Neural network multi-step prediction method based on space-time information conversion

Abstract

The invention provides a neural network multi-step prediction method based on spatio-temporal information conversion. The method comprises the following steps: collecting and preprocessing data; constructing a neural network model for preliminary prediction; model parameters are optimized through a spatio-temporal information conversion equation, and multi-step prediction is carried out; calculating a prediction error and extracting principal component information, and establishing an error estimation model; and finally, outputting a final prediction result by using the optimized model, and carrying out dynamic compensation through an error estimation model. Through spatio-temporal information conversion and an error compensation mechanism, the accuracy and the stability of a prediction result are remarkably improved, and the problem that the precision is attenuated along with the increase of the step length in multi-step prediction is solved. The method is suitable for various industrial processes, and has wide application prospects and remarkable economic benefits.

Description

Technical Field

[0001] The present invention relates to a neural network multi-step prediction method based on spatio-temporal information transformation. Background Art

[0002] In modern industrial processes such as sewage treatment, petrochemical industry, and metal smelting, there are many random factors. If not properly handled, it will lead to low production efficiency, high operating costs, and more seriously, abnormal working conditions, equipment downtime and other problems, causing huge economic losses and endangering social stability. Therefore, real-time monitoring of key variables and dynamically adjusting controllable factors in modern industrial processes based on the monitoring results can ensure their optimal operation.

[0003] Modern industrial processes usually have characteristics such as long process, large scale, non-linearity, and strong coupling. Due to technical conditions and economic costs, it is difficult to monitor key variables in real time. Soft sensing technology has the advantages of low detection cost, fast response speed, and simple maintenance. Although it has been widely studied and applied, due to the large spatio-temporal span of the collected sample data, it is difficult to provide effective information for soft sensing modeling, thus affecting the model prediction performance and resulting in poor prediction accuracy. In addition, most of the existing soft sensing models are single-step predictions, and the timeliness of the prediction results is poor, unable to provide timely working condition information for operators.

[0004] The proposed multi-step prediction method solves the above problems and can predict the changes in operating conditions early, which is helpful for the intelligent management of industrial processes. However, the multi-step prediction model structure is complex, and with the increase of the prediction steps, the prediction accuracy continuously decays, resulting in model inaccuracy. Therefore, mining spatio-temporal information in sample data and improving the stability of model prediction performance is an urgent technical problem to be solved. Summary of the Invention

[0005] The present invention provides a neural network multi-step prediction method based on spatio-temporal information transformation to solve the problems existing in the above-mentioned prior art.

[0006] The technical solutions adopted by the present invention are as follows:

[0007] A neural network multi-step prediction method based on spatio-temporal information transformation, comprising the following steps:

[0008] S1) Collect the original data set in the target industrial process and perform standardization processing on the original data set;

[0009] S2) Construct a neural network model and output a preliminary prediction result of the target variable based on the standardized data;

[0010] S3) Establish a spatio-temporal information conversion equation to convert spatial information into temporal information, optimize the parameters of the neural network, and use the optimized neural network model for multi-step prediction to output the preliminary multi-step prediction results;

[0011] S4) Calculate the preliminary prediction error based on the preliminary prediction results in S2), perform variational mode decomposition on the prediction error, and obtain the principal component information;

[0012] S5) Based on the obtained principal component information, establish a prediction error estimation model based on a bidirectional long short-term memory neural network for subsequent dynamic compensation of the multi-step prediction results;

[0013] S6) Input the test data into the optimized neural network model to output the final multi-step prediction results, and use the prediction error estimation model in S5) to perform dynamic compensation on the final multi-step prediction results to improve the accuracy of the prediction results.

[0014] Furthermore, the original dataset consists of multiple process variables measured by installed sensors t represents the measurement time, and n represents the number of auxiliary variables. Among them, the last column of data represents the target variable y t , that is, the key variable to be predicted, and the remaining columns of data represent the auxiliary variables

[0015] Furthermore, perform standardization processing on the original dataset, specifically:

[0016]

[0017] Among them, x represents the original data, x′ represents the normalized sample data, x min and x max represent the minimum and maximum values in the original dataset respectively. i

[0018] Furthermore, construct a (m + 1)-layer neural network model, which is denoted as F, and transmit data information through each layer of the network;

[0019]

[0020] Among them, x i represents the i-th auxiliary variable data input, h k represents the output of the neural network, f k represents the non-linear activation function, ω k and b k represent the weight matrix and the threshold vector respectively, k represents the k-th layer of the neural network, and l represents the number of neurons.

[0021] Furthermore, the spatio-temporal information conversion equation is:

[0022]

[0023] Among them, represents the auxiliary variable data at time t, and Y t =(y t , y t+1 , …, y t+L-1 ) represents the target variable data at L steps after time t, where t = 1, 2, …, m, m represents the number of training data, and A and B represent conjugate coefficient matrices, that is, AB = I.

[0024] Furthermore, the prediction error is decomposed by Hilbert transform:

[0025]

[0026] Among them, Q represents the total number of modes, δ(t) represents the distribution function at time t, u q (t) represents the q-th mode at time t, ω q is the corresponding center frequency, and e(t) represents the preliminary prediction error at time t.

[0027] Furthermore, the principal component information is obtained as the input data for modeling:

[0028] v(t) = u1(t), u2(t), …, u p (t)

[0029] Among them, p represents the number of principal components obtained;

[0030] A prediction error estimation model based on a bidirectional long short-term memory neural network is established:

[0031] f t = σ(w f · [e′(t - 1), v(t)] + b f )

[0032] In the forget gate, ft t represents the information retention coefficient, e′(t - 1) represents the output information at time (t - 1), v(t) represents the principal component information extracted after variational mode decomposition, σ represents the activation function, w f and b f represent the weight matrix and threshold vector in the forget gate;

[0033] i t = σ(w t [e′(t - 1), v(t)] + b t )

[0034]

[0035] At the input gate, i t represents the output information of the sigmoid layer, w t and b t represent the weight matrix and threshold vector of this layer of the network; represents the output information of the tanh layer, w c and b c represent the weight matrix and threshold vector of this layer of the network;

[0036] O t =σ(w o [e′(t - 1), v(t)] + b o )

[0037] At the output gate, O t represents the output information coefficient, w o and b o represent the weight matrix and threshold vector of this layer of the network;

[0038]

[0039] e′(t)=O t *tanh(C t )

[0040] where C t represents the output information of the current input gate, and e′(t) represents the output information at time t, that is, the estimated value of the prediction error.

[0041] Furthermore, the S6) is specifically: inputting the test data into the optimized neural network model to output the final multi-step prediction results of the key variables:

[0042]

[0043] where represents the test data, and (y t , y t+1 ,…, y t+L-1 ) represents the final multi-step prediction results of the key variables;

[0044] Performing dynamic compensation on the final multi-step prediction results using the prediction error estimation model:

[0045]

[0046] where represents the final multi-step prediction results of the key variables.

[0047] The present invention has the following beneficial effects:

[0048] 1) The complex structure of modern industrial processes results in the acquisition of raw data from different spaces and times. By establishing a spatio-temporal information conversion equation, this invention converts spatial information into temporal information, effectively solving the problem of large spatio-temporal span of data in industrial processes. This conversion method can make full use of the spatio-temporal information in the data and improve the prediction performance of the model.

[0049] 2) Using a prediction error estimation model to dynamically compensate the multi-step prediction results effectively prevents the problem of model performance degradation as the prediction step length increases. The bidirectional long short-term memory neural network can fully consider the influence of data changes at adjacent times, significantly enhancing the stability of model performance and the accuracy of prediction results.

[0050] 3) The multi-step prediction method can predict the changes in operating conditions early, providing timely condition information for operators and contributing to the intelligent management of industrial processes. Compared with single-step prediction, multi-step prediction can provide more comprehensive future trend information to support more scientific decision-making.

[0051] 4) Optimizing the parameters of the neural network through the spatio-temporal information conversion equation improves the training efficiency and prediction accuracy of the model. The optimized model can better fit complex industrial processes, reducing the risks of overfitting and underfitting.

[0052] 5) Extracting principal component information through variational mode decomposition makes full use of the key features in the data and improves the data utilization efficiency. This data processing method can effectively reduce noise interference and enhance the robustness of the model.

[0053] 6) Through the error compensation mechanism and spatio-temporal information conversion, the model shows stronger robustness in the face of complex operating conditions and data noise. Even in the case of poor data quality, it can maintain a high prediction accuracy.

[0054] 7) This technical solution is applicable to various modern industrial processes such as sewage treatment, petrochemical industry, and metal smelting, with broad application prospects. By real-time monitoring of key variables, it can ensure the optimized operation of industrial processes, reduce operating costs, and improve production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of the present invention.

[0056] Figure 2 It is a structure diagram of a certain actual sewage treatment plant.

[0057] Figure 3 It is a neural network prediction model diagram based on spatio-temporal information conversion.

[0058] Figure 4 It is a diagram of a bidirectional long short-term memory neural network.

[0059] Figure 5 This is the prediction result graph of the effluent BOD concentration by the method of the present invention. Specific embodiments

[0060] The present invention will be further described below with reference to the accompanying drawings.

[0061] Figure 1 This is the algorithm flowchart of the method of the present invention. The present invention will be further described below with reference to specific embodiments.

[0062] This embodiment takes an actual sewage treatment plant as an example. The sewage treatment plant adopts the A 2 / O treatment process. Figure 2 This is its structure diagram, including primary physical treatment, secondary biochemical treatment and tertiary advanced treatment. The sewage treatment scale is 150,000 tons per day. A total of 14 process variable sensors are installed at the inlet and outlet to monitor the sewage treatment process, and the detection frequency of each sensor is 1 day.

[0063] The specific implementation steps are as follows:

[0064] Step 1: Collect the original data set of this sewage treatment plant from November 1, 2022 to February 28, 2023, a total of 120 groups, including the pH value, chemical oxygen demand concentration (COD), biochemical oxygen demand concentration (BOD), ammonia nitrogen concentration, total phosphorus, total nitrogen, and suspended solid concentration at the inlet and outlet. Among them, the effluent BOD is used as the target variable y t , and the remaining variables are auxiliary variables. Use the first 90 groups of data as the training data set and the last 30 groups of data as the test data set;

[0065] Step 2: Standardize the 120 groups of original data sets, and the method is as follows:

[0066]

[0067] Among them, x represents the original data, x′ represents the normalized sample data, x min and x max respectively represent the minimum and maximum values in the original data set;

[0068] Step 3: Establish a 5-layer neural network prediction model. This neural network is denoted as F, and data information is transmitted through each layer of the network;

[0069]

[0070] Among them, x i represents the input auxiliary variable data, h k represents the output of the neural network, f krepresents a non - linear activation function, ω k and b k represent the weight matrix and the bias vector respectively, k represents the k - th layer neural network, and l = 50 represents the number of neurons.

[0071] Step 4: Establish a spatio - temporal information conversion equation, determine the parameters of the neural network, and apply the model to the multi - step prediction of the system; The neural network prediction model based on spatio - temporal information conversion is as Figure 3 shown;

[0072]

[0073] Among them, represents the auxiliary variable data at time t, Y r =(y r ,y r+1 ,…,y t+L-1 ) represents the target variable data at L steps after time t, t = 1,2,…,m, m = 90 represents the number of training data, L represents the prediction step length, A and B represent conjugate coefficient matrices, that is, AB = I;

[0074] Step 5: Calculate the preliminary prediction error, perform variational mode decomposition to obtain the principal component information, and establish a prediction error estimation model based on a bidirectional long - short - term memory neural network. The specific steps are as follows:

[0075] Step 5.1: Decompose the prediction error using Hilbert transform;

[0076]

[0077] Among them, Q = 10 represents the total number of modes, δ(t) represents the distribution function at time t, u q (t) represents the q - th mode at time t, ω q is the corresponding central frequency, and e(t) represents the preliminary prediction error at time t;

[0078] Step 5.2: Obtain 5 principal component information as the input data for modeling;

[0079] v(t)=u1(t),u2(t),…,u5(t)

[0080] Step 5.3: Establish a prediction error estimation model based on a bidirectional long - short - term memory neural network; (The input layer has 5 neurons, the BiLSTM layer has 16 neurons, and the output layer has 1 neuron. The bidirectional long - short - term memory neural network is as Figure 4 shown)

[0081] f t =σ(w f ·[e′(t - 1),v(t)]+b f)

[0082] At the forgetting gate, f t represents the information retention coefficient, e′(t - 1) represents the output information at time (t - 1), v(t) represents the principal component information extracted after variational mode decomposition, σ represents the activation function, w f and b f represent the weight matrix and threshold vector in the forgetting gate;

[0083] i t = σ(w t [e′(t - 1), v(t)] + b t )

[0084]

[0085] At the input gate, i t represents the output information of the sigmoid layer, w t and b t represent the weight matrix and threshold vector of this layer of the network; represents the output information of the tanh layer, w c and b c represent the weight matrix and threshold vector of this layer of the network;

[0086] O t = σ(w o [e′(t - 1), v(t)] + b o )

[0087] At the output gate, O t represents the output information coefficient, w o and b o represent the weight matrix and threshold vector of this layer of the network;

[0088]

[0089] e′(t) = O t * tanh(C t )

[0090] where C t represents the output information of the current input gate, e′(t) represents the output information at time t, that is, the estimated value of the prediction error;

[0091] Step 6: Input 30 groups of test data into the multi-step prediction model, output the multi-step prediction results of key variables, and at the same time use the prediction error estimation model to dynamically compensate the multi-step prediction results to improve the accuracy of the prediction results;

[0092] Step 6.1: Input 30 groups of test data into the multi-step prediction model, and output the multi-step prediction results of key variables;

[0093]

[0094] Among them, represents the test data, (y t , y t+1 , …, y t+L-1 ) represents the multi-step prediction results of the key variables, where t = 91, 92, …, 120;

[0095] Step 6.2: Dynamically compensate the multi-step prediction results using the prediction error estimation model to improve the accuracy of the prediction results;

[0096]

[0097] Among them, represents the final multi-step prediction results of the key variables. In this embodiment, the prediction results of the BOD concentration at the outlet are as Figure 5 shown. The X-axis is the number of samples of the test data, the Y-axis is the true value and the predicted value of the BOD concentration at the outlet, with the unit of mg / L. The solid line is the true value of the BOD concentration at the outlet, and the dashed line is the predicted value of the BOD concentration at the outlet.

[0098] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, several improvements can be made without departing from the principle of the present invention, and these improvements should also be regarded as the protection scope of the present invention.

Claims

1. A neural network multi-step prediction method based on spatio-temporal information transformation, characterized in that: It includes the following steps: S1) Collect the original data set in the target industrial process and standardize the original data set; S2) Build a neural network model and output the preliminary prediction result of the target variable based on the standardized data; S3) Establish a spatio-temporal information conversion equation to convert spatial information into temporal information, optimize the parameters of the neural network, and use the optimized neural network model for multi-step prediction to output the preliminary multi-step prediction result; S4) Calculate the preliminary prediction error based on the preliminary prediction result in S2), perform variational mode decomposition on the prediction error, and obtain the principal component information; S5) Based on the obtained principal component information, establish a prediction error estimation model based on a bidirectional long short-term memory neural network for subsequent dynamic compensation of the multi-step prediction result; S6) Input the test data into the optimized neural network model to output the final multi-step prediction result, and use the prediction error estimation model in S5) to dynamically compensate the final multi-step prediction result to improve the accuracy of the prediction result.

2. The neural network multi-step prediction method based on spatio-temporal information transformation according to claim 1, wherein: The specific method for standardizing the original data set is as follows: where x represents the original data, x' represents the normalized sample data, x min and x max represent the minimum and maximum values in the original dataset, respectively.

3. The multi-step prediction method of neural network based on spatio-temporal information transformation according to claim 1, characterized in that: Build a neural network model with (m + 1) layers, which is denoted as F, and transmit data information through each layer of the network; Among them, x i represents the i-th auxiliary variable data of the input, h k represents the output of the neural network, f k represents the non-linear activation function, ω k and b k represent the weight matrix and the threshold vector respectively, k represents the k-th layer neural network, and l represents the number of neurons.

4. The neural network multi-step prediction method based on spatio-temporal information transformation according to claim 1, characterized in that: The spatio-temporal information conversion equation is: Among them, represents the auxiliary variable data at time t, Y t =(y t , y t+1 , …, y t+L-1 ) represents the target variable data at L steps after time t, t = 1, 2, …, m, where m represents the number of training data, and A and B represent conjugate coefficient matrices, i.e., AB = I.

5. The neural network multi-step prediction method based on spatio-temporal information transformation according to claim 1, characterized in that: Use the Hilbert transform to decompose the prediction error: Among them, Q represents the total number of modes, δ(t) represents the distribution function at time t, and u q (t) represents the q-th order mode at time t, and ω q is the corresponding center frequency, and e(t) represents the preliminary prediction error at time t.

6. The neural network multi-step prediction method based on spatio-temporal information transformation according to claim 5, characterized in that: Obtain the principal component information as the input data for modeling: v(t) = u1(t), u2(t), …, u p (t) where p represents the number of principal components obtained; Establish a prediction error estimation model based on a bidirectional long short-term memory neural network: f t = σ(w f ·[e′(t - 1), v(t)] + b f ) At the forgetting gate, f t represents the information retention coefficient, e′(t - 1) represents the output information at time (t - 1), v(t) represents the principal component information extracted after variational mode decomposition, σ represents the activation function, w f and b f represent the weight matrix and threshold vector in the forgetting gate; i t = σ(w t [e′(t - 1), v(t)] + b t ) At the input gate, i t represents the output information of the sigmoid layer, w t and b t represent the weight matrix and threshold vector of this layer of the network; represents the output information of the tanh layer, w c and b c represent the weight matrix and threshold vector of this layer of the network; O t = σ(w o [e′(t - 1), v(t)] + b o ) At the output gate, O t represents the output information coefficient, w o and b o represent the weight matrix and the threshold vector of this layer of the network; e′(t) = O t *tanh(C t ) Among them, C t represents the output information of the current input gate, and e′(t) represents the output information at time t, that is, the estimated value of the prediction error.

7. The neural network multi-step prediction method based on spatio-temporal information transformation according to claim 1, characterized in that: The specific content of S6) is: input the test data into the optimized neural network model to output the final multi-step prediction result of the key variable: Among them, represents the test data, (y t , y t+1 , …, y t+L-1 ) represents the final multi-step prediction results of the key variables; Dynamically compensate the final multi-step prediction result using the prediction error estimation model: Among them, represents the final multi-step prediction result of the key variable.

Images