A multi-feature based smelting product index prediction system and method

Abstract

This invention discloses a multi-feature-based system and method for predicting smelting product indicators, belonging to the field of industrial intelligence and time series prediction technology. The system includes a data preprocessing module, a multi-frequency feature construction module, a label historical feature extraction module, a feature fusion and standardization module, a model training and verification module, and a prediction output module. It simultaneously accesses time series data at multiple granularities (second-level, minute-level, and hour-level) to construct aggregated statistical features; mines historical time series information of prediction labels, extracting historical label values ​​and corresponding statistical features before the prediction time; and trains the model using a time series segmentation method without data leakage after fusing time features and lag features. This invention can fully integrate multi-source heterogeneous time series data information, strengthen the learning of the label's own time series dependencies, effectively improve the prediction accuracy and generalization performance of time series indicators in industrial scenarios, and the method is reasonable and highly implementable.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligence and time-series prediction technology, specifically a system and method for predicting smelting product indicators based on multiple features. Background Technology

[0002] In industrial production, accurate prediction of key production indicators (such as grade, component content, and energy consumption levels) is the core basis for production scheduling optimization, process parameter adjustment, and safety risk prevention and control. Industrial sites typically collect multiple types of time-series data: including high-frequency sampled second- and minute-level process parameters (such as temperature, pressure, and flow rate), low-frequency statistical hourly operating condition data (such as hourly output and average load), and key indicator label data from interval testing. These three types of data have different sampling frequencies and complementary information dimensions.

[0003] Existing time series forecasting techniques have the following technical shortcomings: 1. Single data frequency utilization: Most prediction schemes only build features based on time series data of a single granularity, which cannot simultaneously take into account short-term process fluctuations and long-term operating conditions, resulting in low information utilization of multi-source data; 2. Insufficient tag information mining: Existing feature engineering focuses mainly on the process features of the input side, without fully mining the temporal dependencies of the predicted tags themselves, and without utilizing the changing patterns of historical tags to help improve prediction accuracy; 3. High risk of data leakage: Some solutions introduce future information during feature construction, dataset partitioning and standardization, resulting in excellent offline evaluation performance, but a significant drop in generalization performance when deployed in actual use; 4. Poor adaptability to industrial scenarios: Industrial data generally suffers from problems such as missing values, uneven sampling intervals, and high data noise. Existing methods are not robust enough to complex data quality and are difficult to apply directly.

[0004] Therefore, developing a time series forecasting method that can integrate multi-frequency time series data, deeply mine historical tag information, avoid data leakage, and adapt to complex industrial scenarios is a pressing technical problem in this field. To address this issue, a multi-feature-based smelting product index prediction system and method are proposed. Summary of the Invention

[0005] To address the aforementioned issues, a multi-feature-based system and method for predicting smelting product indicators are presented.

[0006] The specific solution of the present invention is: a multi-feature-based smelting product index prediction system includes: a data preprocessing module, a multi-frequency feature construction module, a label historical feature extraction module, a feature fusion and standardization module, a model training and verification module, and a prediction output module; The data preprocessing module is configured to access high-frequency time-series process data at the second and minute levels, low-frequency operating condition time-series data at the hour level, and smelting product indicator label data from the smelting production scenario. It performs time format parsing, time index sorting, classification of missing values, and time range alignment of multi-source data, outputting a smelting production time-series dataset with a unified time dimension. Among them, the second and minute-level high-frequency data are processed by interpolation to fill missing values, while the hour-level operating condition data and label data are processed by null value removal or labeling to handle missing values. All data are aligned by taking the time intersection. The multi-frequency feature construction module, connected to the data preprocessing module, is configured to use only historical data prior to the prediction time to perform multi-scale feature mining on high-frequency and low-frequency time-series data, and construct multi-frequency fusion features. The multi-frequency feature construction module can be configured with resampling granularity, backtracking window duration, EWMA smoothing coefficient, and lag duration parameters to extract high-frequency statistical features, low-frequency operating condition statistical features, multiple sets of EWMA smoothing trend features, and multiple lag duration features. The label history feature extraction module is configured with a built-in label sequence caching mechanism and a null value completion mechanism. It pre-caches the deduplication and sorting results of the full historical label sequence, extracts historical label data within a preset backtracking window for each prediction time, extracts a fixed number of recent historical label values, calculates the statistical and trend features of the label sequence, and generates a fixed-dimensional label history feature vector to adapt to scenarios with uneven industrial label sampling intervals and missing data. The feature fusion and standardization module, which connects the multi-frequency feature construction module and the label historical feature extraction module, is configured to concatenate multi-frequency fusion features, label historical feature vectors, and time-dimensional features to construct a complete smelting index prediction feature matrix; it only fits a standardizer to the training set data before the time series and performs normalization processing on the full feature matrix to avoid data leakage problems. The model training and verification module, connected to the feature fusion and standardization module, is configured to split the training set and the test set in chronological order, adopt time series cross-validation, train the smelting product index prediction model based on machine learning or neural network algorithms, support hardware-accelerated training, and output multi-dimensional model evaluation indicators and feature importance ranking results. The prediction output module, connected to the model training and verification module, is configured to, for the time to be predicted, call each module to construct the target feature vector, load the trained normalizer and prediction model, and output the prediction results of the smelting product index.

[0007] Furthermore, the multi-frequency feature construction module is specifically used for: For high-frequency data at the second and minute levels, the data is divided into windows according to the set resampling granularity and backtracking window. The mean, standard deviation, minimum and maximum values ​​of the data in each window are calculated. The trend slope is obtained by fitting a first-order polynomial. When the window sub-segment is insufficient, the header is padded with empty values ​​to fix the feature dimension. Multi-scale smoothing features and change slope are extracted by combining multiple sets of EWMA smoothing coefficients. Extract the raw values ​​of low-frequency data from the most recent N hours before the prediction time, and calculate the window mean, standard deviation, and range statistical characteristics. Based on a list of preset lag durations, high- and low-frequency data at corresponding times are matched. When no matching data is available, the most recent historical data is selected as a substitute to construct multi-lag features.

[0008] Furthermore, the tag history feature extraction module is specifically used for: During the initial calculation, the entire label sequence is deduplicated, sorted, and cached, and the cached data is directly called in subsequent calculations; historical labels within a preset backtracking window before the prediction time are extracted, excluding the label value at the current time. Extract the most recent M historical label values ​​from the window. If the number is insufficient, pad the header with empty values ​​to a fixed length. When the number of labels in the window meets the standard, calculate the mean, standard deviation, range, and linear trend slope of the label sequence. If the standard is not met, set the corresponding statistical features to empty and concatenate them to obtain the label historical feature vector.

[0009] Furthermore, the time dimension features include hour, day of the week, date, whether it is a weekend, whether it is a working day, and the number of minutes in a day as time-series attribute features; during the standardization process, the test set and validation set do not participate in the normalizer fitting, and only the fitting parameters of the training set are reused to complete the normalization.

[0010] Furthermore, the evaluation metrics output by the model training and validation module include: mean absolute error (MAE) of the training set and the test set, root mean square error (RMSE), coefficient of determination (R²), and hit rate at a preset threshold; the model training and validation module saves the model file and the normalizer file, and exports a full feature importance ranking table and a TopN core feature table.

[0011] A multi-feature-based method for predicting smelting product indicators, applied to the aforementioned multi-feature-based smelting product indicator prediction system, includes the following steps: S1. Data Preprocessing: Integrate high-frequency time-series data (second-level, minute-level), low-frequency operating condition data (hour-level), and product indicator label data from smelting production. Perform time format conversion and time index sorting, classify high and low frequency data, handle missing values, align the time ranges of multi-source data, and obtain a unified smelting production time-series dataset. S2. Multi-frequency feature construction: Based on the preprocessed dataset, only historical data before the prediction time is used to extract high-frequency multi-scale statistical and smoothing features at the second and minute levels, low-frequency operating condition statistical features at the hour level, and multi-duration lag features, and then fuse them to obtain multi-frequency fusion features. S3. Tag historical feature extraction: Based on the pre-cached historical tag sequence, extract the historical tag data within the backtracking window at each prediction time, complete a fixed number of recent historical tag values, calculate the statistical and trend features of the tag sequence, and generate a tag historical feature vector. S4. Feature Fusion and Standardization: Concatenate multi-frequency fusion features, label historical feature vectors, and time-dimensional features to construct a complete feature matrix; use an independent fitting normalizer for the training set to complete the leakage-free standardization of all features. S5. Model Training and Validation: Split the dataset in chronological order, train the prediction model through time series cross-validation, quantify and output multi-dimensional evaluation metrics, generate feature importance ranking results, and save the model and normalizer. S6. Prediction Output: For the time to be predicted, the target feature vector is generated by reusing the feature construction, fusion and standardization process, and the trained model is called to output the prediction results of the smelting product index.

[0012] Furthermore, in step S1, the high-frequency time series data at the second and minute levels are interpolated to preserve data continuity, while the hourly operating condition data and label data are eliminated or marked to ensure data accuracy. All data are sorted in ascending order by timestamp and aligned by taking the intersection of timestamps.

[0013] Furthermore, step S2 specifically includes: S21. High-frequency feature calculation: Window statistics are performed on high and low frequency data according to configurable parameters. Calculate multi-dimensional statistical features and trend slope of the window, extract multi-scale smoothing features by combining multiple sets of EWMA coefficients, and fill in the fixed feature dimensions with missing values. S22. Low-frequency feature calculation: Extract the raw values ​​of recent hourly data and calculate window statistical features; S23. Lag Feature Calculation: Based on the preset lag duration, historical data is matched, and missing data is filled in from the nearest available data to construct multi-dimensional lag features.

[0014] Furthermore, in step S3, a pre-caching mechanism is used to avoid repeated calculation of the label sequence, and null values ​​are filled in for historical labels that are insufficient to a fixed number. Statistics and trend features are adaptively calculated or set to null based on the number of labels in the window to ensure that the output feature dimensions are consistent.

[0015] Furthermore, in step S5, k-fold time series cross-validation is used, with each fold of the validation set being later than the corresponding training set, simulating real industrial prediction scenarios. Hardware acceleration is used to complete iterative training of the model, thereby achieving model performance optimization and interpretability analysis.

[0016] The present invention has the following beneficial effects: 1. Deep fusion of multi-frequency information: Simultaneously utilize high-frequency process data at the second and minute levels and low-frequency operating condition data at the hour level to construct statistical features and multi-scale smoothing features respectively, taking into account both short-term fluctuation details and long-term trend patterns, and significantly improving the information utilization rate of multi-source time series data; 2. Fully explore the temporal dependence of labels: Introduce historical label values ​​and label statistical features, autonomously learn the temporal change pattern of the prediction target itself, strengthen the model's ability to capture the trend of indicator evolution, and effectively improve prediction accuracy; 3. End-to-end data leakage-free design: Feature construction is strictly limited to using only historical data before the prediction time, the dataset is partitioned in chronological order, and the normalizer is fitted only by the training set, thus avoiding future information leakage throughout the entire process and ensuring the real generalization performance of the model for industrial deployment. 4. Strong adaptability to industrial scenarios: It supports native handling of missing values, non-uniform label spacing, configurable feature windows and model parameters, and is adaptable to scenarios in industrial production with uneven data quality, diverse sampling frequencies and complex operating conditions. It is robust and easy to deploy. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process structure of the present invention; In the diagram: 1. Data preprocessing module; 2. Multi-frequency feature construction module; 3. Label historical feature extraction module; 4. Feature fusion and standardization module; 5. Model training and validation module; 6. Prediction output module; 21. High-frequency features; 22. Low-frequency features; 23. Lag features. Detailed Implementation

[0018] Please refer to Figure 1 A multi-feature-based smelting product index prediction system includes: a data preprocessing module 1, a multi-frequency feature construction module 2, a label historical feature extraction module 3, a feature fusion and standardization module 4, a model training and validation module 5, and a prediction output module 6. The data preprocessing module 1 is configured to access high-frequency time-series process data at the second and minute levels, low-frequency operating condition time-series data at the hour level, and smelting product indicator label data from the smelting production scenario. It performs time format parsing, time index sorting, classification of missing values, and time range alignment of multi-source data, outputting a smelting production time-series dataset with a unified time dimension. Among them, the second and minute-level high-frequency data are processed by interpolation to fill missing values, while the hour-level operating condition data and label data are processed by null value removal or labeling. All data are aligned by taking the time intersection. The multi-frequency feature construction module 2, connected to the data preprocessing module 1, is configured to use only historical data before the prediction time to perform multi-scale feature mining on high-frequency and low-frequency time series data and construct multi-frequency fusion features. The multi-frequency feature construction module 2 can be configured with resampling granularity, backtracking window duration, EWMA smoothing coefficient, and lag duration parameters to extract high-frequency statistical features, low-frequency operating condition statistical features, multiple sets of EWMA smoothing trend features, and multiple lag duration features 23. The label history feature extraction module 3 is configured with a built-in label sequence caching mechanism and a null value completion mechanism. It pre-caches the deduplication and sorting results of the full historical label sequence, extracts historical label data within a preset backtracking window for each prediction time, extracts a fixed number of recent historical label values, calculates the statistical and trend features of the label sequence, and generates a fixed-dimensional label history feature vector to adapt to scenarios with uneven industrial label sampling intervals and missing data. The feature fusion and standardization module 4, which connects the multi-frequency feature construction module 2 and the label historical feature extraction module 3, is configured to splice multi-frequency fusion features, label historical feature vectors and time dimension features to construct a complete smelting index prediction feature matrix; it only fits the standardizer based on the time series training set data before the normalization process to the full feature matrix to avoid data leakage problems. The model training and verification module, connected to the feature fusion and standardization module 4, is configured to split the training set and the test set in chronological order, adopt time series cross-validation, train the smelting product index prediction model based on machine learning or neural network algorithms, support hardware-accelerated training, and output multi-dimensional model evaluation indicators and feature importance ranking results. The prediction output module 6, connected to the model training and verification module, is configured to, for the time to be predicted, call each module to construct the target feature vector, load the trained normalizer and prediction model, and output the prediction result of the smelting product index.

[0019] In this embodiment, the multi-frequency feature construction module 2 is specifically used for: For high-frequency data at the second and minute levels, the data is divided into windows according to the set resampling granularity and backtracking window. The mean, standard deviation, minimum and maximum values ​​of the data in each window are calculated. The trend slope is obtained by fitting a first-order polynomial. When the window sub-segment is insufficient, the header is padded with empty values ​​to fix the feature dimension. Multi-scale smoothing features and change slope are extracted by combining multiple sets of EWMA smoothing coefficients. Extract the raw values ​​of low-frequency data from the most recent N hours before the prediction time, and calculate the window mean, standard deviation, and range statistical characteristics. Based on a list of preset lag durations, high and low frequency data at corresponding times are matched. When no matching data is available, the most recent historical data is selected as a substitute to construct multi-lag features.23

[0020] In this embodiment, the tag history feature extraction module 3 is specifically used for: During the initial calculation, the entire label sequence is deduplicated, sorted, and cached, and the cached data is directly called in subsequent calculations; historical labels within a preset backtracking window before the prediction time are extracted, excluding the label value at the current time. Extract the most recent M historical label values ​​from the window. If the number is insufficient, pad the header with empty values ​​to a fixed length. When the number of labels in the window meets the standard, calculate the mean, standard deviation, range, and linear trend slope of the label sequence. If the standard is not met, set the corresponding statistical features to empty and concatenate them to obtain the label historical feature vector.

[0021] In this embodiment, the time dimension features include hour, day of the week, date, whether it is a weekend, whether it is working hours, and the number of minutes in a day as time-series attribute features; during the standardization process, the test set and validation set do not participate in the normalizer fitting, and only the fitting parameters of the training set are reused to complete the normalization.

[0022] In this embodiment, the evaluation metrics output by the model training and validation module include: mean absolute error (MAE) of the training set and the test set, root mean square error (RMSE), coefficient of determination (R²), and hit rate at a preset threshold; the model training and validation module saves the model file and the normalizer file, and exports the full feature importance ranking table and the TopN core feature table.

[0023] A multi-feature-based method for predicting smelting product indicators, applied to the aforementioned multi-feature-based smelting product indicator prediction system, includes the following steps: S1. Data Preprocessing: Integrate high-frequency time-series data (second-level, minute-level), low-frequency operating condition data (hour-level), and product indicator label data from smelting production. Perform time format conversion and time index sorting, classify high and low frequency data, handle missing values, align the time ranges of multi-source data, and obtain a unified smelting production time-series dataset. S2. Multi-frequency feature construction: Based on the preprocessed dataset, only historical data before the prediction time are used to extract high-frequency multi-scale statistical and smoothing features at the second and minute levels, low-frequency operating condition statistical features at the hour level, and multi-duration lag features23, which are then fused to obtain multi-frequency fusion features. S3. Tag historical feature extraction: Based on the pre-cached historical tag sequence, extract the historical tag data within the backtracking window at each prediction time, complete a fixed number of recent historical tag values, calculate the statistical and trend features of the tag sequence, and generate a tag historical feature vector. S4. Feature Fusion and Standardization: Concatenate multi-frequency fusion features, label historical feature vectors, and time-dimensional features to construct a complete feature matrix; use an independent fitting normalizer for the training set to complete the leakage-free standardization of all features. S5. Model Training and Validation: Split the dataset in chronological order, train the prediction model through time series cross-validation, quantify and output multi-dimensional evaluation metrics, generate feature importance ranking results, and save the model and normalizer. S6. Prediction Output: For the time to be predicted, the target feature vector is generated by reusing the feature construction, fusion and standardization process, and the trained model is called to output the prediction results of the smelting product index.

[0024] In this embodiment, in step S1, the high-frequency time series data at the second and minute levels are interpolated to preserve data continuity, and the hourly operating condition data and label data are eliminated or marked to ensure data accuracy. All data are sorted in ascending order by timestamp and aligned by taking the intersection of timestamps.

[0025] In this embodiment, step S2 specifically includes: S21, Calculation of high-frequency feature 21: Window statistics are performed on high and low frequency data according to configurable parameters, multi-dimensional statistical features and trend slope of the window are calculated, multi-scale smoothing features are extracted by combining multiple sets of EWMA coefficients, and null values ​​are filled to complete fixed feature dimensions. S22, Low-frequency feature 22 calculation: Extract the raw values ​​of recent hourly data and calculate window statistical features; S23, Calculation of Lag Feature 23: Based on the preset lag duration, historical data is matched, and missing data is filled in from the nearest available data to construct multi-dimensional lag feature 23.

[0026] In this embodiment, in step S3, the label sequence is avoided from being calculated repeatedly through a pre-caching mechanism, and null values ​​are filled in for historical labels that are less than a fixed number. The statistical and trend features are calculated or set to null based on the number of labels in the window to ensure that the output feature dimensions are consistent.

[0027] In this embodiment, step S5 uses k-fold time series cross-validation, where the time of each fold validation set is later than the corresponding training set, simulating real industrial prediction scenarios. Hardware acceleration is used to complete iterative training of the model, thereby achieving model performance optimization and interpretability analysis.

[0028] The specific implementation method of this embodiment is as follows: The time series prediction method based on multi-frequency features and label history described in this embodiment is implemented in the following specific process: Step 1: Data Preprocessing Access time-series data: second-level, minute-level feature data, hour-level feature data, and tag data; 1. Time format processing: Parse the time column for each type of data and set the time column as the data index; 2. Missing value classification and handling: High-frequency data at the second and minute levels are imputed or missing values ​​are handled using interpolation methods to maintain the continuity of process data; 3. Time range alignment: Take the intersection of the data time ranges as a unified analysis interval, and extract all data within the corresponding interval to ensure the consistency of the data in the time dimension.

[0029] Step 2: Construction of multi-frequency features For each prediction time, features are constructed using only historical data prior to that time, and consist of three parts: 1. Second-level and minute-level feature calculation Set the backtracking window to t hours and the resampling granularity to x minutes, and extract data windows at the second and minute levels prior to the prediction time: Divide the window into multiple segments with a granularity of x minutes, and calculate the mean, standard deviation, minimum, and maximum of each segment; if the number of segments is less than n, pad with null values ​​at the beginning to n groups to ensure that the feature dimensions are fixed. Based on the sub-segment mean sequence, a first-order polynomial is used to fit and calculate the trend slope within the window; 2. Hour-level feature calculation Set the number of hourly backtracking points to y, and extract the raw values ​​of the most recent y-hourly data points before the prediction time: Based on p data points, calculate the mean, standard deviation, range (maximum value - minimum value), and other feature values ​​for each feature dimension; 3. Calculation of lag characteristic 23 Set the lag time to y hours, find the second, minute and hour data values ​​at the corresponding time; if there is no corresponding data at that time, take the most recent historical data value to form the lag feature 23.

[0030] Step 3: Extracting Historical Features of Tags For each prediction time, perform the following operations: 1. Tag pre-caching: During the initial calculation, the entire tag sequence is deduplicated, sorted, and cached. Subsequent point-by-point calculations directly call the cached sequence to avoid duplicate calculations. 2. Window Capture: Set the label backtracking window to q hours, capture all historical labels within the backtracking window before the prediction time, and strictly exclude the label value at the prediction time itself; 3. Historical value extraction: Extract the most recent v historical tag values; if the number of tags in the window is less than v, pad the header with null values ​​to the specified length; 4. Statistical characteristic calculation: When the number of labels in the window is greater than or equal to a certain value, calculate the mean, standard deviation, range, and slope of the linear fitting trend of the label sequence; when the number is insufficient, set the corresponding statistics to null values. 5. Feature concatenation: Concatenate historical label values ​​with statistical features to form a k-dimensional label historical feature vector.

[0031] Step 4: Feature Fusion and Standardization 1. Feature Fusion: Second-level, minute-level, hour-level, tag history features, and lag features are sequentially concatenated with time-dimensional features to form a complete feature vector; the time-dimensional features include: hour, day of the week, date, whether it is a weekend, whether it is a working day, and the number of minutes in the current day. 2. Standardization Processing: A portion of the samples are taken in chronological order as the training set and another portion as the test set. The training set features are fitted, and then the full feature matrix is ​​standardized. The test set does not participate in the standardizer fitting, thus avoiding data leakage at the source.

[0032] Step 5: Model Training and Validation 1. Dataset partitioning: The dataset is partitioned chronologically, with one part serving as the training set and the other as the test set; k-fold time series cross-validation is performed, with the validation set time of each fold being later than the corresponding training set time, simulating real prediction scenarios; 2. Model Configuration: Mainstream machine learning and deep learning models are selected; hardware-accelerated training is supported. 3. Training and Evaluation: Fit the model on the training set and validate the effect on the test set; calculate and output the MAE, RMSE, and R² metrics for the training and test sets, as well as the hit rate when the squared error is less than a specific metric; 4. Results saving: Save the model file and the normalizer file; at the same time, export the full feature importance ranking table and the Top n important feature table for model interpretability analysis and feature selection.

[0033] Step 6: Single Sample Prediction For a given time to be predicted, repeat the feature construction process of steps 2-4, using only all historical data before that time to generate feature vectors; call the trained normalizer and prediction model to calculate and output the final predicted value.

[0034] The above embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation procedures. However, the scope of protection of the present invention is not limited to the above embodiments. Unless otherwise specified, the algorithms, parameter configurations, etc. used in the above embodiments can be implemented using conventional technical solutions in the art.

Claims

1. A multi-feature-based system for predicting smelting product indicators, characterized in that, include: The module includes a data preprocessing module, a multi-frequency feature construction module, a label historical feature extraction module, a feature fusion and standardization module, a model training and validation module, and a prediction output module. The data preprocessing module is configured to access high-frequency time-series process data at the second and minute levels, low-frequency operating condition time-series data at the hour level, and smelting product indicator label data from the smelting production scenario. It performs time format parsing, time index sorting, classification of missing values, and time range alignment of multi-source data, outputting a smelting production time-series dataset with a unified time dimension. Among them, the second and minute-level high-frequency data are processed by interpolation to fill missing values, while the hour-level operating condition data and label data are processed by null value removal or labeling to handle missing values. All data are aligned by taking the time intersection. The multi-frequency feature construction module, connected to the data preprocessing module, is configured to use only historical data prior to the prediction time to perform multi-scale feature mining on high-frequency and low-frequency time-series data, and construct multi-frequency fusion features. The multi-frequency feature construction module can be configured with resampling granularity, backtracking window duration, EWMA smoothing coefficient, and lag duration parameters to extract high-frequency statistical features, low-frequency operating condition statistical features, multiple sets of EWMA smoothing trend features, and multiple lag duration features. The label history feature extraction module is configured with a built-in label sequence caching mechanism and a null value completion mechanism. It pre-caches the deduplication and sorting results of the full historical label sequence, extracts historical label data within a preset backtracking window for each prediction time, extracts a fixed number of recent historical label values, calculates the statistical and trend features of the label sequence, and generates a fixed-dimensional label history feature vector to adapt to scenarios with uneven industrial label sampling intervals and missing data. The feature fusion and standardization module, which connects the multi-frequency feature construction module and the label historical feature extraction module, is configured to concatenate multi-frequency fusion features, label historical feature vectors, and time-dimensional features to construct a complete smelting index prediction feature matrix; it only fits a standardizer to the training set data before the time series and performs normalization processing on the full feature matrix to avoid data leakage problems. The model training and verification module, connected to the feature fusion and standardization module, is configured to split the training set and the test set in chronological order, adopt time series cross-validation, train the smelting product index prediction model based on machine learning or neural network algorithms, support hardware-accelerated training, and output multi-dimensional model evaluation indicators and feature importance ranking results. The prediction output module, connected to the model training and verification module, is configured to, for the time to be predicted, call each module to construct the target feature vector, load the trained normalizer and prediction model, and output the prediction results of the smelting product index.

2. The multi-feature-based smelting product index prediction system according to claim 1, characterized in that, The time dimension features include hour, day of the week, date, whether it is a weekend, whether it is a working day, and time-series attribute features of the number of minutes on the day; During the standardization process, the test set and validation set do not participate in the normalizer fitting; only the fitting parameters of the training set are reused to complete the normalization.

3. The multi-feature-based smelting product index prediction system according to claim 1, characterized in that, The evaluation metrics output by the model training and validation module include: mean absolute error (MAE) of the training set and the test set, root mean square error (RMSE), coefficient of determination (R²), and hit rate at a preset threshold. The model training and validation module saves the model file and the normalizer file, and exports the full feature importance ranking table and the TopN core feature table.

4. A method for predicting smelting product indicators based on multiple features, characterized in that, The method applied to the prediction system according to any one of claims 1-3 includes the following steps: S1. Data Preprocessing: Integrate high-frequency time-series data (second-level, minute-level), low-frequency operating condition data (hour-level), and product indicator label data from smelting production. Perform time format conversion and time index sorting, classify high and low frequency data, handle missing values, align the time ranges of multi-source data, and obtain a unified smelting production time-series dataset. S2. Multi-frequency feature construction: Based on the preprocessed dataset, only historical data before the prediction time is used to extract high-frequency multi-scale statistical and smoothing features at the second and minute levels, low-frequency operating condition statistical features at the hour level, and multi-duration lag features, and then fuse them to obtain multi-frequency fusion features. S3. Tag historical feature extraction: Based on the pre-cached historical tag sequence, extract the historical tag data within the backtracking window at each prediction time, complete a fixed number of recent historical tag values, calculate the statistical and trend features of the tag sequence, and generate a tag historical feature vector. S4. Feature Fusion and Standardization: Concatenate multi-frequency fusion features, label historical feature vectors, and time-dimensional features to construct a complete feature matrix; use an independent fitting normalizer for the training set to complete the leakage-free standardization of all features. S5. Model Training and Validation: Split the dataset in chronological order, train the prediction model through time series cross-validation, quantify and output multi-dimensional evaluation metrics, generate feature importance ranking results, and save the model and normalizer. S6. Prediction Output: For the time to be predicted, the target feature vector is generated by reusing the feature construction, fusion and standardization process, and the trained model is called to output the prediction results of the smelting product index.

5. The method for predicting smelting product indicators based on multiple features according to claim 4, characterized in that, In step S1, high-frequency time series data at the second and minute levels are interpolated to preserve data continuity, while hourly operating condition data and label data are eliminated or marked to ensure data accuracy. All data are sorted in ascending order by timestamp and aligned by taking the intersection of timestamps.

6. The method for predicting smelting product indicators based on multiple features according to claim 4, characterized in that, Step S2 specifically includes: S21. High-frequency feature calculation: Window statistics are performed on high and low frequency data according to configurable parameters. Calculate multi-dimensional statistical features and trend slope of the window, extract multi-scale smoothing features by combining multiple sets of EWMA coefficients, and fill in the fixed feature dimensions with missing values. S22. Low-frequency feature calculation: Extract the raw values ​​of recent hourly data and calculate window statistical features; S23. Lag Feature Calculation: Based on the preset lag duration, historical data is matched, and missing data is filled in from the nearest available data to construct multi-dimensional lag features.

7. The method for predicting smelting product indicators based on multiple features according to claim 4, characterized in that, In step S3, a pre-caching mechanism is used to avoid repeated calculation of the label sequence, and null values ​​are filled in for historical labels that are less than a fixed number. Statistics and trend features are calculated or left blank according to the number of labels in the window to ensure that the output feature dimensions are consistent.

8. The method for predicting smelting product indicators based on multiple features according to claim 4, characterized in that, In step S5, k-fold time series cross-validation is used, with each fold of the validation set being later than the corresponding training set, simulating real industrial prediction scenarios. Hardware acceleration is used to complete iterative training of the model, thereby achieving model performance optimization and interpretability analysis.

Images