Cement decomposing furnace temperature prediction control method based on KAN enhanced neural network

Abstract

The invention provides a cement decomposing furnace temperature prediction control method based on a KAN enhanced neural network, and belongs to the technical field of cement calcination. A CNN-LSTM-KAN network is established as an f-CaO and Nox prediction model, and a state parameter optimization objective function is established in connection with actual production requirements; and solving an optimal solution and an optimal solution interval of the state parameters of the decomposition furnace through a multi-target particle swarm optimization algorithm. A KAN-Enhanced-Informer network is established to predict the temperature of the decomposition furnace, a target function is established by combining a state optimization result with the more accurate prediction result, the less coal consumption and the reasonable alternative fuel quantity as targets, the optimal solutions of the coal feeding quantity and the alternative fuel quantity are obtained, and the optimal solutions are applied to a control model to ensure that the temperature of the decomposition furnace reaches the optimal set value. The problems that in the actual cement production process, energy consumption is too high, the yield is not ideal, and much sewage is discharged are solved.

Description

Technical Field

[0001] The invention belongs to the technical field of cement calcination, and specifically relates to a cement decomposition furnace temperature prediction and control method based on a KAN enhanced neural network. Background Art

[0002] Compared with conventional cement production, the new dry process cement production introduces pre-decomposition and suspension preheating technology. The decomposition furnace, as the key equipment of the pre-decomposition technology, is located between the preheater and the rotary kiln, and undertakes the carbonate decomposition task that was originally only performed in the rotary kiln. At the same time, the coal consumption of the pre-decomposition link in the decomposition furnace accounts for 60% of the total, which is the core of the entire actual production process. Therefore, the normal operation of the decomposition furnace is a key step in the cement production process. For the preheater system, crusting and blocking are common and serious problems. In order to solve such problems, the selection of fuel and the control of the temperature in the furnace are crucial. In actual industrial production, in addition to using coal as a decomposition furnace fuel, municipal solid waste and the like are also used as alternative fuels for the decomposition furnace (hereinafter referred to as alternative fuels) to achieve the requirements of reducing costs and low pollution. However, the alternative fuels themselves have the characteristics of high water content, low calorific value, high volatile components, and unstable combustion, so that the use of alternative fuels may affect the stability of the temperature in the decomposition furnace.

[0003] In industrial production, the control of feed amount, coal feed amount and alternative fuel blending ratio is all tracked by setting values ​​obtained empirically or adjusted according to production data, which makes the data such as decomposition furnace temperature and pollution emission in cement production fluctuate greatly, and also has a certain impact on cement quality. Therefore, some key methods to keep the decomposition furnace temperature stable and the process running normally require the adjustment of coal feed amount, alternative fuel blending ratio and stable control of decomposition furnace temperature. At present, most cement plants use traditional PID control and manual control, which is difficult to meet production requirements, resulting in excessive energy consumption, unsatisfactory output and more pollution in the actual cement production process. Summary of the invention

[0004] In order to solve the problem that most of the cement plants currently use traditional PID control and manual control, which is difficult to meet production requirements. The present invention provides a cement decomposition furnace temperature prediction and control method based on KAN enhanced neural network, which can stabilize the decomposition furnace temperature and improve the proportion of alternative fuel use under the premise of ensuring cement quality and NOx emission standards, enhance the stability of the control system, thereby meeting production requirements and solving the problems of excessive energy consumption, unsatisfactory output and more pollution in the actual cement production process.

[0005] The technical solution of a cement decomposition furnace temperature prediction and control method based on KAN enhanced neural network of the present invention is:

[0006] A cement decomposition furnace temperature prediction and control method based on KAN enhanced neural network comprises the following steps:

[0007] S1. Establish f-CaO prediction model and Nox prediction optimization model, and establish objective function in combination with the actual needs of cement production. Use multi-objective particle swarm algorithm to find the optimal solution of decomposition furnace state parameters and obtain the optimal interval;

[0008] S2. Establish a decomposition furnace temperature prediction model to predict the decomposition furnace temperature;

[0009] S3. According to the decomposition furnace temperature prediction model and the optimal interval of state parameters, a control objective function is established. The optimal solution of the coal feeding amount and the alternative fuel feeding amount is obtained through the particle swarm algorithm. Combined with rolling optimization and feedback correction, stable control of the decomposition furnace temperature is achieved.

[0010] A further improvement of the technical solution of the present invention is that: the step S1 includes the following steps:

[0011] S1.1. According to the cement burning system, the predicted variables f-CaO emission and NOx content are selected, which are recorded as y1 and y2 respectively. The variables with greater influence on f-CaO content are selected, namely, the decomposition furnace outlet temperature x1, the kiln tail temperature x2, the secondary air temperature x3, the total feed amount x4, the high temperature fan speed x5, and the EP fan speed feedback x6. The variables with greater influence on NOx emission are selected, namely, the decomposition furnace outlet temperature x1, the kiln tail temperature x2, the coal feed amount u1, the alternative fuel amount u2, the kiln head coal feedback u3, and the total ammonia flow u4. The prediction model expression is as follows:

[0012] y1=f(x1,x2,x3,x4,x5,x6)#(1)

[0013] y2=f(x1,x2,u1,u2,u3,u4)#(2)

[0014] Export the above data from the database and pre-process the data;

[0015] S1.2. After data preprocessing, a CNN-LSTM-KAN network was established as the prediction model for f-CaO and Nox, and their values ​​were predicted respectively;

[0016] S1.3. Based on the f-CaO and NOx prediction models in S1.2, an objective function is constructed according to the requirements of cement product quality and energy conservation and emission reduction:

[0017] minJ1=||y1(k)-Y1|| 2

[0018] minJ2=||y2(k)-Y2|| 2+||Δu4(k)|| 2 #(3)

[0019]

[0020] In the above objective functions, J1 is the objective function for optimizing cement quality, y1(k) is the predicted value of f-CaO at time k, and Y1 is the optimal value of f-CaO content in the process requirements; J2 is the objective function for optimizing NOx emission and ammonia consumption, y2(k) is the predicted value of Nox at time k, Y2 is the standard value of Nox content in the process requirements, Δu4(k) is the increment of ammonia at time k compared with the previous moment, and x i is the input to the f-CaO prediction model, x i ,u i is the input quantity to the Nox prediction model;

[0021] S1.4. Use multi-objective particle swarm optimization algorithm to optimize the objective function J1 and J2 parameters.

[0022] A further improvement of the above technical solution of the present invention is that the data preprocessing in step S1.1 includes the following steps:

[0023] S1.1.1. Use the Pearson correlation coefficient method to perform correlation analysis on the variables. The analysis results determine the time delay between the variables and adjust the data based on the results. The Pearson correlation coefficient formula is:

[0024]

[0025] Among them, r is the calculated Pearson correlation coefficient, x i ,y i is the input data for different parameters;

[0026] S1.1.2. According to the 3σ principle, outliers in the data are processed and normalized. The normalization formula is:

[0027]

[0028] Among them, y j is the normalized result.

[0029] A further improvement of the above technical solution of the present invention is that: the step S1.4 specifically includes:

[0030] According to the obtained state parameters, the optimal solution vector is optimized and the state parameter optimization solution vector interval is established as the constraint condition for optimizing the control quantity in the predictive control process:

[0031]

[0032] The optimal solution range is taken as the 10% range around the optimal value.

[0033] A further improvement of the technical solution of the present invention is that: the step S2 includes the following steps:

[0034] S2.1. Predict the temperature of the decomposition furnace during normal operation;

[0035] S2.2. Using the "Encoder-Decoder" structure, a KAN-Enhanced-Informer decomposition furnace temperature prediction model is established, and the decomposition furnace temperature prediction value is output.

[0036] A further improvement of the above technical solution of the present invention is that: the step S2.1 specifically includes:

[0037] The variables that have a greater impact on the change of the decomposition furnace temperature are selected as the input variables of the decomposition furnace temperature prediction model, including the historical value of the decomposition furnace outlet temperature x1, the first-stage drum outlet temperature x2, the first-stage drum oxygen concentration x3, the feed amount x4, the high-temperature fan speed x5, the coal feed amount u1, and the alternative fuel amount u2. The above data are exported from the database and preprocessed.

[0038] A further improvement of the above technical solution of the present invention is that: the step S2.2 specifically includes:

[0039] The "Encoder-Decoder" structure is used to establish the KAN-Enhanced-Informer decomposition furnace temperature prediction model. After the model parameters are initialized, the preprocessed data in S2.1 are processed secondary in batches. Each batch of data is divided into two groups of data with overlapping parts according to the time series, and after encoding, they are input into the Encoder layer and Decoder layer of the prediction model respectively.

[0040] The further improvement of the above technical solution of the present invention is that: the encoder layer is composed of a sparse self-attention mechanism, a distillation layer and a KAN network, and the decoder layer is composed of a sparse self-attention mechanism combined with a mask mechanism, a full attention mechanism and a KAN network; the data entering the decoder layer is calculated by the sparse self-attention, and then enters the full attention mechanism layer with the output result of the encoder layer. The output result is finally integrated and reduced in dimension by the KAN network, and the predicted value of the decomposition furnace temperature is output. The expression is as follows:

[0041] y3=f(x1,x2,x3,x4,x5,u1,u2)#(7).

[0042] A further improvement of the technical solution of the present invention is that the control objective function in step S3 is:

[0043]

[0044] Among them, P and M are the prediction time domain step and control time domain step, respectively, y(k+s|k) is the predicted decomposition furnace temperature value at time k+s at time k, y s (k+s) is the optimal setting value of the decomposition furnace temperature at time k+s obtained by optimizing the state parameters in step S1, Δu1(k+h-1|k) is the increment of the decomposition furnace coal feeding amount predicted at time k+h-1, Q and R are two control coefficient matrices, ε is the value of the alternative fuel exceeding the optimal range defined by the optimal setting value of the alternative fuel amount at time k+h-1 obtained in step S1, and other x i 、u i The constraint range of is also obtained by the particle swarm algorithm in step S1. The constraint range of Δu1 can be changed according to the actual situation. The formula is as follows:

[0045]

[0046] Where u2 is the amount of alternative fuel, z 2min 、z 2max are the constraint boundary vectors Z 2min ,Z 2max The alternative fuel amount constrains the boundary component.

[0047] A further improvement of the technical solution of the present invention is that in step S3, the optimal solution of the coal feeding amount and the alternative fuel feeding amount is obtained by using a particle swarm algorithm, and the stable control of the decomposition furnace temperature is achieved by combining rolling optimization and feedback correction, which specifically includes:

[0048] The particle swarm algorithm is used to solve the optimal setting value of the coal feeding amount and the amount of alternative fuel in the control system. When it is passed to the bottom controller to control the decomposition furnace, the prediction model is used to obtain the predicted value of the decomposition furnace temperature at time k. The corrected prediction value y(k+1) is obtained through feedback correction:

[0049] y(k+1)=y(k)+r×e(k)#(10)

[0050]

[0051] Among them, r is the feedback coefficient, y(k) is the temperature prediction value at time k, is the actual temperature value at time k;

[0052] According to the optimal temperature setting value of the decomposition furnace at time k in the optimization result and the actual temperature value of the decomposition furnace at time k, the optimal temperature setting value is obtained through softening calculation. The formula is as follows:

[0053]

[0054] Where α is the softening coefficient, y s (k), y s (k+1) is the k moment, the predicted value of the decomposition furnace temperature at k+1 moment, and finally the optimal setting value of the decomposition furnace temperature at k+1 moment is obtained. The optimal temperature setting value y s (k+1) and the model prediction value y(k+1) are put into the optimization algorithm to solve the optimal control quantity.

[0055] Due to the adoption of the above technical solution, the technical advances achieved by the present invention include:

[0056] The present invention aims at predicting and controlling the temperature of a decomposition furnace, establishes a predictive optimization model for state parameters of a decomposition furnace and a predictive control model for the temperature of a decomposition furnace, reduces NOx emissions and carbon emissions while ensuring the quality of output cement, increases the proportion of alternative fuels in fuels and thus reduces coal consumption, improves the accuracy of predicting the temperature of the decomposition furnace, accurately controls the temperature of the decomposition furnace according to the state parameter optimization results and the temperature prediction results, enhances the stability of the control system, and solves the problems of excessive energy consumption, unsatisfactory output, and high pollution discharge in the actual cement production process.

[0057] The present invention establishes a decomposition furnace temperature prediction model, fully utilizes the time series characteristics of data, overcomes the inherent nonlinearity, strong coupling and time delay characteristics of the decomposition furnace system, and realizes rapid and accurate prediction of the decomposition furnace temperature. At the same time, the KAN network is used to simplify the complex network structure required for data feature extraction in the informer network and enhance network performance, so that more accurate prediction results can be obtained using fewer levels and parameters.

[0058] The present invention establishes a decomposition furnace state parameter optimization model, uses the KAN network to enhance the LSTM network performance, and predicts and optimizes the relevant influencing parameters of f-CaO and NOx contents respectively, thereby ensuring the quality of cement products and reducing NOx emissions and carbon emissions in production.

[0059] The present invention establishes a temperature prediction and control system for a decomposition furnace, which solves the problems of the fluctuation of the coal feeding amount and the blending ratio of alternative fuel within a specific proportion range in the actual cement production process, the instability caused by sudden changes in working conditions under human intervention and the substandard cement quality. According to the constraint conditions obtained by optimizing the state parameters, the replacement ratio of alternative fuels is controlled, and the coal consumption is effectively reduced while ensuring stable temperature control. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a framework diagram of a cement decomposition furnace temperature prediction and control method based on KAN enhanced neural network of the present invention;

[0061] Figure 2It is a structural diagram of an f-CaO prediction model and a Nox prediction model of a cement decomposition furnace temperature prediction and control method based on a KAN enhanced neural network of the present invention;

[0062] Figure 3 It is a structural diagram of a decomposition furnace temperature prediction model of a cement decomposition furnace temperature prediction and control method based on a KAN enhanced neural network of the present invention;

[0063] Figure 4 The invention discloses a prediction control flow chart of a cement decomposition furnace temperature prediction control method based on a KAN enhanced neural network. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail below in combination with specific implementation methods and with reference to the accompanying drawings. In the following description, the description of the public structure and technology is omitted to avoid unnecessary confusion of the concept of the present invention.

[0065] The present invention provides a cement decomposition furnace temperature prediction and control method based on KAN enhanced neural network, comprising the following steps:

[0066] Step 1: Select parameters related to the f-CaO and NOx content produced by the normal operation of the decomposition furnace from the database as the state parameters to be optimized, select 48096 data for each parameter for preprocessing, use the Pearson correlation coefficient method to perform correlation analysis on each variable, determine the time delay between variables, organize the data set, and then remove outliers in the data according to the 3σ principle, and normalize the processed data. The specific method is as follows:

[0067] (1) Pearson correlation coefficient method:

[0068]

[0069] Among them, r is the correlation coefficient, x i ,y i For the input data of different parameters, the time lag interval between variables is determined according to the absolute value of the calculated correlation coefficient, and the data set is sorted.

[0070] (2) 3σ principle:

[0071] Assume that the data of each parameter in the data set is approximately normally distributed, σ represents the standard value, μ represents the mean value, and x=μ is the symmetry axis of the image. At this time, the probability that the value is distributed in (μ-σ, μ+σ) is 0.6826; the probability that the value is distributed in (μ-2σ, μ+2σ) is 0.9544; the probability that the value is distributed in (μ-3σ, μ+3σ) is 0.9974. It can be considered that the values ​​of y are almost concentrated in the interval (μ-3σ, μ+3σ), and the data outside this range can be marked as outliers.

[0072] (3) Normalization:

[0073]

[0074] Among them, y j is the normalized result, x i Normalization converts different parameters into the same order of magnitude, improves model accuracy, speeds up convergence, and avoids overfitting of model data.

[0075] After normalization, the first 60% of the data set is divided into a training set, the middle 20% is divided into a validation set, and the last 20% is divided into a training set.

[0076] Since the collected data is time series data, the sliding window algorithm is a commonly used algorithm for processing time series. While continuously sliding the window, the time series data is continuously processed and analyzed by calculating the data aggregation within a fixed time window, so as to better obtain the timing information between the data. Therefore, the data is processed by sliding window according to the time series of each parameter of the data set, and 5 data (i.e. 5 time units) are collected in each time window. While reducing the burden of model training with a smaller amount of data in each batch, the model data timing feature processing capability is improved.

[0077] Figure 2 It is a prediction model for f-CaO and Nox content. The present invention uses the same model architecture to train and predict different prediction targets. First, relevant state parameters are obtained from the data input layer, and the input data is integrated through a convolution layer to extract time series features. Then, the data features are further extracted through the LSTM layer to generate a preliminary prediction result. Finally, the data is integrated again through the KAN network layer to obtain the final prediction result. The parameters that need to be updated in the convolution layer during the training process are the weight matrix W c .Bias vector b. The parameters that need to be updated in the LSTM layer are the input gate weight W i , loop weight U i Bias b i , the weight of the forget gate W f , loop weight U i , bias b f, output gate weight W o , loop weight U o , bias b o , cell state weight W c , loop weight U c , bias b c The parameters that need to be updated in the KAN network layer mainly include the amplification coefficients w1 and w2, and the control points (coefficients) c of all spline basis functions that make up the third-order B-spline curve (B-spline). i , bias b k .

[0078] Among them, LSTM layer, h t-1 、h t are the hidden states at time t-1 and time t, C t-1 , C t are the cell states at time t-1 and time t, and the input data x at time t. t and the hidden state h at time t-1 t-1 Enter the forget gate, input gate and output gate respectively. The formula of the forget gate is as follows:

[0079] f t =σ(W f x t +U f h t-1 +b f )#(14)

[0080] x t and h t-1 After weighted summation, the output is Sigmoid nonlinear mapping, and the cell state C at time t-1 is obtained. t-1 Multiplication is equivalent to updating the cell state so that it forgets the parts that are irrelevant to the prediction results.

[0081] The input gate formula is as follows:

[0082] i t =σ(W i x t +U i h t-1 +b i )

[0083]

[0084] The output result is i t and Multiply to get the new cell state that needs to be remembered, and then add it to the old cell state C t-1 Add, the output is C t .

[0085] The output gate formula is as follows:

[0086] o t =σ(W o x t +U o h t-1 +b o )

[0087] h t =o t ×tanh(C t )#(16)

[0088] The result of the output gate output will be used as the hidden state of the next LSTM unit at the previous moment and participate in the operation of the next unit.

[0089] When the last LSTM unit calculates the output h t After that, the hidden state is input into the KAN network for fitting. The KAN network formula is as follows:

[0090] φ(x)=w1b(x)+w2spline(x)+b k #(17)

[0091] in,

[0092]

[0093] b(x)=silu(x)#(18)

[0094] Here B i (x) is the i-th third-order B-spline curve, and the combined curve spline is a linear combination of n third-order B-splines. The hidden state output by LSTM enters the KAN network layer, and is calculated by formula (16) and residually connected through the silu function to obtain the final prediction result. In the whole process, we use the Adam optimizer for model training. The specific method is as follows:

[0095] (1) Initialize network parameters: Randomly initialize the weights and biases of the neural network.

[0096] (2) Forward propagation: The input data is calculated layer by layer through weighted and activated functions to finally generate the output result of the model.

[0097] (3) Calculate the loss: Calculate the model loss function based on the predicted results and the true value to obtain the loss value

[0098] (4) Back propagation and parameter update: Starting from the output layer, the network parameters are differentiated according to the loss function and the gradient is calculated. The gradient is propagated back to each layer, and the gradient of each layer is calculated according to the chain rule. The first and second moments of each layer training parameter are updated using the current gradient and the previous state to correct the deviation, and the corrected estimated value is obtained. The parameter model is updated using the learning rate.

[0099] (5) Repeat steps 2-4 until the prediction result meets the termination condition.

[0100] After the model is trained successfully, the validation set data is input into the prediction model to obtain the output results, and the MAE and MAD test methods are used to test the model fitting results. The formula is as follows:

[0101]

[0102] Among them, y is the predicted value output by the model, y s is the true value, y s is the average of the true values.

[0103] After the model training is completed, the f-CaO and Nox prediction models are combined with the multi-objective particle swarm algorithm to solve the optimal state parameter values ​​according to the following objective function:

[0104] min J1=||y1(k)-Y1|| 2

[0105] min J2=||y2(k)-Y2||+||Δu4(k)|| 2 #(twenty one)

[0106]

[0107] In the above objective function, J1 is the objective function for optimizing cement quality, y1(k) is the predicted value of f-CaO at time k, and Y1 is the optimal value of f-CaO content in the process requirements. This objective function aims to control the f-CaO content generated in the actual production process as close to the optimal value as possible, so as to produce cement with the best quality; J2 is the objective function for optimizing pollution (Nox) emission and ammonia consumption, y2(k) is the predicted value of Nox at time k, Y2 is the standard value of Nox content in the process requirements, Δu4(k) is the increment of ammonia at time k compared with the previous moment, and x i is the input to the f-CaO prediction model, x i ,u i is the input to the NOx prediction model. The objective function aims to control the NOx emissions generated in the actual production process to the specified standard value while reducing the ammonia loss as much as possible, so as to meet the environmental protection requirements of this system.

[0108] Specific method of particle swarm optimization:

[0109] Initialization parameters and particle states: randomly generate a group of particles, assign an initial position and velocity to each particle, and set a fitness function to evaluate the quality of the particles.

[0110] Update speed and position: Update the speed and position of the particle according to the particle's historical best position and the group's best position.

[0111] Evaluate fitness: Calculate the fitness of each particle's new position and update the personal best position and the global best position based on the fitness.

[0112] Iteration: Repeat steps 2 and 3 until a preset stopping condition is reached.

[0113] According to the obtained operating parameters, the optimal solution vector (denoted as Z1, Z2) is optimized, and the state parameter optimization solution vector interval is established as the constraint condition for optimizing the control quantity in the predictive control process:

[0114]

[0115] Here, the 10% range around the optimal value is taken as the optimal solution range, which can be adjusted according to the actual situation in actual use.

[0116] Step 2: Similarly, parameters related to the decomposition furnace temperature are selected from the database as state parameters to be optimized. 48096 data are selected for each parameter for preprocessing. The preprocessing method is the same as that in step 1 and will not be repeated here.

[0117] In order to establish a decomposition furnace temperature prediction model, the present invention improves the Informer prediction model traditionally applied to long-sequence NLP tasks through the KAN network, enhances its model performance, reduces model complexity, and improves model prediction accuracy. The traditional Informer model adopts an "Encoder-Decoder" structure. After the model parameters are initialized, the pre-processed data is processed twice in batches. Each batch of data is divided into two groups of data with overlapping parts according to the time series, and is input into the Encoder layer and Decoder layer of the prediction model respectively. After a group of data is input into the Encoder layer, it first passes through the encoding layer, and the implicit position information between each data in the input data is captured by Positional Embedding. Token Embedding converts the input data into an embedded vector through one-dimensional convolution, increases the data dimension, and extracts more time series information. The parameters that need to be involved in the training of this model are mainly the weights and biases of the convolutional layer and the weights and biases of each linear layer.

[0118] The encoded data enters the multi-head sparse self-attention layer:

[0119] The input data is transformed into a query vector (query), a key vector (key), and a value vector (value) through linear transformation. The formula is as follows:

[0120] q=W q ×

[0121] k=W k ×x#(23)

[0122] v=W v ×

[0123] Previous studies have shown that the traditional attention mechanism is sparse. The advantage dot product causes the attention distribution corresponding to q to be far away from the uniform distribution. If the attention distribution is close to the uniform distribution, the calculation of the attention appears to be redundant.

[0124] Based on the above theory, Informer uses KL divergence to calculate the relative entropy of the attention score of each q. The larger the relative entropy, the further the attention distribution is from the uniform distribution, and the more worthy the attention score is to be calculated. Correspondingly, the attention score with a smaller relative entropy is less important and can be replaced by a uniform distribution. The attention calculation formula for the i-th q is as follows:

[0125]

[0126] Among them, K and V are respectively i ,v i The matrix composed of.

[0127] The sparse metric formula (KL divergence) of the i-th q is as follows:

[0128]

[0129] Among them, d is a scaling factor to avoid gradient vanishing. The KL divergence of each q is calculated by formula (24). The q with higher divergence is considered to have higher activity and contain important information. Its multi-head attention score needs to be calculated according to formula (23). The attention scores of other relatively "lazy" q are replaced by uniform distribution to obtain sparse self-attention scores.

[0130] The sparse self-attention score passes through the linear layer and the distillation layer, which greatly reduces its time dimension and reduces the computational burden of the model. The formula of the distillation layer is as follows:

[0131]

[0132] Among them, AB represents the attention block, They represent the data before entering the attention block and the data after pooling respectively.

[0133] The encoded data goes through the above steps twice and is finally integrated through the fully connected layer as a "reference" for the Decoder layer.

[0134] The second set of data enters from the Decoder layer, and after being processed by the multi-head sparse self-attention layer combined with the mask mechanism, the output result and the "reference" from the Encoder layer are jointly processed by the multi-head full attention layer to calculate the attention score. Finally, the data is reduced in dimension by the fully connected layer to obtain the prediction result.

[0135] The present invention adopts the KAN network to simplify the complex network structure required for data feature extraction in the informer network and enhance the network performance, and can obtain more accurate prediction results with fewer levels and parameters.

[0136] The core calculation formula of the KAN network layer is shown in formula (16)(17). The KAN network is different from the traditional MLP network in which the "weight-bias-nonlinear activation function" neuron model is used between nodes. By using multiple trained B-spline curves with better fitting effects for superposition and combination, the neurons in the network are added with nonlinear characteristics, which can be regarded as a complex curve function. As a result, the KAN network is more suitable for extracting complex feature relationships between input data than the traditional MLP network. It can achieve or even exceed the effect of the complex MLP network with a simpler hierarchical structure. At the same time, the input data is connected with the output result of the silu function and the combined curve to achieve residual connection to avoid gradient disappearance. A trainable amplification factor is added to train the model more stably using automatic scaling.

[0137] like Figure 3 As shown in the figure, the improved model uses the KAN network to extract features from the input data, and obtains the q, k, v vectors required for calculating the attention score through formula (16) and formula (17). The attention score calculation result is also processed by a layer of KAN network for data integration and dimensionality reduction. Based on the above theory applied to practice, the Encoder layer only needs to calculate the sparse self-attention score once, which can be used as the input of the Decoder layer after data integration and distillation. In addition, the dimension required for feature extraction is reduced from the original 512 dimensions of the linear layer to 16 dimensions, which greatly reduces the model complexity, speeds up the model training speed, and improves the model prediction accuracy. Due to its strong ability to extract data features, subsequent experiments have also proved that good results can be achieved with less data.

[0138] The improved model is also trained using the Adam optimizer, and the training and testing methods are the same as in step 1.

[0139] Step 3: Establish a decomposition furnace temperature prediction control model based on the state parameter optimization results and the decomposition furnace temperature prediction model. Figure 4 This is the structure diagram of the control model.

[0140] According to the optimization results, the control objective function is established as follows:

[0141]

[0142] Where P and M are the prediction time domain step and control time domain step, respectively. y(k+s) is the predicted decomposition furnace temperature value at time k+s at time k. s (k+s) is the optimal setting value of the decomposition furnace temperature at time k+s, Δu1(k+h-1|k) is the incremental prediction of the coal feeding amount at time k to time k+h-1, Q and R are two control coefficient matrices. ε is the value of the optimal range defined by the optimal setting value of the alternative fuel amount when the alternative fuel exceeds k+h-1. Other x i ,u i The constraint range of is also obtained from the state parameter optimization result, and the constraint range of Δu1 can be adjusted according to the actual situation.

[0143] The particle swarm algorithm is also used to solve the optimal setting value of the coal feeding amount and the amount of alternative fuel in the control system. When it is passed to the bottom controller to control the decomposition furnace, the prediction model is used to obtain the predicted value of the decomposition furnace temperature at time k, and the corrected prediction value y(k+1) is obtained through feedback correction:

[0144] y(k+1)=y(k+1)+r×e(k+1)#(28)

[0145]

[0146] Among them, r is the feedback coefficient, y(k) is the temperature prediction value at time k, is the actual temperature value at time k.

[0147] According to the optimal temperature setting value of the decomposition furnace at time k in the optimization result and the actual temperature value of the decomposition furnace at time k, the optimal temperature setting value is obtained through softening calculation. The formula is as follows:

[0148]

[0149] Among them, α is the softening coefficient, y s (k+1) is the predicted value of the decomposition furnace temperature at time k+1. Finally, the optimal setting value of the decomposition furnace temperature at time k+1 is obtained, and the optimal temperature setting value y s (k+1) and the model prediction value y(k+1) are put into the optimization algorithm to solve the optimal control quantity.

[0150] In the above embodiments, the present invention provides a cement decomposition furnace temperature prediction and control method based on KAN enhanced neural network. The decomposition furnace temperature prediction and control model established by the present invention adopts advanced control strategies and realizes stable control of the decomposition furnace temperature through three key processes. Rolling optimization ensures that in each control cycle, the control input is dynamically adjusted based on the latest system status and prediction information to adapt to the changing working conditions. Secondly, predictive control uses the dynamic model of the system to accurately predict the temperature changes in the future period, thereby providing a scientific basis for control decisions. The feedback correction mechanism monitors the deviation between the actual temperature and the expected target in real time, and adjusts the control strategy in time to eliminate interference and uncertainty. This comprehensive method effectively improves the control accuracy and stability of the decomposition furnace temperature and provides reliable technical support for industrial applications.

[0151] The above-described embodiments are merely descriptions of preferred implementations of the present invention, and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made by ordinary persons in the art to the technical solution of the present invention should fall within the protection scope of the present invention, and the technical contents for which protection is sought in the present invention have been fully recorded in the claims.

Claims

1. A cement decomposition furnace temperature prediction and control method based on KAN enhanced neural network, characterized in that: The following steps are included: S1. Establish prediction optimization models for f-CaO and Nox respectively, and establish objective functions in combination with the actual needs of cement production. Use multi-objective particle swarm algorithm to find the optimal solution of the state parameters of the decomposition furnace and obtain the optimal interval; S2. Establish a decomposition furnace temperature prediction model to predict the decomposition furnace temperature; S3. According to the decomposition furnace temperature prediction model and the optimal interval of state parameters, a control objective function is established. The optimal solution of the coal feeding amount and the alternative fuel feeding amount is obtained through the particle swarm algorithm. Combined with rolling optimization and feedback correction, stable control of the decomposition furnace temperature is achieved.

2. The method for predicting and controlling cement decomposition furnace temperature based on KAN enhanced neural network according to claim 1 is characterized in that: The step S1 comprises the following steps: S1.

1. According to the cement burning system, the predicted variables f-CaO emission and NOx content are selected, which are recorded as y1 and y2 respectively. The variables with greater influence on f-CaO content are selected, namely, the decomposition furnace outlet temperature x1, the kiln tail temperature x2, the secondary air temperature x3, the total feed amount x4, the high temperature fan speed x5, and the EP fan speed feedback x6. The variables with greater influence on NOx emission are selected, namely, the decomposition furnace outlet temperature x1, the kiln tail temperature x2, the coal feed amount u1, the alternative fuel amount u2, the kiln head coal feedback u3, and the total ammonia flow u4. The prediction model expression is as follows: y1=f(x1,x2,x3,x4,x5,x6)#(1) y2=f(x1,x2,u1,u2,u3,u4)#(2) Export the above data from the database and pre-process the data; S1.

2. After data preprocessing, a CNN-LSTM-KAN network was established as the prediction model for f-CaO and Nox, and their values ​​were predicted respectively; S1.

3. Based on the f-CaO and NOx prediction models in S1.2, an objective function is constructed according to the requirements of cement product quality and energy conservation and emission reduction: min J1=||y1(k)-Y1|| 2 my J2=||y2(k)-Y2|| 2 +||Δu4(k)|| 2 #(3) In the above objective functions, J1 is the objective function for optimizing cement quality, y1(k) is the predicted value of f-CaO at time k, and Y1 is the optimal value of f-CaO content in the process requirements; J2 is the objective function for optimizing NOx emission and ammonia consumption, y2(k) is the predicted value of Nox at time k, Y2 is the standard value of Nox content in the process requirements, Δu4(k) is the increment of ammonia at time k compared with the previous moment, and x i is the input to the f-CaO prediction model, x i ,u i is the input quantity to the Nox prediction model; S1.

4. Use multi-objective particle swarm optimization algorithm to optimize the objective function J1 and J2 parameters.

3. The method for predicting and controlling cement decomposition furnace temperature based on KAN enhanced neural network according to claim 2 is characterized in that: The data preprocessing in step S1.1 includes the following steps: S1.1.

1. Use the Pearson correlation coefficient method to analyze the correlation between variables. The analysis results determine the time delay between variables and adjust the data based on the results. The Pearson correlation coefficient formula is: Among them, r is the calculated Pearson correlation coefficient, x i ,y i is the input data for different parameters; S1.1.

2. According to the 3σ principle, outliers in the data are processed and normalized. The normalization formula is: Among them, y j is the normalized result.

4. The method for predicting and controlling cement decomposition furnace temperature based on KAN enhanced neural network according to claim 2 is characterized in that: The step S1.4 specifically includes: According to the obtained state parameters, the optimal solution vector is optimized and the state parameter optimization solution vector interval is established as the constraint condition for optimizing the control quantity in the predictive control process: The optimal solution range is taken as the 10% range around the optimal value.

5. The method for predicting and controlling cement decomposition furnace temperature based on KAN enhanced neural network according to claim 1, characterized in that: The step S2 comprises the following steps: S2.

1. Predict the temperature of the decomposition furnace during normal operation; S2.

2. Using the "Encoder-Decoder" structure, a KAN-Enhanced-Informer decomposition furnace temperature prediction model is established, and the decomposition furnace temperature prediction value is output.

6. The method for predicting and controlling cement decomposition furnace temperature based on KAN enhanced neural network according to claim 5, characterized in that: The step S2.1 specifically includes: The variables that have a greater impact on the change of the decomposition furnace temperature are selected as the input variables of the decomposition furnace temperature prediction model, including the historical value of the decomposition furnace outlet temperature x1, the first-stage drum outlet temperature x2, the first-stage drum oxygen concentration x3, the feed amount x4, the high-temperature fan speed x5, the coal feed amount u1, and the alternative fuel amount u2. The above data are exported from the database and preprocessed.

7. The method for predicting and controlling cement decomposition furnace temperature based on KAN enhanced neural network according to claim 6, characterized in that: The step S2.2 specifically includes: The "Encoder-Decoder" structure is used to establish the KAN-Enhanced-Informer decomposition furnace temperature prediction model. After the model parameters are initialized, the preprocessed data in S2.1 are processed again in batches. Each batch of data is divided into two groups of data with overlapping parts according to the time series. After encoding, they are input into the Encoder layer and Decoder layer of the prediction model respectively.

8. The method for predicting and controlling cement decomposition furnace temperature based on KAN enhanced neural network according to claim 7, characterized in that: The encoder layer consists of a sparse self-attention mechanism, a distillation layer and a KAN network. The decoder layer consists of a sparse self-attention mechanism combined with a mask mechanism, a full attention mechanism and a KAN network. After the data entering the decoder layer is calculated by the sparse self-attention, it enters the full attention mechanism layer with the output result of the encoder layer. The output result is finally integrated and reduced in dimension by the KAN network, and the predicted value of the decomposition furnace temperature is output. The expression is as follows: y3=f(x1,x2,x3,x4,x5,u1,u2)#(7).

9. The method for predicting and controlling cement decomposition furnace temperature based on KAN enhanced neural network according to claim 1, characterized in that: The control objective function in step S3 is: Among them, P and M are the prediction time domain step and control time domain step, respectively, y(k+s|k) is the predicted decomposition furnace temperature value at time k+s at time k, y s (k+s) is the optimal setting value of the decomposition furnace temperature at time k+s obtained by optimizing the state parameters in step S1, Δu1(k+h-1|k) is the increment of the decomposition furnace coal feeding amount predicted at time k+h-1, Q and R are two control coefficient matrices, ε is the value of the alternative fuel exceeding the optimal range defined by the optimal setting value of the alternative fuel amount at time k+h-1 obtained in step S1, x i 、u i The constraint range of is also obtained by the particle swarm algorithm in step S1. The constraint range of Δu1 can be changed according to the actual situation. The formula is as follows: Where u2 is the amount of alternative fuel, z 2min 、z 2max are the constraint boundary vectors Z 2min ,Z 2max The alternative fuel amount constrains the boundary component.

10. The method for predicting and controlling cement decomposition furnace temperature based on KAN enhanced neural network according to claim 1, characterized in that: In step S3, the optimal solution of the coal feeding amount and the alternative fuel feeding amount is obtained by using the particle swarm algorithm, and the stable control of the decomposition furnace temperature is achieved by combining rolling optimization and feedback correction, which specifically includes: The particle swarm algorithm is used to solve the optimal setting value of the coal feeding amount and the amount of alternative fuel in the control system. When it is passed to the bottom controller to control the decomposition furnace, the prediction model is used to obtain the predicted value of the decomposition furnace temperature at time k. The corrected prediction value y(k+1) is obtained through feedback correction: y(k+1)=y(k+1)+r×e(k+1)#(10) Among them, r is the feedback coefficient, y(k) is the temperature prediction value at time k, is the actual temperature value at time k; According to the optimal temperature setting value of the decomposition furnace at time k in the optimization result and the actual temperature value of the decomposition furnace at time k, the optimal temperature setting value is obtained through softening calculation. The formula is as follows: Where α is the softening coefficient, y s (k+1) is the predicted value of the decomposition furnace temperature at time k+1, and finally the optimal setting value of the decomposition furnace temperature at time k+1 is obtained. The optimal temperature setting value y s (k+1) and the model prediction value y(k+1) are put into the optimization algorithm to solve the optimal control quantity.

Images