Multivariable time sequence prediction method based on channel grouping and prediction correction

Abstract

The invention discloses a multivariable time sequence prediction method based on channel grouping and prediction correction, and the specific scheme comprises the steps: (1) multivariable grouping modeling based on time-frequency joint feature extraction: constructing an optimal clustering mode related to a channel part, and providing an efficient and reliable variable grouping structure; (2) training a model fused with periodic constraints: providing a model loss function fused with the periodic constraints, accurately estimating the length of a data dominant period, and quantifying the numerical value difference of each data segment on a period dimension; and (3) prediction correction based on time sequence structure consistency guidance: calculating an adjustment factor for data correction by quantifying correlation differences among a historical data segment, a current data segment and a prediction data segment, and realizing adaptive correction of model output. According to the method, high-precision prediction of the multivariable time sequence data is realized, an innovative technical scheme is provided for the problem of multivariable time sequence prediction in a process industry system, and the method has important practical value.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of multivariate time series prediction, and particularly relates to a method for multivariate time series prediction based on channel grouping and prediction correction. BACKGROUND

[0002] Multivariate time series in process industry systems, as one of the information carriers of core processes in production lines, reflect the state information of numerous process links and device units in industrial production processes, and provide an important basis for production control and intelligent decision-making. Therefore, accurately predicting multivariate time series data in the future has important industrial application value for early identification of potential risks, dynamic optimization of production processes, and efficient allocation of resources.

[0003] However, modern process industry systems often exhibit characteristics such as multi-region distribution, multi-level interaction, large inertia response, and strong time delay transmission, and industrial production processes are often affected by various periodic factors such as production rhythm, equipment operation cycle, raw fuel supply cycle, and technological innovation cycle, which makes the multivariate time series data have significant periodicity and similar or symmetric time series change patterns, i.e., the values of these variables have similar rising or falling evolution trends, which increases the difficulty of online monitoring of operating states and online modeling of product quality based on data-driven. Therefore, effectively constructing the coupling relationship between multivariate in process industry systems, mining the periodic patterns of multivariate and the structured association between multivariate, and accurately predicting the time series data of multivariate, are key prerequisites for subsequent process modeling, optimization control, and performance evaluation.

[0004] Existing multivariate time series prediction methods can be divided into statistical learning-based methods, signal processing-based methods, machine learning-based methods, and deep learning-based methods. Statistical learning-based methods are based on statistical theory, such as autoregressive moving average models and vector autoregressive models, which achieve prediction by mining the linear relationship between variables, but have limited modeling capability for non-linear data sets. Signal processing-based methods analyze sequence features from the frequency or time domain, such as Fourier transform and wavelet transform, which decompose complex time series data into simpler sequences to better understand their internal structure and effectively handle periodic and trend signals, but have difficulty handling non-stationary data sets. Machine learning-based methods, such as support vector machines, decision trees, and random forests, can mine non-linear relationships in data and achieve prediction through feature engineering and model training, and perform well in small sample scenarios, but have difficulty capturing long-range dependencies in time series data. Deep learning-based methods use neural network architectures to extract high-order features layer by layer, such as convolutional neural networks, long short-term memory networks, and autoencoder models, which have significant advantages in capturing long-range dependencies of time series and complex coupling relationships between multivariate, and are particularly suitable for large-scale data scenarios such as process industry systems.

[0005] Through the comprehensive comparison of the advantages and limitations of the above various time series prediction methods, it can be seen that the method based on deep learning has the most extensive application range and the best model effect. However, the current deep learning method still has several difficulties when applied to multivariate time series prediction. First, the existing methods mainly adopt two modeling strategies of complete independence and complete dependence of channels, respectively assuming that the variables are completely irrelevant or completely relevant, making it difficult for the model to accurately represent the complex variable coupling relationship of process industrial systems. At the same time, the existing methods mainly focus on establishing a direct mapping relationship from the historical data segment and the current data segment to the predicted data segment, and generally ignore the periodic correlation between the historical data segment, the current data segment and the predicted data segment. For example, patent application CN117076885A discloses a process parameter single variable time series prediction method and system based on number knowledge transfer, which performs time series prediction under non-stable working conditions through a pre-trained knowledge embedded DKN network, solving the problems of generalization performance degradation and slow network convergence of traditional single variable time series prediction model; However, this technology does not fully consider the periodicity of variables, ignores the important role of periodic patterns in time series prediction, and makes the accuracy of the prediction result not high. In addition, the existing time series prediction methods generally ignore the possible time series structured association of multivariate in the time series evolution process, that is, there are similar or symmetric time series value change patterns between some variables, such as the approximate upward or downward evolution trend of the values of these variables. For example, patent application CN118228875A discloses an industrial multi-scale time series data dynamic enhancement and prediction method based on attention mechanism, which generates unlabeled pseudo samples and their corresponding labeled pseudo sample data through time series distribution measurement double generative adversarial network, and trains a process industrial time series prediction model to predict process industrial quality index data; However, this technology does not fully exploit the coupling relationship between multivariate and the similar or symmetric time series change patterns between variables, resulting in low accuracy of the time series prediction result. Therefore, ignoring the above characteristics of multivariate time series data in process industrial systems will affect the accuracy of the model prediction result.

[0006] Therefore, in view of the above difficulties and problems existing in the existing methods, it is necessary to develop a new multivariate time series prediction method. SUMMARY

[0007] In order to solve the above-mentioned difficulties and problems existing in the prior art, the present application innovatively proposes a multivariate time series prediction method based on channel grouping and prediction correction. The present application studies a multivariate grouping modeling strategy based on time-frequency joint feature extraction, designs a weight distribution mechanism based on adaptive information entropy, innovatively proposes a model loss function fused with periodicity constraint, evaluates the correlation between the historical data segment, the current data segment and the predicted data segment in the periodic dimension, proposes a prediction correction strategy guided by time series structure consistency, calculates the adjustment factor for data correction, corrects the prediction output of the model, overcomes the difficulty that the periodicity influence of variables and the time series structured correlation between data segments are not considered in time series prediction, and realizes accurate prediction of multivariate time series data.

[0008] The present application aims to propose a multivariate time series prediction method based on channel grouping and prediction correction. In view of the difficulty in analyzing the multivariate coupling relationship caused by the complex characteristics such as multi-region distribution and large inertia response of process industrial systems, the present application designs a multivariate grouping modeling strategy based on time-frequency joint feature extraction, synchronously captures the correlation characteristics of variables in time domain and frequency domain, and proposes a weight distribution mechanism based on adaptive information entropy, to realize the optimal grouping of industrial multivariables with high cohesion within the group and low coupling between groups. In view of the difficulty in modeling the periodic pattern of variables caused by multiple periodic factors such as production rhythm and raw fuel supply in process industry, the present application proposes a model loss function fused with periodicity constraint, accurately estimates the dominant period length of data, and quantifies the numerical difference of each data segment in the periodic dimension. In order to establish the time series structured correlation between multivariables within the group in the time series evolution process, the present application proposes a prediction correction strategy guided by time series structure consistency, quantifies the correlation difference between each data segment, calculates the adjustment factor for data correction, and then adaptively corrects the predicted data. In summary, the method proposed by the present application can accurately predict the future time series data of multivariables in process industry, has the advantage of high accuracy, and provides an innovative idea for multivariate time series prediction in process industrial systems.

[0009] The present application proposes a multivariate time series prediction method based on channel grouping and prediction correction.

[0010] Figure 1 is the implementation flowchart of the method proposed by the present application, including the following steps:

[0011] (1) a multivariate grouping modeling strategy based on time-frequency joint feature extraction is proposed, the correlation characteristics of variables are calculated in time domain, frequency domain and time-frequency domain, and a weight distribution mechanism based on adaptive information entropy is designed, to build an optimal clustering mode of variables for prediction tasks, and provide an efficient and reliable variable grouping structure for multivariate time series modeling and prediction of complex industrial systems;

[0012] (2) We propose a model loss function that incorporates periodic constraints, design an adaptive periodic constraint term, accurately estimate the length of the data-dominant period, and evaluate the numerical differences between each data segment in the periodic dimension, thereby effectively maintaining the periodic pattern of the data while improving the prediction accuracy.

[0013] (3) A prediction correction strategy based on the consistency of the temporal structure was designed to quantify the correlation differences between the historical data segments, the current data segments, and the predicted data segments. The adjustment factors for data correction were calculated, and the model output was adaptively corrected to improve the consistency of the temporal structure between the variables within the group and the accuracy of the prediction results.

[0014] The present invention assumes that the multivariate time series data set of the process industry is ,Include Variable types and In order to meet the needs of subsequent data modeling, the present invention first constructs a data set for model training, and converts the multivariate time series data set into Divide into historical data segments , current data segment and the true value of the predicted data segment Among them, the lengths of the three data segments are 、 and , and the three are closely connected in the time dimension without time interval. The task of this invention is to and the current data segment Modeling is performed to predict the future length Prediction data segment and with the real By comparing the models, the model parameters are updated through the loss function to train a multivariate time series prediction model with strong prediction capabilities. The specific implementation plan is as follows:

[0015] (1) Multivariate grouping modeling based on time-frequency joint feature extraction

[0016] The present invention proposes a multivariate grouping modeling strategy based on time-frequency joint feature extraction, extracting the frequency domain features of the variables through short-time Fourier transform in the frequency domain dimension, and then calculating the Pearson correlation coefficient matrix of the multivariate in the time domain dimension, frequency domain dimension and time-frequency domain dimension. At the same time, the invention also designs a weight distribution mechanism based on adaptive information entropy, and implements variable grouping based on spectral clustering, constructing an optimal clustering pattern of channel partial correlation with "high cohesion within the group and low coupling between groups", providing an efficient and reliable variable grouping structure for multivariate time series modeling and prediction of complex industrial systems. Specifically, it includes the following steps:

[0017] Step1: In order to extract the frequency domain features of the time series variables, the present application performs short-time Fourier transform on the original time series data to extract the energy distribution characteristics of each variable in different frequency bands. Short-time Fourier transform can map the signal to the frequency domain while preserving the local information of time, thereby revealing the frequency domain variation law of the variable, and its calculation formula is:

[0018]

[0019] wherein, is the original signal, wherein is the integral variable, representing each time point in the entire time domain, is the short-time Fourier transform function, the subscript indicates that is transformed, is the time variable, the input signal is , the windowed signal is , is the frequency variable, is the imaginary unit, is the circular constant. By the above method, the frequency domain features of the historical data segment and the current data segment are extracted, and the frequency domain data of the historical data segment and the current data segment are obtained respectively and , is the dimension of the frequency domain feature.

[0020] Step2: In order to calculate the global correlation of each variable in - , - , - and - , Pearson correlation coefficient is used for calculation and analysis.

[0021] Pearson correlation coefficient is a statistical index for measuring the linear correlation degree of two variables, and its value is between-1 and +1. When applied to a multivariate data set, the relationship between these variables can be analyzed by calculating the Pearson correlation coefficient between each pair of variables. For the sample data of two variables and , the Pearson correlation coefficient can be calculated by the following formula:

[0022]

[0023] wherein, and The first and second rows represent the first and second variables, respectively The first and second columns represent the first and second observations, respectively The first and second columns represent the first and second observations, respectively The first and second columns represent the first and second observations, respectively The first and second columns represent the first and second observations, respectively

[0024] In process industries, multivariate time series data are ubiquitous, and there are often complex correlations between parameters such as temperature, pressure, flow, composition, and speed. Pearson correlation coefficient can be used to quantify the degree of linear correlation between these variables, and to identify the synergistic relationship between key variables.

[0025] Therefore, the present application calculates the global correlation of each variable in - 、 - 、 - and - to obtain correlation matrices 、 、 and , respectively, which provides a data basis for subsequent weight allocation based on adaptive information entropy.

[0026] Step 3: Next, based on the correlation matrices 、 、 and , a comprehensive and normalized adaptive information entropy index is calculated in turn to measure the information complexity of the entire matrix, and the positive and negative correlations are weighted by the weight allocation mechanism, and finally an entropy value between 0 and 1 is output.

[0027] First, copy the original matrix (set as ) and clear the diagonal elements. Then, separate the positive and negative value data to construct two sub-matrices, only keep the positive values, i.e. the positive correlation part, the part of the negative value data after taking the absolute value, i.e. the negative correlation part. Next, normalize the positive and negative correlation matrices, and calculate the entropy values of the positive and negative correlations, respectively. The calculation formula is:

[0028]

[0029] wherein is the normalized correlation strength value, and the positive correlation entropy and the negative correlation entropy can be obtained.

[0030] Then, the elements in the matrix are summed after taking the absolute value, and the proportion of positive and negative correlation in the entire correlation matrix is calculated, and it is used as a weight:

[0031]

[0032]

[0033] Thus, the weighted average entropy is calculated:

[0034]

[0035] This is a typical adaptive weight mechanism, which makes the strong correlation (whether positive or negative) have a greater impact on the overall entropy. In order to normalize the final entropy value, so that it falls within the [0, 1] interval, the maximum possible entropy needs to be calculated, which is the maximum uncertainty when all non-diagonal elements are equally distributed, and the calculation formula is:

[0036]

[0037] where, is the number of variables.

[0038] Finally, the matrix is obtained

[0039]

[0040] The normalized adaptive information entropy of the correlation matrix , , and is calculated in turn, and , , and are obtained, and the sum of the entropy values is obtained:

[0041]

[0042] Again, normalization is performed, and the final information entropy is obtained:

[0043]

[0044] Therefore, the correlation matrix after weighting ( ) is:

[0045]

[0046] Step4: Next, we need to determine the optimal number of groups , so that the variable groups have good internal consistency and external separation, that is, "high cohesion within the group, low coupling between groups", which is convenient for subsequent modeling, dimensionality reduction or process analysis. Use Calinski-Harabasz ( ) evaluation index to search under different values, and finally choose the value that makes the largest as the optimal clustering number , The calculation formula is:

[0047]

[0048] Where, is the total number of variables, is the intra-class dispersion matrix, is the inter-class dispersion matrix, denotes the trace of the matrix. In cluster analysis, the intra-class dispersion matrix and the inter-class dispersion matrix are important indicators for evaluating the quality of clustering, which are used to measure the separation degree between different clusters and the tightness of each cluster.

[0049] The intra-class dispersion matrix measures the sum of the distances from each data point in a cluster to the cluster center, reflecting the dispersion of data within each cluster. Let be the th sample point, be the mean (center point) of the th cluster, which can be calculated by:

[0050]

[0051] Where, denotes the number of samples in the th cluster, denotes the set of all samples in the th cluster. If belongs to the th cluster, the intra-class dispersion matrix when the clustering number is can be calculated as follows:

[0052]

[0053] The inter-cluster deviation matrix measures the position difference of each cluster center relative to the sample center of the whole population, and reflects the separation degree between clusters. is the average value of all sample points:

[0054]

[0055] wherein, is the total number of samples. The number of clusters is The inter-cluster deviation matrix when the number of clusters is can be calculated by the following formula:

[0056]

[0057] After obtaining the optimal number of groups , the original correlation matrix is converted into a similarity measure matrix suitable for clustering , wherein is calculated in the following manner:

[0058]

[0059] Finally, variable grouping is performed based on spectral clustering. A graph Laplacian matrix is constructed, which captures the similarity and structure information between data points, thereby reflecting the inherent structure of the data, is a diagonal matrix, and the calculation formula is:

[0060]

[0061] In spectral clustering, by calculating the eigenvalues and eigenvectors of the graph Laplacian matrix, a low-dimensional representation of the data can be found, which preserves the main structural features of the data. Among them, the eigenvector corresponding to the largest eigenvalue can reveal the most significant pattern or structure in the data. Therefore, the first eigenvectors of the graph Laplacian matrix are clustered, and according to the clustering results, the variable index is assigned to different groups, thereby obtaining the optimal variable clustering pattern with “high cohesion within the group and low coupling between the groups”, which provides an efficient and reliable variable grouping structure for subsequent model training and parameter updating.

[0062] (2) Model training with periodic constraint fusion

[0063] In process industries, due to the superposition of various periodic factors such as production rhythm, raw material and fuel supply, it is difficult to effectively model the periodic characteristics of variables, such as Figure 2The time series data of the process industry shown in the dashed box in (a) presents significant and complex periodic patterns. The present application proposes a model loss function that fuses periodic constraints, accurately estimates the dominant period length of the data, and quantifies the numerical differences of each data segment in the periodic dimension. Specifically, the following steps are included:

[0064] Step 1: Model loss function (L) is composed of two parts, namely the Mean Squared Error (MSE) loss function and the periodic constraint term:

[0065]

[0066] wherein, and are the predicted data segment and the true value of the predicted data segment of a certain variable in and respectively, is the length of the predicted data segment, is the period of the variable, and are the MSE loss function and the periodic constraint function respectively, is the weight of the periodic constraint term. Concatenate the historical data segment and the current data segment to obtain the concatenated data segment , wherein, , is the time series data of a certain variable in .

[0067] The MSE loss function measures the prediction performance of the model by calculating the average of the squared differences between the predicted value and the true value. The MSE loss function encourages the model to minimize the difference between the predicted value and the true value as much as possible, thereby improving the accuracy of the model. For a given set of predicted values and corresponding true values , the calculation formula of the MSE loss function is as follows:

[0068]

[0069] wherein, and represent the predicted value and the true value of the th sample respectively.

[0070] To enhance the model's perception of the periodic patterns of time series variables, the present application designs a periodic constraint mechanism and incorporates it into the loss function to achieve reasonable adjustment of the model optimization direction.

[0071] Step2: Next, the autocorrelation function estimation method and the fast Fourier transform method will be used to calculate the period in the periodic constraint term respectively. ).

[0072] First, the data in is standardized using as the data source:

[0073]

[0074] where and are the mean and standard deviation of , respectively, resulting in . Then, the autocorrelation function is used to estimate the period, and the autocorrelation coefficient of each lag step is calculated:

[0075]

[0076] where and are the th and th samples in , respectively, and then normalized:

[0077]

[0078] where represents the sequence completely aligned with itself (no lag). The lag step corresponding to the maximum peak is chosen as the autocorrelation period ( ):

[0079]

[0080] Step3: Calculate the period size using the fast Fourier transform. Convert the time domain signal to the frequency domain, and find the main period by analyzing the frequency components ( ). The data samples in are decentered:

[0081]

[0082] resulting in , and the period is calculated:

[0083]

[0084] where, and are preset minimum and maximum values of the period, denotes the fast Fourier transform, is the mean value of the sequence , denotes the period corresponding to the maximum power, denotes the frequency. This method finds the strongest frequency component by mapping the time series to the frequency domain, converts it to the period length, and thus identifies the main periodic structure in the data.

[0085] Step4: fuse the calculation results of the above two methods, and . First, judge the deviation of the two:

[0086]

[0087] where, denotes the absolute value. If the deviation is less than or equal to 20%, take the average as the final period:

[0088]

[0089] where, denotes the rounding operation. If the deviation is greater than 20%, calculate the Pearson correlation coefficient of the two pieces of data in the data set with a time interval of one period under the two periods respectively, and get and , then:

[0090]

[0091] Step5: Finally, calculate the periodicity constraint term, which measures the periodicity of the predicted sequence ( ) and the spliced data segment ( ) under the given period, that is, the values separated by one period should be approximately. For each time step in , calculate the difference between the current time and the predicted value lagged by one period:

[0092]

[0093] where, denotes the th sample in the predicted sequence , denotes the th sample in the predicted sequence two samples, the time interval of the two samples is .

[0094] The calculated and are combined as a loss function during model training, and the weight of the periodic constraint is .

[0095] (3) Prediction correction based on time sequence structure consistency guidance

[0096] The process industry system has the characteristics of multi-region distribution, multi-level interaction, large inertia response and strong time delay transmission, so that the multivariate may show the internal time sequence structured correlation in the time sequence evolution process, that is, some variables show similar or symmetric time sequence change patterns, such as the values of these variables have approximate rising or falling evolution trends. Figure 2 The time sequence value change shown by the solid line box in each subgraph intuitively presents the similar or symmetric numerical value change trend of the four variables in the local subsystem after being disturbed by a certain specific external disturbance, indicating that these variables have similar time sequence evolution characteristics, so in-depth analysis of the time sequence structure consistency between variables is of great significance to improve the accuracy of time series prediction.

[0097] Therefore, the present application designs a prediction correction strategy based on time sequence structure consistency guidance, calculates the adjustment factor for data correction by quantifying the correlation difference between the historical data segment, the current data segment and the predicted data segment, and realizes adaptive correction of the model output. Specifically, the following steps are included:

[0098] Step 1: First, the prediction model of the present application adopts the Transformer architecture, the input is the concatenated data segment , and the output is the predicted data segment with a length of .

[0099] According to the historical data segment , the current data segment and the true value and the predicted value of the predicted data segment, let , , and be the data samples belonging to the same variable in , , and .

[0100] Step 2: Then, calculate four groups of Pearson correlation coefficients, including and 、 With 、 With 、 With , respectively 、 、 and .

[0101] Next, the correlation difference is calculated:

[0102]

[0103] where and represent the correlation difference between the predicted data segment and the true value of the predicted data segment, respectively.

[0104] Step 3: Use the correlation difference and as the data source to input into a three-layer fully connected feedforward neural network for training, outputting an adjustment factor between 0 and 1. The first layer of the network maps the input to a 64-dimensional space and introduces nonlinearity through the ReLU activation function; the second layer further compresses the features to 32 dimensions and again uses the ReLU activation; the last layer outputs a scalar value and normalizes it to the [0, 1] interval through the Sigmoid function as the final adjustment factor .

[0105] The adjustment factor is a weight coefficient that controls to what extent the model's original prediction should be "corrected" to a prediction that better fits the true target distribution. When is close to 1, it means that the model's prediction deviates significantly from the true trend and needs strong correction; when is close to 0, it tends to preserve the original prediction.

[0106] Step 4: Use the statistical characteristics of the true value of the predicted data segment to correct the model output . Calculate the mean and standard deviation of and , respectively:

[0107]

[0108]

[0109] where, and are the mean and standard deviation of and are the mean and standard deviation of and are the mean and standard deviation of and are the mean and standard deviation of

[0110] Finally, the correction value is calculated by the following formula :

[0111]

[0112] wherein, is a very small positive number to prevent division by zero and ensure numerical stability. Thus, the original prediction value is weighted and fused with the correction value :

[0113]

[0114] is the final prediction value at the th sample. The purpose of this is to make the model prediction value not only close to the true value in value, but also as consistent as possible in distribution characteristics, thereby improving the accuracy of the prediction result.

[0115] In summary, through the above steps, the model is trained and optimized, and when the preset performance threshold condition of the model is met, a trained multivariate time series prediction model for process industry can be obtained, and the time series data in the future period of time can be accurately predicted.

[0116] Compared with the prior art, the present application has the following beneficial effects:

[0117] (1) To cope with the multi-variable coupling relationship analysis challenge brought by the complex characteristics such as multi-region distribution and large inertia response in process industry systems, the present application proposes a multivariate grouping modeling strategy based on time-frequency joint feature extraction, which can simultaneously mine the correlation features between variables in time domain and frequency domain, and further introduces a weight distribution mechanism based on adaptive information entropy, to realize efficient optimization grouping of industrial multivariate systems, and achieve the effect of efficient grouping modeling of "high cohesion within the group and low coupling between groups".

[0118] ​​​​(2) To solve the problem of variable periodic pattern caused by the interweaving of production rhythm, raw fuel supply and other periodic factors in process industry, the application proposes a model loss function fused with periodic constraints. By introducing an adaptive periodic constraint term, the dominant period length of the data is estimated, and the numerical difference between each data segment in the periodic dimension is quantitatively analyzed, effectively maintaining the periodic characteristics of multivariate time series data, making the prediction result more consistent with the periodic change rule of industrial data; thereby improving the ability of the model to describe complex periodic characteristics, providing accurate optimization direction for model training and parameter updating, and improving the accuracy of the prediction result.

[0119] (3) At the same time, to construct the time series structured correlation contained in the time series evolution process of the variables in the group, the application proposes a prediction correction strategy based on time series structure consistency guidance. By quantifying the correlation difference between different data segments of historical data segment, current data segment and predicted data segment, the adjustment factor for data correction is calculated, and adaptive correction of the prediction result is realized. In summary, the application groups the multivariate data set in the process industry system, designs a model loss function fused with periodic constraints, and proposes a prediction correction mechanism based on time series structure consistency guidance, realizing high-precision prediction of multivariate time series data, providing an innovative technical solution for multivariate time series prediction in process industry systems, and having important practical value. BRIEF DESCRIPTION OF DRAWINGS

[0120] Figure 1 is a method flowchart.

[0121] Figure 2 is a time series data graph of process variables in the blast furnace ironmaking process.

[0122] Figure 3 is a data prediction scatter plot.

[0123] Figure 4 is a data prediction curve graph. DETAILED DESCRIPTION

[0124] The application will be described in detail below in combination with the drawings and specific embodiments. The present embodiment is based on the technical solution of the application, and gives a detailed implementation and specific operation process, but the protection scope of the application is not limited to the following examples.

[0125] Example 1

[0126] This embodiment is based on the 2650 blast furnace of a domestic steel plant. First, the process variable time series data set of the blast furnace ironmaking process is collected Figure 2The four kinds of time series data, air permeability index, resistance coefficient, cold air pressure and hot air pressure, are shown in (a), (b), (c) and (d) respectively. Figure 2 As shown in the dashed box in (a), the time series data has significant periodic characteristics. Figure 2 As shown in the solid box in each subgraph, the four kinds of data have similar or symmetric time series change patterns, that is, the values of these variables have approximate rising or falling evolution trends, for example, the resistance coefficient, cold air pressure and hot air pressure decrease synchronously, and the air permeability index increases synchronously.

[0127] In this embodiment, the target task is to predict the data of 96 time steps in the future. Figure 3 and Figure 4 The scatter plot and curve plot of the prediction results of 415 batches of data are shown in (a) and (b) respectively. In terms of the performance indicators of the model, the average mean square error is 0.006, the average absolute percentage error is 0.023, and the average fitting coefficient is 0.9998. The experimental results fully prove that the method of the present application performs well in the specific process industry multivariate time series prediction task and can meet the needs of the field work, providing scientific and reliable decision support for the field workers.

[0128] The above shows and describes the basic principles and main features of the present application and the advantages of the present application. It should be understood by those skilled in the art that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A multivariate time series prediction method based on channel grouping and prediction correction, characterized in that: The specific plan is as follows: (1) Multivariate grouping modeling based on time-frequency joint feature extraction: In the frequency domain, short-time Fourier transforms are used to extract the frequency domain characteristics of variables, and the Pearson correlation coefficient matrix of multiple variables in the time, frequency, and time-frequency domains is calculated. Based on an adaptive information entropy weight allocation mechanism and spectral clustering, variable grouping is achieved, and an optimal clustering pattern with high intra-group cohesion and low inter-group coupling for channel partial correlation is constructed. This provides an efficient and reliable variable grouping structure for multivariate time series modeling and prediction of complex industrial systems. (2) Model training with integrated periodic constraints: A model loss function that integrates periodic constraints is proposed to accurately estimate the length of the data-dominant period and quantify the numerical differences between each data segment in the period dimension. (3) Prediction correction guided by temporal structure consistency: By quantifying the correlation differences among historical data segments, current data segments, and predicted data segments, the adjustment factors used for data correction are calculated to achieve adaptive correction of the model output.

2. The method for multivariate time series prediction based on channel grouping and prediction correction according to claim 1, characterized in that: Before the multivariate grouping modeling based on time-frequency joint feature extraction, it is assumed that the multivariate time series data set of the process industry is ,Include Variable types and time-step samples; construct a dataset for model training, and convert the multivariate time series dataset Divide into historical data segments , current data segment and the true value of the predicted data segment , where the lengths of the three data segments are 、 and , and the three are closely connected in the time dimension without any time interval.

3. The method for multivariate time series prediction based on channel grouping and prediction correction according to claim 1, characterized in that: The multivariate grouping modeling based on time-frequency joint feature extraction (1) specifically includes the following steps: Step 1: Perform short-time Fourier transform on the original time series data to extract the energy distribution characteristics of each variable in different frequency bands. The calculation formula is: ； in, is the original signal, where is the integral variable, representing every time point in the entire time domain, is the short-time Fourier transform function, and the subscript represents Perform the transformation, is a time variable, and the input signal is , the windowed signal is , is the frequency variable, is an imaginary unit, is the ratio of pi; through the above method, extract the historical data segment and the current data segment The frequency domain characteristics of the historical data segment and the current data segment are obtained respectively. and , is the dimension of frequency domain features; Step 2: Calculate and analyze using the Pearson correlation coefficient - 、 - 、 - and - The global correlation of each variable in is obtained by the correlation matrix 、 、 and ; For two variables and Sample data, Pearson correlation coefficient It can be calculated by the following formula: ； in, and Represents the two variables observations, and represent the sample means of the two variables, is the sample size; Step 3: Based on the correlation matrix between variables 、 、 and , calculate a comprehensive, normalized adaptive information entropy index in turn, and weight the positive and negative correlations separately through the weight distribution mechanism, and finally output an entropy value between 0 and 1: Copy the original matrix And clear the diagonal elements to zero; separate the positive and negative data, and construct two sub-matrices. Only the positive values ​​are retained, i.e. the positive correlation part, The negative correlation part is the part after taking the absolute value of the negative data; normalize the positive and negative correlation matrices, and calculate the entropy of positive and negative correlation respectively. The calculation formula is: ； in, is the normalized correlation strength value, and the positive correlation entropy can be obtained and negative correlation entropy ; Take the absolute values ​​of the elements in the matrix and sum them up, then calculate the proportion of positive and negative correlation parts in the entire correlation matrix and use them as weights: ； ； The weighted average entropy is then calculated: ； Calculate the maximum possible entropy, that is, the maximum uncertainty when all off-diagonal elements are equally distributed, using the formula: ； in, is the number of variables; Get the matrix Normalized adaptive information entropy: ； Calculate the correlation matrix sequentially 、 、 and The normalized adaptive information entropy of 、 、 and , and the sum of the entropy values ​​is : ； Normalize again to get the final information entropy: ； Weighted correlation matrix for: ； Step 4: Determine the optimal number of groups , so that the variable grouping has good internal consistency and external separability; using Calinski-Harabasz Evaluation indicators vary Search under the value and finally select The largest The optimal number of clusters , The calculation formula is: ； in, is the total number of variables, is the intra-class dispersion matrix, is the between-class dispersion matrix, represents the trace of the matrix; Within-class dispersion matrix: Let It is Sample points, It is The mean center point of a cluster can be calculated by the following formula: ； in, Indicates the The number of samples in a cluster, Indicates the The set of all samples of a cluster; if Belong to clusters, then the number of clusters is The intra-class dispersion matrix It can be calculated by the following formula: ； Between-class dispersion matrix, let is the average value of all sample points: ； in, is the total number of samples; the number of clusters is The inter-class deviation matrix It can be calculated by the following formula: ； To obtain the optimal number of groups Then, the original correlation matrix Convert to a similarity measure matrix suitable for clustering ,set up , calculated as: ； Finally, variables are grouped based on spectral clustering; the graph Laplacian matrix is ​​constructed. , capturing the similarity and structural information between data points, thus reflecting the intrinsic structure of the data, Is a diagonal matrix, the calculation formula is: ； In spectral clustering, by calculating the eigenvalues ​​and eigenvectors of the graph Laplacian matrix, a low-dimensional representation of the data is found. This representation retains the main structural characteristics of the data, among which the eigenvector corresponding to the largest eigenvalue can reveal the most significant pattern or structure in the data.

4. The method for multivariate time series prediction based on channel grouping and prediction correction according to claim 1, characterized in that: The model training of (2) integrating periodic constraints specifically includes the following steps: Step 1: Model loss function It consists of two parts: the mean square error (MSE) loss function and the periodic constraint term: ； in, and , the two are and The predicted data segment of a variable in and the true value of the predicted data segment, To predict the length of the data segment, is the period of the variable, and are the MSE loss function and the periodic constraint function respectively, is the weight of the periodic constraint; splicing historical data segments and the current data segment , get the spliced ​​data segment ,in, , for Time series data of a variable in ; For a given set of predicted values and the corresponding true value , MSE loss function The calculation formula is as follows: ； in, and Respectively represent The predicted value and true value of the sample; Step 2: Calculate the period in the periodic constraint term by using the autocorrelation function estimation method and the fast Fourier transform method respectively. : by For data source, Normalize the data in: ； in, and They are The mean and standard deviation of ; Use the autocorrelation function to estimate the period and calculate the lag length for each step. The autocorrelation coefficient of : ； in, and They are Middle and samples, and then normalized: ； in, Indicates that the sequence is completely aligned with itself and has no lag; the lag step corresponding to the maximum peak is selected as the autocorrelation period : ； Step 3: Use fast Fourier transform to calculate the period size, convert the time domain signal to the frequency domain, and find the main period by analyzing the frequency components. ;right The data samples in are decentralized: ； get , and calculate the period: ； in, and are the preset minimum and maximum values ​​of the cycle, represents the fast Fourier transform, is a sequence The mean of Indicates the period corresponding to the maximum power extraction, Indicates frequency; Step 4: Combine the calculation results of the above two methods. and ; First, determine the deviation between the two: ； in,' 'Indicates taking the absolute value; If the deviation is less than or equal to 20%, the average value is taken as the final cycle: ； in,' 'Indicates rounding operation; If the deviation is greater than 20%, calculate the Pearson correlation coefficient of the two data segments with a time interval of one cycle in the data set under these two cycles respectively, and get and , then: ； Step 5: Calculate the periodic constraint term and measure the predicted sequence and concatenated data segments The periodic consistency under a given period, that is, the values ​​​​across one period should be similar; for At each time step in , calculate the difference between the current moment and the predicted value after one period: ； in, Represents the predicted sequence Middle samples, express Middle samples, and the time interval between these two samples is ; The calculated and Combined, as the loss function during model training, and the weight of the periodic constraint is .

5. The method for multivariate time series prediction based on channel grouping and prediction correction according to claim 1, characterized in that: The prediction correction guided by (3) temporal structure consistency specifically includes the following steps: Step 1: The prediction model uses the Transformer architecture, and the input is the spliced ​​data segment , the output is of length Prediction data segment ; According to historical data segment , current data segment And the true value of the predicted data segment and predicted value ,set up 、 、 and They are 、 、 and Data samples belonging to the same variable; Step 2: Calculate the Pearson correlation coefficients of four groups, including and 、 and 、 and 、 and , respectively 、 、 and ; Calculate the correlation difference: ； in, and Represents the difference in correlation between the actual value and the predicted value of the predicted data segment for the historical data segment and the current data segment respectively; Step 3: Difference in correlation and As the data source, a three-layer fully connected feedforward neural network is input for training, and an adjustment factor between 0 and 1 is output. ; Step 4: Use the true value of the predicted data segment The statistical characteristics of the model output Perform correction processing and calculate and The mean and standard deviation of : ； ； in, and They are The mean and standard deviation of and They are The mean and standard deviation of and They are and The samples; Finally, the correction value is calculated by the following formula : ； in, is a very small positive number; Using Adjustment Factors The original predicted value and correction value Perform weighted fusion: ； That is the The final predicted value at samples.

6. The method for multivariate time series prediction based on channel grouping and prediction correction according to claim 5, characterized in that: In the three-layer fully connected feedforward neural network in Step 3, the first layer of the network maps the input to a 64-dimensional space and introduces nonlinearity through the ReLU activation function; the second layer further compresses the features to 32 dimensions and uses ReLU activation again; the last layer outputs a scalar value and normalizes it to the [0,1] interval through the Sigmoid function as the final adjustment factor ;when When it is close to 1, it means that there is a large deviation between the model prediction and the real trend, and a strong correction is needed; when When it is close to 0, it is more inclined to keep the original prediction.

7. The method for multivariate time series prediction based on channel grouping and prediction correction according to claim 1, characterized in that: The model is trained and optimized. When the model's preset performance threshold conditions are met, a trained multivariate time series forecasting model for process industries is obtained, and the time series data for a period of time in the future is accurately predicted.

Images