemical mechanisms and data-driven. The device for predicting the combustion condition of garbage incineration that integrates physical and chemical mechanisms and data-driven proposed by the present invention may include an internal communication bus, a processor, a read-only memory (ROM), a random access memory (RAM), a communication port, and a hard disk. The internal communication bus can realize data communication between the components of the device for predicting the combustion condition of garbage incineration that integrates physical and chemical mechanisms and data-driven. The processor can make judgments and issue prompts. In some embodiments, the processor may be composed of one or more processors.

[0238] The communication port enables data transmission and communication between the device for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven processes and external input / output devices. In some embodiments, the device for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven processes can send and receive information and data from a network via the communication port. In some embodiments, the device for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven processes can transmit and communicate data with external input / output devices via a wired connection via the input / output port.

[0239] The waste incineration combustion condition prediction device, integrating physical and chemical mechanisms with data-driven methods, can also include various forms of program storage units and data storage units, such as hard disks, read-only memory (ROM), and random access memory (RAM). These can store various data files used for computer processing and / or communication, as well as possible program instructions executed by a processor. The processor executes these instructions to implement the main components of the method. The processor's processing results are transmitted to an external output device via a communication port and displayed on the output device's user interface.

[0240] For example, the implementation process file of the above-mentioned waste incineration combustion condition prediction method that integrates physical and chemical mechanisms and data-driven can be a computer program, saved in a hard disk, and can be recorded in a processor for execution to implement the method of the present invention.

[0241] When the implementation process file of the method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms with data-driven is a computer program, it can also be stored in a computer-readable storage medium as a product. For example, computer-readable storage media may include, but are not limited to, magnetic storage devices (e.g., hard disks, floppy disks, magnetic strips), optical disks (e.g., compact disks (CDs), digital versatile disks (DVDs)), smart cards, and flash memory devices (e.g., electrically erasable programmable read-only memories (EPROMs), cards, sticks, key drives). In addition, the various storage media described herein can represent one or more devices and / or other machine-readable media for storing information. The term "machine-readable medium" may include, but is not limited to, wireless channels and various other media (and / or storage media) that can store, contain, and / or carry code and / or instructions and / or data.

[0242] The method and device for predicting the combustion conditions of waste incineration, which integrates physical and chemical mechanisms with data-driven methods, have the following beneficial effects:

[0243] 1) A multi-level chain model architecture based on physical and chemical mechanisms is adopted to decompose the complex incineration process into multiple levels, effectively reducing the data dimension and training difficulty of the model, and supporting the independent optimization and upgrading of each mechanism reaction unit, thereby improving the flexibility and engineering applicability of the system;

[0244] 2) By embedding physical laws such as mass conservation and energy balance into neural network training as constraints, the model not only ensures prediction accuracy but also the physical rationality of the results, significantly improving the scientific nature and interpretability of the model.

[0245] 3) Time derivatives and difference calculations are introduced in the feature and target value preprocessing to match the dynamic characteristics of the non-steady-state combustion process, and data and physical constraints are integrated into the loss function to enhance the model's generalization ability and predictive robustness for complex working conditions.

[0246] Although the above methods are illustrated and described as a series of acts for simplicity of explanation, it is to be understood and appreciated that these methods are not limited by the order of the acts, as some acts may occur in a different order and / or concurrently with other acts from those illustrated and described herein or not illustrated and described herein but understandable to those skilled in the art according to one or more embodiments.

[0247] As used in this application and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not intended to refer to the singular but may include the plural. Generally speaking, the terms "comprises" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list. A method or apparatus may also include other steps or elements.

[0248] The above embodiments are provided to persons familiar with the art for implementing or using the present invention. Personnel familiar with the art may make various modifications or changes to the above embodiments without departing from the inventive concept of the present invention. Therefore, the scope of protection of the present invention is not limited to the above embodiments, but should be the maximum scope of the innovative features mentioned in the claims.

Claims

1. A method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven, characterized by: The following steps are involved: Step S1: Based on the physical and chemical mechanisms involved in the waste incineration process, characteristic values ​​are collected and a chain model of a multi-level chain characteristic system is constructed. The multi-level chain characteristic system is a modular architecture formed by decomposing the waste incineration process into a plurality of mechanism reaction units at multiple levels. The prediction results of the previous level are used as input features of the subsequent level. Each mechanism reaction unit includes a characteristic value set and a target value set, and the relationship between the characteristic value set and the target value set satisfies a mapping function driven by the physical and chemical mechanism. Step S2: Preprocessing the eigenvalue data and target value data based on the dynamic and non-steady-state characteristics of the waste incineration combustion condition, constructing a new eigenvalue set and a new target value set that reflect the dynamic changes, wherein the relationship between the new eigenvalue set and the new target value set satisfies a mapping function driven by a physical and chemical mechanism, and performing data normalization processing on the new eigenvalue set and the new target value set; In step S3, for each mechanism reaction unit in the multi-level chain feature system, a hybrid model integrating physical constraints is independently trained. Based on the preprocessed new feature value set and the new target value set, the hybrid model is trained under the physical mechanism constraints to construct a chain prediction model, which is used to predict the combustion conditions of waste incineration.

2. The method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to claim 1 is characterized in that: The step S1 further comprises: Based on the transformation laws of matter and energy during waste incineration, the complex heat and mass transfer process of waste incineration is decomposed into multiple levels. The characteristic value data and target value data corresponding to each process are classified into corresponding levels. The multiple levels are progressive functional modules that reflect the incineration process from the actuator to the system level output. Each level corresponds to the heat and mass transfer behavior of a specific process. For each level, several mechanism reaction units are constructed according to the different physical and chemical mechanisms involved in the waste incineration process; The mechanism reaction units in different levels are connected through a modular architecture, and the output of the previous level drives the input of the subsequent level to build a chain model.

3. The method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to claim 1 is characterized in that: The chain model of the multi-level chain feature system constructed in step S1 includes at least three levels, including a first level, a second level, and a third level: At the first level, the operating parameters of the actuator are used as characteristic values, and the generated material flow rate is used as the target value; The second level uses part or all of the target values ​​predicted by the first level and other independent combustion process parameters as characteristic values, and uses the thermal parameters of the combustion process as target values; The third level uses part or all of the target values ​​predicted by the first and second levels, combined with other independent system operating parameters as characteristic values, and uses the final system-level thermal parameters and emission parameters as target values.

4. The method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to claim 1 is characterized in that: The step of pre-processing the characteristic value data in step S2 further includes: Calculate the first derivative of each eigenvalue with respect to time to obtain the rate of change of the eigenvalue with time f t , the rate of change f t Used to characterize the dynamic evolution trend of characteristic value parameters during the unsteady-state combustion process of waste incineration; The original eigenvalue F t The rate of change of the characteristic value with time f t Combined with the above, a new eigenvalue set is constructed to reflect the static value and dynamic change trend of the eigenvalue {F t ,f t }.

5. The method for predicting the combustion conditions of waste incineration by integrating physical and chemical mechanisms with data-driven methods according to claim 1 or 4, characterized in that: The step of pre-processing the target value data in step S2 further includes: Calculate the difference between the value of each target value at the next time node and the value at the current time node, and use the difference to characterize the dynamic evolution trend of the target value parameter during the unsteady-state combustion process of waste incineration; The difference is constructed as a target value difference set and used as a new target value set ΔT t , which is used to reflect the dynamic change characteristics of the target value in time series.

6. The method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to claim 1 is characterized in that: The hybrid model is a neural network model that integrates physical prior knowledge, and the model training is performed based on the new feature value set and the new target value set preprocessed in step S2 as training data; The neural network model integrating physical prior knowledge also includes a physical information long short-term memory network or a physical information informer model to achieve long time series prediction tasks.

7. The method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to claim 1 is characterized in that: The step S1 further comprises: Collecting original eigenvalue data from the distributed control system of the waste incineration system; Outlier detection is performed on the original eigenvalue data, and data points that are out of the set range are replaced to complete data cleaning.

8. The method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to claim 7 is characterized in that: The step of performing outlier detection on the eigenvalue data and replacing data points that exceed a set range further includes: The interquartile range method was used to detect and clean outliers, and the cubic spline interpolation method was used to smooth and replace data points that exceeded the set range [Q1-1.5IQR, Q3+1.5IQR]. Among them, Q1 is the first quartile of the eigenvalue set, Q3 is the third quartile of the eigenvalue set, and IQR is the interquartile range.

9. The method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to claim 1 is characterized in that: The data normalization processing step further includes: The new feature value set and the new target value set reflecting the dynamic changes are normalized by using a minimum-maximum normalization, a Z-Score normalization or a robust normalization method.

10. The method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to claim 1 is characterized in that: The step S3 of performing hybrid model training under physical mechanism constraints further includes: Construct a dual-objective loss function for model training; The dual-objective loss function Loss is expressed as the weighted sum of the data loss term and the physical loss term, and the corresponding expression is: Loss=w D L D +w P L P Among them, L D is the data loss term, which is used to measure the deviation between the model prediction value and the true value; L P are data loss terms, which are used to measure the deviation between the calculated value and the true value under the constraints of physical characteristics; w D With w P are the weights of data loss item and physical loss item respectively.

11. The method for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms and data-driven methods according to claim 10 is characterized in that: The step S3 further includes: optimizing the weight coefficients in the dual-objective loss function using an adaptive moment estimation optimizer.

12. A device for predicting the combustion conditions of waste incineration that integrates physical and chemical mechanisms with data-driven methods, comprising: a memory for storing instructions executable by the processor; A processor, configured to execute the instructions to implement the method according to any one of claims 1 to 11.

13. A computer-readable medium having computer instructions stored thereon, wherein when the computer instructions are executed by a processor, the method according to any one of claims 1 to 11 is performed.

Images