Multi-objective optimized cement decomposing furnace intelligent prediction control method, device and equipment

Abstract

The invention provides a multi-objective optimization cement decomposing furnace intelligent prediction control method, device and equipment, and relates to the technical field of cement decomposing furnace control. The method comprises the following steps: respectively obtaining an f-CaO prediction value, an NOx prediction value and a decomposition furnace temperature prediction value at the next moment through an f-CaO prediction model, an NOx prediction model and a decomposition furnace temperature prediction model. Wherein the f-CaO prediction model, the NOx prediction model and the decomposition furnace temperature prediction model are respectively constructed on the basis of a CNN-P-sLSTM network, a CNN-minGRU network and a Reformer-KAN network, and are obtained through training of the CNN-P-sLSTM network, the CNN-minGRU network and the Reformer-KAN network. And on the basis of the f-CaO predicted value and the NOx predicted value at the next moment, the optimal solution interval of the state parameters is calculated through a multi-objective genetic algorithm NSGA-II. And performing intelligent control on the cement decomposing furnace based on the optimal solution interval of the state parameters and the temperature predicted value of the decomposing furnace at the next moment. The production cost and pollution emission can be reduced, and the quality of a cement finished product and the resource utilization rate are improved.

Description

Technical Field

[0001] This invention relates to the field of cement decomposition furnace control technology, and in particular to a multi-objective optimized intelligent predictive control method, device and equipment for cement decomposition furnaces. Background Technology

[0002] The new dry-process cement production technology mainly utilizes suspension preheating and pre-decomposition technologies, employing computer and networked information technology to guide the cement production process. It features high efficiency, energy saving, environmental protection, and sustainable development. The pre-decomposition system consists of four subsystems: a preheater, a decomposition furnace, a rotary kiln, and a grate cooler. Each subsystem is primarily responsible for the preheating, decomposition, calcination, and cooling stages of the cement clinker calcination process. Since the production process is a continuous physicochemical reaction, the tasks at each stage are interconnected and mutually restrictive. Therefore, any fluctuation in parameters during the pre-decomposition reaction will affect the stable production of the entire system. In the preheater system, scaling and blockage are not only common but also extremely harmful. To effectively address this problem, fuel selection and furnace temperature control are particularly crucial.

[0003] In actual cement production, besides using coal as the main fuel for the precalciner, alternative fuels such as municipal solid waste are often added to reduce production costs and pollution. However, these alternative fuels often have characteristics such as high moisture content, low calorific value, complex volatile components, and unstable combustion processes, which may affect the temperature stability within the precalciner. Precalcination technology relies primarily on the precalciner, which is located before the rotary kiln and provides space for carbonate decomposition. It is a crucial link in the complete production process. In actual production, the physicochemical reactions within the kiln are controlled by real-time adjustment of key parameters.

[0004] In traditional production processes, the control of coal feed at the kiln head and tail, the adjustment of the blending ratio of alternative fuels, and the stable control of the decomposition furnace temperature are mainly regulated by human experience or the regularity of historical parameter changes. This has a certain degree of subjective influence, resulting in poor cement product quality and resource utilization, as well as high production costs and pollution emissions. Summary of the Invention

[0005] This invention provides a multi-objective optimized intelligent predictive control method, device, and equipment for cement decomposition furnaces to address the problems of poor cement product quality and resource utilization, as well as high production costs and pollution emissions.

[0006] In a first aspect, embodiments of the present invention provide a multi-objective optimized intelligent predictive control method for cement decomposition furnaces, comprising: The f-CaO influence parameters at the current time are input into the f-CaO prediction model to obtain the f-CaO prediction value for the next time step output by the f-CaO prediction model. The f-CaO prediction model is built on a CNN-P-sLSTM network (Convolutional Neural Network-Patch-structured Parallel Long Short-Term Memory Network) and trained. The NOx impact parameters at the current time are input into the NOx prediction model to obtain the NOx prediction value at the next time step output by the NOx prediction model. The NOx prediction model is built on a CNN-minGRU network (Convolutional Neural Network - Minimal Gated Recurrent Unit Network) and trained. Based on the predicted values ​​of f-CaO and NOx at the next time step, the optimal solution range of the state parameters is obtained by calculating using the multi-objective genetic algorithm NSGA-II. The temperature influence parameters at the current moment are input into the decomposer temperature prediction model to obtain the decomposer temperature prediction value for the next moment, which is output by the decomposer temperature prediction model. The decomposer temperature prediction model is built based on the Reformer-KAN network and trained. Based on the optimal solution range of the state parameters and the predicted value of the decomposer temperature at the next moment, determine the required amount of coal and alternative fuel at the current moment. The cement decomposition furnace is intelligently controlled based on the required amount of coal and alternative fuel at the current moment.

[0007] In one possible implementation, the temperature-affecting parameters include: primary cylinder outlet temperature, primary cylinder oxygen concentration, feed rate, high-temperature fan speed, historical value of decomposer outlet temperature, coal feed rate, and alternative fuel quantity. The temperature influence parameters at the current moment are input into the decomposer temperature prediction model to obtain the predicted decomposer temperature value for the next moment, output by the model, including: The temperature-affecting parameters at the current moment are divided into batches, and a first and second sequence with overlapping parts are generated based on the time series. The first sequence is input into the Encoder layer of the decomposer furnace temperature prediction model. The Encoder layer performs the following processing: position encoding and feature mapping on the first sequence to obtain a first encoding result; the first encoding result is split into a first sub-data and a second sub-data along the channel dimension; the first sub-data is processed sequentially through a multi-head locality sensitive hashing layer and a feedforward network layer to obtain a first processing result; the first processing result is concatenated with the second sub-data and layer normalized to obtain the first feature data; the first feature data is distilled to compress the time dimension and output the encoded features. The second sequence is input into the Decoder layer of the decomposer temperature prediction model. The Decoder layer performs the following processing: position encoding and feature mapping on the second sequence to obtain the second encoding result; the second encoding result is split into a third sub-data and a fourth sub-data along the channel dimension; the third sub-data is processed sequentially through a multi-head mask locality-sensitive hashing layer and a feedforward network layer to obtain the second processing result; the second processing result is then residually connected and layer normalized with the fourth sub-data to obtain the second feature data; the second feature data and the encoded features are input into a cross-attention layer for interactive computation to obtain the Decoder output features. The Decoder output features are input into the KAN output layer for feature integration and dimensionality reduction, and the predicted value of the decomposition furnace temperature at the next time step is output.

[0008] In one possible implementation, before inputting the temperature influence parameters at the current moment into the decomposer temperature prediction model and obtaining the predicted decomposer temperature value for the next moment output by the decomposer temperature prediction model, the method further includes: The temperature influence parameters and the actual temperature values ​​of the decomposer under historical time periods are preprocessed, and the first training set, the first validation set and the first test set are constructed based on the preprocessed temperature influence parameters and the actual temperature values ​​of the decomposer under historical time periods. An initial model for predicting the temperature of the decomposer furnace was constructed based on the Reformer-KAN network. The initial decomposer temperature prediction model is trained based on the first training set, the first validation set, and the first test set to obtain the decomposer temperature prediction model.

[0009] In one possible implementation, the parameters affecting f-CaO include: decomposition furnace outlet temperature, kiln tail temperature, secondary air temperature, total feed rate, high-temperature fan speed, and EP fan speed feedback. The f-CaO influence parameters at the current time are input into the f-CaO prediction model to obtain the f-CaO prediction value for the next time step, as output by the f-CaO prediction model, including: The f-CaO influence parameters at the current time are input into the one-dimensional convolutional layer of the f-CaO prediction model. The one-dimensional convolutional layer uses the tanh activation function to perform nonlinear feature extraction on the f-CaO influence parameters at the current time and outputs the first convolutional feature. The first convolutional feature is input into the P-sLSTM layer of the f-CaO prediction model for temporal feature modeling. The P-sLSTM layer performs the following processing: the first convolutional feature is divided into multiple channel-independent segments and linearly projected to obtain projected segments; the projected segments are processed by an xLSTM module containing sLSTM units to capture short-term dependency information and obtain segment features; the segment features are flattened and combined by a linear layer to obtain global temporal features. The global temporal features are input into the fully connected output layer of the f-CaO prediction model for mapping, and the f-CaO prediction value for the next time step is output.

[0010] In one possible implementation, before inputting the f-CaO influence parameters at the current moment into the f-CaO prediction model and obtaining the f-CaO prediction value for the next moment output by the f-CaO prediction model, the method further includes: The f-CaO impact parameters and actual f-CaO emissions for historical periods are preprocessed, and a second training set, a second validation set, and a second test set are constructed based on the preprocessed f-CaO impact parameters and actual f-CaO emissions for historical periods. An initial f-CaO prediction model was constructed based on the CNN-P-sLSTM network; The initial f-CaO prediction model is trained based on the second training set, the second validation set, and the second test set to obtain the f-CaO prediction model.

[0011] In one possible implementation, the NOx-affecting parameters include: decomposer outlet temperature, kiln tail temperature, coal feed rate, alternative fuel quantity, kiln head coal feedback, and total ammonia water flow rate. The NOx impact parameters at the current time are input into the NOx prediction model to obtain the NOx prediction value for the next time step, as output by the NOx prediction model, including: The NOx impact parameters at the current moment are input into the one-dimensional convolutional layer of the NOx prediction model for information fusion and time series feature extraction, and the second convolutional feature is output. The second convolutional feature is input into the minGRU layer of the NOx prediction model to model temporal dependencies. The minGRU layer performs the following processing: based on the second convolutional feature and the hidden state of the previous time step, update weights are calculated through the update gate; combining the update weights and the hidden state of the previous time step, the candidate hidden state of the current time step is calculated; based on the update weights, the hidden state of the previous time step, and the candidate hidden state of the current time step, the hidden state of the current time step is calculated. The hidden state at the current time step is input into the fully connected output layer of the NOx prediction model for mapping, and the NOx prediction value at the next time step is output.

[0012] In one possible implementation, before inputting the NOx impact parameters at the current moment into the NOx prediction model and obtaining the NOx prediction value for the next moment output by the NOx prediction model, the method further includes: The NOx impact parameters and actual NOx emissions for historical periods are preprocessed, and a third training set, a third validation set, and a third test set are constructed based on the preprocessed NOx impact parameters and actual NOx emissions for historical periods. An initial NOx prediction model was constructed based on the CNN-minGRU network; The initial NOx prediction model is trained based on the third training set, the third validation set, and the third test set to obtain the NOx prediction model.

[0013] In one possible implementation, based on the predicted f-CaO and NOx values ​​at the next time step, the optimal solution range for the state parameters is calculated using the multi-objective genetic algorithm NSGA-II, including: Based on the f-CaO prediction model and the NOx prediction model, an objective function is constructed, which includes:

[0014]

[0015] The constraints of the objective function are:

[0016]

[0017] in, This indicates the goal of cement quality optimization. For a moment k The predicted value of f-CaO, This represents the optimal value for f-CaO content required by the process. Indicates time The increase in coal feeding compared to the previous moment. Indicates time The increase in coal volume at the kiln head compared to the previous moment. This represents the optimization target for NOx emissions and ammonia consumption. For the first NOx prediction value at time 10:00 The NOx content limit specified in the process standard. Indicates the first The increase in ammonia usage at any given time compared to the previous time. The variables input to the f-CaO prediction model, , These are the variables input into the NOx prediction model; The objective function is solved using the multi-objective genetic algorithm NSGA-II to obtain the optimal solution vector for the state parameters. The optimal solution vector is optimized based on the state parameters, and the optimal solution interval for the state parameters is constructed.

[0018] Secondly, embodiments of the present invention provide a multi-objective optimized intelligent predictive control device for a cement decomposition furnace, comprising: The f-CaO prediction module is used to input the f-CaO influence parameters at the current time into the f-CaO prediction model to obtain the f-CaO prediction value at the next time step output by the f-CaO prediction model. The f-CaO prediction model is built on a CNN-P-sLSTM network and trained. The NOx prediction module is used to input the NOx impact parameters at the current time into the NOx prediction model to obtain the NOx prediction value at the next time step output by the NOx prediction model. The NOx prediction model is built on a CNN-minGRU network and trained. The calculation module is used to calculate the optimal solution range of state parameters based on the predicted values ​​of f-CaO and NOx at the next time step using the multi-objective genetic algorithm NSGA-II. The temperature prediction module is used to input the temperature influence parameters at the current moment into the decomposer temperature prediction model to obtain the decomposer temperature prediction value for the next moment output by the decomposer temperature prediction model; wherein, the decomposer temperature prediction model is built based on the Reformer-KAN network and obtained through training. The control module is used to determine the required amount of coal and alternative fuel for the current moment based on the optimal solution range of the state parameters and the predicted value of the decomposition furnace temperature at the next moment; and to perform intelligent control of the cement decomposition furnace according to the required amount of coal and alternative fuel for the current moment.

[0019] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0020] In this embodiment of the invention, the predicted values ​​of f-CaO, NOx, and decomposer temperature at the next time step are obtained through an f-CaO prediction model, a NOx prediction model, and a decomposer temperature prediction model, respectively. The f-CaO, NOx, and decomposer temperature prediction models are constructed and trained based on CNN-P-sLSTM, CNN-minGRU, and Reformer-KAN networks, respectively. Based on the predicted values ​​of f-CaO and NOx at the next time step, the optimal solution interval for the state parameters is calculated using the multi-objective genetic algorithm NSGA-II. Based on the optimal solution interval for the state parameters and the predicted decomposer temperature at the next time step, the required coal feed rate and alternative fuel rate for the current time step are determined, and intelligent control of the cement decomposer furnace is performed based on these required amounts.

[0021] The solution of this invention forms a closed loop of prediction-optimization-control through parallel prediction of three models. Based on the predicted values ​​of f-CaO and NOx at the next time step, the optimal solution range of state parameters is obtained by calculating using the multi-objective genetic algorithm NSGA-II. This can ensure the quality of cement products while reducing pollution emissions. Based on the optimal solution range of state parameters and the predicted value of the decomposer temperature at the next time step, the required amount of coal feed and alternative fuel is determined at the current time. Intelligent control of the cement decomposer can stably control the decomposer temperature and increase the proportion of alternative fuels used, thereby improving resource utilization and reducing production costs. Therefore, the solution of this invention can reduce production costs and pollution emissions, and improve the quality of cement products and resource utilization. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the implementation of the intelligent predictive control method for multi-objective optimized cement decomposition furnace provided in this embodiment of the invention. Figure 2 This is a flowchart of the f-CaO prediction model and NOx prediction model provided in the embodiments of the present invention; Figure 3 This is a flowchart of the optimized multi-objective genetic algorithm NSGA-II provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the decomposition furnace temperature prediction model provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the intelligent predictive control device for multi-objective optimization of cement decomposition furnace provided in an embodiment of the present invention. Detailed Implementation

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] Figure 1 This is a flowchart illustrating the implementation of the multi-objective optimized intelligent predictive control method for cement decomposition furnaces provided in this embodiment of the invention. Figure 1 As shown, the method includes the following steps: Step 101: Input the f-CaO influence parameters at the current time into the f-CaO prediction model to obtain the f-CaO prediction value at the next time point output by the f-CaO prediction model; Step 102: Input the NOx impact parameters at the current time into the NOx prediction model to obtain the NOx prediction value at the next time point output by the NOx prediction model. Step 103: Based on the predicted values ​​of f-CaO and NOx at the next time step, the optimal solution interval for the state parameters is obtained by calculating using the multi-objective genetic algorithm NSGA-II. Step 104: Input the temperature influence parameters at the current moment into the decomposer temperature prediction model to obtain the decomposer temperature prediction value for the next moment output by the decomposer temperature prediction model. Step 105: Based on the optimal solution range of the state parameters and the predicted value of the decomposer temperature at the next moment, determine the amount of coal and alternative fuel required at the current moment; Step 106: Intelligent control of the cement decomposition furnace based on the required coal feed and alternative fuel quantities at the current moment.

[0025] In practical applications, the main implementer of this multi-objective optimized intelligent predictive control method for cement decomposition furnaces can be a multi-objective optimized intelligent predictive control device for cement decomposition furnaces. There are various ways to implement this device. For example, it can be implemented through computer programs, such as application software; or, for example, chips. It can also be implemented as a medium storing the relevant computer program, such as a USB flash drive or cloud storage; or, it can be implemented through a physical device that integrates or installs the relevant computer program, such as a server or computer.

[0026] See Figure 1 Step 101 includes: inputting the f-CaO influence parameters at the current time into the f-CaO prediction model to obtain the f-CaO prediction value at the next time time output by the f-CaO prediction model; wherein, the f-CaO prediction model is constructed based on the CNN-P-sLSTM network and obtained through training.

[0027] Among them, f-CaO influencing parameters refer to parameters that have a significant impact on f-CaO content. For example, f-CaO influencing parameters include, but are not limited to, the decomposition furnace outlet temperature. Kiln tail temperature , secondary air temperature Total feed amount High-temperature fan speed EP fan speed feedback .

[0028] Preferably, in one possible implementation, step 101 includes: The f-CaO influence parameters at the current time are input into the one-dimensional convolutional layer of the f-CaO prediction model. The one-dimensional convolutional layer uses the tanh activation function to perform nonlinear feature extraction on the f-CaO influence parameters at the current time and outputs the first convolutional feature. The first convolutional feature is input into the P-sLSTM layer of the f-CaO prediction model for temporal feature modeling. The P-sLSTM layer performs the following processing: the first convolutional feature is divided into multiple channel-independent segments and linearly projected to obtain projected segments; the projected segments are processed by an xLSTM module containing sLSTM units to capture short-term dependency information and obtain segment features; the segment features are flattened and combined by a linear layer to obtain global temporal features. The global temporal features are input into the fully connected output layer of the f-CaO prediction model for mapping, and the f-CaO prediction value for the next time step is output.

[0029] Figure 2 This is a flowchart of the f-CaO prediction model and NOx prediction model provided in the embodiments of the present invention. Combined with... Figure 2 The shape of the input layer is determined by the time window length and the input variables. The input variables are then processed by a one-dimensional convolutional layer, which requires setting the number of input and output channels, the number of convolutional kernels, and their shapes, and employs the tanh activation function to achieve non-linear feature extraction. In the temporal feature modeling stage, the input and output structures of the P-sLSTM layer are set, and finally, a fully connected layer is introduced as the output layer. Through these steps, the key parameters of the model are initialized, and the overall structure is completed.

[0030] Predicted f-CaO value at the next time step The calculation formula is as follows:

[0031] See Figure 2First, the f-CaO influence parameters at the current time step are obtained from the data input layer. Then, a convolutional layer fuses the input data to extract time-series features. Finally, a P-SLSTM layer further extracts these features, producing the final prediction result: the predicted f-CaO value for the next time step. During training, the parameters updated by the convolutional layers include the weight matrix. and bias vector The parameters that need to be updated in the P-SLSTM layer are similar to those in the traditional LSTM layer, including the input gate weights. Circular weights Bias Forget Gate Weight Circular weights bias Output gate weights Circular weights bias and cell state weights Circular weights bias .

[0032] In the P-SLSTM network structure , They are respectively Time and Cellular state at any given moment , These are the hidden states at adjacent time steps, and the input data. Hidden state from the previous moment The updates are performed through the forget gate, input gate, and output gate respectively. The formula for the forget gate is as follows:

[0033] The formulas for the input gate and candidate states are as follows:

[0034]

[0035] Its output and Multiply and add the old cell state To obtain new cellular states that need to be remembered. .

[0036] The normalized state is updated as follows:

[0037] The output gate update formula is as follows:

[0038]

[0039] Finally, the output of the last P-SLSTM unit As a prediction result, that is, the predicted value of f-CaO at the next moment.

[0040] Compared to traditional LSTM, P-SLSTM introduces a parallelization mechanism, which improves computational efficiency and significantly increases training speed when modeling long sequences. Furthermore, through the optimization of the gating structure and parameter sharing strategy, redundant computation is reduced and the risk of overfitting is lowered. Its improved state update method can better capture long-distance dependency information, thus exhibiting stronger stability and generalization ability when processing complex nonlinear time series data, achieving a dual improvement in accuracy and efficiency.

[0041] See Figure 2 , Figure 2 The diagram illustrates the overall structure of the designed P-sLSTM, where multivariate time series data is segmented into different channels. These channels share the same backbone structure but are processed independently. Each channel's univariate sequence is segmented into multiple fragments, processed through linear layers, and after passing through multiple system modules, another linear layer produces the final prediction. Specifically, the size is... sample batch Originally transformed into And undergoes independent channel processing, the result is Then, this data undergoes a sharding process to generate... ,in It is the number of fragments. This refers to the size of the fragment. Next, the data is projected through a linear layer to... The data is then processed using the xLSTM module. After flattening, the data becomes... Finally, it is projected onto a linear layer. The output was reshaped into .

[0042] It should be noted that although the activation function of sLSTM has been adjusted, the model cannot retain long-term memory indefinitely, and therefore it still cannot effectively capture long-term dependencies. Therefore, this application uses a slicing method to manually divide the original time series into multiple components, allowing sLSTM to extract different short-term information; finally, these information can be combined through linear layers to obtain global information.

[0043] Based on the above explanation, the f-CaO prediction value for the next time step can be obtained using the f-CaO prediction model. Furthermore, the NOx prediction value for the next time step can be obtained using the NOx prediction model.

[0044] See Figure 1 Step 102 includes: inputting the NOx impact parameters at the current time into the NOx prediction model to obtain the NOx prediction value at the next time output by the NOx prediction model; wherein, the NOx prediction model is constructed based on the CNN-minGRU network and trained.

[0045] Among them, NOx impact parameters refer to parameters that have a significant impact on NOx emissions. For example, NOx impact parameters include, but are not limited to, the decomposer outlet temperature. Kiln tail temperature Coal feed amount Alternative fuel quantity Coal feedback from the kiln head Total flow rate of ammonia water .

[0046] Preferably, in one possible implementation, step 102 includes: The NOx impact parameters at the current time step are input into the one-dimensional convolutional layer of the NOx prediction model for information fusion and time series feature extraction, outputting the second convolutional feature. The second convolutional feature is then input into the minGRU layer of the NOx prediction model for temporal dependency modeling. The minGRU layer performs the following processing: based on the second convolutional feature and the hidden state at the previous time step, update weights are calculated through an update gate; combining the update weights and the hidden state at the previous time step, candidate hidden states at the current time step are calculated; based on the update weights, the hidden state at the previous time step, and the candidate hidden states at the current time step, the hidden state at the current time step is calculated; the hidden state at the current time step is input into the fully connected output layer of the NOx prediction model for mapping, outputting the NOx prediction value for the next time step.

[0047] Specifically, for the input and output structure of the minGRU layer, the input dimension corresponds to the time window length and the number of process variables, the number of hidden units is determined as h based on experimental results, and the output dimension is consistent with the number of predicted target variables.

[0048] NOx prediction at the next time step The calculation formula is as follows:

[0049] Combination Figure 2 First, the NOx impact parameters at the current time step are obtained from the data input layer. Then, the input data is fused through a one-dimensional convolutional layer to extract time series features. These features are then input into the minGRU model to obtain preliminary prediction results. Finally, a fully connected layer completes the mapping and outputs the final predicted value, which is the NOx prediction value for the next time step.

[0050] During training, the parameters that need to be updated in the convolutional layer include the weight matrix. With bias vector The minGRU layer simplifies the traditional GRU structure by removing the independent reset gate and retaining only an update gate, thereby reducing the number of parameters and accelerating convergence. The parameters required for updating include the update gate weights. Circular weights Bias Candidate state weights Circular weights Bias .

[0051] Specifically, the update gate formula is as follows:

[0052] The formula for candidate hidden states is as follows:

[0053] The final hidden state update formula is:

[0054] in, for Input at any time, and They are respectively Time and The hidden state at all times It is the Sigmoid activation function. It is the hyperbolic tangent function. This indicates element-wise multiplication.

[0055] Through the above calculations, minGRU retains the GRU's ability to capture temporal dependencies while reducing the number of gating units, resulting in a simpler structure, lower computational overhead, and greater suitability for handling long sequence inputs in prediction tasks. The final hidden state is obtained. The input is fully connected layer to complete the mapping, and the output is the prediction result, that is, the NOx prediction value at the next time step.

[0056] Further, see Figure 1 Step 103 includes: based on the predicted values ​​of f-CaO and NOx at the next time step, the optimal solution range of the state parameters is calculated using the multi-objective genetic algorithm NSGA-II.

[0057] Preferably, in one possible implementation, step 103 includes: Based on the f-CaO prediction model and the NOx prediction model, an objective function is constructed, which includes:

[0058]

[0059] The constraints of the objective function are:

[0060]

[0061] in, This indicates the goal of cement quality optimization. For a moment k The predicted value of f-CaO, This represents the optimal value for f-CaO content required by the process. Indicates time The increase in coal feeding compared to the previous moment. Indicates time The increase in coal volume at the kiln head compared to the previous moment. This represents the optimization target for NOx emissions and ammonia consumption. Let k be the predicted NOx value at time k. The NOx content limit specified in the process standard. This indicates the increase in ammonia usage at time k compared to the previous time. The variables input to the f-CaO prediction model, , These are the variables input into the NOx prediction model; The objective function is solved using the multi-objective genetic algorithm NSGA-II to obtain the optimal solution vector for the state parameters. The optimal solution vector is optimized based on the state parameters, and the optimal solution interval for the state parameters is constructed.

[0062] In practical applications, an objective function is constructed based on the f-CaO prediction model and the NOx prediction model. Within this objective function, This indicates the cement quality optimization target, and its corresponding For a moment k The predicted value of f-CaO, This represents the optimal value for f-CaO content required by the process. Indicates time k The increase in coal feeding compared to the previous moment. Indicates time k The incremental change in coal content at the kiln head compared to the previous moment. The objective function aims to stably control the f-CaO content generated during actual production around its optimal value to ensure the best cement quality. On the other hand, This represents the optimization target for pollutant (NOx) emissions and ammonia water consumption, where... For a moment k NOx prediction values, The NOx content limit specified in the process standard. Indicates time k The increase in ammonia usage compared to the previous moment. The variables input to the f-CaO prediction model, , These are the variables input into the NOx prediction model. The objective function is designed to control NOx emissions during actual production within the prescribed standard range while minimizing ammonia consumption, thereby meeting the system's environmental protection requirements.

[0063] Specifically, the multi-objective genetic algorithm NSGA-II is used to evaluate the objective function. , To optimize the parameters, first, a suitable population size is set and individual chromosomes are randomly initialized. Then, the fitness of individuals is evaluated using the objective function. Excellent individuals are selected by fast non-dominated sorting and crowding calculation. A new generation of population is generated by combining crossover and mutation operators. The Pareto front solution set is continuously updated during the iteration process to achieve global optimization as much as possible while satisfying the objective function and constraints.

[0064] Optimize the optimal solution vector (denoted as ) based on the state parameters obtained from NSGA-II. , Construct the corresponding state parameter solution vector interval, which serves as the constraint condition for optimizing the control quantity during predictive control:

[0065] in The value can be either 1 or 2, and will fluctuate around the optimal solution value. The range is used as the range of the optimization solution. In practical applications, the floating range can be flexibly adjusted according to the working conditions.

[0066] Figure 3 This is a flowchart of the optimized multi-objective genetic algorithm NSGA-II provided in an embodiment of the present invention. (Refer to...) Figure 3 For example, the specific steps of NSGA-II are as follows: (1) Initialize the population: Randomly generate a set of candidate solutions as the initial population, assign a parameter vector to each individual, and define a fitness function to evaluate the quality of the solution.

[0067] (2) Non-dominated sorting: Based on the principle of multi-objective optimization, the population is divided into several non-dominated layers, with the first layer being the optimal non-dominated solution set. If the solution... Non-inferior to the solution on all objectives And superior to at least one objective Then it is called non-dominance .

[0068] (3) Crowding Calculation: In each layer, crowding distance is used to measure the uniformity of individual distribution in the target space. For individuals Its crowded distance Defined as:

[0069] in, The number of objective functions, , Individuals In the The function values ​​of adjacent individuals on each target. , These are the maximum and minimum values ​​of the target, respectively.

[0070] (4) Selection, crossover and mutation: A tournament selection mechanism based on fast non-dominated sorting and crowding distance is adopted, and crossover and mutation operators are combined to generate offspring individuals.

[0071] (5) Merging and Elite Retention: Merge the parent and offspring populations, and select the best offspring based on their non-dominant ranking and crowding levels. The best individuals form a new generation of the population.

[0072] (6) Iteration: Repeat the above steps until the preset termination condition is met, and then calculate the optimal solution vector (denoted as ). , Furthermore, a state parameter optimization solution vector interval is established as a constraint condition in the predictive control process.

[0073] Further, see Figure 1 Step 104 includes: inputting the temperature influence parameters at the current moment into the decomposition furnace temperature prediction model to obtain the decomposition furnace temperature prediction value for the next moment output by the decomposition furnace temperature prediction model; wherein, the decomposition furnace temperature prediction model is constructed based on the Reformer-KAN network and obtained through training.

[0074] Among them, temperature-affected parameters refer to parameters that have a significant impact on the temperature changes of the decomposition furnace. In one possible implementation, temperature-affected parameters include, but are not limited to, the primary cylinder outlet temperature. oxygen concentration in the first stage cylinder Feeding amount High-temperature fan speed Historical values ​​of decomposition furnace outlet temperature Coal feed amount Alternative fuel quantity .

[0075] Preferably, in one possible implementation, step 104 includes: The temperature-affecting parameters at the current moment are divided into batches, and a first and second sequence with overlapping parts are generated based on the time series. The first sequence is input into the Encoder layer of the decomposer furnace temperature prediction model. The Encoder layer performs the following processing: position encoding and feature mapping on the first sequence to obtain a first encoding result; the first encoding result is split into a first sub-data and a second sub-data along the channel dimension; the first sub-data is processed sequentially through a multi-head locality sensitive hashing layer and a feedforward network layer to obtain a first processing result; the first processing result is concatenated with the second sub-data and layer normalized to obtain the first feature data; the first feature data is distilled to compress the time dimension and output the encoded features. The second sequence is input into the Decoder layer of the decomposer temperature prediction model. The Decoder layer performs the following processing: position encoding and feature mapping on the second sequence to obtain the second encoding result; the second encoding result is split into a third sub-data and a fourth sub-data along the channel dimension; the third sub-data is processed sequentially through a multi-head mask locality-sensitive hashing layer and a feedforward network layer to obtain the second processing result; the second processing result is then residually connected and layer normalized with the fourth sub-data to obtain the second feature data; the second feature data and the encoded features are input into a cross-attention layer for interactive computation to obtain the Decoder output features. The Decoder output features are input into the KAN output layer for feature integration and dimensionality reduction, and the predicted value of the decomposition furnace temperature at the next time step is output.

[0076] Figure 4 This is a schematic diagram of the structure of the decomposition furnace temperature prediction model provided in an embodiment of the present invention. See also... Figure 4Specifically, the decomposer furnace temperature prediction model adopts an "Encoder-Decoder" structure. The temperature-influencing parameters at the current moment are processed twice. Each batch of data is divided into two sets of sequences with overlapping portions based on the time series, and these are input into the Encoder and Decoder layers of the prediction model, respectively. The Encoder layer first splits the data along the channel dimension, feeding it into a Multi-Head Locality-Sensitive Hashing Layer (LSH) and a Feedforward Neural Network Layer (FFN) to achieve efficient modeling and feature extraction of long sequences. The Decoder layer similarly splits the data, feeding it into a Multi-Head Locality-Sensitive Hashing Layer and a Feedforward Neural Network Layer, then combining the results with the Encoder layer's output into a cross-attention layer. Finally, a KAN layer completes feature integration and dimensionality reduction, outputting the predicted decomposer furnace temperature for the next moment.

[0077] For example, the predicted temperature of the decomposer at the next moment. The expression is as follows:

[0078] To construct a model for predicting the outlet temperature of a decomposer furnace, this invention combines the Reformer prediction model, traditionally used for long-sequence modeling tasks, with KAN, improving the original structure to enhance model performance, reduce overall complexity, and further improve prediction accuracy. The Reformer model also employs an "Encoder-Decoder" framework. After parameter initialization, the preprocessed data is further divided into batches. Each batch of data is segmented into two overlapping data segments based on time-series characteristics, which are then input into the model's Encoder and Decoder, respectively. The data entering the Encoder first undergoes an encoding process. Positional Embedding is used to capture potential positional information in the input sequence, while Token Embedding maps the original data into a high-dimensional embedding vector through one-dimensional convolution, thereby extracting richer temporal features. The main parameters that this model needs to learn during training include the weights and biases of the convolutional layers, as well as the weights and biases of each linear transformation layer.

[0079] The first sequence is input into the Encoder layer, which performs the following processing: positional encoding and feature mapping on the first sequence to obtain the first encoding result; the first encoding result is split into a first sub-data and a second sub-data along the channel dimension; the first sub-data is processed sequentially through a multi-head locality sensitive hashing layer and a feedforward network layer to obtain the first processing result, and the first processing result is concatenated with the second sub-data and layer normalized to obtain the first feature data; the first feature data is distilled to compress the time dimension and output the encoded features.

[0080] In this invention, the concept of reversible residuals is introduced to divide the encoded data into two halves along the channel dimension:

[0081] Calculate using two routes:

[0082]

[0083] Among them, LN( This indicates the layer normalization process. This is the output of a multi-head Locality Sensitive Hashing (LSH) sparse self-attention layer. This is the output of the feedforward neural network layer.

[0084] Regarding the multi-head LSH sparse self-attention layer, firstly, the input data is linearly mapped to generate a query vector, a key vector, and a value vector, calculated as follows:

[0085]

[0086]

[0087] in, , , These are the trainable weight matrices. Traditional self-attention mechanisms require calculating the dot product of the query and all keys over the entire sequence, with a computational complexity of O(n log n). In long sequence prediction tasks, computational and storage overhead is excessive. To reduce computational complexity, this invention utilizes the locality-sensitive hashing mechanism introduced in the Reformer model. By hashing the query vector and key vector, similar vectors are assigned to the same hash bucket, and attention calculation is performed only within the bucket, thereby reducing the complexity to a fraction of the total computation. .

[0088] Within a single hash bucket, the first The formula for calculating the attention of each query vector is as follows:

[0089] in, Representation and Query A set of keys that are in the same hash bucket. is the scaling factor. This method allows the model to significantly reduce redundant computation while maintaining its ability to model long sequences.

[0090] To further enhance the model's feature extraction capabilities, this invention employs a multi-head attention mechanism, performing the aforementioned LSH attention computation in parallel across different subspaces to capture multi-scale temporal dependency information. The outputs of each attention head are concatenated and then restored to the original dimension through a linear mapping.

[0091] The feedforward neural network computes each token in the sequence independently, without involving cross-token interactions. It takes the form of a two-layer MLP:

[0092] in, It is the ReLU function; , It is a bias term.

[0093] The feedforward neural network consists of two linear transformations and an intermediate activation function. First, the channel features of each token are increased in dimensionality, and then compressed back to the original dimension after a nonlinear transformation, which facilitates subsequent splicing processing.

[0094] After the input data is split along the channel dimension, After calculation using the LSH sparse self-attention layer and... Perform residual joins, then normalize to obtain... ; After calculation by the FFN layer and Perform residual joins, then normalize to obtain... The two sets of outputs are concatenated along the channel dimension to obtain the final output of the reversible residual layer. The time dimension of the data is compressed using a distillation layer, and this compressed data serves as a "reference" input to the decoder layer for subsequent processing. The calculation formula for the distillation layer is as follows:

[0095] in, Indicates attention block, and These represent the data before entering the attention block and the data after pooling, respectively. This process effectively reduces data dimensionality while preserving key temporal features.

[0096] Through the above steps, the prediction model of the present invention effectively reduces computational complexity and storage overhead while maintaining the ability to capture global dependencies of sequences, thus demonstrating high computational efficiency and prediction accuracy in the task of predicting the outlet temperature of the decomposer.

[0097] The second sequence is input into the Decoder layer of the decomposer temperature prediction model. The Decoder layer performs the following processing: position encoding and feature mapping on the second sequence to obtain the second encoding result; the second encoding result is split into a third sub-data and a fourth sub-data along the channel dimension; the third sub-data is processed sequentially through a multi-head mask locality-sensitive hash layer and a feedforward network layer to obtain the second processing result; the second processing result is then residually connected and layer normalized with the fourth sub-data to obtain the second feature data; the second feature data and the encoded features are input into a cross-attention layer for interactive computation to obtain the Decoder output features.

[0098] The second sequence is input to the Decoder layer. First, it undergoes a reversible residual process similar to that of the encoder layer (the only difference being the addition of a masking mechanism in the attention layer) to capture key information within the sequence. Then, its output, along with the feature "reference" from the Encoder layer, enters a multi-head cross-attention layer for interactive computation, resulting in a more comprehensive temporal correlation representation. Finally, the Decoder output features undergo dimensionality reduction and information integration using a single-layer KAN to obtain the predicted furnace temperature for the next time step.

[0099] This invention introduces KAN on the basis of Reformer, which simplifies the complex structure of some feature extraction in the original Reformer and enhances the model's information extraction ability and overall modeling ability, enabling the model to achieve high-precision prediction with a more streamlined hierarchical structure and fewer parameters.

[0100] Unlike traditional MLP networks that rely on a combination of discrete linear and nonlinear processes to achieve feature mapping, KAN combines multiple trained B-spline curves with good fitting properties, making the neurons themselves behave as complex nonlinear curve functions. This structure endows the network with stronger nonlinear expressive power, enabling it to effectively extract complex feature relationships from input data with a relatively simple hierarchical structure, even surpassing the performance of traditional deep MLPs. Furthermore, the input data achieves residual connections after passing through the silu function and the combined curve output, thus avoiding the gradient vanishing problem and introducing trainable scaling factors to stabilize model training.

[0101] For example, the core calculation formula for the KAN layer includes:

[0102]

[0103]

[0104] in, It is the first A third-order B-spline curve, the combined spline curve is composed of... It is composed of linear combinations of three-order B-splines.

[0105] like Figure 4 As shown, the Reformer-KAN model replaces the traditional fully connected layer with KAN, and obtains the necessary information for calculating the attention probability distribution by extracting and mapping the encoded data. , , The attention calculation result is then passed through a KAN layer for data dimensionality reduction. Combining the Reformer structure, the Encoder layer only needs to calculate the LSH sparse self-attention score once, which, after data integration and compression, can be used as input to the Decoder layer. Simultaneously, due to KAN's powerful fitting ability and streamlined hierarchical structure, model complexity is significantly reduced, training speed is accelerated, and prediction accuracy is improved. Experimental results with KAN also demonstrate that this method can achieve good prediction results with fewer data samples.

[0106] See Figure 1 Step 105 includes: determining the required coal feed rate and alternative fuel rate for the current moment based on the optimal solution range of the state parameters and the predicted temperature of the decomposition furnace at the next moment. Step 106 includes: intelligently controlling the cement decomposition furnace according to the required coal feed rate and alternative fuel rate for the current moment.

[0107] In this embodiment, the method for determining the required coal feed and alternative fuel amount at the current moment is not specifically limited. In one example, the optimal solution interval of the state parameters is used as a constraint, and the predicted value of the precalciner temperature at the next moment is used as input. The required coal feed and alternative fuel amount at the current moment are obtained through a large model or machine learning model. In another example, the required coal feed and alternative fuel amount at the current moment are determined by using a pre-set empirical mapping strategy, corresponding to the optimal solution interval of the state parameters and the predicted value of the precalciner temperature at the next moment. In yet another example, the optimal solution interval of the state parameters and other state setting parameters are used as constraints, and the predicted value of the precalciner temperature at the next moment is used as input. The required coal feed and alternative fuel amount at the current moment are determined through a particle swarm optimization algorithm.

[0108] Optionally, in one possible implementation, step 105 specifically includes: using the optimal solution interval of the state parameters and other state setting parameters as constraints, using the predicted value of the decomposer temperature at the next moment as input, and receiving feedback parameters, the feedback parameters characterizing the difference between the actual temperature of the decomposer and the predicted value of the decomposer temperature, and determining the required amount of coal and alternative fuel at the current moment through a particle swarm optimization algorithm.

[0109] Optionally, in one possible implementation, step 106 specifically includes: sending the required amount of coal and alternative fuel at the current moment to the bottom-level controller, so that the bottom-level controller generates control commands corresponding to the required amount of coal and alternative fuel at the current moment; and sending control commands to the decomposition furnace through the bottom-level controller to perform intelligent control of the cement decomposition furnace.

[0110] Understandably, based on the optimal solution range of state parameters and the predicted temperature of the decomposer at the next moment, the required coal feed and alternative fuel quantities for the current moment are determined. This process comprehensively considers factors such as production costs, pollution emissions, cement product quality, and resource utilization, ultimately determining the required coal feed and alternative fuel quantities for the current moment that balance production costs, pollution emissions, cement product quality, and resource utilization. Based on this, intelligent control of the cement decomposer according to the required coal feed and alternative fuel quantities for the current moment can reduce production costs and pollution emissions, while also improving cement product quality and resource utilization.

[0111] Based on the above explanation, the decomposer temperature prediction model is built upon and trained using the Reformer-KAN network. Specifically, the construction and training process of the decomposer temperature prediction model will be explained in detail below.

[0112] Preferably, in one possible implementation, prior to step 104 above, the method further includes: The temperature influence parameters and actual decomposer temperature values ​​under historical time periods are preprocessed, and a first training set and a first test set are constructed based on the preprocessed temperature influence parameters and actual decomposer temperature values ​​under historical time periods. An initial decomposer temperature prediction model is constructed based on the Reformer-KAN network. The initial decomposer temperature prediction model is trained based on the first training set and the first test set to obtain the decomposer temperature prediction model.

[0113] Specifically, temperature influence parameters and actual decomposer temperatures during historical periods can be obtained from a database. The actual decomposer temperatures during historical periods serve as reference values ​​for the decomposer temperature prediction model training process. During training, the predicted decomposer temperature at each historical moment is compared with the actual decomposer temperature at each historical moment to determine the accuracy and effectiveness of the model.

[0114] In practical applications, after obtaining the temperature influence parameters and the actual temperature values ​​of the decomposer furnace during historical periods, preprocessing can be performed. Specifically, preprocessing includes: using the Pearson correlation coefficient method to test the correlation between various temperature influence parameters after temporal misalignment, thereby determining the possible time lag relationships between variables, and organizing the raw data accordingly. Subsequently, based on... The criteria remove outliers from the data, and then normalize the obtained valid data to obtain the preprocessed temperature influence parameters and the actual temperature values ​​of the decomposition furnace during the historical period.

[0115] Furthermore, based on the preprocessed temperature influence parameters and the actual temperature values ​​of the decomposition furnace during the historical periods, a first dataset is constructed. The first dataset is then divided proportionally: the first 60% is used as the first training set, the middle 20% as the second validation set, and the last 20% as the third test set.

[0116] Because the collected data has time-series characteristics, a sliding window method is used to process it. The sliding window technique continuously moves a fixed-size time window across the time series, covering data segments at different locations, thus balancing local feature extraction and real-time computation to achieve continuous analysis and modeling of time-series data. According to the requirements of this application, the window size is set to 5 (i.e., each window contains data from 5 time points), which reduces the data size per batch to alleviate the model training burden while enhancing the model's ability to learn and extract time-series features.

[0117] Optionally, in one possible implementation, the Adam optimizer is used to train the model for predicting the outlet temperature of the decomposer. After successful model training, the validation set data is input into the prediction model to obtain the output results, and the MAE and RMSE test methods are used to verify the model fitting results.

[0118] Based on the above explanation, the f-CaO prediction model is built upon and trained using a CNN-P-sLSTM network. Specifically, the construction and training process of the f-CaO prediction model will be explained in detail below.

[0119] Optionally, in one possible implementation, prior to step 101 above, the method further includes: The f-CaO impact parameters and actual f-CaO emissions for historical periods are preprocessed. Based on the preprocessed f-CaO impact parameters and actual f-CaO emissions for historical periods, a second training set, a second validation set, and a second test set are constructed. An initial f-CaO prediction model is constructed using a CNN-P-sLSTM network. The initial f-CaO prediction model is then trained using the second training set, the second validation set, and the second test set to obtain the f-CaO prediction model.

[0120] The training process for the f-CaO prediction model is similar to that for the decomposition furnace temperature prediction model described above, and will not be repeated here. Please refer to the explanation above for details.

[0121] Based on the above explanation, the NOx prediction model is built upon and trained using a CNN-minGRU network. Specifically, the construction and training process of the NOx prediction model will be explained in detail below.

[0122] Optionally, in one possible implementation, prior to step 102 above, the method further includes: The NOx impact parameters and actual NOx emissions for historical periods are preprocessed. Based on the preprocessed NOx impact parameters and actual NOx emissions for historical periods, a third training set, a third validation set, and a third test set are constructed. An initial NOx prediction model is constructed using a CNN-minGRU network. The initial NOx prediction model is then trained using the third training set, the third validation set, and the third test set to obtain the final NOx prediction model.

[0123] The training process for the NOx prediction model is similar to that for the decomposer temperature prediction model described above, and will not be repeated here. Please refer to the explanation above for details.

[0124] The technological advancements achieved by this invention are as follows: This invention addresses the optimization of coal feed rate and waste co-combustion rate in the decomposer, aiming to minimize the coal feed rate and maximize solid waste substitution under suitable process conditions, while ensuring the stability of operating parameters such as decomposer temperature, air volume, NOx emissions, and feed rate. Based on this, a decomposer outlet temperature prediction model is established, fully utilizing the advantages of Reformer networks in long sequence modeling to overcome the inherent nonlinearity, strong coupling, and time delay characteristics of the decomposer system, achieving rapid and accurate prediction of the decomposer outlet temperature. Simultaneously, KAN is introduced to simplify the complex structure of the Reformer feature extraction process and improve network performance, enabling the model to obtain more accurate prediction results with fewer layers and parameters. Furthermore, a decomposer state parameter optimization model is constructed, using P-sLSTM networks, minGRU networks, and the multi-objective genetic algorithm NSGA-II to predict f-CaO emissions and NOx content, and optimize related influencing parameters. This ensures cement product quality while reducing coal consumption and effectively reducing NOx and carbon emissions.

[0125] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0126] Figure 5 This is a schematic diagram of the structure of the intelligent predictive control device for multi-objective optimization of cement decomposition furnace provided in an embodiment of the present invention. See also... Figure 5 The multi-objective optimized intelligent predictive control device for cement decomposition furnace includes the following modules: f-CaO prediction module 51 is used to input the f-CaO influence parameters at the current time into the f-CaO prediction model to obtain the f-CaO prediction value at the next time output by the f-CaO prediction model; wherein, the f-CaO prediction model is built based on the CNN-P-sLSTM network and trained. NOx prediction module 52 is used to input the NOx influence parameters at the current time into the NOx prediction model to obtain the NOx prediction value at the next time output by the NOx prediction model; wherein, the NOx prediction model is built based on the CNN-minGRU network and trained. Calculation module 53 is used to calculate the optimal solution range of state parameters based on the predicted values ​​of f-CaO and NOx at the next time step using the multi-objective genetic algorithm NSGA-II. The temperature prediction module 54 is used to input the temperature influence parameters at the current moment into the decomposer temperature prediction model to obtain the decomposer temperature prediction value at the next moment output by the decomposer temperature prediction model; wherein, the decomposer temperature prediction model is built based on the Reformer-KAN network and obtained through training. The control module 55 is used to determine the required amount of coal and alternative fuel for the current moment based on the optimal solution range of the state parameters and the predicted value of the decomposition furnace temperature at the next moment; and to perform intelligent control of the cement decomposition furnace according to the required amount of coal and alternative fuel for the current moment.

[0127] In one possible implementation, the temperature-influencing parameters include: primary stage outlet temperature, primary stage oxygen concentration, feed rate, high-temperature blower speed, historical value of decomposer outlet temperature, coal feed rate, and alternative fuel quantity; the temperature prediction module 54 is specifically used for: The temperature-influencing parameters at the current moment are divided into batches, and a first sequence and a second sequence with overlapping parts are generated based on the time series. The first sequence is input into the Encoder layer of the decomposition furnace temperature prediction model. The Encoder layer performs the following processing: position encoding and feature mapping on the first sequence to obtain a first encoding result; the first encoding result is split into a first sub-data and a second sub-data along the channel dimension; the first sub-data is processed sequentially through a multi-head locality-sensitive hashing layer and a feedforward network layer to obtain a first processing result, and the first processing result is concatenated with the second sub-data and layer normalized to obtain the first feature data; the first feature data is distilled to compress the time dimension and output the encoded features; the second sequence is input into the decomposition furnace temperature prediction model. The Decoder layer of the furnace temperature prediction model performs the following processing: Position encoding and feature mapping are performed on the second sequence to obtain the second encoding result; the second encoding result is split into third and fourth sub-data along the channel dimension; the third sub-data is processed sequentially through a multi-head mask locality-sensitive hashing layer and a feedforward network layer to obtain the second processing result; the second processing result is then residually connected to the fourth sub-data and layer normalized to obtain the second feature data; the second feature data and the encoded features are input into a cross-attention layer for interactive computation to obtain the Decoder output features; the Decoder output features are input into a KAN output layer for feature integration and dimensionality reduction to output the decomposed furnace temperature prediction value for the next time step.

[0128] In one possible implementation, the device further includes: a first training module, configured to: preprocess the temperature influence parameters and the actual temperature values ​​of the decomposer furnace during historical periods, and construct a first training set, a first validation set, and a first test set based on the preprocessed temperature influence parameters and the actual temperature values ​​of the decomposer furnace during historical periods; construct an initial decomposer furnace temperature prediction model according to the Reformer-KAN network; and train the initial decomposer furnace temperature prediction model based on the first training set, the first validation set, and the first test set to obtain the decomposer furnace temperature prediction model.

[0129] In one possible implementation, the f-CaO influencing parameters include: decomposer outlet temperature, kiln tail temperature, secondary air temperature, total feed rate, high-temperature fan speed, and EP fan speed feedback; the f-CaO prediction module 51 is specifically used for: inputting the f-CaO influencing parameters at the current moment into a one-dimensional convolutional layer of the f-CaO prediction model, the one-dimensional convolutional layer using the tanh activation function to perform nonlinear feature extraction on the f-CaO influencing parameters at the current moment, and outputting the first convolutional feature; inputting the first convolutional feature into the P-sLSTM layer of the f-CaO prediction model for temporal feature modeling; the P-sLSTM layer performs the following processing: dividing the first convolutional feature into multiple channel-independent segments and performing linear projection to obtain projected segments; processing the projected segments through an xLSTM module containing sLSTM units to capture short-term dependency information and obtain segment features; flattening the segment features and combining the segment features through a linear layer to obtain global temporal features; inputting the global temporal features into the fully connected output layer of the f-CaO prediction model for mapping, and outputting the f-CaO prediction value for the next moment.

[0130] In one possible implementation, the device further includes: a second training module, configured to: preprocess the f-CaO impact parameters and actual f-CaO emissions during historical periods, and construct a second training set, a second validation set, and a second test set based on the preprocessed f-CaO impact parameters and actual f-CaO emissions during historical periods; construct an initial f-CaO prediction model using a CNN-P-sLSTM network; and train the initial f-CaO prediction model based on the second training set, the second validation set, and the second test set to obtain the f-CaO prediction model.

[0131] In one possible implementation, the NOx influencing parameters include: decomposer outlet temperature, kiln tail temperature, coal feed rate, alternative fuel quantity, kiln head coal feedback, and total ammonia water flow rate; the NOx prediction module 52 is specifically used for: inputting the NOx influencing parameters at the current time into the one-dimensional convolutional layer of the NOx prediction model for information fusion and time series feature extraction, and outputting the second convolutional feature; inputting the second convolutional feature into the minGRU layer of the NOx prediction model for time-series dependency modeling; the minGRU layer performs the following processing: calculating update weights through an update gate based on the second convolutional feature and the hidden state at the previous time; calculating the candidate hidden state at the current time by combining the update weights and the hidden state at the previous time; calculating the hidden state at the current time based on the update weights, the hidden state at the previous time, and the candidate hidden state at the current time; and inputting the hidden state at the current time into the fully connected output layer of the NOx prediction model for mapping, and outputting the NOx prediction value at the next time.

[0132] In one possible implementation, the device further includes: a third training module, configured to: preprocess the NOx impact parameters and actual NOx emissions during historical periods, and construct a third training set, a third validation set, and a third test set based on the preprocessed NOx impact parameters and actual NOx emissions during historical periods; construct an initial NOx prediction model based on a CNN-minGRU network; and train the initial NOx prediction model based on the third training set, the third validation set, and the third test set to obtain the NOx prediction model.

[0133] In one possible implementation, the calculation module 53 is specifically used to: construct an objective function based on the f-CaO prediction model and the NOx prediction model, the objective function including:

[0134]

[0135] The constraints of the objective function are:

[0136]

[0137] in, This indicates the goal of cement quality optimization. For a moment k The predicted value of f-CaO, This represents the optimal value for f-CaO content required by the process. Indicates time The increase in coal feeding compared to the previous moment. Indicates time The increase in coal volume at the kiln head compared to the previous moment. This represents the optimization target for NOx emissions and ammonia consumption. For the first NOx prediction value at time 10:00 The NOx content limit specified in the process standard. Indicates the first The increase in ammonia usage at any given time compared to the previous time. The variables input to the f-CaO prediction model, , These are the variables input into the NOx prediction model; The objective function is solved using the multi-objective genetic algorithm NSGA-II to obtain the optimal solution vector for the state parameters. The optimal solution vector is optimized based on the state parameters, and the optimal solution interval for the state parameters is constructed.

[0138] Furthermore, embodiments of the present invention also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the above method embodiments.

[0139] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A multi-objective optimized intelligent predictive control method for cement decomposition furnaces, characterized in that, include: The f-CaO influence parameters at the current time are input into the f-CaO prediction model to obtain the f-CaO prediction value at the next time step output by the f-CaO prediction model; wherein, the f-CaO prediction model is constructed based on the CNN-P-sLSTM network and trained. The NOx impact parameters at the current time are input into the NOx prediction model to obtain the NOx prediction value at the next time step output by the NOx prediction model; wherein, the NOx prediction model is built based on the CNN-minGRU network and trained. Based on the predicted values ​​of f-CaO and NOx at the next time step, the optimal solution range of the state parameters is obtained by calculating using the multi-objective genetic algorithm NSGA-II. The temperature influence parameters at the current moment are input into the decomposer temperature prediction model to obtain the decomposer temperature prediction value for the next moment output by the decomposer temperature prediction model; wherein, the decomposer temperature prediction model is constructed based on the Reformer-KAN network and obtained through training. Based on the optimal solution range of the state parameters and the predicted value of the decomposer temperature at the next moment, determine the required amount of coal and alternative fuel at the current moment. The cement decomposition furnace is intelligently controlled based on the required coal feed and alternative fuel quantities at the current moment.

2. The intelligent predictive control method for cement decomposition furnace with multi-objective optimization according to claim 1, characterized in that, The temperature-affecting parameters include: primary cylinder outlet temperature, primary cylinder oxygen concentration, feed rate, high-temperature fan speed, historical value of decomposer outlet temperature, coal feed rate, and alternative fuel rate. The step of inputting the temperature influence parameters at the current moment into the decomposer temperature prediction model to obtain the decomposer temperature prediction value for the next moment output by the decomposer temperature prediction model includes: The temperature-affecting parameters at the current moment are divided into batches, and a first and second sequence with overlapping parts are generated based on the time series. The first sequence is input into the Encoder layer of the decomposition furnace temperature prediction model. The Encoder layer performs the following processing: position encoding and feature mapping on the first sequence to obtain a first encoding result; the first encoding result is split into a first sub-data and a second sub-data along the channel dimension; the first sub-data is processed sequentially through a multi-head locality sensitive hashing layer and a feedforward network layer to obtain a first processing result; the first processing result is concatenated with the second sub-data and layer normalized to obtain first feature data; the first feature data is distilled to compress the time dimension and output encoded features. The second sequence is input into the Decoder layer of the decomposition furnace temperature prediction model; the Decoder layer performs the following processing: position encoding and feature mapping on the second sequence to obtain a second encoding result; the second encoding result is split into a third sub-data and a fourth sub-data along the channel dimension; the third sub-data is processed sequentially through a multi-head mask locality-sensitive hash layer and a feedforward network layer to obtain a second processing result, and the second processing result is residually connected and layer normalized with the fourth sub-data to obtain second feature data; the second feature data and the encoded features are input into a cross-attention layer for interactive calculation to obtain the Decoder output features; The Decoder output features are input into the KAN output layer for feature integration and dimensionality reduction, and the predicted value of the decomposition furnace temperature at the next time step is output.

3. The intelligent predictive control method for cement decomposition furnace with multi-objective optimization according to claim 1, characterized in that, Before inputting the temperature influence parameters at the current moment into the decomposer temperature prediction model to obtain the predicted decomposer temperature value for the next moment output by the decomposer temperature prediction model, the method further includes: The temperature influence parameters and the actual temperature values ​​of the decomposer under historical time periods are preprocessed, and the first training set, the first validation set and the first test set are constructed based on the preprocessed temperature influence parameters and the actual temperature values ​​of the decomposer under historical time periods. An initial model for predicting the temperature of the decomposer furnace was constructed based on the Reformer-KAN network. The initial decomposition furnace temperature prediction model is trained based on the first training set, the first validation set, and the first test set to obtain the decomposition furnace temperature prediction model.

4. The intelligent predictive control method for cement decomposition furnace with multi-objective optimization according to claim 1, characterized in that, The parameters affecting f-CaO include: decomposition furnace outlet temperature, kiln tail temperature, secondary air temperature, total feed rate, high-temperature fan speed, and EP fan speed feedback. The step of inputting the f-CaO influence parameters at the current moment into the f-CaO prediction model to obtain the f-CaO prediction value for the next moment output by the f-CaO prediction model includes: The f-CaO influence parameters at the current time are input into the one-dimensional convolutional layer of the f-CaO prediction model. The one-dimensional convolutional layer uses the tanh activation function to perform nonlinear feature extraction on the f-CaO influence parameters at the current time and outputs the first convolutional feature. The first convolutional feature is input into the P-sLSTM layer of the f-CaO prediction model for temporal feature modeling. The P-sLSTM layer performs the following processing: the first convolutional feature is segmented into multiple channel-independent segments and linearly projected to obtain projected segments; the projected segments are processed by an xLSTM module containing sLSTM units to capture short-term dependency information and obtain segment features; the segment features are flattened and combined by a linear layer to obtain global temporal features. The global temporal features are input into the fully connected output layer of the f-CaO prediction model for mapping, and the f-CaO prediction value for the next time step is output.

5. The intelligent predictive control method for cement decomposition furnace with multi-objective optimization according to claim 1, characterized in that, Before inputting the f-CaO influence parameters at the current moment into the f-CaO prediction model to obtain the f-CaO prediction value for the next moment output by the f-CaO prediction model, the method further includes: The f-CaO impact parameters and actual f-CaO emissions for historical periods are preprocessed, and a second training set, a second validation set, and a second test set are constructed based on the preprocessed f-CaO impact parameters and actual f-CaO emissions for historical periods. An initial f-CaO prediction model was constructed based on the CNN-P-sLSTM network; The initial f-CaO prediction model is trained based on the second training set, the second validation set, and the second test set to obtain the f-CaO prediction model.

6. The intelligent predictive control method for cement decomposition furnace with multi-objective optimization according to claim 1, characterized in that, The NOx-affecting parameters include: decomposer outlet temperature, kiln tail temperature, coal feed rate, alternative fuel quantity, kiln head coal feedback, and total ammonia water flow rate. The step of inputting the NOx impact parameters at the current moment into the NOx prediction model to obtain the NOx prediction value for the next moment output by the NOx prediction model includes: The NOx impact parameters at the current moment are input into the one-dimensional convolutional layer of the NOx prediction model for information fusion and time series feature extraction, and the second convolutional feature is output. The second convolutional feature is input into the minGRU layer of the NOx prediction model to model temporal dependencies; the minGRU layer performs the following processing: based on the second convolutional feature and the hidden state at the previous time step, update weights are calculated through an update gate; combining the update weights with the hidden state at the previous time step, the candidate hidden state at the current time step is calculated; based on the update weights, the hidden state at the previous time step, and the candidate hidden state at the current time step, the hidden state at the current time step is calculated. The hidden state at the current moment is input into the fully connected output layer of the NOx prediction model for mapping, and the NOx prediction value at the next moment is output.

7. The intelligent predictive control method for cement decomposition furnace with multi-objective optimization according to claim 1, characterized in that, Before inputting the NOx impact parameters at the current moment into the NOx prediction model to obtain the NOx prediction value for the next moment output by the NOx prediction model, the method further includes: The NOx impact parameters and actual NOx emissions for historical periods are preprocessed, and a third training set, a third validation set, and a third test set are constructed based on the preprocessed NOx impact parameters and actual NOx emissions for historical periods. An initial NOx prediction model was constructed based on the CNN-minGRU network; The initial NOx prediction model is trained based on the third training set, the third validation set, and the third test set to obtain the NOx prediction model.

8. The intelligent predictive control method for multi-objective optimization of cement decomposition furnace according to any one of claims 1-7, characterized in that, The optimal solution range for the state parameters, calculated using the multi-objective genetic algorithm NSGA-II based on the predicted f-CaO and NOx values ​​at the next time step, includes: Based on the f-CaO prediction model and the NOx prediction model, an objective function is constructed, the objective function comprising: The constraints of the objective function are: in, This indicates the goal of cement quality optimization. For a moment k The predicted value of f-CaO, This represents the optimal value for f-CaO content required by the process. Indicates time The increase in coal feeding compared to the previous moment. Indicates time The increase in coal volume at the kiln head compared to the previous moment. This represents the optimization target for NOx emissions and ammonia consumption. For the first NOx prediction value at time 10:00 The NOx content limit specified in the process standard. Indicates the first The increase in ammonia usage at any given time compared to the previous time. The variables input to the f-CaO prediction model, , These are the variables input into the NOx prediction model; The objective function is solved using the multi-objective genetic algorithm NSGA-II to obtain the optimal solution vector for the state parameters. Based on the state parameters, optimize the optimal solution vector and construct the optimal solution interval for the state parameters.

9. A multi-objective optimized intelligent predictive control device for cement decomposition furnaces, characterized in that, include: The f-CaO prediction module is used to input the f-CaO influence parameters at the current time into the f-CaO prediction model to obtain the f-CaO prediction value at the next time step output by the f-CaO prediction model; wherein, the f-CaO prediction model is built based on the CNN-P-sLSTM network and trained. The NOx prediction module is used to input the NOx impact parameters at the current time into the NOx prediction model to obtain the NOx prediction value at the next time output by the NOx prediction model; wherein, the NOx prediction model is built based on the CNN-minGRU network and trained. The calculation module is used to calculate the optimal solution range of state parameters based on the predicted values ​​of f-CaO and NOx at the next time step using the multi-objective genetic algorithm NSGA-II. The temperature prediction module is used to input the temperature influence parameters at the current moment into the decomposition furnace temperature prediction model to obtain the decomposition furnace temperature prediction value for the next moment output by the decomposition furnace temperature prediction model; wherein, the decomposition furnace temperature prediction model is constructed based on the Reformer-KAN network and obtained through training. The control module is used to determine the required coal feed and alternative fuel amount for the current moment based on the optimal solution range of the state parameters and the predicted value of the decomposition furnace temperature at the next moment; and to perform intelligent control of the cement decomposition furnace according to the required coal feed and alternative fuel amount for the current moment.

10. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 8.

Images