Industrial data correction method and device based on improved LSTM neural network

Abstract

The invention discloses an industrial data correction method and device based on an improved LSTM neural network, and relates to the technical field of industrial process variable data correction, an improved robust estimation function is used as a loss function of an LSTM neural network model, and the situation that measurement data containing significant errors pollute other normal data is inhibited. Meanwhile, kernel width parameters in the loss function are updated, the obtained industrial evaporation time sequence data serve as input of the LSTM neural network model, forward propagation and back propagation are carried out to update the network weight, the improved LSTM neural network model can learn the trend of the industrial evaporation time sequence data, and the learning efficiency of the LSTM neural network model is improved. Influences of random errors and significant errors on measurement data are reduced, so that the measurement data are accurately obtained, and errors of industrial data are corrected.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial process variable data correction, and in particular to an industrial data correction method and device based on an improved LSTM neural network. Background Art

[0002] Alumina is a very important raw material in industrial production. Due to its high melting point and high hardness, it is widely used in traditional refractory materials, ceramics, metallurgy, chemical industry and other fields. The evaporation process of alumina production is a key process in alumina production. It is used to extract and concentrate alumina from aluminum-containing solutions, and the quality of the evaporated product is the core concern of the evaporation process of alumina production. During the evaporation process, the excess water in the feed liquid needs to be evaporated to obtain high-quality sodium aluminate solution. Various measuring instruments are inevitably affected by the high temperature and strong alkaline environment on site, which causes random errors in the measurement data.

[0003] In addition, due to improper operation of on-site workers, instrument leakage and other interference, the measurement data may have significant errors; limited by the measurement cost and measurement environment, and the measurement of intermediate process parameters is relatively dangerous, the evaporation process only detects the flow rate, concentration and other information of the slurry at the process inlet and outlet, and most of the parameters in the middle of the process can only be treated as unknown information; inaccurate and incomplete production information affects the industrial process modeling and optimization control accuracy; therefore, in the evaporation process of alumina production, random errors and significant errors are corrected, and the unknown parameter data information is estimated to improve the accuracy and effectiveness of industrial data, which is of great significance to the smooth operation of industry, efficient production and green energy conservation.

[0004] At present, the error correction of industrial data is usually carried out by using simulation software to predict the obtained flow and concentration information in combination with production parameters. However, since the errors of industrial data usually include random errors and significant errors, there is usually a high degree of uncertainty in the prediction, resulting in inaccurate measurement data and difficulty in correcting the errors of industrial data. Summary of the invention

[0005] The embodiment of the present invention provides an industrial data correction method and device based on an improved LSTM neural network, which can solve the problem in the prior art that the measurement data is inaccurate and it is difficult to correct the errors of the industrial data.

[0006] The embodiment of the present invention provides an industrial data correction method based on an improved LSTM neural network, comprising the following steps:

[0007] Obtain time series data of industrial evaporation during industrial production;

[0008] Construct an improved robust estimation function based on random errors and significant errors of industrial evaporation time series data;

[0009] The improved robust estimation function is used as the loss function of the LSTM neural network to build an LSTM neural network model;

[0010] The industrial evaporation time series data is used as the input of the LSTM neural network model for forward propagation. According to the output results of the forward propagation and the deviation of the industrial evaporation time series data containing random errors and significant errors, the loss function value of the current LSTM neural network model is obtained, and the back propagation is performed based on the loss function value to iteratively update the network weights of the LSTM neural network model to obtain the improved LSTM neural network model.

[0011] The improved LSTM neural network model is used to obtain the correction value of industrial evaporation time series data.

[0012] Preferably, the improved robust estimation function is:

[0013]

[0014] Where: e represents the weighted residual between the contamination value and the correction value; δ represents the kernel width of the loss function;

[0015]

[0016] Where: x m Indicates the pollution value containing gross errors; x r represents the correction value; ε represents the standard deviation of the variable.

[0017] Preferably, after the improved robust estimation function is obtained, the kernel width parameter in the improved robust estimation function is adjusted, including:

[0018] The exp(-e 4 / (2δ 2 )) is expressed as exp(-F). As F increases, the value of exp(-F) continues to decrease;

[0019] Set the F value to a threshold F thre , at the beginning of back propagation, set the F value to F ini , and F ini <F thre , as the number of training times continues to increase, the F value changes from F ini The value continues to increase to F thre value; then the value is:

[0020] F=e 4 / (2δ 2 )=F ini +η*epoch

[0021] δ 2 =e 4 / 2(F ini +η*epoch)

[0022] Where: e represents the weighted residual between the contamination value and the correction value; δ represents the kernel width of the loss function; η represents the iteration step size; epoch represents the number of iterations;

[0023] The kernel width parameter δ is updated by the η*epoch factor.

[0024] Preferably, the determination of the forward propagation of the LSTM neural network model includes:

[0025] According to the input dimension and output dimension of the industrial evaporation time series data in the LSTM neural network model, the input dimension, hidden layer dimension, output dimension, LSTM layer, number of LSTM layers and linear layer of the LSTM neural network model are defined;

[0026] According to the correction target of industrial evaporation time series data in the LSTM neural network model, the forward propagation mode of the LSTM neural network model is defined as follows: the input industrial evaporation time series data is processed by the LSTM layer, and the correction value is obtained after processing by the linear layer.

[0027] Preferably, the iterative updating of the LSTM neural network model network weights comprises:

[0028] After obtaining the loss function value of the LSTM neural network model, the loss function value is used as the input of the LSTM neural network model for back propagation. The improved robust estimation function iteratively updates the network weights of the LSTM neural network model, and the weight parameters are updated by the gradient descent method. The update method is:

[0029]

[0030] Where: W ih and b ih Respectively represent the weight parameter and bias parameter between the input state and the hidden state; W hh and b hh They represent the weight parameters and bias parameters between hidden states respectively; α represents a fixed learning rate, and α>0; Ln represents the loss function.

[0031] Preferably, the step of obtaining the correction value of the industrial evaporation time series data includes:

[0032] The improved LSTM neural network model is used to obtain the correction adjustment value of the industrial evaporation time series data, and the industrial evaporation time series data is denormalized based on the correction adjustment value to obtain the correction value of the industrial evaporation time series data; the denormalization method is:

[0033] x r =x nr *(x maxm -x minm )+x minm

[0034] Where: x r Indicates the final correction value; x nr represents the normalized correction value; Indicates the maximum value containing significant errors; Represents the minimum value containing significant error.

[0035] The embodiment of the present invention also provides an industrial data correction device based on an improved LSTM neural network, comprising:

[0036] Data module, used to obtain industrial evaporation time series data in industrial production process;

[0037] Construct an improved robust estimation function based on random errors and significant errors of industrial evaporation time series data;

[0038] The model module is used to use the improved robust estimation function as the loss function of the LSTM neural network to build an LSTM neural network model;

[0039] The industrial evaporation time series data is used as the input of the LSTM neural network model for forward propagation. According to the output results of the forward propagation and the deviation of the industrial evaporation time series data containing random errors and significant errors, the loss function value of the current LSTM neural network model is obtained, and the back propagation is performed based on the loss function value to iteratively update the network weights of the LSTM neural network model to obtain the improved LSTM neural network model.

[0040] The correction module is used to obtain the correction value of industrial evaporation time series data using the improved LSTM neural network model.

[0041] The embodiment of the present invention provides an industrial data correction method and device based on an improved LSTM neural network. Compared with the prior art, the beneficial effects thereof are as follows:

[0042] The present invention uses an improved robust estimation function as the loss function of the LSTM neural network model, uses the loss function to determine the forward propagation process of the constructed LSTM neural network model, suppresses the influence of significant errors and random errors contained in industrial evaporation time series data on normal data when the LSTM neural network model processes the industrial evaporation time series data, and then uses the industrial evaporation time series data as the input of the LSTM neural network model to perform forward propagation and back propagation to update the network weights, so that the improved LSTM neural network model can learn the trend of the industrial evaporation time series data, reduce the influence of random errors and significant errors on the measurement data, so as to accurately obtain the measurement data and correct the errors of the industrial data. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 An overall flow chart of an industrial data correction method based on an improved LSTM neural network provided by an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of a robust correction structure of industrial evaporation data based on an improved LSTM network, provided by an industrial data correction method based on an improved LSTM neural network in an embodiment of the present invention;

[0045] Figure 3 A schematic diagram of comparing industrial variable measurement values ​​and data correction values ​​of an industrial data correction method based on an improved LSTM neural network provided in an embodiment of the present invention;

[0046] Figure 4 A schematic diagram showing the comparison of the influence functions of an improved robust estimation function, a Welsh function and a Correntropy robust estimation function of an industrial data correction method based on an improved LSTM neural network provided in an embodiment of the present invention;

[0047] Figure 5 A schematic diagram comparing standard deviations of different variables calculated by an improved robust data correction method for industrial data correction method based on an improved LSTM neural network and a correction method based on Welsh and Correntropy robust estimation functions provided in an embodiment of the present invention;

[0048] Figure 6 A schematic diagram comparing the data standard deviation calculated by an improved robust data correction method of an industrial data correction method based on an improved LSTM neural network provided in an embodiment of the present invention and LSTM, NN, and NRE-NN methods. DETAILED DESCRIPTION

[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0050] See also Figure 1 The embodiment of the present invention provides an industrial data correction method based on an improved LSTM neural network, comprising the following steps:

[0051] Step 1: In order to avoid significant errors from contaminating the measurement data containing only random errors, a data correction model based on the novel robust estimator (NRE) is constructed to suppress the impact of significant error propagation.

[0052] Step 2: In order to improve the solution accuracy of the data correction model, the improved robust estimation function is used as the loss function of the LSTM neural network to effectively combine data correction with deep learning.

[0053] Step 3: Train and adjust the parameters of the LSTM neural network based on the improved robust estimation function to learn the trend of evaporation data and improve the robustness of the measurement data to significant errors.

[0054] Step 4: Use evaluation indicators to verify the correction results of the LSTM neural network based on the improved robust estimation function, eliminate significant errors, reduce random errors, and improve the validity and reliability of the measurement data.

[0055] Specifically:

[0056] Step 1: First determine the mathematical expression of the improved robust estimation function. The formula is as follows, and calculate its partial derivative with respect to the deviation: The influence function is obtained, and the formula is as follows. Its convergence is analyzed to effectively eliminate the significant error of evaporation measurement data and improve the robustness of measurement data. The formulas of each part are:

[0057]

[0058] Where: x m Indicates the pollution value containing gross errors; x r represents the correction value; ε represents the standard deviation of the variable; e represents the weighted residual between the contamination value and the correction value; δ represents the kernel width of the loss function.

[0059] Step 2: Using the improved robust estimation function as the loss function of the LSTM neural network can automatically learn the trend of the data to process the industrial evaporation time series data in industrial production; the LSTM network includes an input gate I that reads the data into the unit t , output gate O from the cell output entry t , the forget gate F that resets the cell content t , memory element C used to record additional information t and candidate memory And the hidden state H t , the calculation formula is:

[0060] I t =σ(X t W xi +H t-1 W hi +b i )

[0061] F t =σ(X t W xf +H t-1 W hf +b f )

[0062] O t =σ(X t W xo +H t+1 W ho +b o )

[0063]

[0064] H t =O t ☉tanh(C t )

[0065] Where: σ represents the fully connected layer with activation function, X t Represents the input data at the current moment, H t-1 represents the hidden state of the previous moment, W xi , W hi , W xf , W hf , W xo , W ho , W xc , W hc They represent the weights of the current input, the current hidden state, the weights between the input gate and the forget gate, the weights between the hidden state and the forget gate, the weights between the input and the output, the weights between the hidden state and the output, the weights between the input and the memory element, and the weights between the hidden state and the memory element. i , b f , bo , b c They represent the bias of the input gate, forget gate, output gate and memory cell respectively, and e represents the multiplication of the corresponding elements. The input gate, forget gate, output gate and candidate memory cells are all obtained by the combined effect of the current input and the hidden state of the previous moment.

[0066] Because I t With F t After passing through the fully connected layer with activation function, it is between 0 and 1. When the content of the cell at the previous moment is discarded, F t is set to 0, and when the content in the candidate memory cell at the current moment needs to be retained, I t is set to 1 to determine the content of the current memory cell. However, the content in the memory cell does not necessarily need to be passed to the hidden state, so the output gate controls whether there is content that needs to be passed to the hidden state.

[0067] Step 3: Based on the establishment of the LSTM neural network based on the improved robust estimation function, the training and parameter adjustment of the LSTM network model includes the following steps:

[0068] S1: Initialize the parameters of the improved robust estimation function and set the dynamic adjustment rules of the kernel width; the kernel width of the loss function is an adjustable parameter, and the specific parameter adjustment process is:

[0069] Set exp(-e 4 / (2δ 2 )) is expressed as exp(-F). As F increases, the value of exp(-F) first decreases rapidly, and then the downward trend becomes slow. When the function's downward trend becomes slow, the increase in the F value has a smaller impact on the function. Let the F value at this time be a threshold F thre ; When starting back propagation, first set the F value to F ini , F ini <F thre , so that the model has a larger gradient during training; as the number of training times increases, the gradient gradually decreases, the suppression effect on significant errors gradually increases, and the F value also gradually increases until it reaches F thre At this time, the F value continues to increase, and the suppression effect on significant errors no longer changes significantly. thre Different values ​​are taken according to different confidence levels; from the above principle, it can be concluded that F = e 4 / (2δ 2 )=F ini +η*epoch, so δ is derived 2 =e 4 / 2(F ini +η*epoch), the kernel width δ is adjusted by the η*epoch factor, where η is the iteration step and epoch is the number of iterations.

[0070] S2: According to the data input dimension and output dimension and the purpose of data correction robustness, define the input dimension, hidden layer dimension, output dimension and number of LSTM layers of the LSTM network; define the LSTM layer and linear layer and initialize the weights and biases; define the network forward propagation method, that is, implement the forward propagation process of processing the input data through the LSTM layer, then through the linear layer, and finally obtain the correction value, such as Figure 2 shown.

[0071] According to the actual situation in the industrial production process, the number of variables in the input dimension and output dimension of the LSTM network is 35, the number of hidden layers is set to 100, and the number of LSTM layers is set to 5; the network selects the Adam optimizer and sets the learning rate to 0.001. The learning rate scheduler is used to dynamically adjust the learning rate during training to improve the training effect; the network uses the Xavier initialization method to initialize the weights and biases. This initialization method can help the model converge better and help avoid gradient explosion or gradient disappearance.

[0072] S3: Normalize the industrial evaporation data and use the normalized data as the input of the LSTM network model for forward propagation; calculate the loss function value obtained by the current network coordination based on the deviation between the result of the forward propagation output and the data containing random errors and significant errors; perform back propagation based on the loss function value of the improved robust estimation function, and the LSTM network updates and iterates the weights according to the robust loss function; update and calculate the network weights, and optimize the parameters through gradient descent, and the formula is:

[0073]

[0074] Where: W ih and b ih are the weight parameter and bias parameter between the input and hidden state respectively; W hh and b hh are the weight parameter and bias parameter between hidden states respectively; α represents a fixed learning rate, and α>0; Ln represents the loss function; the current parameter minus the weighted partial derivative of the loss function on the parameter, and the gradient space gradually decreases until the optimal solution is found.

[0075] S4: Get the average loss value in the epoch iteration, and use the network with updated weights to go to step S3 until the number of iterations is completed.

[0076] S5: Using the trained improved LSTM neural network model, the standardized evaporation data correction value is obtained, and the obtained normalized data is denormalized by the following formula to obtain the data correction value; the data denormalization formula is:

[0077]

[0078] Where: x r Indicates the final correction value; x nr represents the normalized correction value; Indicates the maximum value containing significant errors; Represents the minimum value containing significant error.

[0079] Step 4: For the constructed LSTM network based on the improved robust estimation function, the standard deviation of the measured value and the corrected value of the evaporation data of the alumina production evaporation process are compared to measure the data correction effect and verify whether the correction result detects and eliminates significant errors.

[0080] The specific experiments are:

[0081] There are 35 variables in the evaporation process of alumina production, and each variable has 800 sets of experimental samples. In order to analyze the robustness of data correction based on the improved robust estimation function embedded in the LSTM neural network, the comparison results of 200 sets of measured values ​​and corrected values ​​are as follows: Figure 3 As shown, from Figure 3 It can be clearly observed that all the significant errors in the measured values ​​with significant errors were successfully identified and effectively corrected after network training. The overall trend of the evaporation data after correction was more stable, and the significant error interference was also well suppressed. As shown in Table 1, 1 to 35 represent the raw liquid flow rate, feed liquid temperature, steam temperature and condensate temperature in each effect evaporator and each level of flash evaporator. From the comparison of the standard deviation between the measured value and the corrected value, it can be seen that the standard deviation calculated by the corrected value is significantly lower than that calculated by the measured value, with a minimum reduction of 0.41, a decrease of 20%, and a maximum reduction of 3.28, a decrease of 87%.

[0082] Table 1 Comparison of standard deviations between measured and corrected data during evaporation process of alumina production

[0083] variable 1 2 3 4 5 6 7 8 9 10 11 12 Measurements 3.77 2.05 3.74 1.65 1.37 1.65 1.64 1.69 1.68 1.81 1.77 1.89 Correction value 0.49 1.64 1.54 1.04 0.56 0.19 0.41 0.25 0.55 0.21 0.63 0.21 variable 13 14 15 16 17 18 19 20 21 22 23 24 Measurements 1.89 2.41 1.57 1.76 1.63 1.58 1.62 1.78 1.85 1.62 1.72 1.87 Correction value 0.42 0.77 0.11 0.04 0.30 0.41 0.63 0.60 0.52 0.35 0.13 0.37 variable 25 26 27 28 29 30 31 32 33 34 35 Measurements 1.5 1.76 1.68 1.69 1.71 1.76 1.70 1.55 1.69 1.88 1.64 Correction value 0.04 0.43 0.33 0.52 0.78 0.29 0.13 0.36 0.37 0.34 0.035

[0084] Based on the LSTM network, the loss function is replaced by the Welsh and Correntropy robust estimation functions, such as Figure 4 As shown in the figure, a comparative experiment was carried out, and five variables were selected, namely variable 1, variable 22, variable 23, variable 24 and variable 25. The standard deviation of the corrected data is as follows Figure 5 As shown, variable 1 contains significant errors, and the other four variables only contain random errors; Figure 5 It can be seen that the LSTM network based on the improved robust estimation function has good robustness to data containing significant errors, can eliminate significant errors and reduce the impact of random errors.

[0085] NRE-LSTM, LSTM, NN, and NRE-NN were used to correct the data of the alumina production evaporation process, and the standard deviation comparison chart was calculated as follows: Figure 6 As shown in the figure, the standard deviation of the correction result based on NRE-LSTM is smaller than that of the three comparison methods. The NN network using the improved robust estimation function has a better effect on suppressing random errors than the NN network without the improved robust estimation function, and the LSTM network is obviously better at suppressing errors than the NN network.

[0086] The embodiment of the present invention also provides an industrial data correction device based on an improved LSTM neural network, comprising:

[0087] Data module, used to obtain industrial evaporation time series data in industrial production process.

[0088] An improved robust estimation function based on random errors and significant errors of industrial evaporation time series data is constructed.

[0089] The model module is used to use the improved robust estimation function as the loss function of the LSTM neural network and build an LSTM neural network model.

[0090] The industrial evaporation time series data is used as the input of the LSTM neural network model for forward propagation. According to the results of the forward propagation output and the deviation of the industrial evaporation time series data containing random errors and significant errors, the loss function value of the current LSTM neural network model is obtained. Back propagation is performed based on the loss function value, and the network weights of the LSTM neural network model are iteratively updated to obtain the improved LSTM neural network model.

[0091] The correction module is used to obtain the correction value of industrial evaporation time series data using the improved LSTM neural network model.

[0092] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. An industrial data correction method based on an improved LSTM neural network, characterized in that: The following steps are involved: Obtain time series data of industrial evaporation during industrial production; Construct an improved robust estimation function based on random errors and significant errors of industrial evaporation time series data; The improved robust estimation function is used as the loss function of the LSTM neural network to build an LSTM neural network model; The industrial evaporation time series data is used as the input of the LSTM neural network model for forward propagation. According to the output results of the forward propagation and the deviation of the industrial evaporation time series data containing random errors and significant errors, the loss function value of the current LSTM neural network model is obtained, and the back propagation is performed based on the loss function value to iteratively update the network weights of the LSTM neural network model to obtain the improved LSTM neural network model. The improved LSTM neural network model is used to obtain the correction value of industrial evaporation time series data.

2. The industrial data correction method based on the improved LSTM neural network according to claim 1 is characterized in that: The improved robust estimation function is: Where: e represents the weighted residual between the contamination value and the correction value; δ represents the kernel width of the loss function; Where: x m Indicates the pollution value containing gross errors; x r represents the correction value; ε represents the standard deviation of the variable.

3. The industrial data correction method based on the improved LSTM neural network according to claim 2 is characterized in that: After the improved robust estimation function is obtained, the kernel width parameter in the improved robust estimation function is adjusted, including: The exp(-e 4 / (2δ 2 )) is expressed as exp(-F). As F increases, the value of exp(-F) continues to decrease; Set the F value to a threshold F thre , at the beginning of back propagation, set the F value to F ini , and F ini <F thre , as the number of training times continues to increase, the F value changes from F ini The value continues to increase to F thre value; then the value is: F=e 4 / (2δ 2 )=F ini +η*epoch δ 2 =e 4 / 2(F ini +η*epoch) Where: e represents the weighted residual between the contamination value and the correction value; δ represents the kernel width of the loss function; η represents the iteration step size; epoch represents the number of iterations; The kernel width parameter δ is updated by the η*epoch factor.

4. The industrial data correction method based on the improved LSTM neural network according to claim 1 is characterized in that: The determination of the forward propagation of the LSTM neural network model includes: According to the input dimension and output dimension of the industrial evaporation time series data in the LSTM neural network model, the input dimension, hidden layer dimension, output dimension, LSTM layer, number of LSTM layers and linear layer of the LSTM neural network model are defined; According to the correction target of industrial evaporation time series data in the LSTM neural network model, the forward propagation mode of the LSTM neural network model is defined as follows: the input industrial evaporation time series data is processed by the LSTM layer, and the correction value is obtained after processing by the linear layer.

5. The industrial data correction method based on the improved LSTM neural network according to claim 1 is characterized in that: The iterative updating of the LSTM neural network model network weights includes: After obtaining the loss function value of the LSTM neural network model, the loss function value is used as the input of the LSTM neural network model for back propagation. The improved robust estimation function iteratively updates the network weights of the LSTM neural network model, and the weight parameters are updated by the gradient descent method. The update method is: Where: W ih and b ih Respectively represent the weight parameter and bias parameter between the input state and the hidden state; W hh and b hh They respectively represent the weight parameters and bias parameters between hidden states; α represents a fixed learning rate, and α>0; Ln represents the loss function.

6. The industrial data correction method based on the improved LSTM neural network according to claim 1 is characterized in that: The method of obtaining the correction value of the industrial evaporation time series data includes: The improved LSTM neural network model is used to obtain the correction adjustment value of the industrial evaporation time series data, and the industrial evaporation time series data is denormalized based on the correction adjustment value to obtain the correction value of the industrial evaporation time series data; the denormalization method is: Where: x r Indicates the final correction value; x nr represents the normalized correction value; Indicates the maximum value containing significant errors; Represents the minimum value containing significant error.

7. An industrial data correction device based on an improved LSTM neural network, characterized in that: include: Data module, used to obtain industrial evaporation time series data in industrial production process; Construct an improved robust estimation function based on random errors and significant errors of industrial evaporation time series data; The model module is used to use the improved robust estimation function as the loss function of the LSTM neural network to build an LSTM neural network model; The industrial evaporation time series data is used as the input of the LSTM neural network model for forward propagation. According to the output results of the forward propagation and the deviation of the industrial evaporation time series data containing random errors and significant errors, the loss function value of the current LSTM neural network model is obtained, and the back propagation is performed based on the loss function value to iteratively update the network weights of the LSTM neural network model to obtain the improved LSTM neural network model. The correction module is used to obtain the correction value of industrial evaporation time series data using the improved LSTM neural network model.

Images