Manganese electrolysis technological process multi-parameter analysis and electrolysis energy consumption prediction method based on machine learning algorithm
By establishing a gradient boosting regression model based on machine learning, the problem of quantifying the influence of parameter coupling in the electrolytic manganese process was solved, enabling high-precision prediction of current efficiency and energy consumption, and supporting real-time process optimization and energy consumption reduction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TONGREN UNIV
- Filing Date
- 2026-01-12
- Publication Date
- 2026-04-24
AI Technical Summary
Existing technologies struggle to quantify the coupled effects of multiple parameters in the electrolytic manganese process in real time and accurately, leading to unstable electrolysis parameter decisions that rely on human experience, resulting in low current efficiency, high energy consumption, and severe environmental pollution.
A machine learning-based approach is adopted, which obtains a dataset through orthogonal experiments, establishes a machine learning model, and then selects the gradient boosting regression (GBR) model after optimization to predict current efficiency and electrolysis energy consumption, providing data-driven process optimization decisions.
It achieves high-precision prediction of current efficiency and energy consumption, with fast response speed, and can support real-time process adjustment, reduce energy consumption and improve production quality stability.
Smart Images
Figure CN121920603A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electrolytic manganese process optimization and intelligent manufacturing technology, specifically to a method for multi-parameter analysis and electrolytic energy consumption prediction of electrolytic manganese process based on machine learning algorithms, applicable to parameter optimization, energy consumption prediction and quality control of electrolytic manganese production process. Background Technology
[0002] Manganese is a vital strategic resource for my country, and electrolytic manganese metal is a crucial raw material in the metallurgical, energy, and chemical industries. Since 2020, my country has accounted for 98% of global electrolytic manganese production. However, low current efficiency, high energy consumption, and environmental pollution pose significant technical and commercial challenges to the manganese electrolysis process. Because manganese electrolysis involves the combined effects of multiple reactions, including electrochemical, physicochemical, thermal, electrical, and magnetic reactions, it is a nonlinear industrial production system with complex coupling relationships and significant time delays. The instability of parameter measurements and the lack of responsiveness to multiple parameter information (Mn) further exacerbate the challenges. 2+ The comprehensive analytical capabilities of the system are limited, including concentration, (NH4)2SO4 concentration, current density, temperature, and pH value. Therefore, it lacks quantitative description of the electrolysis production process, analysis and prediction of electrolysis product quality and electrolysis energy consumption, and cannot guarantee the safe, efficient and energy-saving operation of the electrolysis process.
[0003] Currently, existing technologies for optimizing electrolytic manganese process parameters and predicting electrolytic energy consumption mainly fall into two categories: The first is the traditional mechanistic model method. Optimization of electrolytic manganese process parameters typically relies on mechanistic models based on electrochemical theories (such as the Butler-Volmer equation and the Nernst-Planck mass transfer model). Key parameters (such as exchange current density and diffusion coefficient) are fitted using experimental data, and then numerical simulations (such as COMSOL) are used to predict energy consumption and deposition effects. The disadvantages of this technique include poor model adaptability, requiring consideration of different electrolyte compositions (such as Mn²⁺). + Recalibrating parameters (such as concentration and H2SO4 content) is time-consuming and costly; the model struggles to achieve complex correlations and quantify the nonlinear coupling effects of multiple parameters (such as temperature, current density, and stirring speed), leading to prediction bias; the model lacks real-time performance, and traditional numerical simulations are computationally intensive and cannot support online process adjustments. The second category of existing technologies is classical statistical modeling methods. Some studies employ response surface methodology (RSM) or multiple linear regression (MLR) to establish empirical models of process parameters and energy consumption / deposition quality through experimental designs (such as Box-Behnken). This type of technology suffers from low-dimensional limitations and overfitting risks; for example, it is only suitable for local optimization of a small number of parameters (usually ≤4) and cannot handle high-dimensional data; it is sensitive to noise and has poor generalization ability under complex industrial conditions.
[0004] In actual manganese electrolysis production, decisions regarding complex electrolysis parameters are mostly made manually, with operators relying on their experience to determine adjustments. Therefore, the decision-making method for electrolysis parameters is highly dependent on the operator's experience, and adjustments can be unstable due to frequent staff turnover. Consequently, product quality often fails to meet requirements due to the limitations of manual operation of electrolysis parameters. Therefore, to reduce energy consumption and pollution while ensuring high-quality and efficient production, real-time monitoring, simulation prediction, and optimized control of electrolysis process parameters are critical issues that urgently need to be addressed. Summary of the Invention
[0005] The present invention aims to provide a method that can quantify the coupled effects of multiple parameters in the electrolytic manganese process in real time and accurately predict current efficiency and electrolysis energy consumption, thereby providing a direct and reliable data-driven decision-making basis for process optimization, and thus stabilizing production and reducing energy consumption.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for multi-parameter analysis and electrolysis energy consumption prediction of electrolytic manganese process based on machine learning, characterized by comprising the following steps: S1. Based on the conventional process parameter range of the electrolytic manganese electrolysis process, design an orthogonal experiment to obtain a dataset of current efficiency and electrolysis energy consumption data; divide the dataset into a training set (70%), a validation set (15%), and a test set (15%) in chronological order. S2. Expand the dataset from step S1 by conducting the same orthogonal experiments on electrolysis processes with different electrolysis durations. S3. Based on step S2, establish a machine learning model for electrolysis process parameters and electrolysis energy consumption indicators; S4. Based on the model determination coefficient R², budgeted accuracy error MSE, and weighted average absolute percentage error. w MAPE and timing response characteristics are used to select a machine learning model. The optimized model is then used to quickly and accurately predict CE and EC for new combinations of process parameters, providing a basis for process optimization and energy-saving control.
[0007] Preferably, as an improvement, the orthogonal experiment in step S1 is L25(5 5 A five-factor, five-level orthogonal experiment was conducted, with the experimental variable being Mn. 2+ Concentration, (NH4)2SO4 concentration, current density, temperature, pH value.
[0008] Preferably, as an improvement, the Mn 2+The concentrations were 25 g / L, 30 g / L, 35 g / L, 40 g / L, and 45 g / L; the (NH4)2SO4 concentrations were 100 g / L, 105 g / L, 110 g / L, 115 g / L, and 120 g / L; and the current densities were 300 A / m. 2 325 A / m 2 350 A / m 2 375 A / m 2 400 A / m 2 The temperatures were 25℃, 30℃, 35℃, 40℃, and 45℃, respectively; the pH values were 6.5, 7.0, 7.5, 8.0, and 8.5, respectively.
[0009] Preferably, as an improvement, the electrolysis time in step S2 is 30 minutes, 60 minutes and 120 minutes respectively.
[0010] Preferably, as an improvement, the machine learning model in step S3 includes: adaptive boosting algorithm, bagging algorithm, decision tree, extremely random tree, gradient boosting regression, K-nearest neighbor algorithm, random forest, support vector regression, extreme gradient boosting, and light gradient boosting decision tree.
[0011] Preferably, as an improvement, the model determination coefficient R in step S4 is... 2 Mean Square Error (MSE) and Weighted Average Absolute Percentage Error (MAS) w The formulas for calculating MAPE are as follows: , , , Where R 2 The coefficient of determination for the model. MSE Mean square error, w MAPE is the weighted average absolute percentage error. n For the number of samples, i For the first i One sample, y i It is the first i The true value of each sample For the predicted values of the machine learning model, This is the average of the true values of all samples.
[0012] Preferably, as an improvement, the calculation formulas for the electrolysis energy consumption EC and current efficiency CE are as follows: , 100%, EC For electrolysis energy consumption, WHAT For current efficiency,U For the electrolytic cell pressure, q The electrochemical equivalent of manganese, WHAT For current efficiency, Δm The mass of electrolytic manganese obtained for each sample test, I The current applied for electrolysis, t This refers to the electrolysis time.
[0013] Preferably, as an improvement, the optimal machine learning model selected in step S4 is a gradient boosting regression model or a random forest model.
[0014] Secondly, the present invention also provides an electrolytic manganese process optimization system, comprising: The data acquisition module is used to acquire electrolysis process parameter data; The prediction module has a built-in machine learning model trained using any of the above methods, which is used to predict current efficiency and electrolysis energy consumption based on the input process parameters. The optimization suggestion module is used to output process parameter adjustment suggestions based on the prediction results.
[0015] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the methods described above.
[0016] This invention constructs a machine learning prediction model for the manganese electrolysis process, systematically revealing the complexity of predicting manganese electrolysis process parameters and the differences in model applicability. The proposed data-driven manganese electrolysis process optimization strategy and performance prediction research approach can also be applied to other electrolysis systems. Compared with existing technologies, this invention has the following advantages: Based on ten representative machine learning algorithms, a multi-algorithm comparison framework was constructed, covering a variety of models from baseline methods (such as KNN and SVR) to advanced ensemble learning (such as XGBoost and LGBM). After optimization using Bayesian optimization and grid search methods, the GBR model improved in prediction... WHAT and EC They performed best in this aspect.
[0017] High prediction accuracy: Through system comparison and optimization, the adopted machine learning models (such as GBR) demonstrate excellent performance in both current efficiency prediction and energy consumption prediction. The R² value for current efficiency prediction is 0.822, and the MSE is 0.0044; the R² value for energy consumption prediction is 0.832, and the MSE is 0.0455. This algorithm not only improves the accuracy and efficiency of current efficiency prediction but also maintains optimal error control in energy consumption prediction. This provides reliable data-driven decision support for precise control, energy consumption reduction, and quality optimization in the electrolytic manganese production process.
[0018] Fast response speed: The optimized model has a fast response speed, meeting the needs of real-time or near-real-time prediction and control in industrial settings, unlike purely academic simulations. This provides a feasible path for the electrolytic manganese industry to introduce an AI-driven quality monitoring system, which is expected to reduce energy consumption and scrap rates, and has significant engineering innovation value.
[0019] Handling complex nonlinear relationships: It can effectively learn and characterize the complex nonlinear coupling relationship between multiple process parameters and target indicators in the electrolysis process, overcoming the limitations of traditional mechanism models and statistical methods.
[0020] Based on the above high-precision model, this invention enables the combination of process parameters at any moment during the production process to be instantly converted into two key performance indicators (KPIs). WHAT and EC This provides a quantitative forecast of the production process. This completely changes the previous reliance on manual qualitative judgment or offline model estimation, and provides a continuous, digital forecasting solution for the production process.
[0021] High generalization and scalability: The data-driven process optimization framework provided by this invention can be transferred to other similar electrolysis process systems and has broad application prospects.
[0022] Guiding production optimization: By accurately predicting energy consumption and efficiency under different parameter combinations, direct decision support can be provided for reducing production costs, improving product quality, and achieving green and energy-saving production. Attached Figure Description
[0023] Figure 1 : Trend charts of current efficiency (Y1) and electrolysis energy consumption (Y2) predicted by RF and GBR models. Detailed Implementation
[0024] The following detailed description illustrates the specific implementation method: Example 1: Dataset Construction Electrolysis experiments were conducted according to the five-factor, five-level orthogonal experimental design table shown in Table 1. Electrolysis experiments were performed for three durations: 30, 60, and 120 minutes, with each experiment repeated three times and the average value taken. The deposited manganese mass and cell voltage were recorded for each experiment, and the current efficiency was calculated using the formula. WHAT and electrolysis energy consumption EC Some experimental results are shown in Table 2 (30 minutes), Table 3 (60 minutes), and Table 4 (120 minutes). The final dataset contains 75 sets of samples (25 sets per duration) and is divided into training set (70%), validation set (15%), and test set (15%) in chronological order.
[0025] Table 1. Factor Levels of Orthogonal Experiment
[0026] Table 2. Results of manganese electrolytic deposition - 30 min
[0027] Table 3. Results of manganese electrolytic deposition - 60 min
[0028] Table 4. Results of manganese electrolytic deposition - 120 min
[0029] Example 2: Model Training and Selection Use Python to build 10 machine learning models: AdaBoost, Bagging, Decision Tree (DTS), Extremely Random Tree (ET), Gradient Boosting Regression (GBR), K-Nearest Neighbors (KNN), Random Forest (RF), Support Vector Regression (SVR), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Decision Tree (LBGM).
[0030] With Mn 2+ Five process parameters—concentration (X1), (NH4)2SO4 concentration (X2), current density (X3), temperature (X4), and pH value (X5)—are used as input features, with current efficiency (Y1) as the input. WHAT and (Y2) EC To achieve the output target, the training time was used to purchase 10 machine learning models. The determination coefficients (R²) of each model were calculated using the validation set. 2 Mean squared error (MSE), weighted average absolute percentage error (...) w Based on the timing response characteristics of MAPE and other performance indicators, the optimal machine learning model is selected; the performance comparison is shown in Table 5.
[0031] , , , Where R 2 The coefficient of determination for the model. MSE Mean square error, w MAPE is the weighted average absolute percentage error. n For the number of samples, i For the first i One sample, y i It is the first i The true value of each sample For the predicted values of the machine learning model, This is the average of the true values of all samples.
[0032] Preferably, as an improvement, the electrolysis energy consumption... EC and current efficiency WHAT The calculation formulas are as follows: , 100%, EC For electrolysis energy consumption, WHAT For current efficiency, U For the electrolytic cell pressure, q The electrochemical equivalent of manganese, WHAT For current efficiency, Δm The mass of electrolytic manganese obtained for each sample test, I The current applied for electrolysis, t This refers to the electrolysis time.
[0033] Table 5. Comparison of performance indicators of different machine learning models for predicting manganese deposition processes.
[0034] Table 5 compares the performance metrics of various machine learning models, listing four evaluation metrics for different model performances, including the coefficient of determination (R²). 2 Mean squared error (MSE), weighted average absolute percentage error (...) w MAPE and response time. Regarding model accuracy, in the Y1 prediction task, the Random Forest (RF) model showed the best performance, with an R² of approximately 0.87, followed by the Gradient Boosting Regression (GBR) model (R² of approximately 0.82). In the Y2 prediction task, the Random Forest (RF) model showed the best performance, with an R² of approximately 0.85, followed by the Gradient Boosting Regression (GBR) model (R² of approximately 0.83). The mean squared error (MSE) of AdaBoost, DTS, and LGBM ranged from 0.004 to 0.008, indicating that they belong to the low-error model category. Most models predicting Y1... w The MAPE values are concentrated in the range of 11% to 16%, highlighting the challenge of current efficiency prediction. For Y2, most models... w The MAPE distribution falls within the range of 6% to 9%. Comprehensive analysis shows that GBR exhibits the lowest performance in both prediction tasks. w MAPE values. XGBoost, Bagging, DTS, ET, GBR, KNN, RF, SVR, and XGBoost were categorized as efficient algorithms with response times below 0.2 seconds, demonstrating good computational efficiency and suitability for real-time prediction needs. After cross-validation across multiple dimensions, the GBR algorithm showed the best overall performance in this study.
[0035] Based on the experimental results in Table 5, the GBR and RF algorithms were selected as the two optimal candidate algorithms for predicting manganese electrodeposition process parameters, and a detailed comparative analysis was conducted. The comparative analysis results of the two models' prediction trends for current efficiency (Y1) and electrolysis energy consumption (Y2) are as follows: Figure 1 As shown.
[0036] Figure 1 The prediction results for the current efficiency Y1 from the two models are shown. For example... Figure 1 As shown in Figure a, both algorithms captured the fluctuation trend of current efficiency, but the gradient boosting regression model (GBR, green curve) showed a better fit between its predicted curve and the true value (blue curve) than the random forest model (RF, orange curve). Particularly near the trough in samples 12-13 and the peak near sample 15, GBR's predictions were closer to the true values, while RF's prediction bias was more pronounced (e.g., RF significantly overestimated in samples 12-13 and significantly underestimated in sample 15). Within the high-frequency fluctuation range of samples 5-10, RF's prediction amplitude was slightly higher than GBR's, indicating slightly weaker stability in fitting local details. In terms of goodness of fit, RF's R² was 0.874, while GBR's was 0.822 (as shown in Table 5), indicating that RF could explain approximately 87.4% of the current efficiency variation, improving the fit by over 100%. This result demonstrates that RF can more effectively learn the complex mapping relationship between process parameters and current efficiency, showing significant advantages in both prediction accuracy and model interpretability. In terms of prediction accuracy, the mean squared error (MSE) of the GBR algorithm is 0.0044, which is slightly lower than that of the RF algorithm (0.0047) (as shown in Table 5), corresponding to an error reduction of about 6.4%.
[0037] In the energy consumption prediction task Y2, both algorithms demonstrated robust modeling capabilities. For example... Figure 1 As shown in b, GBR (green curve) outperforms RF (orange curve) in energy consumption prediction. In the large fluctuation range of samples 15-18 (where the actual value rises rapidly from a trough to a peak), GBR's prediction curve closely matches the actual value, while RF shows a significant lag (e.g., RF underestimates the peak at sample 17). In the relatively stable range of samples 2-10, the prediction differences between the two algorithms are small, but RF's prediction is slightly higher than the actual value, indicating a slight systematic overestimation. GBR's MSE is 0.0455, lower than RF's 0.0503, indicating a performance improvement of approximately 9.5%. In terms of R², GBR scores 0.830, and RF scores 0.855, showing fairly similar goodness of fit. It is noteworthy that although RF's R² is slightly higher than GBR's (an improvement of approximately 6.8%), GBR's superior performance on MSE means its prediction error is smaller and closer to the actual predicted value.
[0038] GBR model optimization parameter configuration: Grid search parameter space param_grid_gbr = { 'n_estimators': [100, 200, 300, 500], 'learning_rate': [0.01, 0.05, 0.1, 0.2], 'max_depth': [3, 5, 7, 9], 'min_samples_split': [2, 5, 10], 'min_samples_leaf': [1, 2, 4], 'subsample': [0.6, 0.8, 1.0], # Sample sampling ratio 'max_features': ['sqrt', 'log2', 0.5, None], 'loss': ['squared_error', 'absolute_error', 'huber'] # Loss function } # Total search space: 4×4×4×3×3×3×4×3 = 10368 combinations Bayesian optimization of parameter space param_space_bayes = { 'n_estimators': Integer(220, 380), 'learning_rate': Real(0.04, 0.12, prior='log-uniform'), 'max_depth': Integer(5, 8), 'min_samples_split': Integer(4, 9), 'subsample': Real(0.7, 0.95), 'max_features': Real(0.5, 0.8), 'alpha': Real(0.1, 0.9) # Quantile parameters of Huber loss } Prior design: learning_rate~LogUniform(0.04,0.12): Log-scaled search learning rate n_estimators~N(300,50): Grid search shows peak values in the 250-350 range. Performance comparison before and after optimization Y1 (Current Efficiency Prediction)
[0039] Y2 (Energy Efficiency Prediction)
[0040] RF model optimization parameter configuration: Grid search parameter space param_grid_rf = { 'n_estimators': [50, 100, 200, 300, 400], # Number of trees 'max_depth': [5, 10, 15, 20, None], # Maximum depth 'min_samples_split': [2, 5, 10, 15], # Minimum number of samples for node splitting 'min_samples_leaf': [1, 2, 4, 8], # Minimum number of samples in the leaf node 'max_features': ['sqrt', 'log2', 0.5, 0.7, None], # Feature sampling strategy 'criterion': ['squared_error', 'absolute_error', 'friedman_mse'], #Splitting criterion 'bootstrap': [True, False], # Bootstrap sampling 'oob_score': [True, False], # Out-of-bag score 'max_samples': [0.7, 0.8, 0.9, None]# Sample sampling ratio } # Total search space: 5×5×4×4×5×3×2×2×4 = 19200 combinations Bayesian optimization of parameter space from skopt.space import Real, Integer, Categorical param_space_bayes = { 'n_estimators': Integer(150, 250), # The valid range defined by the grid 'max_depth': Integer(12, 18), 'min_samples_split': Integer(4, 12), 'min_samples_leaf': Integer(2, 5), 'max_features': Real(0.45, 0.65), # Continuously optimize the feature ratio 'max_samples': Real(0.75, 0.95), 'min_impurity_decrease': Real(1e-5, 1e-3, prior='log-uniform'), # Pruning threshold 'ccp_alpha': Real(0.0, 0.01) # Cost and complexity pruning } Prior distribution design: # Constructing Priors Based on Grid Search Results grid_results = { 'n_estimators': [50, 100, 200, 300], 'R2': [0.82, 0.85, 0.90, 0.92] } # Fitting a Gaussian distribution mu = 200, sigma = 30 prior_n_estimators = Normal(mu=200, sigma=30) # Verification: 200±30 coverage [170,230], including 90% of the high-performance area Performance comparison before and after optimization Y1 (Current Efficiency Prediction)
[0041] Y2 (Energy Efficiency Prediction) 。
Claims
1. A method for multi-parameter analysis and electrolysis energy consumption prediction in a manganese electrolysis process based on machine learning, characterized in that, Includes the following steps: S1. Design an orthogonal experiment to obtain a dataset of current efficiency and electrolysis energy consumption data according to the conventional process parameter range of the electrolytic manganese electrolysis process. S2. Expand the dataset from step S1 by conducting the same orthogonal experiments on electrolysis processes with different electrolysis durations. S3. Based on step S2, establish a machine learning model for electrolysis process parameters and electrolysis energy consumption indicators; S4. Based on the model determination coefficient R², budgeted accuracy error MSE, and weighted average absolute percentage error. w MAPE and time-series response characteristics are used to select machine learning models.
2. The method according to claim 1, characterized in that, The orthogonal experiment mentioned in step S1 is L25(5 5 A five-factor, five-level orthogonal experiment was conducted, with the experimental variable being Mn. 2+ Concentration, (NH4)2SO4 concentration, current density, temperature, pH value.
3. The method according to claim 2, characterized in that, The Mn 2+ The concentrations were 25 g / L, 30 g / L, 35 g / L, 40 g / L, and 45 g / L; the (NH4)2SO4 concentrations were 100 g / L, 105 g / L, 110 g / L, 115 g / L, and 120 g / L; and the current densities were 300 A / m. 2 325 A / m 2 350 A / m 2 375 A / m 2 400 A / m 2 The temperatures were 25℃, 30℃, 35℃, 40℃, and 45℃, respectively; the pH values were 6.5, 7.0, 7.5, 8.0, and 8.5, respectively.
4. The method according to claim 1, characterized in that, The electrolysis times in step S2 are 30 minutes, 60 minutes, and 120 minutes, respectively.
5. The method according to claim 1, characterized in that, The machine learning models mentioned in step S3 include: adaptive boosting algorithm, bagging algorithm, decision tree, extremely random tree, gradient boosting regression, K-nearest neighbor algorithm, random forest, support vector regression, extreme gradient boosting, and light gradient boosting decision tree.
6. The method according to claim 1, characterized in that: The model determination coefficient R mentioned in step S4 2 Mean Square Error (MSE) and Weighted Average Absolute Percentage Error (MAS) w The formulas for calculating MAPE are as follows: ， ， ， Where R 2 The coefficient of determination for the model. MSE Mean square error, w MAPE is the weighted average absolute percentage error. n For the number of samples, i For the first i One sample, y i It is the first i The true value of each sample For the predicted values of the machine learning model, This is the average of the true values of all samples.
7. The method according to claim 1, characterized in that: The electrolysis energy consumption EC and current efficiency CE The calculation formulas are as follows: ， 100%， EC For electrolysis energy consumption, CE For current efficiency, U For electrolytic cell pressure, q The electrochemical equivalent of manganese, CE For current efficiency, Δm The mass of electrolytic manganese obtained for each sample test, I The current applied for electrolysis, t This refers to the electrolysis time.
8. The method according to claim 1, characterized in that, The optimal machine learning model selected in step S4 is either a gradient boosting regression model or a random forest model.
9. A process optimization system for electrolytic manganese, characterized in that, include: The data acquisition module is used to acquire electrolysis process parameter data; The prediction module has a built-in machine learning model trained according to any one of the methods described in claims 1 to 8, which is used to predict current efficiency and electrolysis energy consumption based on the input process parameters. The optimization suggestion module is used to output process parameter adjustment suggestions based on the prediction results.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.