Method for multi-step prediction of carbon oxide content in a methanator based on a multivariate deep reconstruction neural network

Abstract

The application discloses a kind of carbon oxide content multi-step prediction methods of methanation furnace based on multivariate deep reconstruction neural network, the method makes full use of the interaction between historical measurable furnace temperature and other process variable data and historical carbon oxide content, introduces historical process variable guidance, extracts deep correlation characteristics of sequence by stacking structure, realizes more accurate prediction to carbon oxide sequence by forward prediction and backward reconstruction layer in the form of minimizing cumulative prediction error;The application can make full use of the continuity and time sequence of methanation furnace process data and quality data, improve the utilization rate of process data on sequence modeling, enrich sequence characteristics, and improve the accuracy of model prediction, realize the multi-step prediction of carbon oxide content of methanation furnace.

Description

Technical Field

[0001] This invention belongs to the field of industrial process control and soft measurement, specifically relating to a multi-step prediction method for carbon oxide content in methanation furnaces based on a multivariate deep reconstruction neural network. Background Technology

[0002] With the increasing material demands of the people and the progress of global modernization, chemical processes are becoming increasingly complex, and the industrial sector has higher and higher requirements for product quality, production efficiency, and production process safety monitoring. Addressing the high cost and inflexibility of hard sensors in industrial production, soft measurement technology has been proposed and is gradually being applied in the industrial field. The purpose of soft measurement is to utilize all available information to obtain the best estimate of the required variables.

[0003] The ammonia synthesis process is a crucial industrial production process. The methanation furnace is a key component of this process, absorbing excess carbon oxides from upstream processes through methanation. This reduces the impact of these byproducts on subsequent ammonia synthesis steps, ultimately aiming to control the outlet carbon oxide content at a low threshold. Since methanation is a reversible process, industrially, the outlet carbon oxide content is typically maintained by controlling the reactor temperature and inlet gas flow rate. Because carbon oxide concentrations are difficult to detect with hardware, traditional methods primarily rely on laboratory analysis. This involves collecting carbon oxide gas at regular intervals and analyzing the concentration using chemical methods. However, this method suffers from high analytical costs and detection delays. In addition, there are mechanistic modeling-based methods for carbon oxide content detection; however, due to the complexity and nonlinearity of the methanation process, accurate analysis of the process mechanism remains extremely difficult. Therefore, the most widely used methods are data-driven carbon oxide content detection methods, such as Principal Component Regression (PCR), Support Vector Regression (SVR), tree-based GBDT, and XGBoost. However, these methods do not take into account the dynamic nature of the methanation furnace process, i.e., the temporal relationships between the collected data. With the development of dynamic modeling, many studies have improved the original static models into dynamic forms. Early examples include autoregressive models based on probability statistics. In recent years, with the development of deep learning and the in-depth exploration of spatiotemporal sequence modeling, the advantages of deep networks in feature extraction of complex industrial processes have become increasingly significant. However, in the methanation furnace industrial process, due to the low training efficiency of some deep networks such as RNN and LSTM, and the tendency for training degradation, they have not yet been well applied in the detection of carbon oxides in methanation furnaces. Summary of the Invention

[0004] To address the dynamic and temporal nature of methanation furnaces and the need for multi-step quality prediction, this invention proposes a multi-step prediction method for carbon oxide content in methanation furnaces based on a multivariate deep reconstruction neural network. This method fully utilizes process sequences and quality sequences, employs a stacked structure to continuously reduce training residuals, and achieves multi-step prediction of outlet carbon oxide content during the methanation furnace process. It also reduces training difficulty and improves prediction accuracy.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A multi-step prediction method for carbon oxide content in a methanation furnace based on a multivariate deep reconstruction neural network, comprising the following steps:

[0007] Step 1: Collect historical data from the methanation furnace process and construct a dataset F = [X′, Y]. T , X′∈R n×m , Y∈R n Where X′ is a measurable process variable, Y represents the carbon oxide content at the outlet, n represents the number of samples collected, and R represents the set of real numbers.

[0008] Step 2: Normalize each process variable in X′ to the interval [0,1] to obtain X;

[0009] Step 3: Select a reasonable input time window T1 for the methanation furnace process, and at the same time select the time window T2 for the carbon oxide content index that needs to be predicted; perform segmentation and serialization operations on the normalized process variable X and the outlet carbon oxide content Y to obtain a one-to-one corresponding input sequence and prediction sequence.

[0010] Step 4: Construct and train a multivariate deep reconstruction neural network; the multivariate deep reconstruction neural network includes s structural blocks, each structural block including a splicing layer, a shared network, a reverse reconstruction layer, and a forward prediction layer; the shared network includes three nonlinear layers; the splicing layer converts the process variable sequence... and carbon oxide content sequence Connect along the feature dimensions to obtain the input sequence. The first two nonlinear layers of the shared network are paired with the input Z. t Perform feature extraction along the time dimension to obtain Then, feature extraction is performed along the dimensions of the variables to obtain the final sequence feature H extracted by the shared network. k×1 Subsequently, H k×1 The inputs are respectively fed into a reverse reconstruction layer with T1 nodes and a forward prediction layer with T2 nodes; the reverse reconstruction layer reconstructs the input history sequence of the quality variables to obtain the reverse reconstruction output. The forward prediction layer predicts the carbon oxide content at several future moments, resulting in the forward prediction output. Subsequently, the historical quality sequence output by the reverse reconstruction layer is... The input mass sequence Y of this structural block t The difference is calculated to obtain the quality sequence residual. The residual With process sequence X t The sum of the forward prediction outputs of the s structural blocks serves as the input to the next structural block; the sum of these forward prediction outputs is used as the final quality index prediction sequence.

[0011] Step 5: Normalize the process variables to be tested, and then input them into the trained multivariate deep reconstruction neural network to obtain the predicted sequence of outlet carbon oxide content.

[0012] Furthermore, the normalization formula for step two is as follows:

[0013]

[0014] Among them, X min This represents the minimum value of each process variable over time. Correspondingly, X max This represents the minimum value of each process variable over time.

[0015] Furthermore, the feature extraction process of the shared network is represented by the following formula:

[0016]

[0017]

[0018]

[0019] in, This represents the weighting coefficients of the first nonlinear layer. This represents the bias coefficient of the first nonlinear layer. This represents the weighting coefficients of the second nonlinear layer. This represents the bias coefficient of the first nonlinear layer. This represents the weighting coefficients of the third nonlinear layer. This represents the bias coefficient of the third nonlinear layer. ReLU is the activation function; for inputs less than 0, its output is 0, and for inputs greater than 0, the output remains unchanged.

[0020] Furthermore, the inverse reconstruction output and forward prediction output The calculation formula is as follows:

[0021]

[0022]

[0023] in, Indicates the weight coefficients of the reverse reconstruction layer. This represents the bias coefficient of the reverse reconstruction layer. This represents the weight coefficients of the forward prediction layer. This represents the bias coefficient of the forward prediction layer.

[0024] Furthermore, the reconstruction residual of the last structural block is used as the final reconstruction residual.

[0025]

[0026] The training loss function of the multivariate deep reconstruction neural network is denoted as Loss.

[0027]

[0028] The beneficial effects of this invention are as follows:

[0029] Compared to other methods for predicting carbon oxide content in methanation furnaces, this invention fully utilizes the interaction between historical measurable furnace temperature and other process variables and historical carbon oxide content data. By introducing historical process variables as guidance and using a deep reconstruction neural network to deeply reconstruct the quality sequence, the error in carbon oxide content prediction is gradually reduced, thus improving the accuracy of industrial quality prediction. At the same time, this method efficiently solves the multi-step prediction of important quality indicators in industrial processes. It can not only achieve advanced prediction of the carbon oxide content at the outlet, but also solve the problem of instrument detection delay. Attached Figure Description

[0030] Figure 1 This is a flowchart of the multi-step prediction method for carbon oxide content in a methanation furnace based on a multivariate deep reconstruction neural network, according to the present invention.

[0031] Figure 2 This is a schematic diagram of historical time windows and predictive time windows.

[0032] Figure 3 A framework diagram for multivariate deep reconstruction neural networks.

[0033] Figure 4 This is a schematic diagram of a single structural block in a multivariate deep reconstruction neural network.

[0034] Figure 5 This is a process flow diagram of a methanation furnace.

[0035] Figure 6 This is a prediction fit diagram of the multivariate deep reconstruction neural network of the present invention and existing soft measurement models based on MLP and univariate reconstruction (N-Beats).

[0036] Figure 7 This is a graph showing the prediction error of the multivariate deep reconstruction neural network of this invention and existing soft measurement models based on MLP and univariate reconstruction (N-Beats). Detailed Implementation

[0037] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0038] like Figure 1 As shown, the multi-step prediction method for carbon oxide content in a methanation furnace based on a multivariate deep reconstruction neural network of the present invention includes the following steps:

[0039] Step 1: Collect historical data from the methanation furnace process and construct a dataset F = [X′, Y]. T , X′∈R n×m , Y∈R n Where X′ is a measurable process variable, Y represents the carbon oxide content at the outlet, n represents the number of samples collected, and R represents the set of real numbers.

[0040] Step 2: Normalize each process variable in X′ to the interval [0,1]. The normalization formula is as follows:

[0041]

[0042] Among them, X min This represents the minimum value of each process variable over time. Correspondingly, Record X min and X max .

[0043] Step 3: Select a reasonable input time window T1 for the methanation furnace process, and simultaneously select the time window T2 for the carbon oxide content index that needs to be predicted, such as... Figure 2 As shown, the normalized process variable X and the outlet carbon oxide content Y are segmented and serialized to obtain a one-to-one corresponding input sequence and prediction sequence. T1, based on expert experience, is generally between 10 and 30. Taking a serialized sample as an example, it takes the form of... There are a total of n-T1-T2+1 samples.

[0044] Step 4: Construct and train a multivariate deep reconstruction neural network (hereinafter referred to as MDRNN).

[0045] like Figure 3 As shown, the multivariate deep reconstruction neural network consists of s structural blocks. Figure 4 As shown, each structural block includes a stitching layer (cat), a shared bottom network, a backcast layer, and a forward prediction layer. The stitching layer stitches the process variable sequence... and carbon oxide content sequence Connect along the feature dimensions to obtain the input sequence. The shared network consists of three nonlinear layers, with the first two layers having k nodes each. For input Z... t Perform feature extraction along the time dimension, that is, map the time dimension to k dimensions to obtain... The last layer has 1 node, mapping the variable dimension to 1 dimension, thus obtaining the final sequence feature H extracted by this shared network. k×1 :

[0046]

[0047]

[0048]

[0049] in, This represents the weighting coefficients of the first nonlinear layer. This represents the bias coefficient of the first nonlinear layer. This represents the weighting coefficients of the second nonlinear layer. This represents the bias coefficient of the first nonlinear layer. This represents the weighting coefficients of the third nonlinear layer. This represents the bias coefficient of the third nonlinear layer. ReLU is the activation function; for inputs less than 0, its output is 0, and for inputs greater than 0, the output remains unchanged.

[0050] Subsequently, the features H extracted from the shared network are... k×1 The inputs are respectively fed into a backcast layer with T1 nodes and a forward prediction layer with T2 nodes; the backcast layer reconstructs the input history sequence of the quality variables to obtain the backcast output. The forward prediction layer predicts the carbon oxide content at several future time points, yielding the forward prediction output.

[0051]

[0052]

[0053] in, Indicates the weight coefficients of the reverse reconstruction layer. This represents the bias coefficient of the reverse reconstruction layer. This represents the weight coefficients of the forward prediction layer. This represents the bias coefficient of the forward prediction layer.

[0054] Subsequently, the historical quality sequence output by the reverse reconstruction layer is... The input quality sequence of this structure block The difference is calculated to obtain the quality sequence residual.

[0055]

[0056] Among them, the residual With process sequence X t As input for the next structure block.

[0057] The sum of the forward prediction outputs of s structural blocks is used as the final quality index prediction sequence.

[0058]

[0059] Among them, the reconstruction residual of the last structural block is used as the final reconstruction residual.

[0060]

[0061] The training loss function of the model is denoted as Loss. Minimizing this Loss completes the training, and the model parameters are saved.

[0062]

[0063] When new historical process sequences and quality sequences are obtained At that time, first, based on X recorded in the offline phase... min and X max For the process data Normalization is then performed. Subsequently, this process sequence, along with the historical quality sequence, is input into the trained model to obtain the accumulated output of s structural blocks, which is the model's final prediction sequence.

[0064]

[0065] The following specific embodiment demonstrates the effectiveness of the prediction method of the present invention.

[0066] The process flow diagram of the methanation furnace is as follows: Figure 5As shown, this embodiment uses 10 process variables as part of the soft-sensor modeling input, including flow rate, pressure, temperature, and liquid level. For this process, approximately 210,000 data points of measurable process variables and carbon oxide content in the furnace were sampled continuously at equal time intervals. Subsequently, the dataset was divided into training, validation, and test sets in a 7:1:2 ratio. To predict in advance that the concentrations of CO and CO2 at the methanation furnace outlet will not exceed the threshold range, the method of this invention predicts the concentrations at the next 5 time points. Here, an input time window size of 30 is selected, that is, the carbon oxide concentration at the next 5 time points is predicted based on the process sequence and mass sequence of 30 historical sampling time points.

[0067] Simultaneously, to verify the effectiveness of the method of the present invention, soft measurement models based on MLP and univariate reconstruction (N-Beats) were constructed. In the multivariate deep reconstruction network of the present invention, the first two shared base layers have 15-15 nodes respectively, transformed along the time dimension; the last shared layer has 1 node, transformed along the feature dimension; the subsequent reverse reconstruction layer has 30 nodes, and the forward prediction layer has 5 nodes. The MLP-based network flattens the sequence data, followed by three layers with the number of nodes being 330-80-5 respectively. The univariate reconstruction (N-Beats) soft measurement model does not input process variables, and its parameters are consistent with those of the multivariate deep reconstruction network. Figure 6 As shown, the three methods fit the predictions for different future times. Figure 7 The prediction error of the model is shown. It can be seen that the prediction method based on Multivariate Deep Reconstruction Network (MDRNN) of this invention achieves better prediction accuracy and lower error at any time. This demonstrates that by stacking parallel reconstruction residual structures and forward prediction structures, the residuals of the historical quality sequences input to each structural block can be continuously reduced, thus more accurately approximating the quality index to be predicted. Table 1 shows the prediction R values ​​of three deep network-based soft measurement methods at different time points. 2 and mse.

[0068] Table 1. Prediction performance of three deep network-based soft measurement methods at different time points.

[0069]

[0070]

[0071] As can also be seen from Table 1, the R of the method of the present invention at different future times... 2 The values ​​are all greater than those of MLP and N-Beats, while the values ​​of mse are all less than those of MLP and N-Beats. This indicates that MDRNN is better able to capture the sequential nature of variables and conforms to the dynamic nature of the process in the task of predicting carbon oxide content in methanation furnaces, thus having better prediction accuracy.

[0072] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for multi-step prediction of carbon oxides content in a methanator based on a multivariate deep reconstruction neural network, characterized in that, The method comprises the following steps: Step one: Collecting historical data from the methanator process to build a dataset ; wherein, is a measurable process variable, Y represents the outlet carbon oxide content; n represents the number of collected samples, and R represents a real number set. Step two: Normalize each process variable in the set to the interval [0, 1] to obtain ; Step three: select a reasonable input time window for the methanation furnace process At the same time, select the time window of the carbon oxide content index to be predicted ; split and sequence the normalized process variables X and outlet carbon oxide content Y to obtain one-to-one input sequence and prediction sequence; Step 4: Construct and train a multivariate deep reconstruction neural network; the multivariate deep reconstruction neural network includes s structural blocks, each structural block including a splicing layer, a shared network, a reverse reconstruction layer, and a forward prediction layer; the shared network includes three nonlinear layers; the splicing layer converts the process variable sequence... and carbon oxide content sequence Connect along the feature dimensions to obtain the input sequence. The first two nonlinear layers of the shared network are related to the input. Perform feature extraction along the time dimension to obtain Then, feature extraction is performed along the dimensions of the variables to obtain the final sequence features extracted by the shared network. Subsequently, Each input to a node number is The reverse reconstruction layer and a node number are The forward prediction layer; the reverse reconstruction layer reconstructs the input history sequence of the quality variables to obtain the reverse reconstruction output. The forward prediction layer predicts the carbon oxide content at several future time points, obtaining the forward prediction output. Subsequently, the historical quality sequence output by the reverse reconstruction layer is... The input quality sequence of this structure block The difference is calculated to obtain the quality sequence residual. The residual With process sequence The sum of the forward prediction outputs of the s structural blocks serves as the input to the next structural block; the sum of these forward prediction outputs is used as the final quality index prediction sequence. ; Step five: normalizing the process variable to be tested, and then inputting the trained multivariate deep reconstruction neural network to obtain a predicted sequence of the outlet carbon oxide content.

2. The multistep prediction method of carbon oxides content of a methanator based on a multivariate deep reconstruction neural network according to claim 1, characterized in that, The normalization formula of step two is as follows: ； wherein denotes the minimum value of each process variable in the time dimension, , respectively, denotes the maximum value of each process variable in the time dimension, .

3. The multistep prediction method of carbon oxides content of a methanator based on a multivariate deep reconstruction neural network according to claim 1, characterized in that, The process of feature extraction of the shared network is expressed by the following formula: ； ； ； wherein, represents a weight coefficient of the first layer nonlinear layer, represents a bias coefficient of the first layer nonlinear layer, represents a weight coefficient of the second layer nonlinear layer, represents a bias coefficient of the first layer nonlinear layer, represents a weight coefficient of the third layer nonlinear layer, represents a bias coefficient of the third layer nonlinear layer, ReLU is an activation function, which outputs 0 for inputs less than 0 and leaves the output unchanged for inputs greater than 0.

4. The multistep prediction method of carbon oxides content of a methanator based on a multivariate deep reconstruction neural network according to claim 1, characterized in that, The reverse reconstruction output and the forward prediction output The calculation formula is as follows: ; ; wherein, denotes a weight coefficient of the backward reconstruction layer, denotes a bias coefficient of the backward reconstruction layer, denotes a weight coefficient of the forward prediction layer, denotes a bias coefficient of the forward prediction layer.

5. The multistep prediction method of carbon oxides content of a methanator based on a multivariate deep reconstruction neural network according to claim 1, characterized in that, the reconstructed residual of the last structure block as the final reconstructed residual : ; The training loss function of the multivariate deep reconstruction neural network is denoted as Loss 。

Images