A physical information-based circulating fluidized bed boiler sulfur dioxide emission interval prediction method and device

Abstract

The present application relates to a kind of based on physical information's circulating fluidized bed boiler sulfur dioxide emission interval prediction method and device, method includes: the operating control parameter of circulating fluidized bed boiler is acquired, operating control parameter is preprocessed, and training data set is constructed;Interval prediction model is trained using training data set, while model parameters are optimized through joint optimization function;In the training process, the feature extraction of training data set is carried out by deep learning feature extraction module, obtain high-dimensional hidden layer space feature matrix, and input to the multiple-site quantile prediction head of interval prediction model to obtain the accurate derivative of state variable to time, the accurate derivative and the prediction state input sulfur dioxide emission mechanism submodel, obtain physical mechanism deviation, to build the joint loss function of fusion physical constraint;The interval prediction model after training optimization is compressed and model deployment, to predict sulfur dioxide emission interval value, obtain sulfur dioxide emission interval prediction value.

Description

Technical Field

[0001] This invention relates to the field of industrial process modeling and environmental emission control technology, and in particular to a method and apparatus for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information. Background Technology

[0002] Circulating fluidized bed (CFB) combustion technology has been widely used in coal-fired power generation due to its wide load adaptability and high fuel flexibility. However, during operation, CFB boilers exhibit highly nonlinear and random fluctuations in sulfur dioxide (SO2) generation and emissions due to load fluctuations, fuel inhomogeneity, and multi-scale coupled reactions of combustion and desulfurization. Accurate prediction of SO2 concentration is not only crucial for achieving ultra-low emission standards but also essential for improving the stability and economy of power plant operation.

[0003] Existing prediction models are mainly divided into mechanistic models and data-driven models. Traditional mechanistic models (such as one-dimensional steady-state models or computational particulate fluid dynamics (CPFD)) predict emissions by simulating the physicochemical mechanisms of desulfurization and combustion, but they often rely on a large number of parameter assumptions, resulting in high computational costs and difficulty in adapting to transient conditions. Data-driven models (such as LSTM, SVR, Transformer, or decision tree algorithms) can effectively capture the temporal dependence of data, but they have significant limitations: First, pure data-driven models are highly susceptible to abnormal fluctuations in multi-step predictions, leading to severe error accumulation; second, these models can only provide predictions at a single deterministic point, failing to quantify the uncertainty of the prediction and providing confidence intervals for emission fluctuations for actual operators; finally, they lack physical mechanism constraints, making them prone to overfitting when encountering unseen operating conditions, and even outputting prediction results that violate the law of mass conservation.

[0004] Physical Information Neural Networks (PINNs) improve data efficiency and physical consistency by embedding control equations into network training. However, existing PINN-based prediction models often focus on deterministic point predictions, making it difficult to quantify the uncertainties in sulfur dioxide emissions from circulating fluidized bed boilers under load fluctuations, fuel composition changes, and combustion-desulfurization coupling reactions. Furthermore, existing deep learning models generally suffer from large parameter sizes and high inference computational complexity. In the real industrial control environment of thermal power plants, factors such as cloud communication latency and limited storage and computing power of edge control devices can affect the online deployment effectiveness of these models.

[0005] Therefore, there is an urgent need for a method and device for predicting sulfur dioxide emissions from circulating fluidized bed boilers that can simultaneously integrate sulfur dioxide emission mechanism constraints, quantile interval prediction, and lightweight deployment mechanisms, in order to improve the physical consistency, interval reliability, and engineering application adaptability of the prediction results. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for predicting sulfur dioxide emissions from circulating fluidized bed boilers based on physical information. By integrating mechanism differential equations, neural network approximation, and quantile regression theory, this invention addresses the problems of severe error accumulation, lack of uncertainty quantification capability, and lack of physical consistency in the output results of existing methods for predicting sulfur dioxide emissions from circulating fluidized bed boilers.

[0007] To achieve the above objectives, the present invention provides the following solution: A method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information includes: Obtain the operating control parameters of the circulating fluidized bed boiler, preprocess the operating control parameters, and construct a training dataset; The interval prediction model is trained using the training dataset, and the model parameters are optimized through a joint optimization function. During the training process, the training dataset is used to extract features through a deep learning feature extraction module to obtain a high-dimensional hidden layer spatial feature matrix, which is then input into the multi-point quantile prediction head of the interval prediction model to obtain the precise derivative of the state variable with respect to time. The precise derivative and the predicted state are then input into the sulfur dioxide emission mechanism sub-model to obtain the physical mechanism bias, so as to construct a joint loss function that integrates physical constraints. The trained and optimized interval prediction model is lightweighted, compressed, and deployed to predict sulfur dioxide emission interval values, thus obtaining the predicted sulfur dioxide emission interval values.

[0008] Optionally, the operating control parameters include: furnace temperature, oxygen concentration, coal feed rate, limestone injection rate, and air flow rate or exhaust volume per unit time; the training dataset also includes sulfur dioxide concentration as an observation label.

[0009] Optionally, the interval prediction model includes: A sulfur dioxide emission mechanism sub-model is used to describe the evolution of combustion carbon mass, effective desulfurizer mass, and sulfur dioxide concentration using the aforementioned operating control parameters; A multi-point quantile prediction head is used to set a quantile set; the quantile set includes: low quantile points, median quantile points, and high quantile points; Based on the set of quantiles, the predicted state at the corresponding quantile is output in parallel through the evolution results; the predicted state includes: predicted values ​​of combustion carbon state variables, predicted values ​​of desulfurizing agent state variables, and predicted values ​​of sulfur dioxide concentration state variables.

[0010] Optionally, using the aforementioned operating control parameters to describe the evolution of combustion carbon mass, effective desulfurizer mass, and sulfur dioxide concentration includes: Describe the evolution of combustion carbon mass: ; Evolution of Effective Desulfurizer Quality: ; Evolution of sulfur dioxide concentration: ; in, This refers to the concentration of unburned carbon in the furnace. For coal feed rate, For furnace temperature, Oxygen concentration, To fix the carbon formation efficiency coefficient. The carbon oxidation rate coefficient, The rate constant for carbon oxidation is temperature-dependent. For effective desulfurizer quality, Limestone injection rate, This represents the volume fraction of sulfur dioxide. The desulfurizer addition coefficient is given. The coefficient of sulfur dioxide absorption reaction. The desulfurization reaction constant is affected by temperature. This refers to the physical loss rate of the desulfurizing agent. This is the initial sulfur dioxide formation rate coefficient. The ratio of sulfur dioxide produced by carbon combustion. The efficiency of desulfurization reaction. This is the exhaust dilution emission factor. Exhaust volume per unit time It is a time variable.

[0011] Optionally, obtaining the high-dimensional hidden layer spatial feature matrix includes: The time variables and preprocessed operation control parameters in the training dataset are used as multi-source heterogeneous inputs to the input layer. The multilayer perceptron architecture containing multiple fully connected layers and nonlinear activation operators in the deep learning feature extraction module is used to perform spatial dimensionality upscaling and layer-by-layer nonlinear feature transformation on the multi-source heterogeneous inputs of the input layer. This is to capture the multi-scale temporal coupling dependency of combustion reaction and desulfurization reaction in circulating fluidized bed boiler under transient conditions and extract a high-dimensional hidden layer spatial feature matrix to characterize the dynamic evolution of the furnace state.

[0012] Optionally, the exact derivative includes: ; in, Indicates quantiles, Represents the set of quantiles. quantiles The next state variable at time t The predicted value.

[0013] Optionally, the physical mechanism deviation includes: Physical residuals of carbon mass evolution during combustion: ; Physical residuals of desulfurizer quality evolution: ; Physical residuals of sulfur dioxide concentration evolution: ; in, The physical residuals of the evolution of carbon mass from combustion. The physical residuals of the desulfurizer quality evolution, The physical residual of sulfur dioxide concentration evolution. For coal feed rate, For furnace temperature, Oxygen concentration, To fix the carbon formation efficiency coefficient. The carbon oxidation rate coefficient, The rate constant for carbon oxidation is temperature-dependent. Limestone injection rate, The desulfurizer addition coefficient is given. The coefficient of sulfur dioxide absorption reaction. The desulfurization reaction constant is affected by temperature. This refers to the physical loss rate of the desulfurizing agent. This is the initial sulfur dioxide formation rate coefficient. The ratio of sulfur dioxide produced by carbon combustion. The efficiency of desulfurization reaction. This is the exhaust dilution emission factor. Exhaust volume per unit time , and Quantities are respectively represented Predicted values ​​of combustion carbon state variables, desulfurizer state variables, and sulfur dioxide concentration state variables.

[0014] Optionally, the lightweight compression and deployment of the trained and optimized interval prediction model include: The multi-bit floating-point parameters containing weights and biases in the trained and optimized interval prediction model are converted to eight-bit integers by a layer-by-layer dynamic quantization algorithm. A sensitivity-guided structured pruning strategy is used to remove redundant connections and feature channels in the interval prediction model. At the same time, a hardware-aware operator fusion technology is used to merge and compile adjacent linear layers and activation layers in the interval prediction model into a composite operator, resulting in a lightweight interval prediction model. The lightweight interval prediction model is deployed to a computing device, which includes an edge hardware-accelerated inference engine deployed in an industrial field to receive real-time operating control parameters and input them into the deployed interval prediction model for real-time inference calculation.

[0015] To achieve the above objectives, the present invention also provides a device for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information, comprising: The acquisition module is used to acquire the operating control parameters of the circulating fluidized bed boiler, preprocess the operating control parameters, and construct a training dataset. An optimization module is used to train an interval prediction model using the training dataset and optimize the model parameters through a joint optimization function. During the training process, a deep learning feature extraction module extracts features from the training dataset to obtain a high-dimensional hidden layer spatial feature matrix, which is then input into the multi-point quantile prediction head of the interval prediction model to obtain the precise derivative of the state variable with respect to time. The precise derivative and the predicted state are then input into the sulfur dioxide emission mechanism sub-model to obtain the physical mechanism bias, so as to construct a joint loss function that integrates physical constraints. The prediction deployment module is used to compress the trained and optimized interval prediction model into a lightweight form and deploy the model to predict the interval values ​​of sulfur dioxide emissions, thus obtaining the predicted values ​​of sulfur dioxide emissions.

[0016] The beneficial effects of this invention are as follows: This invention utilizes a deep learning output head architecture based on quantile regression to not only output the predicted median but also simultaneously output the upper and lower bounds of the confidence interval. This quantifies the emission uncertainties caused by fuel heterogeneity and load fluctuations, providing a safe operating margin for on-site chemical dosing control. The mechanistic equation embedding module incorporates the combustion carbon kinetics of sulfur dioxide generation and the desulfurizer reaction kinetics as physical residual constraints into the neural network optimization. This improves the consistency between multi-step prediction trajectories and physicochemical evolution laws, reduces the risk of error accumulation and non-physical fluctuations that are easily generated in long-sequence predictions by pure data-driven models, and enhances the model's cross-scenario generalization ability under unseen operating conditions.

[0017] This invention reduces the model's parameter size, computational complexity, and memory overhead by introducing model lightweighting techniques such as layer-by-layer dynamic quantization, sensitivity-guided structured pruning, and hardware-aware operator fusion, while maintaining prediction interval coverage and accuracy. Furthermore, based on a cloud-edge collaborative deployment architecture, it improves the model's real-time predictive adaptability in dynamic industrial environments by combining offline cloud training with hardware-accelerated inference at the edge. This helps thermal power plants optimize desulfurizer dosing strategies and reduce the risk of excessive emissions, thus possessing environmental benefits, engineering implementation value, and economic value. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.

[0019] Figure 1 This is a flowchart of a method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information, according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown in the figure, this embodiment discloses a method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information, including the following steps: S1. Obtain continuous operating data of the circulating fluidized bed boiler and preprocess it to construct training and test datasets; S2. Based on the preprocessed dataset, construct a quantile regression-based physical information neural network (QR-PINN) interval prediction model. The interval prediction model includes a sulfur dioxide emission mechanism model, a deep learning feature extraction module, a multi-site quantile prediction head, and a multi-site physical residual constraint module. S3. Construct a global joint optimization loss function that integrates physical mechanism constraints based on the interval prediction model, and iteratively train and optimize the interval prediction model. S4. Lightweight compression is performed on the trained and optimized interval prediction model, and the compressed model is deployed to the computing device. Combined with real-time operation control parameters, the predicted values ​​of sulfur dioxide emissions intervals are output, consisting of low quantile prediction values, median prediction values, and high quantile prediction values.

[0023] Furthermore, S1 includes: Acquire continuous time series sensor data, which includes sulfur dioxide concentration as an observation label, as well as operating control parameters such as furnace temperature, oxygen concentration, coal feed rate, limestone injection rate, air flow rate, and exhaust volume per unit time. The time-series sensor data is smoothed using a moving average method to filter out high-frequency random noise; By using a pruning method based on the statistical sigma rule, abnormal peak values ​​caused by impulse interference or transient conditions are compressed; The preprocessed time series data is divided into training and test sets according to a strict time order and a preset ratio to prevent data leakage. During model training, the preprocessed sulfur dioxide concentration is used as the observation label and together with the predicted sulfur dioxide concentration output by the multi-site quantile prediction head, it is used to calculate the data regression loss. During the model inference process, the time variable and the aforementioned operation control parameters are used as model inputs, and the range prediction value of sulfur dioxide concentration is output.

[0024] Furthermore, S2 includes: Based on the pre-treated operating control parameters, a sulfur dioxide emission mechanism model was constructed, which includes the physical evolution laws of combustion carbon, desulfurizing agent, and sulfur dioxide concentration. Based on the Physical Information Neural Network (PINN) framework, a deep learning feature extraction module is constructed to approximate the time evolution of state variables related to sulfur dioxide emissions. Based on the output of the deep learning feature extraction module, a multi-site quantile prediction head is constructed to simultaneously output multiple confidence lower bounds, medians, and upper bounds. Based on the sulfur dioxide emission mechanism model and the output of the multi-site quantile prediction head, a multi-site physical residual constraint module based on an automatic differentiation mechanism is constructed.

[0025] Furthermore, methods for constructing sulfur dioxide emission mechanism models include: A first-order differential equation for the evolution of combustion carbon based on fixed carbon generation and carbon oxidation consumption affected by furnace temperature and oxygen concentration is established. Establish a differential equation for the evolution of desulfurizing agent based on limestone addition, sulfur dioxide absorption and consumption affected by furnace temperature, and physical losses; A differential equation for the evolution of sulfur dioxide concentration was established based on coal input, combustion carbon oxidation generation, desulfurizer absorption and removal, and gas exhaust dilution.

[0026] Furthermore, methods for constructing deep learning feature extraction modules to approximate the time evolution of state variables related to sulfur dioxide emissions include: The time variable and the pre-processed operation control parameters are used as multi-source heterogeneous inputs for the input layer. The operation control parameters include at least coal feed rate, limestone injection rate, furnace temperature, oxygen concentration and exhaust volume per unit time. By utilizing a multilayer perceptron architecture containing multiple fully connected layers and nonlinear activation operators, spatial dimensionality upscaling and layer-by-layer nonlinear feature transformation are performed on the multi-source heterogeneous inputs of the input layer to capture the multi-scale temporal coupling dependency of combustion reaction and desulfurization reaction in circulating fluidized bed boiler under transient conditions, and extract a high-dimensional hidden layer spatial feature matrix to characterize the dynamic evolution of the furnace state. The high-dimensional hidden layer spatial feature matrix is ​​used as a shared feature input and synchronously transmitted to the subsequent multi-site quantile prediction head. This is used to approximate the time evolution of sulfur dioxide emission-related state variables under the real physicochemical manifold through the iteration of network weights.

[0027] Furthermore, methods for constructing multi-site quantile prediction heads include: Multiple quantile points are set, and each quantile point corresponds to an independent output branch of the neural network; The outputs of different quantile branches are used to characterize the predicted values ​​of sulfur dioxide emission-related state variables at the corresponding quantiles. The lower quantile, median, and upper quantile are used to determine the lower bound of the confidence interval, the median of the expected trend, and the upper bound of the confidence interval for the sulfur dioxide emission interval prediction results, respectively. The time derivative of the state variable output by each quantile prediction branch is obtained through the automatic differentiation operation inside the neural network.

[0028] Furthermore, the method for constructing a multi-site physical residual constraint module includes: Substitute the quantile state variables and their automatic time derivatives output by the multi-site quantile prediction head into the sulfur dioxide emission mechanism model to calculate the physical residual of the combustion carbon mass evolution at each quantile. Based on the quantile state variables, automatic differential derivatives, and state variables of limestone injection rate, furnace temperature, and sulfur dioxide concentration, calculate the physical residual of the desulfurizing agent mass evolution at each quantile. Based on the quantile state variables, automatic differential derivatives, coal feed rate, exhaust volume per unit time, and desulfurizer state variables, the physical residuals of sulfur dioxide concentration evolution at each quantile are calculated.

[0029] Furthermore, S3 includes: Based on the predicted values ​​and actual observed values ​​at different quantiles, an asymmetric penalized pinball data regression loss is constructed. Based on the multi-site physical residual constraint module, a multi-site physical residual loss is constructed to constrain quantile state variables to satisfy the sulfur dioxide emission mechanism model. Based on the relationship between the high and low quantiles, a maximum value function is used to construct an interval consistency loss that prevents quantile curves from crossing. The data regression loss, multi-site physical residual loss, initial condition loss, boundary condition loss, and interval consistency loss are combined to obtain the overall joint optimization loss function. Based on the overall joint optimization loss function, the interval prediction model is backpropagated and the parameters are iteratively updated to improve the consistency between the model prediction results and the physical evolution law of sulfur dioxide generation and desulfurization.

[0030] Furthermore, S4 includes: Computing devices include edge hardware-accelerated inference engines deployed in industrial settings; The cloud-edge collaborative architecture is used to deploy the execution model. Offline training and parameter optimization of the physical information neural network model are performed on the cloud server, and the trained, optimized and lightweight compressed model is sent to the edge hardware-accelerated inference engine. The edge hardware-accelerated inference engine receives real-time operating control parameters and performs real-time inference calculations based on the distributed model, outputting predicted values ​​for sulfur dioxide emission ranges.

[0031] Furthermore, the trained and optimized physical information neural network model is subjected to lightweight compression, including: The 32-bit floating-point parameters containing weights and biases in the physical information neural network model are converted into 8-bit integer parameters through a layer-by-layer dynamic quantization algorithm. A sensitivity-guided structured pruning strategy is adopted to remove redundant connections and feature channels in the network model; Hardware-aware operator fusion technology is used to merge adjacent linear layers and activation layers in the model into a composite operator.

[0032] This embodiment also provides a device for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information, including: The data acquisition module is used to acquire continuous operating data of the circulating fluidized bed boiler and preprocess it to build training and test datasets. The module is used to build a physical information neural network interval prediction model based on quantile regression based on the preprocessed dataset. The interval prediction model includes a sulfur dioxide emission mechanism model, a deep learning feature extraction module, a multi-site quantile prediction head, and a multi-site physical residual constraint module. The optimization module is used to construct a total joint optimization loss function based on the interval prediction model, which includes Pinball data regression loss, multi-site physical residual loss, initial condition loss, boundary condition loss and interval consistency loss, and to iteratively train and optimize the interval prediction model. The prediction deployment module is used to perform lightweight compression on the trained and optimized interval prediction model, and deploy the compressed lightweight model to the edge hardware-accelerated inference engine in the cloud-edge collaborative architecture, and output the predicted value of sulfur dioxide emission intervals in combination with real-time running control parameters.

[0033] In one possible implementation, this embodiment discloses a method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information, the steps of which include: S1. Obtain the continuous emission operation data of the circulating fluidized bed boiler, preprocess it, and then construct a dataset.

[0034] Multimodal operating data of the circulating fluidized bed boiler was acquired, preprocessed, and then used to construct a dataset. The operating data was collected from the Continuous Emission Monitoring System (CEMS) of the 330MW circulating fluidized bed boiler. Continuous time-series sensor data were sampled at 10-second intervals, including sulfur dioxide concentration as the observation label, as well as operating control parameters such as furnace temperature, oxygen concentration or air supply rate, coal feed rate, limestone injection rate, and air flow rate and / or exhaust volume per unit time. The preprocessing process included: S11. The time series sensor data is smoothed using the moving average method to effectively filter out high-frequency random noise and reveal the deep trend of the data. S12. By using a shaving and constraint method based on the statistical Sigma rule, abnormal data peaks caused by pulse electromagnetic interference or instantaneous system start-up and shutdown are compressed, thereby improving the stability and reliability of the data. S13. Extract the first 300 to 400 sampling points from the preprocessed data in strict chronological order as the training set, and use the remaining samples as the test set and validation set, in order to avoid data leakage caused by using data from future time points during the training phase.

[0035] S2. Based on the preprocessed dataset, construct a quantile regression-based physical information neural network (QR-PINN) interval prediction model.

[0036] A QR-PINN prediction model was constructed, which includes a deep learning network feature extraction module, a sulfur dioxide emission mechanism model, a multi-site quantile prediction head, and a multi-site physical residual constraint module.

[0037] Furthermore, the specific implementation of the deep learning network feature extraction module includes: In time and control variables (coal feed rate) limestone injection rate ,temperature Oxygen concentration Exhaust volume Using the network input, a multilayer perceptron containing a fully connected layer is used to extract hidden layer features, resulting in a high-dimensional hidden layer spatial feature matrix that characterizes the coupling state of combustion reaction and desulfurization reaction in a circulating fluidized bed boiler. The high-dimensional hidden layer spatial feature matrix is ​​used to approximate the time evolution of sulfur dioxide emission-related state variables.

[0038] Furthermore, the specific implementation of the sulfur dioxide emission mechanism model includes: Based on the kinetics of carbon fixation oxidation and desulfurization reactions in CFB boilers, a first-order differential equation describing the evolution of combustion carbon mass is established: ; in, The concentration of unburned carbon in the furnace; Coal feed rate; The furnace temperature; Oxygen concentration; To fix the carbon generation efficiency coefficient; The carbon oxidation rate coefficient; is the temperature-dependent rate constant for carbon oxidation.

[0039] Establish a differential equation describing the evolution of the effective desulfurizer's quality: ; in, For effective desulfurizer quality; Limestone injection rate; This represents the volume fraction of sulfur dioxide. The desulfurizing agent addition coefficient; The coefficient of sulfur dioxide absorption reaction; The desulfurization reaction constant is affected by temperature; This represents the physical loss rate of the desulfurizing agent.

[0040] Establish a differential equation describing the evolution of sulfur dioxide concentration: ; in, This is the initial sulfur dioxide formation rate coefficient; The coefficient for the proportion of sulfur dioxide produced by carbon combustion; The desulfurization reaction efficiency; This refers to the exhaust dilution emission factor; This refers to the exhaust volume per unit time.

[0041] The sulfur dioxide emission mechanism model is used to constrain the training of the interval prediction model. Specifically, through the multi-site physical residual constraint module, the predicted values ​​of combustion carbon state variables, desulfurizer state variables, and sulfur dioxide concentration state variables at each quantile output by the multi-site quantile prediction head, along with their automatic time derivatives, are substituted into the corresponding combustion carbon mass evolution equation, effective desulfurizer mass evolution equation, and sulfur dioxide concentration evolution equation to obtain the corresponding physical residuals. Based on these physical residuals, a multi-site physical residual loss is constructed, and this loss is added to the overall joint optimization loss function to train and optimize the interval prediction model.

[0042] Furthermore, the specific implementation of the multi-site quantile prediction head and automatic differentiation mechanism includes: Set quantile set ,in , and These correspond to the lower quantile, median quantile, and upper quantile, respectively. The network branches output the predicted state at the corresponding quantile in parallel: ; in, This represents the operating control parameters consisting of coal feed rate, limestone injection rate, furnace temperature, oxygen concentration, and exhaust volume per unit time. , and Quantities are respectively represented The predicted values ​​for the state variables of combustion carbon, desulfurizer, and sulfur dioxide concentration are given. The lower quantile, median, and upper quantile predicted values ​​of the sulfur dioxide concentration state variable constitute the lower bound, median of the expected trend, and upper bound of the confidence interval, respectively.

[0043] By utilizing the automatic differentiation mechanism of deep learning frameworks, the exact derivative of the state variable with respect to time can be obtained directly without numerical differencing: ; in, Indicates quantiles, Represents the set of quantiles. quantiles The next state variable at time t The predicted values, wherein the state variables include combustion carbon state variables, desulfurizer state variables, and sulfur dioxide concentration state variables.

[0044] Furthermore, the specific implementation of the multi-site physical residual constraint module includes: Substituting the automatic differential derivative and the network state output into the above mechanism model, we calculate each quantile. The following physical mechanism deviation: Calculate the physical residuals of the combustion carbon mass evolution : ; Calculate the physical residuals of desulfurizer quality evolution : ; Calculate the physical residuals of sulfur dioxide concentration evolution : ; S3. Based on the preprocessed dataset and the constructed model, construct a joint loss function that incorporates physical constraints, and train and optimize the model.

[0045] Construct a global joint loss function that includes asymmetric quantile regression loss (Pinball loss), physical mechanism residual loss, initial and boundary condition loss, and interval consistency constraints.

[0046] The specific formula is as follows: quantile regression data loss. Defined as: ; ; in, For the time set of observation data; To observe the true value; For prediction error residuals; This is a pinball penalty function that assigns asymmetric weights to positive and negative errors.

[0047] Multi-site physical residual loss Defined as: ; in, This is the collocation set used to calculate the physical residuals.

[0048] Initial conditional loss for each quantile Defined as: ; in, The corresponding known initial values.

[0049] Interval consistency regularization loss (Used to prevent the lower quantile prediction curve from being larger than the higher quantile prediction curve) is defined as: in, These are the corresponding high and low quantile points.

[0050] Final joint optimization loss function The combinations are as follows: in, The balancing weight coefficients contributing to each loss are used. The gradient of the overall joint optimization loss function with respect to the network parameters is calculated using the AdamW optimizer, and the parameters of the interval prediction model are iteratively updated through backpropagation to improve the consistency between the model prediction results and the physical evolution of sulfur dioxide generation and desulfurization.

[0051] S4. Lightweight compression is performed on the trained and optimized physical information neural network model, and the compressed model is deployed to the computing device. Combined with real-time operation control parameters, the predicted value of sulfur dioxide emission range with statistical confidence is output.

[0052] In this embodiment, the computing device includes an edge hardware-accelerated inference engine deployed in the industrial field. In actual industrial scenarios, model deployment is performed based on a cloud-edge collaborative architecture: offline training and parameter optimization of the physical information neural network model are performed on the cloud server, and the optimized and compressed model is then distributed to the edge hardware-accelerated inference engine for real-time inference computation.

[0053] Specifically, before deploying the model to the edge, a hardware-aware optimization pipeline is used to perform lightweight compression on the model, including: First, the 32-bit floating-point parameters containing weights and biases in the physical information neural network model are converted into 8-bit integer parameters through a layer-by-layer dynamic quantization algorithm, thereby reducing the overall size of the model while maintaining the interval prediction coverage and prediction accuracy.

[0054] Secondly, a sensitivity-guided structured pruning strategy is adopted. By evaluating the sensitivity of each layer channel of the network to the prediction results, redundant connections and feature channels in the network model are eliminated, thereby reducing memory bandwidth usage.

[0055] Finally, hardware-aware operator fusion technology is used to merge adjacent linear layers and activation layers in the model into a single composite operator at the bottom layer, which significantly reduces inference latency.

[0056] The optimized lightweight inference model is deployed on an edge coprocessor array consisting of a real-time physical control system and multiple FPGA hardware acceleration boards. The field real-time control system continuously acquires online operational data streams from multiple industrial sensors, including the continuous emission monitoring system, via digital and analog interfaces. After execution signal synchronization and preprocessing, the data is streamed to the tightly coupled FPGA array. The FPGA employs a parallel hardware pipeline architecture to perform parallel inference of the lightweight model, synchronously outputting the lower quantile bound, median expected trend, and upper quantile bound of sulfur dioxide concentration. These prediction intervals are ultimately sent to the distributed control system of the power plant to guide the closed-loop control of the on-site desulfurizer dosing strategy.

[0057] This embodiment also provides a physical information-based circulating fluidized bed boiler sulfur dioxide emission interval prediction device. This device is used to implement the methods described in the embodiment, including: a data acquisition module for acquiring continuous operating data of the circulating fluidized bed boiler and preprocessing it to construct a training dataset and a test dataset; a construction module for constructing a physical information neural network interval prediction model based on quantile regression according to the preprocessed dataset. The interval prediction model includes a sulfur dioxide emission mechanism model, a deep learning feature extraction module, a multi-site quantile prediction head, and a multi-site physical residual constraint module; an optimization module for constructing a comprehensive joint optimization loss function including Pinball data regression loss, multi-site physical residual loss, initial condition loss, boundary condition loss, and interval consistency loss according to the interval prediction model, and iteratively training and optimizing the interval prediction model; and a prediction deployment module for lightweight compression of the trained and optimized interval prediction model, and deploying the compressed lightweight model to the edge hardware-accelerated inference engine in a cloud-edge collaborative architecture, combining real-time operating control parameters to output sulfur dioxide emission interval prediction values.

[0058] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information, characterized in that, include: Obtain the operating control parameters of the circulating fluidized bed boiler, preprocess the operating control parameters, and construct a training dataset; The interval prediction model is trained using the training dataset, and the model parameters are optimized using a joint optimization function. During training, the training dataset is used to extract features through the deep learning feature extraction module to obtain a high-dimensional hidden layer spatial feature matrix, which is then input into the multi-site quantile prediction head of the interval prediction model to obtain the precise derivative of the state variable with respect to time. The precise derivative and the predicted state are then input into the sulfur dioxide emission mechanism sub-model to obtain the physical mechanism bias, so as to construct a joint loss function that integrates physical constraints. The trained and optimized interval prediction model is lightweighted, compressed, and deployed to predict sulfur dioxide emission interval values, thus obtaining the predicted sulfur dioxide emission interval values.

2. The method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information as described in claim 1, characterized in that, The operating control parameters include: furnace temperature, oxygen concentration, coal feed rate, limestone injection rate, and air flow rate or exhaust volume per unit time; the training dataset also includes sulfur dioxide concentration as an observation label.

3. The method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information according to claim 1, characterized in that, The interval prediction model includes: A sulfur dioxide emission mechanism sub-model is used to describe the evolution of combustion carbon mass, effective desulfurizer mass, and sulfur dioxide concentration using the aforementioned operating control parameters; A multi-point quantile prediction head is used to set a quantile set; the quantile set includes: low quantile points, median quantile points, and high quantile points; Based on the set of quantiles, the predicted state at the corresponding quantile is output in parallel through the evolution results; the predicted state includes: predicted values ​​of combustion carbon state variables, predicted values ​​of desulfurizing agent state variables, and predicted values ​​of sulfur dioxide concentration state variables.

4. The method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information according to claim 3, characterized in that, The operational control parameters described above include: the evolution of combustion carbon mass, the evolution of effective desulfurizer mass, and the evolution of sulfur dioxide concentration. Describe the evolution of combustion carbon mass: ； Evolution of Effective Desulfurizer Quality: ； Evolution of sulfur dioxide concentration: ； in, This refers to the concentration of unburned carbon in the furnace. For coal feed rate, For furnace temperature, Oxygen concentration, To fix the carbon formation efficiency coefficient. The carbon oxidation rate coefficient, The rate constant for carbon oxidation is temperature-dependent. For effective desulfurizer quality, Limestone injection rate, This represents the volume fraction of sulfur dioxide. The desulfurizer addition coefficient is given. The coefficient of sulfur dioxide absorption reaction. The desulfurization reaction constant is affected by temperature. This refers to the physical loss rate of the desulfurizing agent. This is the initial sulfur dioxide formation rate coefficient. The ratio of sulfur dioxide produced by carbon combustion. The efficiency of desulfurization reaction. This is the exhaust dilution emission factor. Exhaust volume per unit time It is a time variable.

5. The method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information according to claim 1, characterized in that, The high-dimensional hidden layer space feature matrix is ​​obtained by: The time variables and preprocessed operation control parameters in the training dataset are used as multi-source heterogeneous inputs to the input layer. The multilayer perceptron architecture containing multiple fully connected layers and nonlinear activation operators in the deep learning feature extraction module is used to perform spatial dimensionality upscaling and layer-by-layer nonlinear feature transformation on the multi-source heterogeneous inputs of the input layer. This is to capture the multi-scale temporal coupling dependency of combustion reaction and desulfurization reaction in circulating fluidized bed boiler under transient conditions and extract a high-dimensional hidden layer spatial feature matrix to characterize the dynamic evolution of the furnace state.

6. The method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information according to claim 1, characterized in that, The exact derivative includes: ； in, Indicates quantiles, Represents the set of quantiles. quantiles The next state variable at time t The predicted value.

7. The method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information according to claim 1, characterized in that, The physical mechanism deviations include: Physical residuals of carbon mass evolution during combustion: ； Physical residuals of desulfurizer quality evolution: ； Physical residuals of sulfur dioxide concentration evolution: ； in, The physical residuals of the evolution of carbon mass from combustion. The physical residuals of the desulfurizer quality evolution, The physical residual of sulfur dioxide concentration evolution. For coal feed rate, For furnace temperature, Oxygen concentration, To fix the carbon formation efficiency coefficient. The carbon oxidation rate coefficient, The rate constant for carbon oxidation is temperature-dependent. Limestone injection rate, The desulfurizer addition coefficient is given. The coefficient of sulfur dioxide absorption reaction. The desulfurization reaction constant is affected by temperature. This refers to the physical loss rate of the desulfurizing agent. This is the initial sulfur dioxide formation rate coefficient. The ratio of sulfur dioxide produced by carbon combustion. The efficiency of desulfurization reaction. This is the exhaust dilution emission factor. Exhaust volume per unit time , and Quantities are respectively represented Predicted values ​​of combustion carbon state variables, desulfurizer state variables, and sulfur dioxide concentration state variables.

8. The method for predicting sulfur dioxide emission ranges from a circulating fluidized bed boiler based on physical information according to claim 1, characterized in that, The lightweight compression and deployment of the trained and optimized interval prediction model include: The multi-bit floating-point parameters containing weights and biases in the trained and optimized interval prediction model are converted to eight-bit integers by a layer-by-layer dynamic quantization algorithm. A sensitivity-guided structured pruning strategy is used to remove redundant connections and feature channels in the interval prediction model. At the same time, a hardware-aware operator fusion technology is used to merge and compile adjacent linear layers and activation layers in the interval prediction model into a composite operator, resulting in a lightweight interval prediction model. The lightweight interval prediction model is deployed to a computing device, which includes an edge hardware-accelerated inference engine deployed in an industrial field to receive real-time operating control parameters and input them into the deployed interval prediction model for real-time inference calculation.

9. A device for predicting the range of sulfur dioxide emissions from a circulating fluidized bed boiler based on physical information, implemented according to any one of claims 1-8, characterized in that, include: The acquisition module is used to acquire the operating control parameters of the circulating fluidized bed boiler, preprocess the operating control parameters, and construct a training dataset. The optimization module is used to train the interval prediction model using the training dataset and optimize the model parameters through a joint optimization function. During training, the training dataset is used to extract features through the deep learning feature extraction module to obtain a high-dimensional hidden layer spatial feature matrix, which is then input into the multi-site quantile prediction head of the interval prediction model to obtain the precise derivative of the state variable with respect to time. The precise derivative and the predicted state are then input into the sulfur dioxide emission mechanism sub-model to obtain the physical mechanism bias, so as to construct a joint loss function that integrates physical constraints. The prediction deployment module is used to compress the trained and optimized interval prediction model into a lightweight form and deploy the model to predict the interval values ​​of sulfur dioxide emissions, thus obtaining the predicted values ​​of sulfur dioxide emissions.

Images