A clean coal ash content online prediction method based on bayesian hyperparameter optimization lightgbm regression

Abstract

The present application relates to a kind of based on the Light GBM regression of bayesian hyperparameter optimization clean coal ash online prediction method, the purpose is to solve the technical problem of existing clean coal ash prediction inefficiency, technical scheme is: real-time data acquisition, data reading and cleaning, feature selection, normalization processing, using Light GBM regression algorithm and bayesian hyperparameter optimization constructs training model;Data tailings ash prediction value and clean coal ash measured value are compared and calculated, and compensation model is obtained;Using compensation model to evaluate the training model after optimization, according to the evaluation result, further model adjustment is carried out;The present application combines Light GBM model iteration and bayesian hyperparameter optimization, improves performance from three levels of data processing, model progression and algorithm tuning, so that the model can most comprehensively extract all relevant data learning and improve performance, so that the model can accurately predict the clean coal ash under different working conditions, so that production personnel can timely and accurately make reagent adjustment.

Description

Technical Field

[0001] This invention belongs to the field of coal slime flotation technology, specifically relating to an online prediction method for clean coal ash content based on LightGBM regression with Bayesian hyperparameter optimization. Background Technology

[0002] Driven by the rapid development of artificial intelligence technology and policies promoting intelligent coal mining, intelligent coal sorting is a growing trend. However, in the core process of flotation, namely the sorting of fine coal slime (-0.5mm), the progress of intelligent research has been relatively slow. This is mainly attributed to the difficulty in predicting the ash content of coal slime in real time online.

[0003] Coal slime flotation is a mineral processing method that separates fine-grained associated or mixed minerals based on differences in the physical and chemical properties of mineral surfaces. Since the commercialization of flotation technology in the early 20th century, it has been responsible for separating fine-grained coal slime (-0.5 mm). In recent years, with the sharp decline in high-quality coal reserves and the widespread use of hydrocyclones, the proportion of flotation in coal washing processes has been increasing year by year.

[0004] Ash content is the most important indicator for evaluating coal slime flotation products, used directly or indirectly to calculate flotation performance indicators and product quality. Currently, the national standard for ash content testing (i.e., ash content analysis) has a complex and time-consuming process. In China, most ash content testing methods are offline, involving sampling of flotation concentrate and tailings at fixed intervals, followed by sample preparation, ash burning, weighing, and calculation in the laboratory. This is typically done every half hour to an hour, delaying the adjustment window for flotation process control. As a result, flotation operators mainly rely on the visual characteristics of flotation froth / tailings to adjust process variables in the flotation machine. However, this experience-based method has a large margin of error, easily leading to losses of flotation concentrate and reagents. Furthermore, because ash content data is provided too slowly, changes after the addition of flotation reagents cannot be measured in real time. This large lag process can only be estimated by waiting for time and visual inspection, significantly reducing the research and advancement of intelligent flotation processes.

[0005] Radioactive isotope technology has also been used in some studies to detect ash content in slurries, but the safety and environmental constraints of radioactive sources limit the application prospects of this method.

[0006] The prediction of ash content in tailings images is also constantly advancing. In flotation production, the operator often relies on the flotation operator's judgment of the flotation foam and tailings color. However, during industrial site image acquisition, factors such as vibration, natural light, ambient light sources, temperature, and humidity inevitably cause a series of problems in the acquired foam images, affecting the accuracy of extracting flotation foam image features and ultimately impacting the accuracy of flotation foam image classification and ash content prediction. However, since the flotation feed conditions can be obtained through sensors, reagent consumption can be quantified by the dosing device, and tailings ash content can be inferred from the tailings ash content using the aforementioned methods, a suitable mathematical model can be established to predict the clean coal ash content after knowing the total amount of coal slime fed into the flotation, the amount of reagent added, and the tailings condition. This allows for online prediction of clean coal ash content.

[0007] Furthermore, the realization of intelligent control of the flotation system also heavily relies on the feedback adjustment of the ash content of the flotation clean coal / tailings. With the establishment of the goal of intelligent construction of coal preparation plants, high efficiency, environmental protection and online prediction of ash content in the flotation process are the top priorities for the intelligent construction of coal slime flotation process. Summary of the Invention

[0008] The purpose of this invention is to solve the above-mentioned problems and provide an online prediction method for clean coal ash content based on Bayesian hyperparameter optimization and LightGBM regression.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] An online prediction method for clean coal ash content based on Bayesian hyperparameter optimization and Light GBM regression includes the following steps:

[0011] Step S1: Real-time data acquisition: Collect real-time data on the PLC as a training set. The data includes ash content setpoint, feed concentration, flow rate, aeration valve opening, clean water input, dry coal slime quantity, collector input, frother input, diesel input, tailings ash content, and flotation cell level.

[0012] The AI ​​slurry ash analyzer was used to obtain the predicted ash content of tailings, and the predicted ash content of tailings was used as the training set and test set.

[0013] When the coal feed is stable, the required clean coal ash content is taken every 20 minutes for laboratory testing. When the coal feed is unstable, the clean coal ash content is taken every 5 minutes. The samples are numbered, stored, and tested in real time. The results are used as a test set.

[0014] Step S2, Data Reading and Cleaning: The predicted tailings ash content obtained by the AI ​​slurry ash analyzer in step S1 is used as the output value, and the real-time collected training set data is used as the input value for cleaning and removing outliers.

[0015] Step S3, Feature Selection: Perform correlation analysis and sort the data input values ​​obtained in step S2, and then perform feature selection to select the top five variables with the highest correlation as input values.

[0016] Step S4, Normalization: Normalize the data that has undergone feature selection in step S3;

[0017] Step S5: Construct the training model: Construct the training model using the Light GBM regression algorithm and Bayesian hyperparameter optimization;

[0018] The specific steps for building and training a model are as follows:

[0019] Step S5.1: At the beginning of training, the value range of each feature is discretized to form a histogram;

[0020] Step S5.2: The histogram divides the continuous values ​​of the feature into multiple discrete bins. Each bin stores the number of samples of the feature in the training set, the gradient, and the sum of squared gradients.

[0021] Step S5.3: Construct a decision tree: Directly manipulate the feature histogram using the Light GBM regression algorithm; calculate the gradient and sum of squared gradients for each bin, and then evaluate the gain of each bin as a split point based on these statistics.

[0022] Step S5.4: Define the hyperparameter space: After determining and selecting the corresponding split nodes and parameters based on the above architecture, define the hyperparameter space. Each hyperparameter has a search range or value range.

[0023] Step S5.5: Define the objective function of Bayesian optimization: The input of this function is a combination of hyperparameters, and the output is the performance index of the model on the training data. This objective function will be called in each iteration of Bayesian optimization. The purpose is to train the Light GBM model according to the given combination of hyperparameters and return the performance index.

[0024] Step S5.6: After defining the objective function, in each iteration, the hyperparameters of the Light GBM regression algorithm are tuned using the Bayesian optimization algorithm, and the performance of the optimal hyperparameter combination is verified by cross-validation and validation set evaluation methods.

[0025] Step S5.7: Set the stopping condition: Stop the Bayesian optimization process when the model performance can no longer be improved during the optimization process;

[0026] Step S5.8: Retrain the Light GBM model using the final optimized hyperparameter configuration obtained in step S5.7;

[0027] Step S6: Compare and calculate the predicted tailings ash content and the measured clean coal ash content of the test set data obtained in step S1 to obtain the compensation model.

[0028] Step S7: Evaluate the optimized training model using the compensation model;

[0029] Step S8: Make further model adjustments based on the evaluation results. When the model's evaluation index R-squared reaches 0.9 or higher and no longer improves, output the results.

[0030] Furthermore, the hyperparameters in step S5.4 include the learning rate, the number of trees, the number of leaf nodes, and the depth of the trees.

[0031] Furthermore, the performance metrics of the model on the training data in step S5.5 include accuracy and AUC.

[0032] Furthermore, in step S6, if the difference between the two values ​​is greater than 5%, the measured value of the clean coal ash content is selected; if the difference between the two values ​​is 2% ≤ the difference value ≤ 5%, the average value of the two values ​​is taken; if the difference between the two values ​​is less than 2%, the predicted value of the tailings ash content is directly selected.

[0033] Furthermore, the evaluation and comparison indicators in step S7 include mean square error, root mean square error, and coefficient of determination.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] 1. This invention combines Light GBM model iteration and Bayesian hyperparameter optimization to comprehensively improve performance from three aspects: data processing, model advancement and algorithm tuning. This allows the model to extract and learn all relevant data in the most comprehensive way and improve performance, enabling the model to accurately predict the ash content of clean coal under different working conditions, allowing production personnel to make timely and accurate adjustments to the reagents.

[0036] 2. The model trained based on real-time on-site data can adapt to the ever-changing data conditions on-site. The ash content value is determined based on the real-time collected concentration flow rate, reagent dosage, and air valve opening, etc. In addition, the compensation model calculated by comparing the ash content given by the AI ​​slurry ash analyzer with the measured ash content can improve the accuracy of the predicted value of clean coal ash content. It can also make the prediction of tailings ash content more accurate through feedback, so that laboratory tests are not required so frequently in daily production. It can realize the rapid and accurate prediction of clean coal ash content on the flotation site and guide the on-site decision-making changes in the coal slime flotation process.

[0037] 3. The prediction method used in this invention can quickly process large amounts of data and change the number of variables in real time according to the on-site production conditions. It is a dynamic adaptive process. In addition, the continuous iteration of Light GBM and the continuous optimization of Bayesian hyperparameters make the model always dynamically updated. Moreover, the biggest advantage of Light GBM is its efficient training and prediction speed, which can minimize the loss of other useful variables, making the model's prediction more accurate and timely, and minimizing the economic losses caused by anomalies in production. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the clean coal ash content prediction process in the flotation process of the present invention; Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] The Light GBM regression algorithm gradually reduces the prediction error by iteratively training multiple weak models and combining their prediction results, thereby building a more powerful overall model to achieve online prediction of clean coal ash content. Light GBM is an efficient gradient boosting framework, mainly used to handle large-scale datasets and efficient training. It improves the performance of Gradient Boosting Decision Tree (GBDT) by reducing training time and memory consumption.

[0041] Bayesian hyperparameter optimization methods involve selecting the hyperparameters that best perform on the surrogate model to evaluate the actual objective function, then continuously updating the surrogate model to obtain new hyperparameter results, and re-evaluating the objective function until the maximum number of iterations is reached.

[0042] like Figure 1 As shown, an online prediction method for clean coal ash content based on Bayesian hyperparameter optimization and Light GBM regression includes the following steps:

[0043] Step S1: Real-time data acquisition: Use Kep Server to collect real-time data on the PLC of the coal preparation plant as a training set. The data includes ash content setpoint, feed concentration, flow rate, aeration valve opening, clean water input, dry coal slime quantity, collector input, frother input, diesel input, tailings ash content, and flotation cell level.

[0044] The AI ​​slurry ash analyzer was used to obtain the predicted ash content of tailings, and the predicted ash content of tailings was used as the training set and test set.

[0045] When the coal feed is stable, the required clean coal ash content is taken every 20 minutes for laboratory testing. When the coal feed is unstable, the clean coal ash content is taken every 5 minutes. The samples are numbered, stored, and tested in real time. The results are used as a test set.

[0046] The flotation feed flow rate and concentration, reagent dosage, aeration valve opening, and tailings ash content prediction values ​​from the AI ​​slurry analyzer all have varying degrees of impact on the prediction of clean coal ash content. Specifically:

[0047] The tailings ash content measured by the AI ​​slurry ash analyzer exhibits a very strong linear relationship when the feed ash content remains relatively stable. The specific relationship is as follows:

[0048]

[0049] Among them, Ad γ Ad J Ad W These represent the ash content of raw coal, clean coal, and tailings from flotation, respectively. γ J and γ W These represent the yield of clean coal and the yield of tailings from flotation, respectively. During the flotation process, the ash content of the raw coal generally does not fluctuate significantly, while the yields of clean coal and tailings can be monitored in real time using belt scales. Therefore, after simplification, the following linear relationship can be derived:

[0050] Ad J =K × Ad W + B

[0051] As can be seen from the above, there is a correlation between the ash content of tailings and the ash content of clean coal, and this correlation can be evaluated by analyzing the grayscale feature values ​​of tailings images.

[0052] (1) The flotation feed concentration and flotation feed flow rate reflect the current processing capacity of the flotation system in real time. When the processing capacity is within the range that the flotation machine can handle, the ash content of the clean coal is relatively low. In actual implementation, if it exceeds the processing capacity of the flotation machine (i.e., the processing capacity of feed concentration and feed flow rate is too large), it will lead to the loss of flotation clean coal and the ash content of clean coal will also increase. When the processing capacity is too small, it is easy to mix minerals into the clean coal, which will also lead to the increase of the ash content of clean coal.

[0053] (2) The opening degree of the aeration valve is related to the amount of aeration, which is directly related to the amount of bubbles in the flotation tank and the amount of clean coal carried out. In actual implementation, if the amount of aeration is too large, some minerals that should have settled will be scraped out with the foam, which will increase the ash content of the clean coal. If the amount of aeration is too small, the clean coal will not be fully carried out by the foam, resulting in the loss of clean coal with the tailings.

[0054] (3) The dosage of reagents includes the dosage of collector and frother. Generally, there is a certain proportional relationship between the two. Excessive collector may cause excessive coverage of the ore surface, resulting in excessive adhesion between ore particles, making it difficult to separate the ore particles, reducing the amount of useful minerals in the flotation product, and increasing the ash content of the clean coal. Insufficient collector will also prevent the clean coal from adhering sufficiently and will also make the foam unstable, leading to an increase in the ash content of the clean coal. Excessive frother will produce a large number of sticky and stable bubbles, making the foam exist in the flotation cell for a long time and difficult to break. A large amount of gangue minerals will also adhere to the foam, leading to an increase in the ash content of the clean coal. Similarly, insufficient frother will result in bubbles that are too small and uneven, affecting the contact effect between the bubbles and the ore, leading to insufficient recovery of clean coal and an increase in the ash content of the clean coal.

[0055] Step S2, Data Reading and Cleaning: The predicted tailings ash content obtained by the AI ​​slurry ash analyzer in step S1 is used as the output value, and the real-time collected training set data is used as the input value for cleaning and removing outliers.

[0056] Step S3, Feature Selection: Perform correlation analysis and sort the data input values ​​obtained from Step S2, and then perform feature selection to select the top five variables with the highest correlation as input values. The collected variables include ash content setpoint, concentration, flow rate, aeration valve opening, clean water setpoint, dry coal slime quantity, collector setpoint, frother setpoint, diesel setpoint, tailings ash content, and flotation cell level.

[0057] Step S4, Normalization: Normalize the data that has undergone feature selection in step S3;

[0058] Step S5: Construct the training model: Construct the training model using the Light GBM regression algorithm and Bayesian hyperparameter optimization;

[0059] The specific steps for building and training a model are as follows:

[0060] Step S5.1: At the start of training, the value range of each feature is discretized. Light GBM uses a histogram to represent the discretization of features.

[0061] Step S5.2: The histogram divides the continuous values ​​of the feature into multiple discrete bins. Each bin stores the number of samples of the feature in the training set, the gradient, and the sum of squared gradients.

[0062] Step S5.3: Constructing a decision tree: The Light GBM regression algorithm is used to directly manipulate the feature histogram; the gradient and sum of squared gradients of each bin are calculated, and then the gain of each bin as a split point is evaluated based on these statistical information; due to the characteristics of Light GBM, when constructing a decision tree, it is not necessary to sort the original data, but to directly manipulate the feature histogram. The characteristic of the Light GBM algorithm is that it automatically selects the feature with the largest current gain and the best split point to split the node.

[0063] Step S5.4: Define the hyperparameter space: After determining and selecting the corresponding split nodes and parameters based on the above architecture, define the hyperparameter space, which includes the learning rate, the number of trees, the number of leaf nodes, and the depth of the trees. Each hyperparameter has a search range or value range.

[0064] Step S5.5: Define the objective function of Bayesian optimization: The input of this function is a combination of hyperparameters. The objective function will use minimizing the mean squared error (MSE) and maximizing the coefficient of determination (R-squared). The output is the performance index of the model on the training data. This objective function will be called in each iteration of Bayesian optimization. The purpose is to train the Light GBM model according to the given combination of hyperparameters and return performance indexes such as accuracy and AUC.

[0065] Step S5.6: After defining the objective function, in each iteration, the hyperparameters of the Light GBM regression algorithm are tuned using the Bayesian optimization algorithm, such as the learning rate, tree depth, minimum number of samples in the leaf nodes, etc., and the performance of the optimal hyperparameter combination is verified by cross-validation and validation set evaluation methods.

[0066] Step S5.7: Set the stopping condition: Stop the Bayesian optimization process when the model performance can no longer be improved during the optimization process;

[0067] Step S5.8: Retrain the Light GBM model using the final optimized hyperparameter configuration obtained in step S5.7;

[0068] Step S6: Compare and calculate the predicted tailings ash content and the measured clean coal ash content of the test set data obtained in Step S1 to obtain the compensation model: Since the sample of measured clean coal ash content is small and the sampling time of PLC is three seconds, there will be nearly one hundred identical values, which is not conducive to training the model. If the difference between the two is >5%, the measured clean coal ash content will be selected. If the difference between the two is 2% ≤ 5%, the average value of the two will be taken. If the difference between the two is <2%, the predicted tailings ash content will be selected directly.

[0069] Step S7: Use the compensation model to evaluate the optimized training model. The evaluation and comparison indicators include mean square error, root mean square error, coefficient of determination, etc.

[0070] Step S8: Make further model adjustments based on the evaluation results. When the model's evaluation index R-squared reaches 0.9 or higher and no longer improves, output the results.

[0071] The working principle of this invention is as follows:

[0072] In the flotation process, the variables are numerous and the data is complex, exhibiting intricate nonlinear relationships, and the data is characterized by irregularity and high noise. Prediction under these conditions is a typical problem that Light GBM can handle. The fundamental challenge in predicting the ash content of refined coal lies in learning from the complex data to build a comprehensive model. However, general regression prediction models often only address a specific type of data. For large, disorganized data with complex nonlinear relationships, traditional decision tree algorithms (such as CART) require sorting the data to find the optimal split point at each node. Sorting on large-scale datasets is extremely time-consuming, especially under memory constraints. Histogram-based algorithms bin continuous feature values ​​into discrete integer values ​​and construct a histogram. Each bin stores the cumulative gradient statistics (such as the sum of gradients and the sum of squared gradients) for that feature value. Thus, when building the decision tree, there is no need to sort the original data; instead, these histograms are directly manipulated, significantly reducing computation and memory consumption.

[0073] LightGBM achieves fast and efficient training on large-scale datasets through techniques such as histogram-based decision tree algorithms, leaf-wise leaf growth strategies, feature parallelization, histogram parallelization, optimizers, and split search algorithms. Its design goal is to reduce training time and memory consumption while maintaining model performance. This makes LightGBM an ideal choice for handling large datasets and efficient training. Furthermore, parameter optimization during model iteration can be achieved through Bayesian hyperparameter optimization.

[0074] This invention combines model iteration and Bayesian hyperparameter optimization to comprehensively improve performance from three levels: data processing, model advancement, and algorithm tuning. This allows the model to extract and learn from all relevant data in the most comprehensive way, thereby improving performance. The model can accurately predict the ash content of clean coal under different operating conditions, enabling production personnel to make timely and accurate adjustments to reagents.

[0075] Specifically:

[0076] The first step is to solve the data processing problem. Kep ​​Server is used to collect relevant data on the PLC of the coal preparation plant, classify the massive data, select the inputs and outputs, clean up outliers and outliers, and then normalize them.

[0077] Based on the above data processing, Light GBM uses histograms to accelerate the decision tree construction process;

[0078] At the start of training, the range of values ​​for each feature is discretized to form a histogram. This histogram divides the continuous values ​​of a feature into multiple discrete bins, and each bin stores the number of samples for that feature in the training set, the gradient, and the sum of squared gradients.

[0079] The decision tree is then constructed. During this process, Light GBM does not need to sort the original data; instead, it directly manipulates the histograms of the features. For each feature, Light GBM calculates the gradient and the sum of squared gradients for each bin, and then uses these statistics to evaluate the gain of each bin as a split point. Instead of iterating through all possible feature values, Light GBM selects the feature with the largest gain and the optimal split point.

[0080] After determining the architecture of the Light GBM regression model and selecting the appropriate split nodes and parameters, a hyperparameter space is defined, containing all the hyperparameters of the Light GBM model we wish to optimize. These hyperparameters can include the learning rate, the number of trees, the number of leaf nodes, the tree depth, etc. Each hyperparameter has a search range or value range.

[0081] We also need to define an objective function. This function takes a combination of hyperparameters as input and outputs performance metrics of the model on the training data, such as accuracy or AUC. The objective function aims to train the Light GBM model based on a given combination of hyperparameters and return the performance metrics.

[0082] Next, we use Bayesian optimization algorithms, such as Gaussian processes or tree-based methods, to iteratively search for the optimal combination of hyperparameters in the hyperparameter space. Initially, the Bayesian optimization algorithm randomly selects an initial set of hyperparameter combinations and calculates their performance metrics using the objective function.

[0083] By integrating Bayesian hyperparameter optimization with Light GBM regression, the optimal combination of hyperparameters can be automatically selected, thereby improving the performance and generalization ability of the Light GBM regression model. This ensemble approach can more effectively search the hyperparameter space and find the best-performing hyperparameter combination more quickly, thus accelerating the model optimization process. Based on the currently known hyperparameter combinations and performance metrics, the Bayesian optimization algorithm infers the next most promising hyperparameter combination for evaluation. This process continues for multiple rounds until a preset number of iterations is reached or a certain convergence condition is met.

[0084] As an optimization process, Bayesian optimization is an iterative process, requiring control of the number of iterations to avoid overfitting or wasting computational resources. It is generally recommended to set an appropriate maximum number of iterations and dynamically adjust the search direction and scope of the next iteration based on observations in each iteration. Furthermore, the optimal hyperparameter combination found by Bayesian optimization should be validated and interpreted to ensure its effectiveness and generalization ability in practical problems. The performance of the optimal hyperparameter combination can be verified through methods such as cross-validation and validation set evaluation, and its impact on the model and improvement can be analyzed.

[0085] Through the above process, the Light GBM regression model based on Bayesian hyperparameter optimization is trained. The above method is then applied to the online prediction of ash content in flotation clean coal. When the operating conditions change or anomalies occur, the ash content of clean coal can be accurately predicted, enabling production personnel or automatic dosing systems to make judgments and adjust the given amount of reagents, thereby minimizing economic losses caused by abnormal situations.

Claims

1. A method for online prediction of clean coal ash content based on Light GBM regression with Bayesian hyperparameter optimization, characterized in that, Includes the following steps: Step S1: Real-time data acquisition: Collect real-time data on the PLC as a training set. The data includes ash content setpoint, feed concentration, flow rate, aeration valve opening, clean water input, dry coal slime quantity, collector input, frother input, diesel input, tailings ash content, and flotation cell level. The AI ​​slurry ash analyzer was used to obtain the predicted ash content of tailings, and the predicted ash content of tailings was used as the training set and test set. When the coal feed is stable, the required clean coal ash content is taken every 20 minutes for laboratory testing. When the coal feed is unstable, the clean coal ash content is taken every 5 minutes. The samples are numbered, stored, and tested in real time. The results are used as a test set. Step S2, Data Reading and Cleaning: The predicted tailings ash content obtained by the AI ​​slurry ash analyzer in step S1 is used as the output value, and the real-time collected training set data is used as the input value for cleaning and removing outliers. Step S3, Feature Selection: Perform correlation analysis and sort the data input values ​​obtained in step S2, and then perform feature selection to select the top five variables with the highest correlation as input values. Step S4, Normalization: Normalize the data that has undergone feature selection in step S3; Step S5: Construct the training model: Construct the training model using the Light GBM regression algorithm and Bayesian hyperparameter optimization; The specific steps for building and training a model are as follows: Step S5.1: At the beginning of training, the value range of each feature is discretized to form a histogram; Step S5.2: The histogram divides the continuous values ​​of the feature into multiple discrete bins. Each bin stores the number of samples of the feature in the training set, the gradient, and the sum of squared gradients. Step S5.3: Construct a decision tree: Directly manipulate the feature histogram using the Light GBM regression algorithm; calculate the gradient and sum of squared gradients for each bin, and then evaluate the gain of each bin as a split point based on these statistics. Step S5.4: Define the hyperparameter space: After determining and selecting the corresponding split nodes and parameters based on the above architecture, define the hyperparameter space. Each hyperparameter has a search range or value range. Step S5.5: Define the objective function of Bayesian optimization: The input of this function is a combination of hyperparameters, and the output is the performance index of the model on the training data. This objective function will be called in each iteration of Bayesian optimization. The purpose is to train the Light GBM model according to the given combination of hyperparameters and return the performance index. Step S5.6: After defining the objective function, in each iteration, the hyperparameters of the LightGBM regression algorithm are tuned using the Bayesian optimization algorithm, and the performance of the optimal hyperparameter combination is verified by cross-validation and validation set evaluation methods. Step S5.7: Set the stopping condition: Stop the Bayesian optimization process when the model performance can no longer be improved during the optimization process; Step S5.8: Retrain the Light GBM model using the final optimized hyperparameter configuration obtained in step S5.7; Step S6: Compare and calculate the predicted tailings ash content and the measured clean coal ash content of the test set data obtained in step S1 to obtain the compensation model. Step S7: Evaluate the optimized training model using the compensation model; Step S8: Make further model adjustments based on the evaluation results. When the model's evaluation index R-squared reaches 0.9 or higher and no longer improves, output the results.

2. The online prediction method for clean coal ash content based on Bayesian hyperparameter optimization and Light GBM regression according to claim 1, characterized in that, The hyperparameters in step S5.4 include the learning rate, the number of trees, the number of leaf nodes, and the depth of the trees.

3. The online prediction method for clean coal ash content based on Bayesian hyperparameter optimization and Light GBM regression according to claim 1, characterized in that, In step S5.5, the model's performance metrics on the training data include accuracy and AUC.

4. The online prediction method for clean coal ash content based on Bayesian hyperparameter optimization and Light GBM regression according to claim 1, characterized in that, In step S6, if the difference between the two values ​​is greater than 5%, the measured value of the clean coal ash content is selected; if the difference between the two values ​​is 2% ≤ 5%, the average value of the two values ​​is taken; if the difference between the two values ​​is less than 2%, the predicted value of the tailings ash content is directly selected.

5. The online prediction method for clean coal ash content based on Bayesian hyperparameter optimization and Light GBM regression according to claim 1, characterized in that, The evaluation and comparison indicators in step S7 include mean square error, root mean square error, and coefficient of determination.

Images