A Joint Prediction Method for Power Consumption and Fineness of Cement Vertical Mill Based on Hierarchical Multi-Task Graph Attention Network

By combining a hierarchical multi-task graph attention network with the vertical mill mechanism, the problems of single-objective fragmentation and poor generalization of small sample prediction of cement vertical mill power consumption and fineness prediction are solved, realizing joint prediction with high accuracy and high generalization, which is suitable for intelligent control of cement production equipment.

CN123116203APending Publication Date: 2026-11-17SINOMA SUZHOU CONSTR
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611195768.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-07
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for predicting the power consumption and fineness of cement vertical mills suffer from problems such as fragmented coupling of single objectives, poor generalization with small samples, and lack of physical meaning in the graph structure, making it difficult to achieve joint prediction with high accuracy and high generalization.

Method used

A method based on hierarchical multi-task graph attention network is adopted. Through data acquisition and preprocessing, hierarchical graph structure construction, multi-task prediction head and dynamic weight loss function, combined with the mechanical mechanism of vertical mill and data driving, a physically meaningful model is constructed to jointly predict power consumption and fineness.

Benefits of technology

It achieves high-precision and high-generalization joint prediction of vertical mill power consumption and fineness, with strong physical consistency, applicable to small sample scenarios, and improves the interpretability and prediction accuracy of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116203A_ABST
    Figure CN123116203A_ABST
Patent Text Reader

Abstract

This invention discloses a method for jointly predicting the power consumption and fineness of a cement vertical mill based on a hierarchical multi-task graph attention network. The method collects and preprocesses historical operating data of the vertical mill, while simultaneously constructing a mechanism simulation dataset to expand the sample. The vertical mill is decomposed into four subsystems: feeding, grinding, grading, and ventilation, constructing a hierarchical graph structure. A hierarchical multi-task graph attention network model is built, embedding mechanism prior weights in the attention layer, and training the model using a dynamic weighted multi-task loss function with mechanism constraint regularization terms. Real-time data is input into the trained model, outputting predicted power consumption and fineness values. This invention combines the advantages of mechanism and data-driven approaches, solving the problem of overfitting with small samples, achieving high prediction accuracy and strong interpretability, and is suitable for optimizing the operation and saving energy in cement vertical mills.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for cement production equipment, and specifically relates to a method for jointly predicting the power consumption and raw material fineness of a cement vertical mill raw material grinding system. Background Technology

[0002] Cement raw meal grinding is one of the most energy-intensive processes in cement production, accounting for 30%-40% of total electricity consumption. Raw meal fineness is a key indicator affecting the quality of subsequent clinker calcination; insufficient fineness leads to decreased clinker strength, while excessive grinding causes a significant increase in electricity consumption. Therefore, achieving high-precision joint prediction of vertical mill electricity consumption and fineness is crucial for balancing product quality and energy consumption.

[0003] Existing methods for predicting the performance of vertical mills are mainly divided into two categories.

[0004] The first category is the mechanism modeling method, which establishes a mathematical model based on theories such as grinding dynamics and fluid mechanics. The physical meaning is clear, but the model parameters are difficult to obtain accurately, the calculation is complex, and it is difficult to adapt to fluctuations in operating conditions.

[0005] The second category is data-driven methods, which use machine learning algorithms such as neural networks and support vector machines to build predictive models. These methods do not require complex mechanistic derivations, but they have the following shortcomings: First, most of them are single-objective predictions, which sever the inherent coupling relationship between power consumption and fineness, and cannot guide the coordinated optimization of production parameters. Second, pure data-driven models have poor generalization ability and are prone to overfitting in industrial small sample data scenarios, resulting in a sharp drop in prediction accuracy. Third, existing graph neural network methods mostly use fully connected graph structures, which do not consider the mechanical structural characteristics of the vertical mill system, resulting in redundant parameters, and the attention weights have no physical meaning, leading to poor interpretability.

[0006] Therefore, there is an urgent need for a dual-objective prediction method for power consumption and fineness that can combine the mechanical mechanism of vertical mills with the advantages of data-driven approaches and is suitable for small sample scenarios. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a dual-objective prediction method for power consumption and fineness of cement vertical mills based on hierarchical multi-task graph attention networks. This method solves the problems of single-objective fragmented coupling relationships, poor generalization in small samples, and lack of physical meaning in graph structures in existing methods, thereby achieving high-precision and high-generalization joint prediction of power consumption and fineness of vertical mills.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for jointly predicting the power consumption and fineness of a cement vertical mill based on a hierarchical multi-task graph attention network includes the following steps: S1): Data acquisition, preprocessing and simulation expansion: Collect historical operating data of the vertical mill system, perform outlier processing, time series smoothing and standardization to obtain a preprocessed dataset; at the same time, construct a mechanism simulation dataset based on the grinding mechanism of the vertical mill to expand the training samples; Data acquisition includes collecting vertical mill operating parameters (feed rate, grinding pressure, classifier speed, ventilation volume), status parameters (vibration, pressure difference), and performance indicators (power consumption, fineness). Anomaly removal employs a combination of hard removal based on mechanistic principles (direct deletion of values ​​exceeding physical limits) and the 3σ statistical criterion. (determined to be abnormal) Time series smoothing uses a moving average method (window k) to filter out noise: ; Standardization uses Z-score (based on the mean and standard deviation of the training set) to eliminate dimensions: ; Mechanism simulation expansion is based on grinding kinetics and the law of conservation of energy. It constructs a simplified mechanism model of vertical mill, generates simulation data covering the entire operating range, and mixes it with real industrial data to solve the problem of small sample size.

[0009] S2): Construct a physically meaningful hierarchical graph structure: Decompose the vertical mill system into a feeding subsystem, a grinding subsystem, a grading subsystem, and a ventilation subsystem. Use the key operating parameters of each subsystem as graph nodes, and establish edge connections only within the same subsystem and between adjacent subsystems to construct a hierarchical graph structure. Define a directed graph Nodes represent runtime parameters, and edges represent physical connections; Hierarchical subsystem division: ; in, This refers to the node set of the feeding subsystem (such as feed rate and belt speed). This refers to the node set of the grinding subsystem (such as grinding pressure, material layer thickness, and mill vibration). This refers to the node set of the classification subsystem (such as the classifier speed and the amount of powder returned). For the ventilation subsystem node set (such as hot air volume, system pressure difference, outlet temperature); Sparse edge construction rules: Internally: Nodes within the same subsystem are fully connected; Adjacent: Edge connections are established only between core nodes of adjacent subsystems (such as grinding pressure and mill current); Non-adjacent: Nodes that cross non-adjacent subsystems (such as feeding and grading) are not directly connected; Result: The adjacency matrix is ​​a block diagonally dominant matrix, which greatly reduces parameter redundancy and improves interpretability.

[0010] S3): Construct a hierarchical multi-task graph attention network model, including an input layer, a hierarchical graph attention layer, a dual-task prediction head, and a dynamic weighted multi-task loss function layer containing mechanistic constraint regularization terms; The input layer receives standardized data and generates an initial node feature matrix: ; Layered graph attention layer for the first Layers, nodes The output features are: ; Mechanism-based prior attention weight calculation (including physical embedding): in, For learnable attention vectors, This indicates a feature concatenation operation. The prior weight of the mechanism is determined in advance by the grinding mechanism of the vertical mill, and the value range is [0,1]. The stronger the physical correlation, the larger the weight value. The dual-task prediction head consists of two parallel fully connected layers, one for power consumption and the other for finer-grained prediction. The output features of the last graph attention layer are globally pooled and then input into the two fully connected layers to obtain the power consumption predictions. and fineness prediction value : in, , This is the weight matrix of the fully connected layer. , For bias vectors, The feature vector after global pooling; The loss function layer employs a dynamic weighted multi-task loss function, simultaneously optimizing the prediction errors for power consumption and granularity, and incorporating a mechanism-constrained regularization term. The total loss function is: in, This is the mean square error loss in power consumption prediction. This represents the mean square error loss for fineness prediction. and The uncertainty parameters for the two tasks are learned automatically by the model. The coefficient of the regularization term, The loss is constrained by the mechanism.

[0011] Mechanism-constrained loss The penalty for predictions that violate the core physical laws of the vertical mill consists of the sum of penalty functions for the following five constraints: in, For the first The violation degree function of the mechanistic constraint. The input feature matrix is ​​used. The violation degree function corresponding to each constraint is the constraint of negative correlation between classifier speed and fineness. The positive correlation between grinding pressure and power consumption is a constraint. Optimal feed rate range constraint: Fineness is positively correlated with power consumption, constrained by this. Constraint that ventilation volume is positively correlated with particle size: ; S4): Mixed dataset training and hyperparameter optimization: The preprocessed real dataset is mixed with the mechanism simulation dataset and divided into training set, validation set and test set. The model is trained and hyperparameters are optimized, and the model performance is verified. Data partitioning: The real + simulation mixed dataset is divided into training set, validation set and test set according to the proportion.

[0012] Optimizer: Adam is used for backpropagation, and an early stopping mechanism is introduced (the system stops if the validation set loss does not decrease for several consecutive rounds) to combat overfitting; Hyperparameter optimization: Multi-objective Bayesian optimization is employed to minimize power consumption and validation error of fineness, seeking the optimal learning rate, number of layers, and number of hidden units; Evaluation metrics: The regression performance was evaluated using three dimensions: RMSE (root mean square error), MAE (mean absolute error), and R² (coefficient of determination). in, The number of samples in the test set. For the true value, For predicted values, This is the mean of the true values.

[0013] S5): Real-time deployment and application: Input the real-time operating data of the vertical mill system into the trained model, and output the predicted values ​​of power consumption and raw material fineness.

[0014] The beneficial effects of this invention are: Physical consistency guarantee: This invention ensures that the model prediction results strictly follow the physical laws of vertical mill grinding through a triple mechanism of hierarchical graph structure, mechanistic prior attention weights, and physical constraint regularization terms. In the embodiment, the physical constraint violation rate is only 2.1%, far lower than the 15.8% of the comparison method, solving the pain point of "mathematically sound but physically contradictory" prediction results of purely data-driven models.

[0015] 2. Significantly Improved Generalization Ability with Small Samples: This invention effectively solves the overfitting problem caused by the scarcity of training samples in industrial scenarios by expanding the mechanistic simulation data (generating 20,000 simulation data points in the example) and reducing the number of parameters (58.7% less than fully connected graphs) through a sparse hierarchical graph structure. In a small-sample ablation experiment (5,000 training data points), the RMSE of this invention is 0.537, while the RMSE of the comparative method is 1.234, representing an accuracy improvement of 56.5%.

[0016] 3. High accuracy of dual-objective joint prediction: This invention employs a dual-task prediction head combined with a dynamically uncertain weighted loss function to adaptively balance the training of the two tasks: power consumption and granularity. Example test set power consumption RMSE = 0.421 (R 2 =0.963), Fineness RMSE=0.297 (R 2 =0.951), compared with the single-task model, the power consumption accuracy is improved by 24.4% and the fineness accuracy is improved by 31.9%, which verifies the superiority of dual-objective collaborative learning.

[0017] 4. High model interpretability: Each attention weight in this invention... All are multiplied by the physical prior weights. This allows the variables that the model focuses on to have clear physical meaning. Engineers can directly interpret the basis of the model's decisions (e.g., whether the attention weights for grinding pressure and power consumption are reasonable), overcoming the "black box" defect of deep neural networks and facilitating technical trust and promotion in industrial settings.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 This is a schematic diagram of the layered structure of the vertical mill subsystem; Figure 3 This is a diagram showing the overall architecture of the Hierarchical Multi-Task Graph Attention Network (HMT-GAT) model. Figure 4 This is a schematic diagram illustrating the computational principle of the attention layer. Figure 5 This is a schematic diagram illustrating the principle of the dynamic weighted multi-task loss function. Detailed Implementation

[0020] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention. Specific Implementation Example 1: This embodiment uses an HRM-type vertical roller mill in a cement production line as the application example. The mill is designed to produce 200 t / h, and its matching air classifier is model SLK7000, with a main motor power of 3800 kW. The data collection period in this embodiment is 30 consecutive days, with a sampling interval of 1 minute, resulting in 43,200 valid samples. The following is combined with... Figures 1 to 5 The specific implementation process of the present invention will be described in detail.

[0022] I. Data Acquisition and Preprocessing The raw data collected in this embodiment comprises a total of 28 operating parameters, which are divided into three categories: Operating parameters (15): Feed rate (t / h), grinding pressure (MPa), classifier speed (rpm), ventilation volume (Nm³ / h). 3 / h), hot air valve opening (%), circulating air valve opening (%), exhaust valve opening (%), grinding roller hydraulic station pressure (MPa), grinding roller hydraulic station flow rate (L / min), tensioning pressure (MPa), classifier current (A), bucket elevator current (A), belt scale speed (r / min), main motor current (A), classifier guide blade angle (°).

[0023] Status parameters (8): mill vibration value (mm / s), mill pressure difference (Pa), mill outlet temperature (°C), mill inlet temperature (°C), mill disc material layer thickness (mm), powder return rate (t / h), circulating load rate (%), and classifier inlet and outlet pressure difference (Pa).

[0024] Performance indicators (5): power consumption per unit hour (kWh / t), raw material fineness (percentage of residue on 80μm sieve, %), finished product output (t / h), mill operating rate (%), and system air pressure (Pa).

[0025] The specific preprocessing steps are as follows: (1) Outlier removal First, perform mechanistic hard filtering: delete data with a feed rate of less than 50t / h or greater than 300t / h (exceeding the design operating range of the vertical mill); delete data with a grinding pressure of less than 2MPa or greater than 20MPa; delete data with a raw material fineness of less than 1% or greater than 30% (exceeding the physical limit).

[0026] Then, a second screening is performed using the 3σ statistical criterion: for any feature sequence If satisfied If a value is found to be an outlier, it will be removed. This is the mean of the characteristic sequence. This represents the standard deviation of the feature sequence. After two-step elimination, 40,800 valid samples remain.

[0027] (2) Temporal smoothing A moving average method is used to filter and reduce noise in the data after outlier removal. For the eigenvalues ​​at any time t... Its smoothed value is: In this embodiment, the sliding window size is set to k=6 (corresponding to a 6-minute window, which can effectively filter out high-frequency noise while retaining the trend of operating condition changes).

[0028] (3) Standardization process The Z-score normalization method is used to map all features to a standard normal distribution with a mean of 0 and a variance of 1. The formula is as follows: in, The mean of this feature for the training set, This represents the standard deviation of this feature in the training set. It is important to note that all standardized parameters are calculated based on the training set and then applied to the validation and test sets to prevent data leakage.

[0029] (4) Construction of mechanism simulation dataset Based on the grinding kinetics of vertical mills and the law of conservation of energy, a simplified mechanism simulation model is constructed.

[0030] Using the above mechanism model, within the normal operating parameters of the vertical mill (feed rate 120~280t / h, grinding pressure 3~15MPa, classifier speed 60~120rpm, ventilation volume 180000~320000Nm), 3 The simulation data, consisting of 20,000 records, was generated using the Latin hypercube sampling method, covering the entire operating range. These simulation data were then fused with preprocessed real industrial data (40,800 records) to form a hybrid dataset of 60,800 records.

[0031] II. Construction of Hierarchical Graph Structure Based on the system structure and energy flow law of the vertical mill, a layered diagram structure is constructed. Node partitioning: In this embodiment, 28 operating parameters are mapped to 28 graph nodes, which are divided into four subsets according to the subsystem: (Feeding subsystem, 5 nodes): feed rate, belt scale speed, bucket elevator current, external discharge valve opening, and mill inlet temperature.

[0032] (Grinding subsystem, 9 nodes): grinding pressure, mill vibration value, mill pressure difference, grinding disc material layer thickness, main motor current, grinding roller hydraulic station pressure, grinding roller hydraulic station flow rate, tension pressure, and mill operating rate.

[0033] (Classification subsystem, 7 nodes): classifier speed, classifier current, returned powder amount, circulating load rate, classifier guide vane angle, classifier inlet and outlet pressure difference, finished product output.

[0034] (Ventilation subsystem, 7 nodes): ventilation volume, hot air valve opening, circulating air valve opening, mill outlet temperature, system air pressure, circulating air volume, exhaust air volume.

[0035] Edge connection rules: Fully connected within a subsystem: Any two nodes within the same subsystem are connected by bidirectional edges. For example, within V2, all C(9,2)=36 edges are established, including those related to grinding pressure and mill vibration, grinding pressure and material layer thickness, and mill vibration and material layer thickness.

[0036] Adjacent subsystems are connected through core nodes: the feeding subsystem V1 and the grinding subsystem V2 are connected only through the core node between the feeding amount and the grinding pressure; the grinding subsystem V2 and the classifying subsystem V3 are connected only through the core node between the material layer thickness and the classifier speed, and the mill pressure difference and the amount of powder returned; the classifying subsystem V3 and the ventilation subsystem V4 are connected only through the core node between the classifier inlet and outlet pressure difference and the ventilation volume.

[0037] Non-core nodes in cross-subsystems are not directly connected: for example, there is no direct edge connection between the belt scale speed of the feeding subsystem V1 and the classifier current of the grading subsystem V3.

[0038] After construction, the adjacency matrix A in this embodiment is a 28×28 sparse block diagonally dominant matrix with a total of only 156 edges (a fully connected graph would require 28×27 / 2=378 edges), reducing the number of parameters by approximately 58.7%.

[0039] III. HMT-GAT Model Construction (1) Input layer The input layer receives standardized 28-dimensional node feature data and substitutes it to generate the initial node feature matrix. , where 64 is the initial feature embedding dimension.

[0040] (2) Layered graph attention layer This embodiment employs a stacked two-layer hierarchical graph attention layer. The first layer maps the 64-dimensional input to 128-dimensional hidden features, and the second layer maps the 128-dimensional input to 64-dimensional features. Each layer uses a multi-head attention mechanism with four heads, and the output is averaged.

[0041] For the Layered graph attention layer, nodes The output features are calculated as follows: Attention weight The calculation introduces the mechanism of prior weights : In this embodiment, the mechanism prior weight The default rules are: Grinding pressure → Power consumption (within the same subsystem): w_ij = 0.95; Feed rate → Power consumption (adjacent subsystem): w_ij = 0.85; Air classifier speed → fineness (within the same subsystem): w_ij = 0.90; Ventilation volume → Fineness (adjacent subsystems): w_ij=0.80; Belt scale speed → fineness (indirect correlation across two subsystems): w_ij=0.30; Between nodes that are not adjacent subsystems and have no physical connection: w_ij=0.

[0042] (3) Dual-task prediction head The output features of the second graph attention layer are subjected to global average pooling to obtain the global feature vector. Then, two independent single-layer fully connected networks are input in parallel: Power consumption prediction branch: , The output dimension is 1, corresponding to the predicted power consumption per unit hour (kWh / t).

[0043] Fineness prediction branch: , The output dimension is 1, corresponding to the predicted value of raw material fineness (percentage of residue on an 80μm sieve).

[0044] Both branches have linear activation functions (regression tasks do not require nonlinear transformations).

[0045] (4) Dynamic weighted multi-task loss function layer The total loss function is: in: L1 is the mean squared error loss for power consumption prediction: L1 = (1 / M)Σ(y1 - ŷ1) 2 ; L2 is the mean squared error loss for fineness prediction: L2 = (1 / M)Σ(y2 - ŷ2) 2 ; σ1 and σ2 are learnable uncertainty parameters, both initially set to 1.0, and softplus activation is used to ensure they are always positive; λ is the regularization coefficient, and in this embodiment, λ = 0.05 is taken through cross-validation.

[0046] The mechanistic constraint loss L_physics consists of five penalty functions: The violation degree function of each constraint is approximated using numerical gradient (central difference method, step size ε=0.001): Constraint 1 (Air classifier speed is negatively correlated with fineness): ; Constraint 2 (Grinding pressure is positively correlated with power consumption): ; Constraint 3 (There is an optimal range for feed rate): ; Constraint 4 (Fineness is positively correlated with power consumption): ; Constraint 5 (ventilation volume is positively correlated with particle size): .

[0047] IV. Model Training and Hyperparameter Optimization (1) Dataset partitioning The 60,800 mixed dataset entries were divided into a training set (42,560 entries), a validation set (9,120 entries), and a test set (9,120 entries) in a ratio of 7:1.5:1.5. To ensure the balance of the working conditions, stratified sampling was performed based on two key parameters: feed rate and grinding pressure, before partitioning.

[0048] (2) Training configuration The optimizer used is Adam, with an initial learning rate of 0.001, an exponential decay rate β1 = 0.9 for the first moment estimate, an exponential decay rate β2 = 0.999 for the second moment estimate, and a weight decay coefficient set to 1 × 10⁻⁶. -5 The batch size is set to 128. The maximum number of training epochs is set to 500, and an early stopping mechanism is introduced: training is terminated early when the total loss of the validation set does not decrease for 30 consecutive epochs. In this embodiment, training actually stopped early on the 247th epoch.

[0049] (3) Hyperparameter multi-objective Bayesian optimization A multi-objective Bayesian optimization algorithm is employed to optimize key hyperparameters, with the optimization objective being to simultaneously minimize the validation set power consumption RMSE and the fineness RMSE. The optimization space is set as follows: Number of attention layers in the graph: {1, 2, 3} Hidden layer dimensions: {32, 64, 128, 256} Number of attention heads: {2, 4, 6, 8} Learning rate: [0.0001, 0.01] (log space sampling) Regularization coefficient λ: [0.001, 0.5] (log space sampling) Dropout rate: [0.05, 0.5] A Tree-structured Parzen Estimator (TPE) was used as the surrogate model, and the EHVI (Expected Hypervolume Improvement) function was employed. Initially, 20 random samples were collected, followed by 80 samples in subsequent iterations, for a total of 100 Bayesian optimization iterations. The final optimal hyperparameter combination was: number of layers = 2, hidden dimension = 128, number of attention heads = 4, learning rate = 0.0007, λ = 0.05, and dropout rate = 0.15.

[0050] (4) Evaluation indicators Calculate three core evaluation metrics on the test set: Root mean square error: Mean absolute error: Coefficient of determination: The final performance of this embodiment on the test set is as follows: V. Comparative Experiment Verification To verify the technical advantages of the present invention, this embodiment sets up four sets of comparative experiments (all evaluated on the same test set): The results of the comparative experiment show that: 1. Compared with the single-task model, this invention reduces the power consumption RMSE by 24.4% and the granularity RMSE by 31.9% through dual-task joint prediction, proving that dual-objective collaborative learning can effectively improve prediction accuracy.

[0051] 2. Compared with the fully connected GAT method without hierarchical structure and physical regularization, the present invention improves power consumption and fineness prediction by 21.7% and 25.4% respectively, and reduces the physical constraint violation rate from 15.8% to 2.1%, proving the effectiveness of physical mechanism embedding.

[0052] 3. Compared with traditional BP neural networks, the RMSE of this invention is reduced by more than 50%, and the generalization performance is significantly improved in small sample scenarios.

[0053] 4. In the small-sample ablation experiment, when the training data was reduced to only 5,000 samples, the power consumption RMSE of the method of this invention was 0.537, while the RMSE of pure data-driven GAT was 1.234, proving that the mechanism simulation data augmentation strategy plays a decisive role in solving the small-sample overfitting problem.

[0054] VI. Industrial Deployment and Application Effects The trained HMT-GAT model was solidified in ONNX format and deployed to the edge computing server of the industrial control system (configuration: Intel Xeon Silver 4210, 32GB RAM, NVIDIA T4 GPU). The average time for a single inference is 12.6ms, which fully meets the real-time prediction requirements of the vertical mill with a 1-minute sampling cycle.

[0055] In actual production operation, the system guides the optimization of operating parameters based on model prediction results: when the model predicts a fineness lower than the control target (e.g., <12%), the system prompts to appropriately reduce the classifier speed to reduce over-grinding and lower power consumption; when the predicted fineness is higher than the target upper limit (e.g., >16%), the system prompts to increase grinding pressure or reduce feed rate. After a month of industrial application verification, while ensuring a raw material fineness qualification rate of ≥98.5%, the system's hourly power consumption was reduced by an average of 1.8 kWh / t. Based on a production capacity of 200 t / h, this translates to an annual power saving of approximately 3.15 million kWh, demonstrating significant energy-saving and consumption-reducing effects.

[0056] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for jointly predicting the power consumption and fineness of a cement vertical mill based on a hierarchical multi-task graph attention network, characterized in that: Includes the following steps: S1): Data acquisition, preprocessing and simulation expansion: Collect historical operating data of the vertical mill system, perform outlier processing, time series smoothing and standardization to obtain a preprocessed dataset; at the same time, construct a mechanism simulation dataset based on the grinding mechanism of the vertical mill to expand the training samples; S2): Construct a physically meaningful hierarchical graph structure: Decompose the vertical mill system into a feeding subsystem, a grinding subsystem, a grading subsystem, and a ventilation subsystem. Use the key operating parameters of each subsystem as graph nodes, and establish edge connections only within the same subsystem and between adjacent subsystems to construct a hierarchical graph structure. S3): Construct a hierarchical multi-task graph attention network model, including an input layer, a hierarchical graph attention layer, a dual-task prediction head, and a dynamic weighted multi-task loss function layer containing mechanistic constraint regularization terms; S4): Mixed dataset training and hyperparameter optimization: The preprocessed real dataset is mixed with the mechanism simulation dataset and divided into training set, validation set and test set. The model is trained and hyperparameters are optimized, and the model performance is verified. S5): Real-time deployment and application: Input the real-time operating data of the vertical mill system into the trained model, and output the predicted values ​​of power consumption and raw material fineness.

2. The method for jointly predicting the power consumption and fineness of a cement vertical mill based on a hierarchical multi-task graph attention network according to claim 1, characterized in that: In step S3, the dual-task prediction head is divided into a power consumption prediction branch and a detail prediction branch.

3. The method for jointly predicting the power consumption and fineness of a cement vertical mill based on a hierarchical multi-task graph attention network according to claim 1, characterized in that: The outlier handling in step S1 uses a combination of mechanistic hard removal and the 3σ principle; the time series smoothing uses the moving average method; and the standardization process uses the Z-score standardization method.

4. The method for jointly predicting the power consumption and fineness of a cement vertical mill based on a hierarchical multi-task graph attention network according to claim 1, characterized in that: The adjacency matrix of the hierarchical graph structure described in step S2 is a block-diagonally dominant matrix. All nodes within the same subsystem are interconnected, adjacent subsystems are connected through core nodes, and non-core nodes across subsystems do not establish direct connections.

5. The method for jointly predicting the power consumption and fineness of a cement vertical mill based on a hierarchical multi-task graph attention network according to claim 1, characterized in that: The hierarchical graph attention layer described in step S3 incorporates mechanistic prior weights into the standard graph attention calculation.

6. The method for jointly predicting the power consumption and fineness of a cement vertical mill based on a hierarchical multi-task graph attention network according to claim 1, characterized in that: The hyperparameter optimization described in step S4 employs a multi-objective Bayesian optimization algorithm, and the model performance is evaluated using root mean square error, mean absolute error, and coefficient of determination.