Methods and systems for predicting the operating conditions of waste incinerators and for reverse optimization

Abstract

This invention provides a method and system for predicting and reversing the operating conditions of a waste incinerator, relating to the field of intelligent control technology. The method utilizes L... 25 (5) 5 Orthogonal experiments can significantly reduce the number of simulation case groups. Combined with a two-dimensional planar CFD model, it can quickly generate high-quality datasets, significantly reducing data acquisition costs and cycles. It adopts sliding window MLP prediction technology, which has high prediction accuracy and lightweight deployment. The latency of forward inference and backward optimization is extremely low. It can automatically differentiate gradient-based backward optimization based on the forward model, upgrading from "post-event adjustment" to "proactive prediction + optimal control". Multi-parameter collaborative optimization has low error and fast response, fully meeting the needs of industrial real-time control.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, specifically to a method and system for predicting and reversing the operating conditions of a waste incinerator. Background Technology

[0002] In recent years, municipal solid waste incineration power generation has become a core technological approach to achieving waste reduction, harmlessness, and resource recovery. Accurate prediction and proactive control of key operating parameters such as the average furnace temperature and flue gas pollutant concentration are crucial for ensuring the stable operation of the incineration system and achieving emission standards.

[0003] The existing technology has the following significant engineering flaws:

[0004] 1. Low data acquisition efficiency and high cost: The 5-factor, 5-level full factorial experiment requires 3,125 sets of simulations / experiments, which is time-consuming and extremely costly; the calculation cycle for a single working condition in 3D CFD simulation can be as long as several weeks, making it difficult to quickly generate sufficient training data; the axisymmetric 2D model does not match the actual geometric structure of the rectangular furnace, resulting in serious distortion of simulation results, which cannot be used for engineering guidance.

[0005] 2. Conflict between accuracy and deployability in operating condition prediction: Time series deep learning models such as LSTM and GRU have a large number of parameters and high inference latency, making it difficult to achieve lightweight deployment at the edge of power plants; Traditional ensemble learning models such as XGBoost cannot capture the strong coupling nonlinear characteristics of multiple parameters in the combustion process, resulting in insufficient prediction accuracy for low-concentration pollutants such as SO2.

[0006] 3. Lagging control methods and poor coordination: The mainstream single-parameter PID feedback control can only achieve post-event adjustment and cannot predict combustion fluctuations in advance; the existing intelligent optimization scheme and the operating condition prediction model are independent of each other and belong to open-loop control, which cannot achieve multi-parameter coordinated optimization, and the optimization results deviate significantly from the actual situation on site. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a method and system for predicting and reversing the operating conditions of waste incinerators, solving the problems of poor prediction results and lack of reverse optimization function.

[0009] (II) Technical Solution

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A method for predicting and reverse-optimizing the operating conditions of a waste incinerator, the method comprising:

[0012] S1. Through L 25 (5)5 The orthogonal experimental design uses 5 factors and 5 levels for each factor to simulate 25 sets of working conditions, and combines them with a two-dimensional planar CFD model to quickly generate a high-quality training dataset.

[0013] S2. The training dataset is used to construct time-series samples through sliding window technology to train a lightweight MLP model, thereby achieving real-time high-precision prediction of furnace average temperature, SO2 concentration, and combustion power.

[0014] S3. Load the trained MLP model, fix all weight parameters of the model, and use the gradient descent algorithm to back-calculate the optimal control parameters for outputting incinerator control commands.

[0015] Preferably, S1 specifically includes:

[0016] S1.1. Select five factors: primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate. Set five levels for each factor, using L... 25 (5) 5 ) Design 25 sets of simulation conditions using orthogonal experimental tables;

[0017] S1.2. Take the central section of the incinerator along the length of the grate and build a two-dimensional planar geometric model including the grate, furnace, primary air inlet, secondary air inlet and outlet flue.

[0018] S1.3. According to the parameter combinations of the above 25 working conditions, set the boundary conditions for primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate one by one; generate transient data for multiple time steps for each working condition; extract the input and output parameters for each time step;

[0019] S1.4. Remove invalid data from the initial ignition stage of each working condition, use the IQR method to remove numerical oscillation anomalies, fill in a small number of missing values ​​through linear interpolation, normalize all parameters to the [0,1] interval, and obtain a standardized time-series training dataset adapted to the sliding window MLP.

[0020] Preferably, in S1.2, the two-dimensional planar geometric model is a physical model set up based on the ANSYS Fluent 2025 R1 platform: a pressure-based transient solver is used, and gravity is enabled; the turbulence model uses the standard k-ε model, combined with the standard wall function; the radiation model uses the P1 model; the component transport and reaction model and the discrete phase model are enabled, and the waste particle injection source injection-0 is defined; and the non-essential sub-models such as multiphase flow, NOx, and soot are disabled.

[0021] Preferably, in step S1.3, the input parameters include: primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate;

[0022] The output parameters include: average furnace temperature, SO2 concentration, and combustion power.

[0023] Preferably, in step S1.4, if supplementary data is needed for a specific working condition, a single-point simulation is carried out within the parameter range covered by the orthogonal experiment.

[0024] Preferably, S2 specifically includes:

[0025] S2.1. A sliding window of size 3 is used to generate training samples. The input parameters at times t-3, t-2, and t-1 are concatenated in chronological order to generate a 15-dimensional static feature vector. The furnace average temperature, SO2 concentration, and combustion power at time t are used as output parameters. Starting from the third time step, the window is slid sequentially to generate multiple valid time-series samples. The first 70% of the samples are divided into a training set and the last 30% into a test set in chronological order.

[0026] S2.2. Construct a simplified MLP model with a single hidden layer, the structure of which is: 15 neurons in the input layer → 16 neurons in the hidden layer → 3 neurons in the output layer; use mean squared error as the loss function, Adam optimizer, learning rate 0.001, batch size=16, maximum number of iterations 200; set an early stopping strategy with patience=30, stop training early when the validation set loss does not decrease for 30 consecutive iterations, and restore the optimal weights;

[0027] S2.3. Input the test set into the trained model, obtain the prediction results, and inversely normalize them to the original physical quantity scale.

[0028] Preferably, if it is necessary to add prediction metrics, the number of neurons in the output layer can be modified and corresponding training data can be added.

[0029] Preferably, S3 specifically includes:

[0030] S3.1. The objective function is the mean square error between the predicted and target operating conditions; the constraints are the actual operating ranges of each control parameter, consistent with the level range of the orthogonal experiment.

[0031] The primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate are set as optimizable tensor variables, and the initial values ​​are taken as the median values ​​of each factor in the orthogonal experiment.

[0032] S3.2. The Adam optimizer is used to calculate the gradient of the objective function with respect to the control variables through the automatic differentiation mechanism of the PyTorch framework, and the control variables are updated iteratively. After each iteration, the control variables are restricted to the physical range verified by orthogonal experiments through hard constraint functions until the objective function converges or the maximum number of iterations is reached. In this way, the optimal control parameters can be obtained by back-deriving the input parameters from the target output parameters, which are used to output the incinerator control commands.

[0033] A waste incinerator operating condition prediction and reverse optimization system, the system comprising: a training dataset generation module, an operating condition forward prediction module, and a reverse optimization module;

[0034] The training dataset generation module is used to generate L 25 (5) 5 The orthogonal experimental design uses 5 factors and 5 levels for each factor to simulate 25 sets of working conditions, and combines them with a two-dimensional planar CFD model to quickly generate a high-quality training dataset.

[0035] The positive prediction module for operating conditions is used to construct time-series samples from the standardized time-series training dataset using the sliding window technique, train a lightweight MLP model, and achieve real-time high-precision prediction of furnace average temperature, SO2 concentration, and combustion power.

[0036] The reverse optimization module is used to load the trained MLP model, fix all weight parameters of the model, and reverse-engineer the optimal control parameters through the gradient descent algorithm to output incinerator control commands.

[0037] (III) Beneficial Effects

[0038] This invention provides a method and system for predicting and reverse-optimizing the operating conditions of a waste incinerator. Compared with existing technologies, it has the following advantages:

[0039] In this invention, the method is achieved through L 25 (5) 5 The orthogonal experiment reduces the existing 3125 sets of full factorial simulation conditions to 25 sets. Combined with a two-dimensional planar CFD model, it can quickly generate high-quality datasets, improving the data acquisition cycle by more than 60 times and significantly reducing data collection costs and cycles. It adopts sliding window MLP prediction technology, which has high prediction accuracy and lightweight deployment, with extremely low latency in both forward inference and backward optimization. It can automatically differentiate gradient-based backward optimization based on the forward model, upgrading from "post-event adjustment" to "proactive prediction + optimal control". Multi-parameter collaborative optimization results in low error and fast response, fully meeting the needs of real-time industrial control. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of the method described in an embodiment of the present invention.

[0042] Figure 2 L in the embodiments of the present invention 25 (5) 5 Table of orthogonal experimental factor levels and working condition combinations.

[0043] Figure 3 This is a schematic diagram of the two-dimensional planar CFD simulation geometric model and mesh division of a waste incinerator in an embodiment of the present invention.

[0044] Figure 4 This is a schematic diagram illustrating the construction process of the sliding window timing sample in an embodiment of the present invention.

[0045] Figure 5 This is a comparison chart of the operating condition prediction results and actual values ​​of the sliding window MLP model in this embodiment of the invention.

[0046] Figure 6 This is a schematic diagram of the convergence curve of the objective function in the gradient-based backward optimization process in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] This application provides a method and system for predicting and reversing the operating conditions of waste incinerators, which solves the problems of poor prediction results and lack of reverse optimization function.

[0049] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0050] Example:

[0051] like Figures 1-6 As shown, this invention provides a method for predicting and reverse-optimizing the operating conditions of a waste incinerator, the method comprising:

[0052] S1. Based on L 25 (5) 5 Orthogonal experiments combined with two-dimensional planar CFD models generate training datasets.

[0053] Based on the mainstream 500t / d forward-pushing mechanical grate waste incinerator in China, the simulation workload was reduced through orthogonal experimental design, and a training dataset was generated by combining a two-dimensional planar CFD model.

[0054] S1.1. Orthogonal experimental design

[0055] Five core control parameters—primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate—were selected as experimental factors. Five levels were set for each factor, using L... 25 (5) 5 The orthogonal experimental design includes 25 simulation conditions. The orthogonal experimental design is as follows: Figure 2 As shown;

[0056] S1.2. Construction of a 2D Planar CFD Model

[0057] A two-dimensional planar geometric model of the incinerator, including the grate, furnace, primary air inlet, secondary air inlet, and outlet flue, is constructed by taking a central section along the length of the grate. Figure 3 As shown, a structured quadrilateral mesh is used for partitioning;

[0058] The physical model was set up using the ANSYS Fluent 2025 R1 platform: a pressure-based transient solver was used, with gravity (y-direction -9.81 m / s²) enabled. 2 The standard k-ε turbulence model is used, along with the standard wall function; the P1 radiation model is used; the component transport and reaction model and the discrete phase (DPM) model are enabled, and the waste particle injection source injection-0 is defined; the multiphase flow, NOx, soot and other unnecessary sub-models are disabled.

[0059] S1.3. Simulation Calculation and Data Extraction

[0060] Based on the parameter combinations of the above 25 working conditions, boundary conditions such as primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate were set one by one; the total simulation time for a single working condition was set to 25s, with a time step of 0.01s, and 2500 time steps of transient data were generated for each working condition; the input parameters (primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate) and output parameters (average furnace temperature, SO2 concentration, and combustion power) for each time step were extracted.

[0061] S1.4. Data Preprocessing

[0062] Invalid data from the first 50 time steps of the initial ignition phase for each working condition were removed, leaving 2450 valid time steps for each working condition. A total of 61250 valid time steps were obtained from the 25 working conditions. Numerical oscillation outliers were removed using the IQR method, and a small number of missing values ​​were filled in by linear interpolation. All parameters were normalized to the [0,1] interval to obtain a standardized time series training dataset adapted to the sliding window MLP. Based on this, if data needs to be supplemented for a specific working condition, single-point simulations can be carried out within the parameter range covered by the orthogonal experiment without having to repeat the full factorial experiment, further reducing the cost of data acquisition.

[0063] S2. Transient condition forward prediction based on sliding window MLP

[0064] A standardized time-series training dataset is used to construct time-series samples using the sliding window technique, and a lightweight MLP model is trained to achieve real-time prediction of operating conditions.

[0065] S2.1. Construction of Sliding Window Time Series Dataset

[0066] Training samples are generated using a sliding window with a window size of 3. The construction process is as follows: Figure 4 As shown: The five input parameters at times t-3, t-2, and t-1 are concatenated in chronological order to generate a 15-dimensional static feature vector; the average furnace temperature, SO2 concentration, and combustion power at time t are used as output parameters; starting from the third time step, the parameters are sequentially slid to generate a total of 61,247 valid time-series samples; the first 70% are divided into a training set (42,873 samples) and the last 30% into a test set (18,374 samples) in chronological order to avoid data leakage;

[0067] S2.2. Lightweight MLP Model Construction and Training

[0068] A simplified MLP model with a single hidden layer is constructed, with the following structure: 15 neurons in the input layer (corresponding to 15-dimensional sliding window features) → 16 neurons in the hidden layer (ReLU activation function, with L2 regularization and coefficient 0.001) → 3 neurons in the output layer (linear activation function, corresponding to three operating parameters: furnace average temperature, SO2 concentration, and combustion power). Based on experimental verification, these three core parameters with the best prediction performance are selected as the model output to improve prediction accuracy and engineering practicality. The mean squared error (MSE) is used as the loss function, with the Adam optimizer (learning rate 0.001), batch size=16, and a maximum number of iterations of 200. An early stopping strategy with patience=30 is set, which stops training early when the validation set loss does not decrease for 30 consecutive iterations, restoring the optimal weights.

[0069] S2.3. Model Validation and Result Output

[0070] The test set is input into the trained model to obtain prediction results, which are then inversely normalized to the original physical quantity scale; the prediction results are, for example, Figure 5 As shown, the test set validation metrics are as follows:

[0071] Average furnace temperature: coefficient of determination R 2 =0.87, mean absolute error (MAE) = 15.3K;

[0072] SO2 concentration: coefficient of determination R 2 =0.78, Mean Absolute Error (MAE) = 0.00025 mg / m 3 ;

[0073] Combustion power: Determinant coefficient R 2 =0.72, mean absolute error (MAE) =0.0012W;

[0074] Forward inference latency per sample: 6.8ms;

[0075] Based on this, if it is necessary to add prediction indicators such as volatile concentration and reaction exothermic power, it is only necessary to modify the number of neurons in the output layer and supplement the corresponding training data, without changing the overall structure of the model and the training process.

[0076] S3. Gradient-based backward optimization based on automatic differentiation of the forward model

[0077] Load the trained MLP model, fix all weight parameters of the model, and use the gradient descent algorithm to back-calculate the optimal control parameters.

[0078] S3.1. Setting Optimization Objectives and Constraints

[0079] The objective function is the mean square error between the predicted and target operating conditions; the constraints are the actual operating ranges of each control parameter, consistent with the level range of the orthogonal experiment.

[0080] Primary air temperature ≤ 900K ≤ 1100K;

[0081] 4m / s≤ Primary wind velocity≤8m / s;

[0082] 1100K≤Secondary air temperature≤1300K;

[0083] 0.4 m / s ≤ secondary wind velocity ≤ 1.2 m / s;

[0084] 0.1 kg / s ≤ feed rate ≤ 0.5 kg / s;

[0085] The primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate were set as optimizable tensor variables, with the initial values ​​taken as the median level of each factor in the orthogonal experiment (primary air temperature = 1000K, primary air velocity = 6m / s, secondary air temperature = 1200K, secondary air velocity = 0.8m / s, feed rate = 0.3kg / s).

[0086] S3.2. Gradient Iterative Optimization and Convergence Verification

[0087] The Adam optimizer (learning rate 0.1) is used, and the gradient of the objective function with respect to the control variables is calculated through the automatic differentiation mechanism of the PyTorch framework. The control variables are then updated iteratively. After each iteration, a hard constraint function is used to restrict the control variables to the physical range verified by orthogonal experiments, until the objective function converges or the maximum number of iterations (500) is reached. The convergence curve of the objective function in the reverse optimization process is shown in the figure. Figure 6 As shown: In the early stage of iteration (0~100 times), the objective function decreases exponentially and rapidly, and more than 80% of the loss is reduced in just 100 iterations, which verifies the correctness and efficiency of the optimization direction; in the middle stage of iteration (100~300 times), the loss converges slowly, and the process enters the stage of fine-tuning the control parameters; after 300 iterations, the objective function converges completely and stabilizes at an extremely low level, without oscillation or rebound, which verifies the global convergence and robustness of the optimization algorithm;

[0088] S3.3. Verification of Optimization Results

[0089] Taking a target furnace average temperature of 1200K as an example, the optimal control parameters obtained through optimization are: primary air temperature 985K, primary air velocity 6.2m / s, secondary air temperature 1230K, secondary air velocity 0.75m / s, and feed rate 0.32kg / s. When these parameters are input into the forward model, the predicted temperature is 1198.7K, with a deviation of only 1.3K from the target. The total time for a single reverse optimization is ≤10ms, which fully meets the real-time control requirements of industrial sites.

[0090] This invention provides a system for predicting and reversing the operating conditions of a waste incinerator, the system comprising: a training dataset generation module, a forward operating condition prediction module, and a reverse optimization module;

[0091] The training dataset generation module is used to generate L 25 (5) 5 The orthogonal experimental design uses 5 factors and 5 levels for each factor to simulate 25 sets of working conditions, and combines them with a two-dimensional planar CFD model to quickly generate a high-quality training dataset.

[0092] The positive prediction module for operating conditions is used to construct time-series samples from the standardized time-series training dataset using the sliding window technique, train a lightweight MLP model, and achieve real-time high-precision prediction of furnace average temperature, SO2 concentration, and combustion power.

[0093] The reverse optimization module is used to load the trained MLP model, fix all weight parameters of the model, and reverse-engineer the optimal control parameters through the gradient descent algorithm.

[0094] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0095] 1. In this embodiment of the invention, L 25 (5) 5 Orthogonal experiments reduce the existing 3125 full factorial simulation scenarios to 25 scenarios. Combined with a two-dimensional planar CFD model, high-quality datasets can be generated quickly. The calculation cycle for a single scenario is ≤2 hours, which improves the total data acquisition cycle by more than 60 times compared to the traditional three-dimensional full factorial method, significantly reducing data acquisition costs and cycle time.

[0096] 2. In this embodiment of the invention, a sliding window MLP prediction technique is adopted, which takes into account both prediction accuracy and lightweight deployment requirements. The prediction coefficient of furnace average temperature is ≥0.87, the prediction coefficient of SO2 concentration is ≥0.78, and the prediction coefficient of combustion efficiency is ≥0.72. The total number of model parameters is 307, which is about 0.5% of that of an LSTM model with the same accuracy. The forward inference latency of a single sample is ≤7ms, and the backward optimization latency of a single test is ≤10ms. It can be directly deployed on the existing edge computing platform of the power plant without additional hardware investment.

[0097] 3. In this embodiment of the invention, gradient-based backward optimization based on automatic differentiation of the forward model upgrades from "post-event adjustment" to "active prediction + optimal control", with multi-parameter collaborative optimization, low error and fast response, fully meeting the needs of industrial real-time control.

[0098] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0099] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting and reverse-optimizing the operating conditions of a waste incinerator, characterized in that, The method includes: S1. Through L 25 (5) 5 The orthogonal experimental design uses 5 factors and 5 levels for each factor to simulate 25 sets of working conditions, and combines them with a two-dimensional planar CFD model to quickly generate a high-quality training dataset. S2. The training dataset is used to construct time-series samples through sliding window technology to train a lightweight MLP model, thereby achieving real-time high-precision prediction of furnace average temperature, SO2 concentration, and combustion power. S3. Load the trained MLP model, fix all weight parameters of the model, and use the gradient descent algorithm to back-calculate the optimal control parameters for outputting incinerator control commands.

2. The waste incinerator operating condition prediction and reverse optimization method as described in claim 1, characterized in that, S1 specifically includes: S1.

1. Select five factors: primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate. Set five levels for each factor, using L... 25 (5) 5 ) Design 25 sets of simulation conditions using orthogonal experimental tables; S1.

2. Take the central section of the incinerator along the length of the grate and build a two-dimensional planar geometric model including the grate, furnace, primary air inlet, secondary air inlet and outlet flue. S1.

3. According to the parameter combinations of the above 25 working conditions, set the boundary conditions for primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate one by one; generate transient data for multiple time steps for each working condition; extract the input and output parameters for each time step; S1.

4. Remove invalid data from the initial ignition stage of each working condition, use the IQR method to remove numerical oscillation anomalies, fill in a small number of missing values ​​through linear interpolation, normalize all parameters to the [0,1] interval, and obtain a standardized time-series training dataset adapted to the sliding window MLP.

3. The waste incinerator operating condition prediction and reverse optimization method as described in claim 2, characterized in that, In S1.2, the two-dimensional planar geometric model is a physical model set up based on the ANSYS Fluent 2025 R1 platform: a pressure-based transient solver is used, and gravity is enabled; the turbulence model uses the standard k-ε model, combined with the standard wall function; the radiation model uses the P1 model; the component transport and reaction model and the discrete phase model are enabled, and the waste particle injection source injection-0 is defined; the multiphase flow, NOx, and soot non-essential sub-models are disabled.

4. The method for predicting and reverse-optimizing the operating conditions of a waste incinerator as described in claim 2, characterized in that, In S1.3, the input parameters include: primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate; The output parameters include: average furnace temperature, SO2 concentration, and combustion power.

5. The method for predicting and reverse-optimizing the operating conditions of a waste incinerator as described in claim 2, characterized in that, In S1.4, if supplementary data is needed for a specific working condition, a single-point simulation is carried out within the parameter range covered by the orthogonal experiment.

6. The method for predicting and reverse-optimizing the operating conditions of a waste incinerator as described in claim 1, characterized in that, S2 specifically includes: S2.

1. A sliding window of size 3 is used to generate training samples. The input parameters at times t-3, t-2, and t-1 are concatenated in chronological order to generate a 15-dimensional static feature vector. The furnace average temperature, SO2 concentration, and combustion power at time t are used as output parameters. Starting from the third time step, the window is slid sequentially to generate multiple valid time-series samples. The first 70% of the samples are divided into a training set and the last 30% into a test set in chronological order. S2.

2. Construct a simplified MLP model with a single hidden layer, the structure of which is: 15 neurons in the input layer → 16 neurons in the hidden layer → 3 neurons in the output layer; use mean squared error as the loss function, Adam optimizer, learning rate 0.001, batch size=16, maximum number of iterations 200; set an early stopping strategy with patience=30, stop training early when the validation set loss does not decrease for 30 consecutive iterations, and restore the optimal weights; S2.

3. Input the test set into the trained model, obtain the prediction results, and inversely normalize them to the original physical quantity scale.

7. The waste incinerator operating condition prediction and reverse optimization method as described in claim 6, characterized in that, Based on this, if it is necessary to add prediction metrics, simply modify the number of neurons in the output layer and supplement the corresponding training data.

8. The method for predicting and reverse-optimizing the operating conditions of a waste incinerator as described in claim 1, characterized in that, S3 specifically includes: S3.

1. The objective function is the mean square error between the predicted and target operating conditions; the constraints are the actual operating ranges of each control parameter, consistent with the level range of the orthogonal experiment. The primary air temperature, primary air velocity, secondary air temperature, secondary air velocity, and feed rate are set as optimizable tensor variables, and the initial values ​​are taken as the median values ​​of each factor in the orthogonal experiment. S3.

2. The Adam optimizer is used to calculate the gradient of the objective function with respect to the control variables through the automatic differentiation mechanism of the PyTorch framework, and the control variables are updated iteratively. After each iteration, the control variables are restricted to the physical range verified by orthogonal experiments through hard constraint functions until the objective function converges or the maximum number of iterations is reached. In this way, the optimal control parameters can be obtained by back-deriving the input parameters from the target output parameters, which are used to output the incinerator control commands.

9. A system for predicting and reverse-optimizing the operating conditions of a waste incinerator, characterized in that, The system includes: a training dataset generation module, a working condition forward prediction module, and a reverse optimization module; The training dataset generation module is used to generate L 25 (5) 5 The orthogonal experimental design uses 5 factors and 5 levels for each factor to simulate 25 sets of working conditions, and combines them with a two-dimensional planar CFD model to quickly generate a high-quality training dataset. The positive prediction module for operating conditions is used to construct time-series samples from the standardized time-series training dataset using the sliding window technique, train a lightweight MLP model, and achieve real-time high-precision prediction of furnace average temperature, SO2 concentration, and combustion power. The reverse optimization module is used to load the trained MLP model, fix all weight parameters of the model, and reverse-engineer the optimal control parameters through the gradient descent algorithm to output incinerator control commands.

Images