Method for predicting and adjusting temperature of low-temperature catalytic desulfurization and denitrification device based on deep learning

Abstract

The invention provides a temperature prediction and adjustment method for a low-temperature catalytic desulfurization and denitrification device based on deep learning, and belongs to the technical field of temperature prediction based on deep learning. The method comprises the following steps: firstly, acquiring nine types of multi-source time series data such as flue gas inlet parameters, ammonia injection rate, heating power and the like based on a physical decoupling idea, and constructing a temperature characteristic data set; secondly, designing a feature extraction module based on physical decoupling, and respectively modeling convective heat transfer, catalytic heat release and thermal inertia processes through three branch networks; thirdly, constructing a space-time decoding prediction module, fusing multi-physical process characteristics and predicting a future temperature field; and finally, designing a multi-stage risk early warning and intelligent adjusting module, and generating a control instruction in combination with gradient optimization to realize temperature regulation and control. According to the method, the temperature prediction precision and the physical interpretability are improved, advanced risk prediction and adjustment are realized, and the engineering applicability is high.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of temperature prediction based on deep learning, and particularly relates to a low-temperature catalytic desulfurization and denitrification device temperature prediction and regulation method based on deep learning. BACKGROUND

[0002] With the increasingly stringent environmental protection regulations and the deepening of industrial green transformation, precise control of flue gas purification equipment has become the key. As the core equipment for achieving ultra-low emission, the stability of the temperature field of the low-temperature catalytic desulfurization and denitrification device directly determines the catalytic reaction efficiency, removal rate and catalyst life. The low-temperature catalyst is extremely sensitive to temperature changes: too low temperature will cause catalytic inactivation, and too high temperature will cause sintering risk, and sulfur oxides are easy to cause catalyst poisoning. Therefore, precise prediction and dynamic regulation of the catalyst bed temperature field are the core technology to achieve efficient removal, prolong catalyst life and reduce energy consumption. The existing low-temperature catalytic device temperature control methods mainly include the following three types: methods based on thermodynamic mechanism model, methods based on traditional feature engineering and statistical modeling, and methods based on deep learning.

[0003] Methods based on thermodynamic mechanism model: this method establishes partial differential equation sets of flue gas convection heat transfer, catalytic reaction heat release and system thermal inertia to simulate the temperature distribution of the catalyst bed. For example, the local temperature field is predicted by coupling the energy conservation equation with the reaction kinetics model. This method has a clear physical meaning, but relies on simplified assumptions such as ignoring bed heterogeneity, has high modeling complexity for multi-physical field strong coupling process convection-reaction-thermal compensation, and is difficult to calibrate model parameters in real time, so its applicability is limited in actual variable working condition scenarios.

[0004] Methods based on traditional feature engineering and statistical modeling: this method extracts static features such as temperature mean, fluctuation variance or time series features such as sliding window trend from historical operation data, and combines regression models such as support vector regression or rule systems for temperature prediction. For example, the average bed temperature is predicted by linear combination of flue gas flow and inlet temperature. This method reduces the modeling complexity, but its feature design relies on expert experience, and it is difficult to effectively capture the deep coupling relationship between multi-source heterogeneous data temperature field, gas concentration, ammonia injection rate, etc., and the prediction accuracy of nonlinear dynamics such as thermal inertia hysteresis effect and local hot spot mutation is insufficient.

[0005] Deep learning based methods: In recent years, with the development of industrial internet of things technology, data-driven methods have become a research hotspot. This method uses deep neural networks such as LSTM and CNN to automatically learn the spatio-temporal correlation of multi-source sensor data. For example, GRU network is used to model the temporal dependence of temperature, or CNN is used to process the spatial temperature distribution of thermocouple array. Deep learning models show advantages in complex pattern learning, avoiding the limitations of manual feature design. However, existing methods still have significant shortcomings: 1) Lack of decoupling of multiple physical processes: Most models do not explicitly separate convection, reaction, thermal inertia and other core physical processes, resulting in mixed feature learning; 2) Insufficient dynamic modeling capability: traditional RNN is difficult to accurately describe the long-term evolution of thermal inertia; 3) Weak early warning mechanism: the prediction result lacks linkage with risk level. SUMMARY

[0006] To solve the problems existing in the prior art, the present application provides a deep learning based low-temperature catalytic desulfurization and denitrification device temperature prediction and regulation method, comprising the following steps: S1, collecting convective heat transfer process feature data, catalytic reaction heat release process feature data, system thermal inertia and heat compensation process feature data related to the temperature of the low-temperature catalytic desulfurization and denitrification device; S2, extracting features from the three kinds of feature data collected in S1 through a catalytic reaction temperature feature extraction module, including a time series feature extraction network and a convective heat transfer process feature extraction branch network, a catalytic reaction heat release process feature extraction branch network and a system thermal inertia and heat compensation process feature extraction branch network shared with the network, obtaining a convective heat transfer depth feature vector, a catalytic reaction heat release depth feature vector and a system thermal inertia depth feature vector; S3, inputting the three kinds of feature vectors obtained in S2 into a physical process depth feature interaction and fusion network for feature fusion and nonlinear transformation to obtain a comprehensive thermodynamic state vector, and then inputting the comprehensive thermodynamic state vector into a comprehensive thermodynamic reasoning network for spatio-temporal feature evolution reasoning to obtain a temperature field hidden state feature map; inputting the temperature field hidden state feature map into a temperature field decoding regression network based on a spatial attention mechanism for temperature field key region feature capture and regression mapping to obtain a catalyst bed internal temperature field prediction matrix at a future preset time step; S4, generating a level warning based on the catalyst bed internal temperature field prediction matrix and triggering risk evolution trend calculation, and generating an optimal control instruction vector including an ammonia injection rate adjustment value and a heating system power adjustment value based on the intelligent adjustment strategy of model gradient optimization.

[0007] Preferably, the convective heat transfer process characteristic data includes flue gas inlet temperature data, flue gas inlet flow rate data, and inlet oxygen concentration data; the catalytic reaction exothermic process characteristic data includes inlet nitrogen oxide concentration data, inlet sulfur oxide concentration data, inlet oxygen concentration data, ammonia injection rate data, and multi-point temperature field data inside the catalyst bed; and the system thermal inertia and heat compensation process characteristic data includes heating system power data and ambient temperature data. At the inlet of the device flue gas main, a temperature sensor is installed to collect flue gas inlet temperature data in real time, a Pitot tube flowmeter is installed to collect flue gas inlet flow rate data in real time, and a gas analyzer is installed to collect inlet nitrogen oxide concentration data, inlet sulfur oxide concentration data, and inlet oxygen concentration data in real time; the frequency signal or flowmeter signal of the ammonia water injection pump of the denitration system is collected, which is recorded as ammonia injection rate data; the power signal of the electric heater of the flue gas reheating system is collected, which is recorded as heating system power data; an array of armored thermocouple sensors is arranged at different depths and radial positions of the catalyst bed to collect multi-point temperature field data inside the catalyst bed in real time; and ambient temperature data is collected near the device shell for quantifying system heat loss.

[0008] Preferably, the time sequence feature extraction network is composed of a feature alignment layer and two serially connected gated recurrent unit (GRU) networks; first, all the data collected in S1 is input into the feature alignment layer, and the timestamps of the temperature field data with the highest sampling frequency are used as the reference to resample the remaining data using a cubic spline interpolation algorithm; then, the aligned multi-source data is input into the double-layer GRU network, which uses its gating mechanism to capture the long-term time dependence of each working condition parameter, and finally outputs a set of high-dimensional universal time sequence deep features containing all the input information.

[0009] Preferably, the convective heat transfer process feature extraction branch network first converts the inlet temperature time sequence deep features, inlet flow rate time sequence deep features, and oxygen concentration time sequence deep features output by the time sequence feature extraction network into continuous time sequence deep features through Gaussian process interpolation, and defines the continuous deep features as the control signal path of the convective heat transfer process; secondly, defines the initial convective heat transfer process hidden state as a zero vector with 32 dimensions; thirdly, uses a multilayer perceptron with three hidden layers as a vector field function network, and inputs the initial convective heat transfer process hidden state and the control signal path of the convective heat transfer process as inputs to output the rate of change of the convective heat transfer process hidden state, i.e., the derivative of the convective heat transfer process hidden state, which is parameterized by the vector field function network; fourthly, integrates the kinetic equation parameters describing the evolution of the convective heat transfer process hidden state over time and the control signal path of the convective heat transfer process parameterized by the vector field function network within the observation period through a differentiable ordinary differential equation solver to obtain the convective heat transfer deep feature vector.

[0010] Preferably, the catalytic reaction exothermic process feature extraction branch network splices the inlet nitrogen oxide concentration time series deep features, the inlet sulfur oxide concentration time series deep features, and the inlet oxygen concentration time series deep features output by the time series feature extraction network to form a reactant feature sequence; secondly, the ammonia injection rate time series deep features as the core control variable are spliced with the catalyst bed temperature field time series deep features as the key reaction condition to form a condition control feature sequence; and thirdly, a dynamic parameter generation network is used, with the reactant feature sequence as the input of a single hidden layer multilayer perceptron and the condition control feature sequence as the input of a linear activation layer, a second linear activation layer, and a third linear activation layer for outputting a set of state space parameters for controlling the feature extraction process of the reactant feature sequence; the state space parameters include a selection step, an input modulation matrix, an output selection matrix, and a state transition matrix. The dynamic parameter generation network performs selective scanning processing on the reactant feature sequence along the time dimension to obtain a reactant process representation sequence composed of instantaneous output feature vectors at each time; the reactant process representation sequence is integrated by a global average pooling layer to generate a reactant comprehensive process feature vector; finally, the reactant comprehensive process feature vector is further feature-extracted and dimension-adjusted by an output mapping network composed of a fully connected layer and a ReLU activation function, and a catalytic reaction exothermic deep feature vector that can finely represent the complex coupling relationship between reactants and the dynamic exothermic effect under different operating conditions is output.

[0011] Preferably, the system thermal inertia and thermal compensation process feature extraction branch network inputs the heating system power time series deep features and the ambient temperature time series deep features output by the time series feature extraction network, and is composed of multiple stacked long short-term memory networks (LSTMs) with inter-layer skip connections, aiming to perform hierarchical deep modeling on time series features; each long short-term memory network layer contains an input gate, a forget gate, and an output gate, as well as a memory cell running through the entire time sequence, and a skip connection is introduced between the stacked LSTM layers, the input features of a certain LSTM layer are directly added to the output features of the next LSTM layer through a bypass channel, and by stacking multiple LSTM layers with skip connections, the complex dynamic rules of system heat charging and discharging and internal and external heat exchange can be learned from shallow to deep and hierarchically from the deep features of heating power and ambient temperature, and finally, a system thermal inertia deep feature vector balancing the overall heat budget of the system is output.

[0012] Preferably, the physical process deep feature interaction and fusion network first constructs and embeds feature sequences, splices the three feature vectors obtained by S2 into a physical process fusion feature sequence with a length of 3, and respectively superimposes a learnable embedding vector on each feature vector in the physical process fusion feature sequence; secondly, the input sequence is input into a multi-head self-attention layer, the input sequence is projected into multiple different representation subspaces in parallel through the multi-head self-attention mechanism, the attention weights between features are calculated independently in each sub-space, and after attention weighting and multi-head result splicing, a set of physical process fusion feature sequences that have undergone deep interaction and fused context information are output; thirdly, the physical process fusion feature sequence is nonlinearly transformed through a position-by-position feedforward network composed of two fully connected layers and a ReLU activation function to obtain a physical process fusion enhanced feature sequence; finally, output aggregation is performed: the physical process fusion enhanced feature sequence is aggregated through a global average pooling layer to compress and fuse all dimensions of information into a fixed-length vector; finally, the physical process deep feature interaction and fusion network outputs a comprehensive thermodynamic state vector containing complex dynamic interaction information between physical processes; the multi-head self-attention mechanism layer and the position-by-position feedforward network are both followed by layer normalization operations to normalize the output, and residual connections are used, both of which reduce the risk of gradient explosion.

[0013] Preferably, the comprehensive thermodynamic inference network first maps the comprehensive thermodynamic state vector to a high-dimensional latent space through a fully connected layer, and reshapes it into an initial comprehensive thermodynamic core feature tensor; then, a series of transpose convolutional layers are used to gradually upsample the comprehensive thermodynamic core feature tensor in space, and learn to expand it into an initial feature map of a three-dimensional temperature field with a preset spatial resolution; secondly, the initial feature map of the temperature field is input into a sequence decoding network composed of multiple stacked ConvLSTM layers, the recurrent update process of each layer of ConvLSTM transmits timing information while the convolution kernel inside it also performs sliding calculation in the spatial dimension of the feature map, effectively simulating the conduction and diffusion process of heat between adjacent physical points in the catalyst bed, and finally outputting a temperature field hidden state feature map containing rich spatio-temporal information at future time.

[0014] Preferably, the temperature field decoding regression network firstly performs spatial importance evaluation, inputs the output temperature field hidden state feature map into a spatial attention subnetwork, and simultaneously passes through a global maximum pooling layer and a global average pooling layer to capture significant features and global statistical features of the feature map respectively; then, the significant features and the global statistical features are processed through a shared multilayer perceptron, and finally a three-dimensional spatial temperature field attention weight map consistent with the spatial dimension of the initial feature map of the temperature field is generated through a Sigmoid activation function; secondly, the temperature field attention weight map is applied for feature enhancement: the generated spatial attention weight map is subjected to element-by-element multiplication operation with the input original hidden state feature map, so that a temperature field focused feature map optimized by attention weighting is obtained; finally, deep decoding regression is performed: the temperature field focused feature map is input into a deep decoding network composed of two serial convolutional layers, the first convolutional layer is used for deep nonlinear integration and information extraction of the weighted features, and the second convolutional layer is responsible for accurately mapping the extracted high-dimensional features to the single-channel temperature value, and finally outputting the catalyst bed internal temperature field prediction matrix at the future preset time step.

[0015] Preferably, the specific process of S4 comprises: Based on the output catalyst bed internal temperature field prediction matrix, the future temperature value of each point in the prediction matrix is compared with the preset rule base in parallel, and the corresponding risk level is triggered according to the preset time window; Based on the risk level identification, the risk evolution trend calculation of the corresponding point is triggered, and its dynamic evolution trend is divided into three stages of incubation period, stable development period and accelerated deterioration period, corresponding to the first, second and third adjustment priority; Define the target loss function, the mean square error of the predicted temperature and the target safety temperature, secondly, utilize the end-to-end differentiable characteristics of the whole deep learning model, calculate the partial derivative gradient of the target loss function with respect to the ammonia injection rate and the heating system power through a virtual gradient back propagation, then the system sets a basic adjustment step, and defines an adjustment scaling factor according to the output adjustment priority, and dynamically scales the step, higher priority corresponds to larger adjustment scaling factor; finally, combine the calculated gradient direction with the dynamically scaled step to generate an optimal control instruction vector including ammonia injection rate adjustment value and heating system power adjustment value.

[0016] The innovation points of the present application include: (1) Physics-decoupled feature extraction architecture: Independent branch networks are designed for the three major physical processes of convective heat transfer, catalytic exothermic, and thermal inertia. The convective heat transfer branch uses a neural ODE solver to model the continuous dynamics of flue gas heat transfer, parameterizes the hidden state evolution through a vector field function; the catalytic exothermic branch designs a dynamic parameter generation network, and through a selective scanning mechanism, it adaptively regulates the dynamic coupling relationship between reactants and conditions; the thermal inertia branch constructs a stacked LSTM network, and through inter-layer skip connections, it models long-term thermal storage and compensation effects; (2) Spatiotemporal collaborative prediction mechanism: Through the Transformer encoder layer, the features of multiple physical processes are interactively fused to generate a comprehensive thermodynamic state vector; combined with transpose convolution and ConvLSTM network, the thermodynamic state is decoded into a three-dimensional temperature field and the spatiotemporal evolution process is simulated; the spatial attention mechanism is introduced to focus on key areas to improve prediction accuracy; (3) Closed-loop regulation strategy based on gradient optimization: Establish a three-level threshold early warning rule to trigger differentiated response; use the end-to-end differentiable characteristics to calculate the gradient of the loss function with respect to the control parameters; combined with the risk evolution trend, dynamically scale the adjustment step size to generate optimal adjustment instructions for ammonia injection rate and heating power.

[0017] Compared with the prior art, the present application has the following beneficial effects: (1) Improve temperature prediction accuracy and physical interpretability: Physics-decoupled branch network models the core physical process, overcoming the limitations of traditional single model in capturing multi-scale thermodynamic effects; dynamic parameter network deeply integrates physical mechanisms and data-driven, significantly improving the generalization ability of temperature field prediction; (2) Realize early risk intervention: Three-level early warning mechanism combined with gradient optimization regulation can avoid low-temperature inactivation and high-temperature sintering risk in advance; spatiotemporal attention mechanism strengthens key area prediction, supporting precise positioning and regulation; (3) Strong engineering applicability: Full-automatic closed-loop control mode directly connects PLC actuator through industrial protocol; lightweight trend classifier and gradient adjustment strategy are suitable for real-time edge computing deployment. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The figure is the overall technical route flowchart of the present application.

[0019] Figure 2 The figure is the catalytic reaction temperature feature extraction module based on physics-decoupling of the present application.

[0020] Figure 3 The figure is the spatiotemporal decoding prediction module of the present application.

[0021] Figure 4 The figure is the CNN-LSTM temperature prediction error confusion matrix in the embodiment of the present application.

[0022] Figure 5 Temperature prediction error confusion matrix chart in the embodiment of the application. DETAILED DESCRIPTION

[0023] The application mainly proposes a low-temperature catalytic desulfurization and denitrification device temperature prediction and regulation method based on deep learning. First, based on the idea of physical decoupling, nine types of multi-source time series data such as flue gas inlet parameters, ammonia injection rate and heating power are collected to construct a temperature feature data set. Second, a feature extraction module based on physical decoupling is designed, and three branch networks are used to model the convection heat transfer, catalytic heat release and thermal inertia process respectively. Third, a spatio-temporal decoding prediction module is constructed to fuse multi-physical process features and predict the future temperature field. Finally, a multi-level risk early warning and intelligent regulation module is designed to generate control instructions combined with gradient optimization to realize temperature regulation. The overall process is shown in Figure 1

[0024] The application will be further described below in conjunction with specific embodiments.

[0025] I. Multi-source working condition data collection of low-temperature catalytic desulfurization and denitrification device The core physical process features related to the temperature of the low-temperature catalytic desulfurization and denitrification device include: convection heat transfer process feature, catalytic reaction heat release process feature, system thermal inertia and heat compensation process feature. The convection heat transfer process feature mainly shows that the flue gas as a heat carrier brings the upstream heat into and through the entire catalyst bed. The convection heat transfer process feature shows strong positive correlation with the flue gas inlet temperature, flue gas inlet flow rate and inlet oxygen concentration; the catalytic reaction heat release process feature mainly shows that the catalytic reaction heat release such as denitrification directly heats the catalyst bed from inside. The catalytic reaction heat release process feature is mainly affected by the inlet nitrogen oxide concentration, inlet sulfur oxide concentration, inlet oxygen concentration, catalytic aid ammonia injection rate and catalyst bed internal temperature field; the system thermal inertia and heat compensation process feature mainly shows that the catalyst, equipment shell and other components have a large heat capacity, resulting in slow response of the system to temperature rise and fall process, which is manifested as long-term heat storage and release, i.e. thermal inertia. At the same time, the system continuously dissipates heat to the environment through the shell, while the heating system actively compensates for the heat as needed. The system thermal inertia and heat compensation process feature is mainly related to the heating system power and environmental temperature; ​Based on the core physical characteristics affecting the temperature of the low-temperature catalytic desulfurization and denitrification device, the application constructs a low-temperature catalytic desulfurization and denitrification device temperature feature data set, deploys multiple sensors at key positions of the device, and synchronously collects multi-source heterogeneous time series data to comprehensively reflect the temperature changes in the running process of the low-temperature catalytic desulfurization and denitrification device. At the inlet of the flue gas main pipe of the device, a temperature sensor is installed to collect flue gas inlet temperature data in real time, a pitot tube flowmeter is installed to collect flue gas inlet flow rate data in real time, and a gas analyzer is installed to collect inlet nitrogen oxide concentration data, inlet sulfur oxide concentration data and inlet oxygen concentration data in real time; the frequency signal or flowmeter signal of the denitrification system ammonia water injection pump is collected, which is recorded as ammonia injection rate data; the power signal of the electric heater of the flue gas reheating system is collected, which is recorded as heating system power data; armored thermocouple sensor arrays are arranged at different depths and radial positions of the catalyst bed layer to collect multi-point temperature field data inside the catalyst bed layer in real time; the ambient temperature data near the device shell are collected for quantifying the heat loss of the system.

[0026] II. Design a catalytic reaction process temperature feature extraction module based on physical decoupling to extract multi-dimensional depth time series features of the device temperature The catalytic reaction temperature feature extraction module based on physical decoupling designed by the application is aimed at three temperature-related core physical processes of the low-temperature catalytic desulfurization and denitrification device, including convective heat transfer process feature data, catalytic reaction heat release process feature data, and system thermal inertia and heat compensation process feature data. Among them, the convective heat transfer process feature data includes flue gas inlet temperature data, flue gas inlet flow rate data and inlet oxygen concentration data; the catalytic reaction heat release process feature data includes inlet nitrogen oxide concentration data, inlet sulfur oxide concentration data, inlet oxygen concentration data, ammonia injection rate data and multi-point temperature field data inside the catalyst bed layer; the system thermal inertia and heat compensation process feature data includes heating system power data and ambient temperature data; the model structure is as shown in Figure 2 .

[0027] The catalytic reaction process temperature feature extraction module based on physical decoupling designed by the application includes a time series feature extraction network and a feature extraction architecture of three catalytic reaction branch networks. All branch networks share the basic time series features extracted by the time series feature extraction network.

[0028] 1. Time series feature extraction network: To effectively extract the common time-dependent relationship of multi-source heterogeneous data affecting the temperature of low-temperature catalytic desulfurization and denitrification device, a time series feature extraction network was constructed. The network consists of a feature alignment layer and two serial gated recurrent unit (GRU) networks. First, all nine kinds of data collected, including flue gas inlet temperature data, flue gas inlet flow rate data, inlet oxygen concentration data, inlet nitrogen oxide concentration data, inlet sulfur oxide concentration data, ammonia injection rate data, multi-point temperature field data inside the catalyst bed, heating system power data and environmental temperature data, are input into the feature alignment layer. With the timestamp of the temperature field data with the highest sampling frequency as the benchmark, the remaining data are resampled using the cubic spline interpolation algorithm to solve the problem of data misalignment in the time dimension caused by different sensor sampling periods. Subsequently, the aligned multi-source data are input into the double-layer GRU network, which captures the long-term time dependence of each working condition parameter using its gating mechanism. Finally, a set of high-dimensional universal time series deep features containing all the input information is output, including flue gas inlet temperature time series deep features, flue gas inlet flow rate time series deep features, inlet oxygen concentration time series deep features, inlet nitrogen oxide concentration time series deep features, inlet sulfur oxide concentration time series deep features, ammonia injection rate time series deep features, catalyst bed internal temperature field time series deep features, heating system power time series deep features and environmental temperature time series deep features.

[0029] 2. Convective heat transfer process feature extraction branch network: The application constructs a heat exchange process feature extraction branch network, which models the process of heat carried by the main flue gas stream from the bottom physical law level. The input is the inlet temperature time series deep feature, the inlet flow rate time series deep feature and the oxygen concentration time series deep feature output by the time series feature extraction network. First, the discrete multivariate time series input is converted into continuous inlet temperature time series deep feature, inlet flow rate time series deep feature and oxygen concentration time series deep feature through Gaussian process interpolation. The continuous three features are defined as the convection heat transfer process control signal path. Secondly, the initial convection heat transfer process hidden state is defined as a 32-dimensional zero vector. Thirdly, a multilayer perceptron with three hidden layers is used as a vector field function network, and the initial convection heat transfer process hidden state and the convection heat transfer process control signal path are used as inputs. The output is the rate of change of the convection heat transfer process hidden state, i.e. the derivative of the convection heat transfer process hidden state, which is parameterized by the vector field function network. Fourthly, through a differentiable ordinary differential equation solver, the dynamics equation parameters (vector field function network parameters) parameterized by the vector field function network, which describe the evolution of the convection heat transfer process hidden state with time and the convection heat transfer process control signal path, are integrated and solved within the observation time period (the length of time during which the data is collected in the application) to obtain the final convection heat transfer process hidden state. The final convection heat transfer process hidden state is the convection heat transfer deep feature vector, which provides a key basis for subsequent temperature field prediction.

[0030] 3. Catalytic reaction heat release process feature extraction branch network: To simulate the time sequence dynamic interaction and nonlinear exothermic effect of each component in the catalytic reaction in depth, the application constructs a feature extraction branch network of the catalytic reaction exothermic process. The inputs are the time sequence depth features of the inlet nitrogen oxide concentration, the time sequence depth features of the inlet sulfur oxide concentration, the time sequence depth features of the inlet oxygen concentration, the time sequence depth features of the ammonia injection rate and the time sequence depth features of the internal temperature field of the catalyst bed. First, the time sequence depth features of the inlet nitrogen oxide concentration, the time sequence depth features of the inlet sulfur oxide concentration and the time sequence depth features of the inlet oxygen concentration are spliced to form a reactant feature sequence. Second, the time sequence depth features of the ammonia injection rate as the core control variable and the time sequence depth features of the catalyst bed temperature field as the key reaction condition are spliced to form a condition control feature sequence. Third, to further realize adaptive modeling of the reaction process under variable operating conditions, the application designs a dynamic parameter generation network composed of a single hidden layer multi-layer perception, a first linear activation layer, a second linear activation layer and a third linear activation layer. The network takes the reactant feature sequence and the condition control feature sequence as inputs, wherein the reactant feature sequence is the input of the single hidden layer multi-layer perception, and the condition control feature sequence is the input of the linear activation layer, the second linear activation layer and the third linear activation layer for outputting a set of state space parameters to control the feature extraction process of the reactant feature sequence.

[0031] The state space parameters specifically include: a selective step length that controls the attention degree of the current time reactant input feature in the discretization process, the selective step length is a vector that stores the step length of each dimension of the hidden layer state vector of the dynamic parameter generation network, which allows each dimension of the hidden layer to be processed with different sensitivity. An input modulation matrix that dynamically adjusts the influence of the current time reactant feature on the hidden state, the input modulation matrix is a two-dimensional matrix, the first dimension is the dimension of the hidden layer state vector of the dynamic parameter generation network, and the second dimension is the dimension of the input feature of the dynamic parameter generation network; this matrix determines how the current input affects and updates each dimension of the hidden state; an output selection matrix that controls the extraction of the output information most related to the exothermic effect from the hidden state, the input modulation matrix is a two-dimensional matrix, the first dimension is the dimension of the hidden layer state vector of the dynamic parameter generation network, and the second dimension is the dimension of the output feature of the dynamic parameter generation network; this matrix selectively reads and combines the multi-dimensional information most related to the current exothermic effect from the complete hidden state as the output. A state transition matrix that maintains the stability of state evolution as a fixed trainable parameter, the state transition matrix is a diagonal matrix composed of the model parameters of the single hidden layer multi-layer perception, the first linear activation layer, the second linear activation layer and the third linear activation layer in the dynamic parameter generation network; then, the dynamic parameter generation network performs selective scanning processing on the reactant feature sequence along the time dimension.

[0032] The following operations are sequentially performed at each time step in the scanning process to update the hidden layer state vector of the dynamic parameter generation network: 1) generating a corresponding selective step, input modulation matrix and output selection matrix according to the conditional control features of the current time step, specifically inputting the conditional control features of the current time step into the first linear activation layer, the second linear activation layer and the third linear activation layer to obtain the selective step, the input modulation matrix and the output selection matrix; 2) using the selective step to discretize the fixed state transition matrix and the dynamically generated input modulation matrix to obtain a discretized state transition matrix and a discretized input modulation matrix; 3) updating the hidden layer state vector by a recursive method: the update result is composed of two parts of linear combination, one part is the last time hidden state processed by the discretized state transition matrix, which reflects the autonomous evolution of temperature in the low-temperature catalytic process; the other part is the result of the current time reactant feature processed by the discretized input modulation matrix, which reflects the influence of external input on the temperature in the device; after scanning the entire reactant feature sequence, a reactant process representation sequence composed of instantaneous output feature vectors at each time is obtained.

[0033] To aggregate the time series dynamic information and form a fixed-dimension overall feature representation, the present application integrates the reactant process representation sequence through a global average pooling layer to generate a reactant comprehensive process feature vector. Finally, the reactant comprehensive process feature vector is further refined and dimension adjusted through an output mapping network composed of a fully connected layer and a ReLU activation function, and a catalytic reaction exothermic deep feature vector is output, which can finely represent the complex coupling relationship between reactants and the dynamic exothermic effect under different operating conditions.

[0034] 4, system thermal inertia and thermal compensation process feature extraction branch network To precisely model the slow dynamic process of the overall heat storage, heat dissipation and external heating compensation of the device, the time sequence feature extraction network is constructed. The input is the heating system power time sequence deep feature and the environmental temperature time sequence deep feature output by the time sequence feature extraction network. The branch network is composed of multiple stacked long short-term memory networks (LSTM) with interlayer jump connection, aiming at hierarchical deep modeling of time sequence features. Each long short-term memory network layer contains an input gate, a forgetting gate and an output gate, and a memory cell running through the entire time sequence. To further enhance the network's modeling ability for long sequences and optimize the training process, a jump connection is introduced between the stacked LSTM layers. Specifically, the input features of a certain layer of LSTM are directly added to the output features of the next layer of LSTM through a bypass channel. By stacking multiple LSTM layers with jump connections, the model can learn the complex dynamic rules of system heat charging and discharging and internal and external heat exchange from shallow to deep and hierarchically from the deep features of heating power and environmental temperature. Finally, the branch network output can represent the system thermal inertia deep feature vector of the system overall heat balance.

[0035] III. Designing a space-time decoding prediction module to predict the temperature field data of the low-temperature catalytic desulfurization and denitrification device After extracting the deep features containing different physical processes, this step aims to efficiently fuse these deep features and achieve precise prediction of the internal temperature field of the catalyst bed in the low-temperature catalytic desulfurization and denitrification device for multiple time steps in the future. The model structure is as shown in Figure 3

[0036] 1. Physical process deep feature interaction and fusion network: ​To realize the dynamic and adaptive fusion of the extracted convection heat transfer deep feature vector, catalytic reaction heat release deep feature vector and system thermal inertia deep feature vector, the application designs a deep feature interaction and fusion network based on the Transformer encoder layer. First, the feature sequence construction and embedding are carried out: the convection heat transfer deep feature vector, the catalytic reaction heat release deep feature vector and the system thermal inertia deep feature vector are spliced into a physical process fusion feature sequence with a length of 3. In order to enable the model to distinguish the physical sources of different features, a learnable embedding vector is superimposed on each feature vector in the physical process fusion feature sequence, including the convection heat transfer deep feature vector, the catalytic reaction heat release deep feature vector and the system thermal inertia deep feature vector. This step converts three independent feature vectors into an input sequence containing special information that can be processed by the multi-head self-attention mechanism. Second, the input sequence is input into the multi-head self-attention layer, and the input sequence is projected into multiple different representation subspaces in parallel through the multi-head self-attention mechanism. The attention weights between features are calculated independently in each sub-space. After attention weighting and multi-head result splicing, a set of physical process fusion feature sequences that have undergone deep interaction and fused context information are output. Third, the physical process fusion feature sequence is further nonlinearly transformed through a position-by-position feedforward network composed of two fully connected layers and a ReLU activation function to obtain a physical process fusion enhanced feature sequence. Finally, the output aggregation is carried out: the physical process fusion enhanced feature sequence is aggregated through a global average pooling layer to compress and fuse all dimensions of information into a fixed length vector. Finally, the physical process deep feature interaction and fusion network outputs a comprehensive thermodynamic state vector containing complex dynamic interaction information between physical processes; the multi-head self-attention mechanism layer and the position-by-position feedforward network are both normalized after the output is normalized, and residual connection is adopted, which together reduces the risk of gradient explosion.

[0037] 2、Initial space state generation and spatio-temporal evolution process reasoning: To convert the output comprehensive thermodynamic state vector into a temperature field spatial state and perform time evolution, the comprehensive thermodynamic reasoning network is constructed. The specific working process of the comprehensive thermodynamic reasoning network is as follows: first, the comprehensive thermodynamic state vector is mapped to a high-dimensional latent space through a fully connected layer, and is reshaped into an initial comprehensive thermodynamic core feature tensor. Then, the initial comprehensive thermodynamic core feature tensor is gradually upsampled in space through a series of transpose convolutional layers, and is learned to expand into an initial feature map of a three-dimensional temperature field with a preset spatial resolution. Secondly, the initial feature map of the temperature field is input into a sequence decoding network composed of multiple stacked ConvLSTM layers. In this network, the cyclic update process of each layer of ConvLSTM not only transmits time sequence information, but also performs sliding calculation on the spatial dimension of the feature map, effectively simulating the conduction and diffusion process of heat between adjacent physical points in the catalyst bed. This step finally outputs a temperature field hidden state feature map containing rich spatio-temporal information at future time.

[0038] 3. Temperature field decoding regression network based on spatial attention mechanism: To accurately decode and regress the output temperature field hidden state feature map containing rich spatio-temporal information at future time into temperature values with clear physical meaning, and further improve the prediction accuracy of key areas, the present application designs a decoding regression network based on spatial attention mechanism. The specific working process of this module is as follows: first, spatial importance evaluation: input the output temperature field hidden state feature map into a spatial attention subnetwork. This subnetwork parallelly captures the significant features and global statistical features of the feature map through a global max pooling layer and a global average pooling layer, respectively. Then, these two feature vectors are processed through a shared multi-layer perceptron, and finally a three-dimensional spatial temperature field attention weight map with the same spatial dimension as the initial feature map is generated through a Sigmoid activation function. Each element value of the temperature field attention weight map quantifies the importance of the corresponding spatial position to the formation of the final temperature field, enabling the model to automatically focus on areas with more intense or more critical heat changes. Secondly, apply the temperature field attention weight map for feature enhancement: multiply the generated spatial attention weight map with the input original hidden state feature map element by element to obtain a temperature field focus feature map optimized by attention weighting. Finally, deep decoding regression: input the temperature field focus feature map into a deep decoding network composed of two convolutional layers in series. The first convolutional layer is used for deep nonlinear integration and information extraction of the weighted features, and the second convolutional layer is responsible for accurately mapping the extracted high-dimensional features to single-channel temperature values, finally outputting the catalyst bed internal temperature field prediction matrix at the future preset time step.

[0039] IV. Design of risk early warning and adjustment mode After outputting the catalyst bed internal temperature field prediction matrix, in order to realize the adjustment of the low-temperature catalytic desulfurization and denitrification device from the prediction of the future temperature field, the present application designs a complete intelligent adjustment and risk management module from multi-stage risk early warning, risk evolution trend deep analysis and adjustment priority judgment, intelligent adjustment strategy generation to adaptive closed loop.

[0040] Early warning level calculation module: This module aims to convert the output catalyst bed internal temperature field prediction matrix into an early warning signal with a clear risk level in real time, as the core trigger mechanism of the subsequent intelligent adjustment process. The system has a three-level threshold rule base based on expert knowledge, which presets the best efficiency window, low-temperature deactivation threshold and high-temperature sintering threshold of the catalyst. The early warning module compares the future temperature value of each point in the prediction matrix with the rule base in parallel, and triggers the corresponding risk level according to the preset time window, which is divided into first-level attention early warning, second-level warning early warning and third-level danger alarm: first-level attention early warning is triggered when it is predicted that the temperature of any measuring point will reach the boundary of the best efficiency window within 15 minutes in the future. The system will increase the monitoring frequency of the relevant data and prompt the operator to pay attention; second-level warning early warning is triggered when it is predicted that the temperature of any measuring point will exceed the boundary of the best efficiency window within 10 minutes in the future. This level will automatically activate the intelligent adjustment strategy generation module. Third-level danger alarm is triggered immediately when it is measured that the temperature of any measuring point will reach the low-temperature deactivation or high-temperature sintering threshold within 5 minutes in the future, and the highest level of sound and light alarm is activated, and the emergency adjustment strategy generated by the intelligent adjustment strategy generation module is forced to execute.

[0041] Risk evolution trend calculation module: After identifying the potential risk, this module aims to deeply analyze the dynamic evolution trend of the risk, so as to provide decision basis for the urgency of the adjustment strategy. This module receives the output catalyst bed internal temperature field prediction matrix. First, the time gradient of the prediction temperature sequence is calculated to quantify the change rate of the risk point temperature. Second, the change rate is input into a trend classifier based on historical working condition data and simulation data pre-trained by a lightweight one-dimensional convolution network, which divides the dynamic evolution trend into three stages: incubation period, stable development period and accelerated deterioration period, corresponding to first-level, second-level and third-level adjustment priority.

[0042] Intelligent adjustment strategy generation module based on model gradient optimization: This module is the core of intelligent adjustment, responsible for generating specific, quantitative optimal control parameter adjustment scheme. When the adjustment process is triggered, the module will perform the following operations according to the output risk evolution trend: First, define the target loss function, which is the mean square error of the predicted temperature and the target safety temperature. Second, use the end-to-end differentiable characteristics of the entire deep learning model to calculate the partial derivative gradient of the loss function with respect to the ammonia injection rate and the heating system power through a virtual gradient backpropagation. Then, the system sets a basic adjustment step and defines an adjustment scaling factor according to the output adjustment priority, and dynamically scales the step, with higher priority corresponding to larger adjustment scaling factor. Finally, combine the calculated gradient direction and dynamically scaled step to generate an optimal control instruction vector including ammonia injection rate adjustment value and heating system power adjustment value.

[0043] Low-temperature catalytic desulfurization and denitrification device catalytic reaction temperature control: This step is responsible for executing decisions; first, control instruction receiving and issuing: the generated optimal control instruction vector can be executed through human-machine collaboration mode and full-automatic mode. In full-automatic mode, the instruction will be directly transmitted to the underlying programmable logic controller of the device through the industrial communication protocol, realizing full-automatic closed-loop adjustment of the actuator.

[0044] Five, experimental result comparison In order to verify the effectiveness of the method of the present application, a test experiment is designed. The experimental object is a set of low-temperature catalytic desulfurization and denitrification devices with a flue gas treatment capacity of 10,000 Nm³ / h, and the core of the device includes a catalyst bed, a flue gas inlet system and a heating compensation module. According to the S1 step of the present application, multi-source sensors are deployed at key positions of the device: Flue gas main inlet: install temperature sensors to collect flue gas inlet temperature, pitot tube flow meters to collect flue gas inlet flow rate, and gas analyzers to collect inlet nitrogen oxide, sulfur oxide and oxygen concentration; Catalyst bed: arrange armored thermocouple sensor arrays at different depths and radial positions on the bed to collect internal multi-point temperature field data; Auxiliary system: collect ammonia injection rate, heating system power and ambient temperature sensor data near the shell.

[0045] The experiment collected 72 hours of continuous operation data, covering normal working conditions, temperature fluctuations, and abnormal scenes such as heating compensation failure and catalytic reaction; referring to the above process, the data is aligned using the cubic spline interpolation algorithm, and resampled at the highest sampling frequency (temperature field data sampling frequency is 10Hz) as the reference to solve the time dimension inconsistency problem. After data preprocessing, the data is divided into training set (first 50 hours of data), validation set (51 to 62 hours) and test set (63 to 72 hours); The comparison model includes: SVR: as a traditional machine learning benchmark, using a radial basis function kernel; 1D-CNN: one-dimensional convolutional neural network, containing 3 convolutional layers and pooling layers, used to extract local time sequence features; LSTM: single layer hidden unit 128, capturing long-term dependencies; CNN-LSTM: combined model, first use CNN to extract features, then input LSTM for sequence prediction; the experimental results are shown in Table 1: Table 1 Experimental results

[0046] The experimental results shown in Table 1 verify the superiority of the present application in temperature prediction accuracy, response speed and risk adjustment compared with conventional machine learning methods and deep learning methods. Further, Figure 4 and Figure 5 The prediction error distribution of each spatial position of the catalyst bed of CNN-LSTM and the present application is compared. The average error of CNN-LSTM in the edge area of the bed is as high as 2.5 degrees Celsius, and the probability of false identification of hot spot area is 18%. The present application focuses on the key area through the spatial attention mechanism of S3-3, and the average error is reduced to 0.9 degrees Celsius, and the error probability is only 5%, which can better cope with the complex thermodynamic environment of multiple concurrent faults.

[0047] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

[0048] Although the specific embodiments of the present application have been described above, they are not intended to limit the scope of protection of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A deep learning-based low-temperature catalytic desulfurization and denitrification device temperature prediction and adjustment method, characterized in that, The method comprises the following steps: S1, collecting the convection heat transfer process characteristic data related to the temperature of the low-temperature catalytic desulfurization and denitrification device, the catalytic reaction heat release process characteristic data, and the system thermal inertia and heat compensation process characteristic data; S2, extracting features from the three types of characteristic data collected in S1 through a catalytic reaction temperature feature extraction module, including a time series feature extraction network and three branch networks shared with the network: a convection heat transfer process feature extraction branch network, a catalytic reaction heat release process feature extraction branch network, and a system thermal inertia and heat compensation process feature extraction branch network, to obtain a convection heat transfer depth feature vector, a catalytic reaction heat release depth feature vector, and a system thermal inertia depth feature vector; S3, inputting the three types of feature vectors obtained in S2 into a physical process depth feature interaction and fusion network for feature fusion and nonlinear transformation to obtain a comprehensive thermodynamic state vector, and then inputting the vector into a comprehensive thermodynamic reasoning network for time-space feature evolution reasoning to obtain a temperature field hidden state feature map; inputting the feature map into a temperature field decoding regression network based on a spatial attention mechanism for temperature field key region feature capture and regression mapping to obtain a catalyst bed internal temperature field prediction matrix at a future preset time step; S4, generating a level warning based on the catalyst bed internal temperature field prediction matrix and triggering risk evolution trend calculation, and generating an optimal control instruction vector including an ammonia injection rate adjustment value and a heating system power adjustment value based on an intelligent adjustment strategy optimized by model gradient.

2. The low-temperature catalytic desulfurization and denitrification device temperature prediction and adjustment method based on deep learning according to claim 1, characterized in that: The convection heat transfer process characteristic data includes flue gas inlet temperature data, flue gas inlet flow rate data, and inlet oxygen concentration data; the catalytic reaction heat release process characteristic data includes inlet nitrogen oxide concentration data, inlet sulfur oxide concentration data, inlet oxygen concentration data, ammonia injection rate data, and catalyst bed internal multi-point temperature field data; and the system thermal inertia and heat compensation process characteristic data includes heating system power data and ambient temperature data. At the inlet of the device flue gas main pipe, a temperature sensor is installed to collect real-time flue gas inlet temperature data, a pitot tube flowmeter is installed to collect real-time flue gas inlet flow rate data, and a gas analyzer is installed to collect real-time inlet nitrogen oxide concentration data, inlet sulfur oxide concentration data, and inlet oxygen concentration data; the frequency signal or flowmeter signal of the denitrification system ammonia water injection pump is collected, which is recorded as the ammonia injection rate data; the power signal of the electric heater of the flue gas reheating system is collected, which is recorded as the heating system power data; an armored thermocouple sensor array is arranged at different depths and radial positions of the catalyst bed to collect real-time catalyst bed internal multi-point temperature field data; and ambient temperature data is collected near the device shell for quantifying system heat loss.

3. The low-temperature catalytic desulfurization and denitrification device temperature prediction and adjustment method based on deep learning according to claim 1, characterized in that: The time sequence feature extraction network is composed of a feature alignment layer and two serial gate recurrent unit (GRU) networks. First, all data collected in S1 is input into the feature alignment layer, and the timestamps of the temperature field data with the highest sampling frequency are used as the reference to resample the remaining data using a cubic spline interpolation algorithm. Then, the aligned multi-source data is input into the double-layer GRU network, which uses its gating mechanism to capture the long-term time dependence of each working condition parameter, and finally outputs a set of high-dimensional universal time sequence deep features containing all input information.

4. The low-temperature catalytic desulfurization and denitrification device temperature prediction and adjustment method based on deep learning of claim 1, characterized in that: The convection heat transfer process feature extraction branch network first converts the inlet temperature time sequence deep features, the inlet flow rate time sequence deep features and the oxygen concentration time sequence deep features output by the time sequence feature extraction network into continuous time sequence deep features through Gaussian process interpolation, and defines the continuous deep features as the convection heat transfer process control signal path. Second, the initial convection heat transfer process hidden state is defined as a zero vector with a dimension of 32. Third, a multilayer perceptron with three hidden layers is used as a vector field function network, and the initial convection heat transfer process hidden state and the convection heat transfer process control signal path are input into the network to output the rate of change of the convection heat transfer process hidden state, i.e., the derivative of the convection heat transfer process hidden state, which is parameterized by the vector field function network. Fourth, a differentiable ordinary differential equation solver is used to integrate and solve the kinetic equation parameters describing the evolution of the convection heat transfer process hidden state over time and the convection heat transfer process control signal path within the observation time period, obtaining the convection heat transfer deep feature vector.

5. The low-temperature catalytic desulfurization and denitrification device temperature prediction and adjustment method based on deep learning according to claim 1, characterized in that: The catalytic reaction exothermic process feature extraction branch network concatenates the inlet nitrogen oxide concentration time sequence deep features, the inlet sulfur oxide concentration time sequence deep features and the inlet oxygen concentration time sequence deep features output by the time sequence feature extraction network to form a reactant feature sequence. Second, the ammonia injection rate time sequence deep features as the core control variable are concatenated with the catalyst bed temperature field time sequence deep features as the key reaction condition to form a condition control feature sequence. Third, a dynamic parameter generation network is used to input the reactant feature sequence and the condition control feature sequence, wherein the reactant feature sequence is input into a single hidden layer multilayer perceptron, and the condition control feature sequence is input into a linear activation layer, a second linear activation layer and a third linear activation layer to output a set of state space parameters for controlling the feature extraction process of the reactant feature sequence. The state space parameters include a selectivity step, an input modulation matrix, an output selection matrix and a state transition matrix. The dynamic parameter generation network performs selective scanning processing on the reactant feature sequence along the time dimension to obtain a reactant process representation sequence composed of instantaneous output feature vectors at each time; the reactant process representation sequence is integrated through a global average pooling layer to generate a reactant comprehensive process feature vector; finally, the reactant comprehensive process feature vector is further feature extracted and dimension adjusted through an output mapping network composed of a fully connected layer and a ReLU activation function, and a catalytic reaction exothermic deep feature vector capable of finely representing the complex coupling relationship and dynamic exothermic effect between reactants under different operating conditions is output.

6. The low-temperature catalytic desulfurization and denitrification device temperature prediction and adjustment method based on deep learning according to claim 1, characterized in that: The system thermal inertia and thermal compensation process feature extraction branch network inputs the heating system power time series deep feature and the environmental temperature time series deep feature output by the time series feature extraction network, and is composed of multiple stacked long short-term memory networks (LSTM) with interlayer jump connections, aiming to perform hierarchical deep modeling on the time series features; each long short-term memory network layer internally includes an input gate, a forgetting gate and an output gate, as well as a memory cell running through the entire time sequence, and a jump connection is introduced between the stacked LSTM layers; the input features of a certain LSTM layer are directly added to the output features of the next LSTM layer through a bypass channel; by stacking multiple LSTM layers with jump connections, the complex dynamic rules of system heat charging and discharging and internal and external heat exchange can be learned from the depth features of heating power and environmental temperature, from shallow to deep and hierarchically, and finally, the system thermal inertia deep feature vector of the overall heat balance of the system is output.

7. The low-temperature catalytic desulfurization and denitrification device temperature prediction and adjustment method based on deep learning according to claim 1, characterized in that: The physical process deep feature interaction and fusion network first performs feature sequence construction and embedding, splices the three feature vectors obtained in S2 into a physical process fusion feature sequence with a length of 3, and respectively superimposes a learnable embedding vector for each feature vector in the physical process fusion feature sequence; secondly, the input sequence is input into a multi-head self-attention layer, and the input sequence is projected into multiple different representation subspaces in parallel through the multi-head self-attention mechanism, and the attention weights between the features are calculated independently in each sub-space; after attention weighting and multi-head result splicing, a set of physical process fusion feature sequences that have undergone deep interaction and have fused context information are output; Thirdly, the physical process fusion feature sequence is nonlinearly transformed through a position-by-position feedforward network composed of two fully connected layers and a ReLU activation function to obtain a physical process fusion enhanced feature sequence; finally, output aggregation is performed: the physical process fusion enhanced feature sequence is aggregated through a global average pooling layer to compress and fuse the information in all dimensions into a fixed-length vector; finally, the physical process deep feature interaction and fusion network outputs a comprehensive thermodynamic state vector containing complex dynamic interaction information between physical processes; both the multi-head self-attention mechanism layer and the position-by-position feedforward network have layer normalization operations to normalize the output, and residual connections are used, which together reduce the risk of gradient explosion.

8. The low-temperature catalytic desulfurization and denitrification device temperature prediction and adjustment method based on deep learning according to claim 7, characterized in that: The comprehensive thermodynamic inference network first maps the comprehensive thermodynamic state vector to a high-dimensional latent space through a fully connected layer, and reshapes it into an initial comprehensive thermodynamic core feature tensor; then, through a series of transpose convolutional layers, the initial comprehensive thermodynamic core feature tensor is gradually spatially upsampled to learn to expand it into an initial feature map of a three-dimensional temperature field with a preset spatial resolution; secondly, the initial feature map of the temperature field is input into a sequence decoding network composed of multiple stacked ConvLSTM layers, and the cyclic update process of each ConvLSTM layer not only passes on the timing information, but also performs sliding calculation on the spatial dimensions of the feature map, effectively simulating the conduction and diffusion process of heat between adjacent physical points in the catalyst bed, and finally outputting a temperature field hidden state feature map containing rich spatio-temporal information at future time. 9.The low-temperature catalytic desulfurization and denitrification device temperature prediction and adjustment method based on deep learning of claim 8, wherein: The temperature field decoding regression network first performs spatial importance evaluation, inputs the output temperature field hidden state feature map into a spatial attention subnetwork, and simultaneously passes through a global maximum pooling layer and a global average pooling layer to capture significant features and global statistical features of the feature map, respectively; Then, the significant features and global statistical features are processed through a shared multilayer perceptron, and finally a Sigmoid activation function is used to generate a three-dimensional spatial temperature field attention weight map consistent with the spatial dimensions of the initial feature map of the temperature field; secondly, the temperature field attention weight map is applied for feature enhancement: the generated spatial attention weight map is multiplied element by element with the input original hidden state feature map to obtain a temperature field focusing feature map optimized by attention weighting; finally, deep decoding regression is performed: the temperature field focusing feature map is input into a deep decoding network composed of two convolutional layers in series, the first convolutional layer is used to perform deep nonlinear integration and information extraction on the weighted features, and the second convolutional layer is responsible for accurately mapping the extracted high-dimensional features to single-channel temperature values, and finally outputting the catalyst bed internal temperature field prediction matrix at the future preset time step.

10. The low-temperature catalytic desulfurization and denitrification device temperature prediction and adjustment method based on deep learning according to claim 1, characterized in that: The specific process of S4 includes: Based on the output catalyst bed internal temperature field prediction matrix, the future temperature value of each point in the prediction matrix is compared with the preset rule base in parallel, and the corresponding risk level is triggered according to the preset time window; Based on risk level identification, trigger the risk evolution trend calculation of the corresponding point, and divide its dynamic evolution trend into three stages: incubation period, stable development period and accelerated deterioration period, corresponding to first, second and third adjustment priorities; The target loss function is defined as the mean square error between the predicted temperature and the target safe temperature. Then, the partial derivative gradients of the target loss function with respect to the ammonia injection rate and the heating system power are calculated through a virtual gradient backpropagation, taking advantage of the end-to-end differentiable characteristics of the entire deep learning model. Then, the system sets a basic adjustment step size and defines an adjustment scaling factor according to the output adjustment priority. The step size is dynamically scaled, and a higher priority corresponds to a larger adjustment scaling factor. Finally, an optimal control instruction vector including the ammonia injection rate adjustment value and the heating system power adjustment value is generated by combining the calculated gradient direction and the dynamically scaled step size.

Images