Industrial time series data prediction method and device based on multistage feature selection
By employing multi-level feature selection and heterogeneous stacking integration models, the problems of lag and information loss in industrial time series data prediction are solved, improving prediction accuracy and system reliability. This approach is suitable for multi-source heterogeneous time series data in industrial environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BOYA TRIZ (TIANJIN) TECH CO LTD
- Filing Date
- 2026-03-02
- Publication Date
- 2026-05-19
AI Technical Summary
Existing industrial time series data prediction technologies suffer from insufficient time-dependent modeling, traditional single-step iterative methods tend to ignore lag effects, feature selection relies on a single statistical indicator leading to information loss, model homogenization is severe, it is difficult to extract long-term memory time series features and discrete structured features, and generalization ability is limited.
A multi-level feature selection method is adopted, and the problem of weak feature deletion is solved by gating attention mechanism. The heterogeneous stacking model improves the generalization ability, including rolling truncated mean denoising and smoothing, adaptive comprehensive correlation measurement, multi-level feature screening and heterogeneous stacking ensemble regression model.
It effectively alleviates the lag problem of traditional single-step prediction, retains highly correlated features and identifies weakly correlated features with potential synergistic effects, improves the model's ability to capture trend reversals and prediction accuracy, reduces information loss, and enhances the interpretability and reliability of the system.
Smart Images
Figure CN122065288A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent data management technology, specifically to a method and apparatus for predicting industrial time-series data based on multi-level feature selection. Background Technology
[0002] In modern industrial production, the fluctuations of key indicators exhibit complex characteristics such as nonlinearity, time-varying nature, and multi-source coupling. Accurate prediction of their time-series data is crucial for optimizing production, controlling costs, and supporting decision-making. With the development of the Industrial Internet, enterprises can acquire massive amounts of multi-source heterogeneous data, but existing prediction technologies still have significant shortcomings: First, time-dependent modeling is insufficient, and traditional "single-step iteration" methods easily ignore lag effects, making it difficult to predict sudden trend changes; second, feature selection often relies on single statistical indicators, which may eliminate weakly correlated features with potential complementarity, resulting in information loss; third, model homogenization is severe, making it difficult to simultaneously and effectively extract long-term memory time-series features and discrete structured features influenced by policies and processes, resulting in limited generalization ability. Summary of the Invention
[0003] To address the problem of prediction lag, this invention provides the following technical solution: an industrial time-series data prediction method based on multi-level feature selection, which solves the problem of false deletion of weak features through a gating attention mechanism and addresses the problem of weak model generalization ability through heterogeneous stacking, including:
[0004] Historical time series data and multi-source influencing factor data of industrial time series data are obtained. The time series data is then subjected to denoising and smoothing processing based on the rolling truncated mean. The specific formula for the rolling truncated mean denoising and smoothing processing is as follows:
[0005] ;
[0006] in, Indicates the current time point, The length of the historical observation window. For the percentage of trimming, This indicates the floor function. Indicates within the time window Sort by numerical value The observation values at each time point;
[0007] Construct a shift prediction dataset and establish a supervised mapping relationship between historical observation window data and future target time data;
[0008] The adaptive comprehensive relevance measurement module calculates the comprehensive relevance between each feature and the target value. All the aforementioned features are set as a feature set, and the overall correlation is determined based on a preset threshold. The comparison results divide the feature set into strongly correlated subsets, moderately correlated subsets, and weakly correlated subsets;
[0009] The strongly correlated subset, the moderately correlated subset, and the weakly correlated subset are subjected to multi-level screening, including at least stratification based on comprehensive correlation, screening of strongly correlated and moderately correlated features, and supplementary screening of weakly correlated features.
[0010] Based on the final feature subset after screening, a heterogeneous stacked ensemble regression model is constructed and trained to output the predicted values of industrial time series data.
[0011] As a preferred embodiment of the present invention, the length of the historical observation window is set. Predicted displacement step size ,Establish Time and before Data on the length of a historical observation window and the future The supervised mapping relationship between the target value labels at different times; the construction of the shift prediction dataset is defined by the mapping function. Its specific mathematical expression is as follows:
[0012] ;
[0013] ;
[0014] in, For feature dimension, Indicates the first Each feature at time The value, To predict the shift step size; by adjusting The model explicitly learns from historical windows. To future lagging targets Nonlinear mapping across time steps.
[0015] As a preferred technical solution of the present invention, the calculation feature With target value absolute value of Pearson correlation coefficient ;
[0016] The absolute value of the Pearson correlation coefficient The sequence and the normalized mutual information Normalize the sequence and Calculate the entropy value using the information entropy formula. and The details are as follows:
[0017] ;
[0018] ;
[0019] in, For the total number of features, For the first The observations at each time point; the dynamic weights are calculated based on the entropy values. and :
[0020] ;
[0021] ;
[0022] Calculate the overall relevance :
[0023] .
[0024] As a preferred technical solution of the present invention, the first-level screening of the strongly correlated subset is performed on the maximum correlation and minimum redundancy class to obtain the core feature set; the second-level screening based on complementary information measurement is performed on the moderately correlated subset to obtain the complementary feature set.
[0025] As a preferred technical solution of the present invention, a third-level feature supplementation screening based on gated attention units is performed on the weakly correlated subset, the interaction attention score between the weak features and the selected core feature set is calculated, and the weak features with scores higher than the activation threshold are included in the final feature subset.
[0026] As a preferred embodiment of the present invention, the weak feature vector to be evaluated is defined as... The selected core feature set vector is ;
[0027] Construct a single-layer perceptron as a gating network and calculate the gating coefficients. :
[0028] ;
[0029] in, The sigmoid activation function is used, and [,] represents the feature concatenation operation; weights and paranoia Obtained through joint optimization; interaction gain score calculated. :
[0030] ;
[0031] in, The feature transformation weight matrix, Represents the joint mutual information metric; If the preset positive activation threshold is... > If the weak feature has significant interaction value, it is determined to be retained.
[0032] As a preferred embodiment of the present invention, the final feature subset is input into a first-layer heterogeneous model group containing a deep temporal base learner and a structured base learner to generate a primary prediction vector, and then the primary prediction vector is input into a second-layer probabilistic meta-learner to output the final industrial time-series data prediction value; the first-layer heterogeneous model group includes:
[0033] The final feature subset is denoted as the input matrix. Construct the following prediction branch:
[0034] ;
[0035] ;
[0036] ;
[0037] in, This represents a one-dimensional convolution operation. This represents a prediction operation based on gradient boosting of decision trees;
[0038] The deep temporal basis learner uses a CNN-LSTM model and outputs vectors. It is used to extract temporal fluctuation features; the structured base learner adopts the CatBoost model, and the output vector is... This is used to extract categories and discrete structured features based on symmetric tree splitting; the primary prediction vector is represented as... .
[0039] As a preferred embodiment of the present invention, the second-layer probabilistic learner employs a Bayesian ridge regression model, wherein the weights of the Bayesian ridge regression model can be set. Follows a Gaussian prior distribution And by maximizing the logarithmic marginal likelihood function Automatic iterative optimization of hyperparameters and Output the final prediction result:
[0040] ;
[0041] ;
[0042] in, The input vector for the meta-learner. To conform to a mean of 0 and a variance of The noise term of the Gaussian distribution is used to predict the target value and provide the confidence interval of the prediction result.
[0043] As a preferred technical solution of the present invention, the data acquisition and preprocessing module is used to acquire multi-source heterogeneous data, perform rolling truncation mean smoothing, and reconstruct the time series into shift prediction samples.
[0044] The multi-level feature selection module is equipped with an entropy weight calculation unit, a mutual information-based redundancy removal unit, and a gated attention supplementary selection unit, which are used to output the optimal feature subset.
[0045] The heterogeneous stacked integrated prediction module includes parallel deep time series base model units, tree model integrated base model units, and Bayesian regression meta-model units.
[0046] Compared with existing technologies, this invention provides a method and apparatus for predicting industrial time-series data based on multi-level feature selection, which has the following beneficial effects: This method adopts a hierarchical and phased feature selection strategy, dividing features into strong, medium, and weakly correlated subsets through comprehensive correlation calculation, and employing differentiated screening algorithms for different subsets. This mechanism not only retains highly correlated features but also identifies and recalls weakly correlated features with potential synergistic effects, effectively avoiding the information loss problem that may be caused by traditional single-step screening.
[0047] Introducing a rolling truncated mean for denoising and smoothing effectively suppresses impulse noise and outlier interference commonly found in industrial data. By constructing a shift mapping relationship from "historical window to future target," the dependencies across time steps are explicitly modeled, significantly alleviating the lag problem of traditional single-step prediction in scenarios of abrupt changes in time series and improving the model's ability to capture trend reversals.
[0048] A heterogeneous base learner group, employing CNN-LSTM and CatBoost in parallel, extracts information from temporal fluctuation features and structured discrete features, respectively, and then uses Bayesian ridge regression as a meta-learner for probability fusion. This structure combines the fitting ability of deep learning to complex patterns with the advantages of tree models in handling categorical features. Furthermore, the Bayesian framework provides confidence intervals for the prediction results, enhancing the interpretability and reliability of the system.
[0049] The system is equipped with corresponding modules, including data preprocessing, multi-level feature filtering, and heterogeneous integrated prediction. The system structure is clear and highly scalable. Verification through real-world scenarios such as energy price prediction demonstrates that this method is significantly effective in reducing prediction errors and lag times, making it suitable for high-precision prediction of multi-source heterogeneous time-series data in industrial environments. Attached Figure Description
[0050] Figure 1This is a schematic diagram of the industrial time-series data prediction method of the present invention;
[0051] Figure 2 This is a flowchart illustrating the specific implementation of industrial time-series data prediction according to the present invention.
[0052] Figure 3 This is a logical framework diagram of the multi-level feature filtering and heterogeneous stacking integrated regression model based on gated attention of the present invention;
[0053] Figure 4 This is a functional block diagram and a schematic diagram of the electronic device hardware structure of the industrial time-series data prediction device of the present invention. Detailed Implementation
[0054] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0055] This embodiment relates to a method and apparatus for predicting industrial time-series data based on multi-level feature selection, which addresses problems such as noise interference, feature redundancy, and insufficient model generalization ability in time-series data prediction in industrial scenarios. This invention achieves high-precision and robust prediction of industrial time-series data by constructing shift prediction samples, designing a multi-level feature selection mechanism, and employing a heterogeneous stacked ensemble model.
[0056] like Figure 1 As shown, an industrial time-series data prediction method based on multi-level feature selection includes the following steps:
[0057] S1: Obtain historical time series data and multi-source influencing factor data of industrial time series data, and perform noise reduction and smoothing processing based on rolling truncated mean;
[0058] S2: Construct a shift prediction dataset and set the historical observation window length. and predicted shift step size ,Establish Time and before Data on the length of a historical observation window and the future The supervision mapping relationship between the target value labels at different times;
[0059] S3: Calculate the overall correlation between each feature and the target value using the adaptive comprehensive correlation measurement module. And based on preset thresholds and comprehensive relevance Compare the features; set all features as a feature set, and divide the feature set into strongly correlated subsets, moderately correlated subsets, and weakly correlated subsets based on the comparison results;
[0060] S4: Perform a first-level screening on the strongly correlated subsets using the maximum correlation and minimum redundancy class to obtain the core feature set; perform a second-level screening on the moderately correlated subsets based on complementary information metrics to obtain the complementary feature set;
[0061] S5: Perform third-level feature supplementation screening based on gated attention units on the weakly correlated subset, calculate the interaction attention score between the weak features and the selected core feature set, and include the weak features with scores higher than the activation threshold into the final feature subset.
[0062] S6: Construct a heterogeneous stacked ensemble regression model, input the final feature subset into the first layer of the heterogeneous model group containing a deep time series learner and a structured base learner to generate a primary prediction vector, and then input the primary prediction vector into the second layer of the probabilistic meta-learner to output the final industrial time series data prediction value.
[0063] Specifically, such as Figure 2 As shown, the original time series data is first obtained from a multi-source heterogeneous database. Considering the inconsistencies in industrial data dimensions and noise interference, the following processing is required:
[0064] Data normalization, in order to eliminate the influence of different feature units on the convergence speed of subsequent gradient descent, uses max-min normalization to map the data to the [0,1] interval. For the th... Maintaining characteristics Observations at time Its normalization formula is:
[0065] ;
[0066] in, and These represent the minimum and maximum values of the feature over the entire historical period, respectively. In this formula, a small constant is used to prevent the denominator from being zero.
[0067] To address spike noise caused by sensor malfunctions or extreme market sentiment, this embodiment abandons simple moving averages and employs a more robust rolling truncated mean. The historical observation window length is set to... The pruning percentage is (For example First, open the window. The data within is sorted by numerical value to obtain an ordered sequence:
[0068] ;
[0069] Calculate the mean after removing extreme values at the end. ;
[0070] ;
[0071] in, Indicates the current time point, Indicates within the time window Sort by numerical value The observation value at each moment, , This indicates a floor operation, which effectively smooths out non-Gaussian noise while preserving information about trend inflection points.
[0072] Specifically, a shift prediction dataset is constructed; traditional single-step prediction typically learns... This approach is prone to lag when data fluctuates drastically. This embodiment constructs a shift mapping to force the model to learn across time steps. Dependency relationships.
[0073] Define the feature dimension as The historical observation window length is The predicted displacement step size is Constructing the first Input matrix of training samples and scalar labels :
[0074] ;
[0075] In this formula, Let represent the input feature matrix of the i-th shift prediction sample; This represents the normalized observation value of the m-th feature at time t; This refers to the length of the historical observation window; This represents the time span for each sample; This represents the dimensional space of the matrix. The matrix is constructed by extracting heterogeneous data from multiple sources using a sliding window.
[0076] ;
[0077] In this formula, This represents the true value of the target prediction corresponding to the i-th sample; This indicates that at time 1 in the original target time series... The actual observed value; i is the sample index; This refers to the length of the historical observation window; To predict the shift step size (i.e., to predict the future lag step size) (data at each moment) ≥1; This represents the real number field. The model input is explicitly defined here. With output Crossing between Nonlinear mapping relationship at each time step.
[0078] Through the above construction, the model predicts "the future..." The numerical value of "moment", not "next moment".
[0079] Specifically, such as Figure 3 As shown, in order to take into account both linear and nonlinear correlations, this embodiment adopts an adaptive comprehensive correlation index.
[0080] Linear and nonlinear measures for any feature With the goal Calculate the absolute value of the Pearson correlation coefficient. and normalized mutual information Pearson coefficient The calculation method is as follows:
[0081] ;
[0082] in, This represents the absolute value of the Pearson correlation coefficient. The total number of samples used to calculate the correlation (corresponding to the length of the historical observation window in this case). ); and These represent the feature sequence and the target sequence within the window, respectively. The observation values at each time point; This represents the mean of the feature sequence within the window; This represents the mean of the target sequence within the window. This formula is used to measure the degree of linear correlation between a single feature and the target value.
[0083] Mutual Information Defined as:
[0084] ;
[0085] in, Representing characteristic variables With target variable Mutual information values between them; Representing characteristic variables Values And the target variable Values Joint probability distribution at time; and They represent the characteristic variables respectively. and target variable The marginal probability distribution; and These represent the specific values of the feature variable and the target variable after discretization, respectively.
[0086] To eliminate the influence of entropy magnitude, normalized mutual information is calculated. :
[0087] ;
[0088] in ;
[0089] The entropy weighting method dynamically assigns weights to construct the evaluation matrix, and uses information entropy to determine... and The weights. Specifically, assume there are One feature, for Sequence normalization yields Calculate its information entropy :
[0090] ;
[0091] Similarly, calculate ,
[0092] ;
[0093] Calculate the difference coefficient based on the entropy value. and This leads to the dynamic weights:
[0094] , ;
[0095] Finally, the overall relevance was calculated. :
[0096] ;
[0097] Based on preset thresholds and comprehensive relevance Perform numerical comparisons, and base the results on the feature set. The distribution of the feature set Divided into strongly correlated subsets Medium-related subsets and weakly correlated subsets .
[0098] Specifically, targeting The maximum correlation minimum redundancy algorithm is adopted, and its objective function is to maximize :
[0099] ;
[0100] in, The maximum correlation and minimum redundancy evaluation function; For the selected feature set These are candidate features to be evaluated; This represents the mutual information (correlation) between candidate features and the target variable. Representing candidate features and selected features Mutual information (redundancy) between them.
[0101] against The joint mutual information algorithm is adopted, which emphasizes the gain after feature combination. The selection criterion is to maximize the following formula:
[0102] ;
[0103] in, For joint mutual information score; This indicates the subset of features that have been selected into the core feature set.
[0104] After the above screening, a preliminary core feature set is obtained. .
[0105] Weak feature recall based on gated attention units; in order to mine Potential effective features are identified, and GAU cells are designed to compute weak features. With core feature set The interactive value.
[0106] Specifically, define The aggregation vector is The weak features to be evaluated are .
[0107] First, calculate the gating coefficient. This is used to determine whether the interaction is activated.
[0108] ;
[0109] in, core feature set Aggregate vectors (e.g., mean vectors); This represents a vector concatenation operation; ; Here is the weight matrix of the gated network; The bias vector of the gated network
[0110] Secondly, calculate the potential value after feature transformation. :
[0111] ;
[0112] in, This is the potential value vector after feature transformation; It is the hyperbolic tangent activation function; The feature transformation weight matrix; This is the feature transformation bias vector.
[0113] Finally, the interaction attention score for this weak feature is calculated. :
[0114] ;
[0115] in, Assign a score to the interaction's attention. This indicates the original relevance between weak features and the target; In this formula, represents a preset positive activation threshold (e.g., 0.05). If At the preset threshold This indicates that the weak feature is "activated" in a specific context and should be added to the final feature set. .
[0116] A heterogeneous stacked ensemble regression model is constructed. Specifically, a two-layer stacking model is constructed, with Level-0 extracting heterogeneous features and Level-1 performing probability fusion.
[0117] Among them, Level-0: Deep Temporal Basis Model (CNN-LSTM). The CNN layer is used to extract local short-term fluctuation features. For input X, the output of the k-th convolutional kernel is... for:
[0118] ;
[0119] in, Indicates the first Each convolutional kernel at time step The output feature map; To modify the activation function of the linear unit; For the first The weight matrix of each convolutional kernel; This represents a one-dimensional convolution operation; The time dimension size of the convolution kernel; For the first The bias scalar of each convolutional kernel; Indicates the input sequence at time step to A partial fragment.
[0120] Then it is connected to the max pooling layer:
[0121] ;
[0122] in, This represents the output after pooling; This is the pooling window size of the max pooling layer.
[0123] LSTM layers are used to extract long-term dependent features, and their internal state updates follow the following set of equations:
[0124] Forgotten Gate: ;
[0125] Input Gate: ;
[0126] Candidate cell status: ;
[0127] Cell status update: ;
[0128] Output gate: ;
[0129] Hidden layer output: ;
[0130] Final output predicted value .
[0131] in, These are the weight matrices for the forget gate, input gate, candidate state, and output gate, respectively. These are the corresponding bias vectors; This represents the Sigmoid activation function; Represents the hyperbolic tangent activation function; Indicates the hidden state in the previous time step. Input at the current time Vector concatenation; This refers to the final predicted output value of the LSTM layer. and This is used to output the weights and biases of the fully connected layer.
[0132] Level-0: The structured base model (Cat Boost) utilizes Cat Boost to handle discrete features (such as holidays and policy codes). It employs an ordered boosting strategy to reduce prediction bias and obtain the predicted values. .
[0133] Level-1: The Bayesian Ridge Regression meta-learner constructs meta-feature vectors from the base model output. Assuming the target value Follows a Gaussian distribution:
[0134] ;
[0135] In this formula This represents the regression coefficient vector of the Bayesian ridge regression model; Indicate its Gaussian prior distribution; It is a vector with zero mean; It is the identity matrix; To control the precision parameters of the prior distribution variance.
[0136] Introducing weights Prior distribution:
[0137] ;
[0138] in, Regression coefficients The Gaussian prior distribution; This represents a multivariate Gaussian distribution; in this formula... This represents the regression coefficient vector of the Bayesian ridge regression model; Indicates and Zero-mean vectors of the same dimension; To control the precision hyperparameter of the prior distribution variance; It is the identity matrix; Let represent the covariance matrix of the prior distribution.
[0139] The optimal hyperparameters are found by maximizing the logarithmic marginal likelihood function. and The final predicted mean for:
[0140] ;
[0141] in, This is the final predicted mean; Let be the covariance matrix of the posterior distribution; The meta-feature input matrix; and These are hyperparameters that are automatically optimized by maximizing the logarithmic marginal likelihood function.
[0142] in, This result is the final predicted value of the industrial time-series data output by the system.
[0143] like Figure 4 As shown, an industrial time-series data prediction device based on multi-level feature selection includes:
[0144] The data acquisition and preprocessing module is used to collect multi-source heterogeneous data, perform rolling truncation mean smoothing, and reconstruct the time series into shift prediction samples.
[0145] The multi-level feature selection module is equipped with an entropy weight calculation unit, a mutual information-based redundancy removal unit, and a gated attention supplementary selection unit, which are used to output the optimal feature subset.
[0146] The heterogeneous stacked integrated prediction module includes parallel deep time series base model units, tree model integrated base model units, and Bayesian regression meta-model units.
[0147] This embodiment constructs a heterogeneous integrated prediction framework that fuses temporal features and structured features. Specifically, the framework includes the following key modules:
[0148] Shift Window Data Construction and Robust Smoothing Module: Addressing the prevalent non-Gaussian noise in industrial sensor data, a rolling truncated mean algorithm is first employed to remove spikes. Based on this, a shift prediction paradigm is constructed. This module constructs the input historical event window. and a specific point in the future The direct mapping between them forces the model to learn across time gaps. The long-range nonlinear law.
[0149] A multi-level feature backtracking and recall module based on gated attention: At the end of the multi-level feature screening, this embodiment does not directly discard weakly relevant features, but designs a gated attention network. This gated attention network uses weak features as "queries" and selected core features as "context". Through joint training or pre-training, weak features are only re-evaluated and retained when they show significant information gain in the current "context". Significant information gain is the gated activation.
[0150] Specifically, the first level: based on the entropy weight method, the linear correlation coefficient and nonlinear mutual information are fused to adaptively divide the feature set into three subsets: strong correlation, moderate correlation and weak correlation.
[0151] Level 2: For strongly correlated features, the maximum correlation minimum redundancy algorithm is used, and for moderately correlated features, the joint mutual information algorithm is used to form a preliminary "core feature set".
[0152] Level 3: For weakly correlated features, calculate their interaction attention scores with core features to identify features with potential synergistic effects.
[0153] Heterogeneous stacked integrated regression prediction module:
[0154] Base learner layer (Level-0): Deep temporal networks (such as CNN-LSTM) and gradient boosting decision trees (such as CatBoost) are deployed in parallel. Deep temporal networks use convolutions and recurrent units to capture temporal fluctuations, while gradient boosting decision trees use tree structures to process discrete features.
[0155] Meta-learner layer (Level-1): Bayesian ridge regression is employed. By introducing a Gaussian prior distribution, the fusion weights are automatically adjusted based on the prediction variance of the base learners, giving the system extremely strong noise resistance.
[0156] Specific application scenarios and examples:
[0157] This embodiment uses the energy procurement price forecast of an energy-intensive industrial enterprise as an example to verify the practical application effect of the method in this embodiment.
[0158] 1. Data preparation and parameter setting
[0159] Data source: Historical transaction price time series of target industrial energy is collected as the prediction target, and multi-source heterogeneous data, including international market correlation index, industry supply and demand indicators, macroeconomic indicators and related meteorological data, are acquired as input features.
[0160] Parameter settings: Set the historical observation window according to the actual procurement cycle requirements. Each time step predicts the shift step size. The construction ratio parameter of the rolling truncation mean at each time step. Set to 0.05.
[0161] 2. Feature selection process execution
[0162] Layered processing: Based on the comprehensive correlation calculation results, the system automatically divides the historical price autoregressive characteristics into a "strongly correlated subset", the macro market fluctuation characteristics into a "mediumly correlated subset", and some micro-environment indicators into a "weakly correlated subset".
[0163] GAU Recall Mechanism: In the final stage of multi-level screening, the gated attention unit identifies a significant interaction gain relationship between some weak meteorological features (such as extreme weather indicators) and core logistics features. Although these features have low individual correlations, they have strong explanatory power for price fluctuations and are therefore successfully recalled and retained in the final feature set.
[0164] 3. Model Predictions and Results
[0165] The final feature set is then input into the heterogeneous stacked ensemble model.
[0166] Base model performance: The CNN-LSTM model branch effectively captures the long-term cyclical fluctuations of price series; the Cat-Boost model branch accurately responds to the step changes caused by changes in market rules.
[0167] Fusion Results and Validation: The Bayesian ridge regression at Level-1 dynamically adjusts the weights based on the output variance of the base model, outputting the final predicted value including the confidence interval. Compared with traditional single prediction models, the method of this invention significantly reduces the prediction lag in periods of sharp price fluctuations, and the mean absolute percentage error is significantly optimized, effectively improving the scientific nature of enterprise procurement decisions.
[0168] Performance verification: Compared with the traditional ARIMA model, the method of this invention shortens the prediction lag time of price change points from 3 days to 0 days, that is, it achieves complete advance prediction. The average absolute percentage error is reduced from 5.8% to 1.9%, which effectively assists enterprises in locking in procurement strategies at low prices.
[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting industrial time-series data based on multi-level feature selection, characterized in that, The steps of the method include: Historical time series data and multi-source influencing factor data of industrial time series data are obtained. The time series data is then subjected to denoising and smoothing processing based on the rolling truncated mean. The specific formula for the rolling truncated mean denoising and smoothing processing is as follows: ； in, Indicates the current time point, The length of the historical observation window. For the percentage of trimming, This indicates the floor function. Indicates within the time window Sort by numerical value The observation values at each time point; Construct a shift prediction dataset and establish a supervised mapping relationship between historical observation window length data and future target time data; The adaptive comprehensive relevance measurement module calculates the comprehensive relevance between each feature and the target value. All the aforementioned features are set as a feature set, and the overall correlation is determined based on a preset threshold. The comparison results divide the feature set into strongly correlated subsets, moderately correlated subsets, and weakly correlated subsets; The strongly correlated subset, the moderately correlated subset, and the weakly correlated subset are subjected to multi-level screening, including at least stratification based on comprehensive correlation, screening of strongly correlated and moderately correlated features, and supplementary screening of weakly correlated features. Based on the final feature subset after screening, a heterogeneous stacked ensemble regression model is constructed and trained to output the predicted values of industrial time series data.
2. The industrial time-series data prediction method based on multi-level feature selection according to claim 1, characterized in that: Set the length of the historical observation window Predicted displacement step size ,Establish Time and before Data on the length of a historical observation window and the future The supervised mapping relationship between the target value labels at different times; the construction of the shift prediction dataset is defined by the mapping function. Its specific mathematical expression is as follows: ； ； in, For feature dimension, Indicates the first Each feature at time The value, To predict the shift step size; by adjusting The model explicitly learns from historical windows. To future lagging targets Nonlinear mapping across time steps.
3. The industrial time-series data prediction method based on multi-level feature selection according to claim 1, characterized in that, The overall relevance The calculation process includes: Calculation features With target value absolute value of Pearson correlation coefficient and normalized mutual information ; The absolute value of the Pearson correlation coefficient The sequence and the normalized mutual information The probability distribution is obtained by normalizing the sequence. and Calculate the entropy value using the information entropy formula. and The details are as follows: ； ； in, For the total number of features, For the first The observations at each time point; the dynamic weights are calculated based on the entropy values. and : ； ； Calculate the overall relevance : 。 4. The industrial time-series data prediction method based on multi-level feature selection according to claim 3, characterized in that: The first-level screening of the strongly correlated subset is performed using the maximum correlation and minimum redundancy class to obtain the core feature set; the second-level screening based on complementary information metric is performed on the moderately correlated subset to obtain the complementary feature set.
5. The industrial time-series data prediction method based on multi-level feature selection according to claim 4, characterized in that: A third-level feature supplementation screening based on gated attention units is performed on the weakly correlated subset, the interaction attention score between the weak features and the selected core feature set is calculated, and the weak features with scores higher than the activation threshold are included in the final feature subset.
6. The industrial time-series data prediction method based on multi-level feature selection according to claim 5, characterized in that, The feature supplementation and filtering based on gated attention units includes the following calculation steps: Define the weak feature vector to be evaluated as: The selected core feature set vector is ; Construct a single-layer perceptron as a gating network and calculate the gating coefficients. : ； in, The sigmoid activation function is used, and [,] represents the feature concatenation operation; weights and paranoia Obtained through joint optimization; interaction gain score calculated. : ； in, The feature transformation weight matrix, This represents the joint mutual information metric; a preset positive activation threshold is set to... ,like > If the weak feature has significant interaction value, it is determined to be retained.
7. The industrial time-series data prediction method based on multi-level feature selection according to claim 6, characterized in that: The final feature subset is input into the first layer of a heterogeneous model group containing a deep temporal base learner and a structured base learner to generate a primary prediction vector. The primary prediction vector is then input into the second layer of a probabilistic meta-learner to output the final industrial time series data prediction value. The first layer of heterogeneous models includes: The final feature subset is denoted as the input matrix. Construct the following prediction branch: ； ； ； in, This represents a one-dimensional convolution operation. This represents a prediction operation based on gradient boosting of decision trees; The deep temporal basis learner uses a CNN-LSTM model and outputs vectors. It is used to extract temporal fluctuation features; the structured base learner adopts the CatBoost model, and the output vector is... This is used to extract categories and discrete structured features based on symmetric tree splitting; the primary prediction vector is represented as... .
8. The industrial time-series data prediction method based on multi-level feature selection according to claim 7, characterized in that: The second-layer probabilistic learner employs a Bayesian ridge regression model, the weights of which can be set. Follows a Gaussian prior distribution And by maximizing the logarithmic marginal likelihood function Automatic iterative optimization of hyperparameters and Output the final prediction result: ； ； in, The input vector for the meta-learner. To conform to a mean of 0 and a variance of The noise term of the Gaussian distribution is used to predict the target value and provide the confidence interval of the prediction result.
9. An industrial time-series data prediction device based on multi-level feature selection, characterized in that, include: The data acquisition and preprocessing module is used to collect multi-source heterogeneous data, perform rolling truncation mean smoothing, and reconstruct the time series into shift prediction samples. The multi-level feature selection module is equipped with an entropy weight calculation unit, a mutual information-based redundancy removal unit, and a gated attention supplementary selection unit, which are used to output the optimal feature subset. The heterogeneous stacked integrated prediction module includes parallel deep time series base model units, tree model integrated base model units, and Bayesian regression meta-model units.