CN-121682449-B - Variable working condition industrial temperature prediction method based on KWNet model

CN121682449BCN 121682449 BCN121682449 BCN 121682449BCN-121682449-B

Abstract

The invention discloses a variable working condition industrial temperature prediction method based on KWNet model, belonging to the technical field of industrial process parameter prediction. The method adds a novel double-flow wavelet transformation module, integrates approximate and detail components in a time sequence after wavelet decomposition in a double-flow decomposition and reconstruction mode, and can reconstruct all components without difference compared with the traditional wavelet transformation module, so that the approximate and detail components can be independently processed in different branches to prevent cross-level interference and ensure that the unique characteristics of each frequency band are reserved. And then, the characteristic components of seasonal and trending subsequences of the reconstructed time domain are effectively extracted through a learnable dynamic gate mechanism, and the learnable dynamic gate mechanism is innovatively used in pyramid multiscale expansion convolution and Fourier modulation attention, so that the sequence information of the time domain and the frequency domain is effectively balanced.

Inventors

LIU XIANG
WANG YAN
WANG ZIBIN
ZHANG XIAO
ZHAO SIZHE
JI ZHICHENG

Assignees

江南大学

Dates

Publication Date: 20260508
Application Date: 20260210

Claims (9)

1. A method for predicting variable-working-condition industrial temperature based on KWNet model, which is characterized by comprising the following steps: step 1, acquiring historical parameters in a specific industrial process to construct an industrial time sequence data set; Step 2, expert knowledge is formulated for the specific industrial process; step 3, constructing a KWNet-based prediction model by combining the formulated expert knowledge; Step 4, training the KWNet-based predictive model by utilizing the industrial time series data set; Step 5, completing multi-step prediction of variable working condition industrial temperature parameters by using a trained KWNet-based prediction model; The KWNet-based prediction model comprises a double-flow wavelet transformation module, a learnable dynamic gate mechanism module, a domain knowledge analysis and formulation module, an adaptive time-knowledge fusion loss function module and a time-frequency fusion decoder, wherein the domain knowledge analysis and formulation module is used for formulating formulated expert knowledge and constructing an expert knowledge loss function ; The dual-flow wavelet transformation module is used for reconstructing the original input data after wavelet transformation to obtain seasonal components and trending components, the leachable dynamic door machine module is used for determining the weights of the seasonal components and the trending components so as to obtain overall reconstruction data, then inputting the overall reconstruction data into the time-frequency fusion decoder to obtain a preliminary prediction result, and obtaining a time domain loss function based on the preliminary prediction result The adaptive time-knowledge fusion loss function module is used for integrating the expert knowledge loss function And the time domain loss function Obtaining the total loss function of the KWNet-based prediction model 。
2. The method of claim 1, wherein the dual stream wavelet transform module comprises a wavelet decomposition unit, a frequency processing unit, and a time-frequency joint feature extractor; First, the original input data is subjected to m-level discrete wavelet decomposition by a wavelet decomposition unit to generate a group of approximation coefficients And detail coefficient ,j=2,3,...,m; Details coefficients generated by the wavelet decomposition unit The input frequency processing unit obtains And approximate coefficient Co-computing seasonal components ; Approximation coefficients generated by the wavelet decomposition unit Input time-frequency joint characteristic extractor processing to obtain And detail coefficient Co-computing trending components 。
3. The method of claim 2, wherein the step of determining the position of the substrate comprises, the learnable dynamic gate mechanism module obtains seasonal components from the double-current wavelet transformation module And trend component Splicing the final dimension d to obtain a reconstructed signal A reconstructed signal obtained by splicing a linear layer pair added with a Sigmoid function Projection, generating gating tensor Then obtaining output data through gate weighted summation 。
4. A method according to claim 3, wherein the time-frequency fusion decoder designs a fixed core size in the time branch And three different expansion rates Is used for hierarchically capturing local time mutation and long-range context dependence; and the pyramid multiscale expansion convolution and the multiscale Fourier modulation attention are adaptively fused through a learnable dynamic gating mechanism, so that uniform time-frequency representation is finally obtained.
5. The method of claim 4, wherein the adaptive time-knowledge fusion loss function module integrates the resulting total loss function of the KWNet-based predictive model The method comprises the following steps: Wherein the method comprises the steps of And Is an adaptive weight coefficient that is used to determine, 。
6. The method according to claim 5, wherein the specific industrial process in the step 1 is an extraction process, and the historical parameters of the extraction process include the conditions of temperature in the extraction tank, pressure in the extraction tank, flow rate of extraction solvent, secondary heat-preserving steam pressure, secondary temperature rise, temperature of the storage tank and valve opening.
7. The method according to claim 6, wherein the expert knowledge defined for the extraction process in step 2 comprises: Expert knowledge rules one, extracting tank temperature predictions if the second steam valve changes from closed to open In minimum increments Monotonically increasing, otherwise, predicting violating expert knowledge rule one, punishing with additional knowledge loss; expert knowledge rule II, if the working step is changed from the end of the first liquid discharge to the second water addition, extracting the predicted value of the tank temperature Monotonically decreasing with a minimum decrease of Otherwise, the prediction violates expert knowledge rule two, punishing with additional knowledge loss.
8. The method of claim 7, wherein the expert knowledge loss function is derived based on the first expert knowledge rule and the second expert knowledge rule The method comprises the following steps: Wherein, the Wherein the method comprises the steps of Representing a predicted tank temperature of the tank, And The two are respectively used for predicting the tank temperature difference in front and back twice based on the expert knowledge rule I and the expert knowledge rule II; Representing a prediction time domain; And For the indication function of the trigger conditions of the expert knowledge rule I and the expert knowledge rule II, The total number of the first expert knowledge rule and the second expert knowledge rule is violated in the batch; in order to calculate the number of batches, For predicting the length of time of the sequence.
9. The method of claim 8, wherein the time domain loss function The method comprises the following steps: Wherein the method comprises the steps of , And N represents the tank temperature true value, the tank temperature predicted value, and the sample amount, respectively.

Description

Variable working condition industrial temperature prediction method based on KWNet model Technical Field The invention relates to a variable working condition industrial temperature prediction method based on KWNet model, belonging to the technical field of industrial process parameter prediction. Background The accurate prediction of the industrial process parameters is crucial to the monitoring, control and optimization of the industrial process, especially the prediction (i.e. multi-step prediction) is performed for the parameter changes of a plurality of time steps in the future, however, in practical application, due to the complex coupling relation among the multiple variables under different working conditions, the prediction model is difficult to accurately capture the dynamic turning characteristics and the transient fluctuation rules generated by the complex coupling relation, so that the prediction accuracy is affected. For example, the above-mentioned problems are presented in the multi-step prediction of the internal temperature of extraction tanks in industrial extraction processes, an important separation technique widely used in the fields of pharmaceutical manufacturing, food processing, chemical industry, etc., for separating active ingredients from natural or synthetic sources. In this process, the extraction tank internal temperature is a critical parameter controlling the extraction kinetics and reliability, but its temperature is difficult to measure due to the encapsulation of high viscosity solutions and the complex sealed environment. Therefore, at present, the industrial temperature prediction mostly adopts a soft measurement technology, and the deep learning method is very important for the soft measurement technology, so that the finding of a suitable deep learning method model has great significance. The traditional deep learning method for temperature parameter prediction comprises a Convolutional Neural Network (CNN), a long and short term memory network (LSTM), a gate control circulation unit (GRU) and a Transformer method, wherein the CNN can slide on a time sequence by utilizing a one-dimensional convolutional kernel to obtain detailed information of each part in the time sequence, but the detailed information cannot be captured by the CNN, the time information is lacking, the LSTM and the GRU can store the time information to reflect the sequence of data, but due to the sequence characteristics, the great complexity exists when the long-distance time sequence is captured, the calculation and the expansibility are limited to a certain extent, the Transformer is used as a parallel processing framework, more effective long-distance dependent modeling and the expansibility on a large data set are realized, but the Transformer still faces challenges when the complex time sequence mode is captured, the self-attention mechanism of the Transformer possibly suffers information loss when the complicated time dependent relation is processed, the point-by-point structure cannot always keep consistent distribution characteristics on the long sequence, and the inaccuracy of a prediction result can be caused. In recent studies, many new predictive models have emerged. If Autoformer integrates a sequence decomposition module for separating a time sequence into seasonal and trend periodic components, but when the seasonal and trend periodic components are constructed, a trend item is obtained by adopting a fixed sliding window to average, multi-frequency band information cannot be distinguished, high frequency, medium frequency and noise are mixed together due to simple subtraction, a seasonal item is compressed, which frequency band and which period are more important in the future cannot be judged according to the time length, so that the model performance is reduced, the size of the sliding average window is also super-parameter and is difficult to obtain, DLinear proves that the distribution offset can be reduced by adopting simple linear decomposition combined with explicit trend-residual part separation through normalization and sliding window averaging, so that the method is superior to a complex system structure, but the fixed sliding window size selection and the static decomposition mode cannot adapt to different kinds of data. These methods verify the feasibility of the decomposition concept in industry prediction problems. The Physical Information Neural Network (PINNs) encodes physical information as soft constraints during training, which method shows particular effectiveness when run under conditions of sparse data or uncertain measurements. However, existing PINNs rely on accurate characterization of physical principles or control equations, which inherently limit their applicability to strongly nonlinear, coupled and multivariable real world industrial processes, so there is room for further improvement in the prediction accuracy of existing predictive models. Disclosure o