Dam rockfill material compaction quality prediction method based on THO-XGBoost
By using the THO-XGBoost model and combining it with the Tent chaotic mapping to improve the Hippo optimization algorithm, the problems of model accuracy and stability in the prediction of rockfill compaction quality were solved, achieving high-precision compaction quality prediction and providing reliable theoretical guidance for dam construction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-12
- Publication Date
- 2026-04-21
AI Technical Summary
In existing technologies, single machine learning models for predicting the compaction quality of rockfill have low prediction accuracy and insufficient generalization ability. Ensemble learning models are difficult to optimize hyperparameters, and swarm intelligence optimization algorithms lack population diversity, resulting in unstable and inaccurate compaction quality predictions.
The THO-XGBoost model is adopted, and the hyperparameters of the XGBoost model are optimized by introducing the Tent chaotic mapping to improve the hippo optimization algorithm. Combined with the integration of multi-source heterogeneous information and spatiotemporal correlation, a compaction quality prediction model is constructed to achieve high-precision compaction quality prediction.
It improves the accuracy and stability of compaction quality prediction, provides rapid and accurate compaction quality assessment, and supports subsequent rolling construction control and project quantity optimization.
Smart Images
Figure CN121902268A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of construction technology for water conservancy and hydropower engineering, and particularly to the field of dam compaction construction technology in water conservancy and hydropower engineering. Specifically, it relates to a method for predicting the compaction quality of dam rockfill based on THO-XGBoost. Background Technology
[0002] During dam construction, the compaction characteristics of the rockfill are crucial to ensuring the quality of the project. Compaction quality analysis is a key step in the rolling process, and dry density is an important indicator for evaluating compaction quality. High-precision prediction of dry density can not only optimize compaction techniques and improve construction efficiency, but also effectively guide subsequent rolling operations, thereby comprehensively improving the quality and reliability of the dam project.
[0003] During the construction of the dam body, the traditional method for dry density testing is mainly to conduct test pit experiments at randomly selected points on the working face. However, this on-site test method has the following limitations: (1) the number of random sampling points is often small, and it is unreliable to use a limited number of local permeability coefficients to reflect the permeability characteristics of the entire surface; (2) the lengthy sampling and testing procedures consume a lot of time and may affect the progress of the project; (3) test pit experiments are mostly carried out after the construction is completed, making it difficult to make timely remedies for substandard test results.
[0004] Compared with traditional detection methods, establishing a prediction model for the compaction quality of riprap and studying the nonlinear mapping relationship between relevant parameters and compaction quality can achieve real-time assessment of compaction quality. With the rapid development of artificial intelligence technology, machine learning algorithms such as neural networks, fuzzy logic, random forests, and support vector machines are widely used to predict compaction quality. However, single machine learning models have poor stability in prediction and are easily affected by random factors in generalization performance. Ensemble learning, by integrating information from multiple models and using ensemble strategies to construct more diverse models, can effectively reduce overfitting and prediction uncertainty, producing more accurate and robust prediction results. Extreme Gradient Boosting (XGBoost) is an ensemble learning algorithm based on gradient boosting. Its basic idea is to add a new weak learner to fit the residual of the previous training, obtain the prediction score for each sample at the end of training, and finally sum the prediction scores from all weak learners to obtain the predicted value of the sample. Due to its fast computation speed and strong generalization ability, it has been widely used in many fields both domestically and internationally.
[0005] Furthermore, the selection of different hyperparameters has a significant impact on the model's prediction results, making the choice of hyperparameters crucial. Swarm optimization algorithms are optimization algorithms that mimic the behaviors of various biological groups in nature, such as foraging, encircling, reproducing, and escaping. They have advantages such as strong solution capabilities and fast convergence speed, and are widely used in parameter optimization for machine learning algorithms. The Hippo Optimization Algorithm (HO), proposed by Amiri et al. in 2024, is a novel swarm optimization algorithm inspired by three prominent behavioral patterns in hippos: young hippos' tendency to deviate from the group due to curiosity; hippo defensive behavior; and hippos' behavior of escaping predators. The first two behaviors correspond to the exploration phase of the optimization algorithm, while the third corresponds to the development phase. Experimental verification shows that HO performs better than current mainstream algorithms in terms of search accuracy, convergence speed, and stability. However, HO uses a simple random method to initialize the population, which leads to a decrease in population diversity during the search process and a tendency to get trapped in local optima. Therefore, a more efficient optimization algorithm with both global and local search capabilities is urgently needed.
[0006] In summary, among existing compaction quality prediction models, single machine learning models are easily affected by random factors and exhibit poor prediction stability. Although ensemble learning models can achieve better prediction accuracy, different combinations of hyperparameter configurations significantly impact model performance. Therefore, to achieve high-precision compaction quality prediction, it is necessary to conduct research on constructing a THO-XGBoost compaction quality prediction model. This model improves the Hippo Optimization Algorithm (THO) by introducing Tent chaotic mapping to efficiently optimize XGBoost hyperparameters, thereby enhancing prediction accuracy and generalization ability, and providing support for subsequent compaction quality control. Summary of the Invention
[0007] This invention addresses the limitations of existing technologies by disclosing a method for predicting the compaction quality of rockfill in dams based on THO-XGBoost. The purpose of this invention is to address the problems of low prediction accuracy and insufficient generalization ability of single machine learning models in existing compaction quality prediction research. It provides a compaction quality prediction model based on THO-XGBoost to achieve rapid and accurate prediction of rockfill compaction quality during the rolling process, providing effective and reliable theoretical guidance for subsequent rolling construction control and engineering quantity optimization.
[0008] This invention is achieved through the following technical solution:
[0009] A method for predicting the compaction quality of dam rockfill based on THO-XGBoost, characterized by the following steps:
[0010] S1. Obtain the input and output parameters of the dam rockfill compaction process. The input parameters include at least material source characteristic parameters and compaction characteristic parameters, and the output parameter is the compacted dry density.
[0011] S2. By integrating multi-source heterogeneous information and spatiotemporal correlation, the material source characteristic parameters, compaction characteristic parameters and compaction dry density are transformed into a unified spatiotemporal reference coordinate system to achieve spatiotemporal registration of input and output data and generate a compaction quality dataset.
[0012] S3. Based on the compaction quality dataset, construct an XGBoost ensemble learning model, and use the Hippo optimization algorithm with Tent chaotic mapping to globally optimize the hyperparameters of the XGBoost model, thereby establishing a compaction quality prediction model based on THO-XGBoost to achieve high-precision prediction of compaction quality.
[0013] Furthermore, the process of step S1 is as follows:
[0014] 1.1 Obtaining material source characteristic parameters: According to the "Code for Construction of Rolled Earth-Rock Dams" (DL / T5129—2013), the moisture content of the rockfill was obtained through moisture content test, and the P5 content, non-uniformity coefficient and curvature coefficient were obtained through particle size analysis test.
[0015] 1.2 Obtaining compaction characteristic parameters: Relying on the intelligent monitoring system for earth-rock dam compaction, and based on GPS positioning and vibration force monitoring equipment, the spatial location, vibration status and time information of the compaction machine group are collected in real time, and compaction characteristic parameters such as total number of compaction passes, number of vibratory compaction passes, compaction thickness and compaction speed are calculated.
[0016] 1.3 Output parameter acquisition: The compacted dry density at the corresponding location is obtained through in-situ density test.
[0017] 1.4 Data aggregation: The material source characteristic parameters, compaction characteristic parameters and compaction dry density are aggregated and stored in the same database through the PDA real-time information acquisition system, forming the original data source required for subsequent modeling.
[0018] Furthermore, the process of step S2 is as follows:
[0019] 2.1 Time Synchronization: Based on the construction reference time axis, the material source characteristic parameters, compaction characteristic parameters and compaction dry density are matched according to the collection time, and the data from different sources are unified to the same time series.
[0020] 2.2 Spatial alignment: The coordinates of the compaction trajectory, the grid coordinates of the dam body surface, and the coordinates of the test pit location are unified into a reference coordinate system through coordinate transformation.
[0021] 2.3 Neighborhood compaction information integration: Taking the grid of the dam body surface as the spatial unit, the compaction characteristic parameters within the grid are weighted and integrated based on the compaction rate and weight of the neighboring points of the compaction trajectory, so as to obtain the representative compaction characteristic index of each grid.
[0022] 2.4 Spatiotemporal Registration: Based on the completion of time synchronization and spatial alignment, the material source characteristic parameters, compaction characteristic parameters and corresponding compaction dry density of the same grid and time period are matched one by one to generate a spatiotemporal registration data sample containing multi-dimensional input features and target output.
[0023] 2.5 Dataset Construction and Partitioning: The spatiotemporally registered samples are integrated to generate a compaction quality dataset, which is then divided into a training set and a test set according to a preset ratio.
[0024] Furthermore, the process of step S3 is as follows:
[0025] 3.1 XGBoost Model Construction: Using material source characteristic parameters and rolling characteristic parameters as input features, and rolling dry density as the prediction target, a compaction quality prediction model based on XGBoost is constructed.
[0026] 3.2 Fitness Function Design: The prediction error of the XGBoost model on the training set is used as the fitness function to characterize the quality of the hyperparameter combination.
[0027] 3.3 Hippo Optimization Algorithm Modeling: A swarm intelligence optimization algorithm simulating hippo group behavior is used to search for the hyperparameters of XGBoost. The hippo optimization algorithm includes: a first-stage position update behavior, which is used to perform local search in the vicinity of the current solution and global exploration based on the group center; a second-stage defense behavior, which enhances the ability to jump out of local optima by introducing predator positions and Lévy perturbations to achieve large-scale jumps; and a third-stage escape behavior, which performs detailed search in the vicinity of the current optimum to improve local convergence accuracy.
[0028] 3.4 Tent Chaotic Map Initialization: Tent chaotic mapping sequence is introduced into the population initialization process of the Hippo optimization algorithm to generate and optimize the distribution of the initial individuals in the population, so as to improve the population diversity and global search capability.
[0029] 3.5 Hyperparameter optimization and model determination: During the iterative update process, the new and old solutions are compared according to the fitness function, the optimal hyperparameter combination is retained, and finally the optimized THO-XGBoost compaction quality prediction model is obtained.
[0030] Features and benefits of this invention: Compared with existing technologies, this invention employs the XGBoost ensemble learning algorithm, which effectively reduces overfitting and lowers prediction uncertainty, thereby producing more accurate prediction results. Simultaneously, by introducing an improved hippo optimization algorithm, the hyperparameters of the XGBoost model are optimized, ensuring the uniformity and diversity of the initial population distribution, thus establishing a high-precision compaction quality prediction model.
[0031] This invention overcomes the limitations of traditional single machine learning models (insufficient prediction accuracy), ensemble learning models (difficulty in hyperparameter optimization), and swarm intelligence optimization algorithms (insufficient population diversity). Compared to single machine learning prediction models, this invention significantly improves prediction performance, enabling rapid and accurate prediction of compaction quality, thus obtaining more accurate and reliable prediction results. This provides reliable methodological support for ensuring dam compaction quality and has significant engineering application value. Furthermore, the model proposed in this invention also provides new ideas for predicting compaction quality in other types of dam projects, showing promising prospects for engineering applications. Attached Figure Description
[0032] Figure 1 This describes the modeling process of the compaction quality prediction model based on THO-XGBoost in this invention. Detailed Implementation
[0033] The present invention will be further described below with reference to specific embodiments. These specific embodiments are further explanations of the principles of the present invention and are not intended to limit the present invention in any way. Any technology that is the same as or similar to the present invention does not exceed the scope of protection of the present invention.
[0034] In practical engineering, accurate and reliable compaction quality prediction is of great significance for the control of the compaction construction process. Addressing the shortcomings of existing compaction quality prediction research, such as insufficient prediction accuracy of traditional single machine learning models, difficulties in hyperparameter optimization of ensemble learning models, and insufficient diversity of swarm intelligence optimization algorithms, this paper adopts a compaction quality prediction model based on THO-XGBoost. This model can quickly and accurately predict compaction quality under complex compaction conditions, thus providing effective and reliable theoretical guidance for subsequent compaction construction control and engineering quantity optimization.
[0035] This invention relates to a method for predicting the compaction quality of dam rockfill based on THO-XGBoost, comprising the following steps:
[0036] S1. Obtain the input and output parameters of the dam rockfill compaction process. The input parameters include at least material source characteristic parameters and compaction characteristic parameters, and the output parameter is the compacted dry density.
[0037] S2. By integrating multi-source heterogeneous information and spatiotemporal correlation, the material source characteristic parameters, compaction characteristic parameters and compaction dry density are transformed into a unified spatiotemporal reference coordinate system to achieve spatiotemporal registration of input and output data and generate compaction quality dataset.
[0038] S3. Based on the compaction quality dataset, construct an XGBoost ensemble learning model, and use the Hippo optimization algorithm with Tent chaotic mapping to globally optimize the hyperparameters of the XGBoost model, thereby establishing a compaction quality prediction model based on THO-XGBoost to achieve high-precision prediction of compaction quality.
[0039] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0040] 1. Obtain input and output parameters.
[0041] According to the "Code for Construction of Roller-Compacted Earth-Rock Dams" (DL / T 5129—2013), the moisture content of the material source is obtained through moisture content testing, while the P5 content, uniformity coefficient, and curvature coefficient are obtained through particle size analysis. The compaction characteristics are obtained using the dam's intelligent compaction monitoring system. This system collects real-time information on the spatial location, excitation force status, and time of the compaction rollers through GPS positioning and vibration force monitoring equipment. The data is then processed by a database server and application server to calculate the compaction characteristics. The compacted dry density is obtained through in-situ density testing. The collected data is further aggregated into a database using a PDA (personal digital assistant) real-time information acquisition system.
[0042] 2. Generate the dataset.
[0043] By utilizing multi-source heterogeneous information integration technology, the characteristic parameters of multi-source heterogeneous materials, compaction characteristic parameters, and compaction dry density are transformed to the same spatiotemporal reference through spatiotemporal correlation, achieving spatiotemporal registration of input and output data. This includes both time synchronization and spatial alignment. Time synchronization refers to unifying the compaction process information onto a reference standard time axis; spatial alignment requires spatially sequencing the compaction process with the material source characteristic information and compaction characteristic information, unifying them from different coordinate systems to a reference coordinate system through coordinate transformation. The matching criterion is as follows:
[0044] (1)
[0045] Where N is the total number of compaction passes, N2 is the number of vibratory compaction passes, H is the compaction thickness, and V is the compaction speed; f PDA Real-time information collection for PDA; d() represents the spatial coordinates of the information; Ω represents the grid of the dam body surface; V i λ represents the compaction velocity at points adjacent to the rolling trajectory, in m / s;i The weights are those of neighboring points.
[0046] The spatiotemporally registered input and output data are integrated to generate a dataset, which is then further divided into training and test sets according to the proportions.
[0047] 3. Construct a compaction quality prediction model based on THO-XGBoost.
[0048] The main principle of establishing a compaction quality index prediction model based on XGBoost is as follows. Assume the compaction quality dataset obtained above is... Each sample There are m features, each corresponding to a target value. Then for the i-th sample The predicted value of the compaction quality index in the t-th iteration It can be represented as follows:
[0049] (2)
[0050] in This represents the prediction result of the k-th regression decision tree for the i-th sample. Furthermore, the loss function for each iteration... The difference between the actual compaction quality and the model's current predictions was considered, and overfitting was prevented by adding a regularization term. The loss function is shown in the following equation:
[0051] (3)
[0052] (4)
[0053] in Used to measure the actual label y i The loss between the model's predictions and the predictions after adding the t-th tree, Ω(f) t Let ) be the regularization term for the t-th tree, and γ and λ be the regularization coefficients for the number of leaf nodes and their corresponding leaf node weights, respectively.
[0054] The parameters of the XGBoost prediction model hyper are optimized using the hippo optimization algorithm improved by Tent chaotic mapping. The hippo optimization algorithm uses a three-stage model that combines hippo position update, defense strategy against predators, and escape method; its main principle is as follows.
[0055] The first phase simulated the location-updating behavior of hippopotamus groups, which is the HO (Household Occurrence) search phase. During this process, the location of the male hippos... Based on the current solution continuously moving closer to the position of the dominant hippopotamus, achieving local exploration, the formula is:
[0056] (5)
[0057] in, Indicates the position of the dominant hippo (current optimal solution); A random number in the range [0,1]; I1 is an integer of 1 or 2, controlling the proximity. Female or immature hippos... During this stage, the algorithm will continuously move, a process that enhances its global search capability. The formula is:
[0058] (6)
[0059] (7)
[0060] (8)
[0061] (9)
[0062] Where T is used to control the exploration range to gradually shrink with each iteration, and the algorithm gradually transitions to the next stage; MG i Let h1 be the average position of several hippos randomly selected from the hippopotamus herd, and h2 be the reference local group center; h1 and h2 are randomly selected perturbation functions; r is a random number in the range [0,1].
[0063] The second stage simulates the hippo's defensive behavior when facing predators, further enhancing its global search capabilities and enabling it to escape local optima. In this stage, the hippo updates its position based on the predator's threat level. When the threat is high, the hippo makes large jumps to escape local optima. When the threat is low, it maintains a certain distance through moderate perturbations, balancing global and local search. The formula is:
[0064] The predator's position is initialized to a random value within the solution space:
[0065] (10)
[0066] (11)
[0067] (12)
[0068] in, The hippopotamus's position facing the predator; Indicates the location of the predator; This indicates the distance between the hippopotamus and its predator; Simulate large jump behavior using random vectors that follow a Levy distribution. This indicates a random perturbation operation.
[0069] The third stage simulates the hippo's behavior of escaping predators, enhancing the HO's local search capabilities. During this process, the hippo generates new random solutions near its current location:
[0070] (13)
[0071] (14)
[0072] (15)
[0073] in, s1 represents the location where the hippo escaped its predator, and it is searched to find the nearest safe location. s1 is randomly selected from three possible scenarios.
[0074] In the three stages of the Hippo Optimization (HO) algorithm, after each solution update, the fitness value of the new solution is compared with the fitness value of the current solution, and a position replacement is performed based on the comparison result. This process ensures that the algorithm can gradually approach the global optimum.
[0075] The Tent chaotic mapping sequence is used to optimize the population initialization process of HO (Hogwarts) to enhance the connections between different individuals in the HO population, increase the overall randomness of the population, enrich the population diversity, and improve the search capability of the algorithm. The expression for the Tent mapping is:
[0076] (16).
Claims
1. A method for predicting the compaction quality of dam rockfill based on THO-XGBoost, characterized in that... Includes the following steps: S1. Obtain the input and output parameters of the dam rockfill compaction process. The input parameters include at least material source characteristic parameters and compaction characteristic parameters, and the output parameter is the compacted dry density. S2. By integrating multi-source heterogeneous information and spatiotemporal correlation, the material source characteristic parameters, compaction characteristic parameters and compaction dry density are transformed into a unified spatiotemporal reference coordinate system to achieve spatiotemporal registration of input and output data and generate a compaction quality dataset. S3. Based on the compaction quality dataset, construct an XGBoost ensemble learning model, and use the Hippo optimization algorithm with Tent chaotic mapping to globally optimize the hyperparameters of the XGBoost model, establish a compaction quality prediction model based on THO-XGBoost, and achieve high-precision prediction of compaction quality.
2. The method for predicting the compaction quality of dam rockfill based on THO-XGBoost according to claim 1, characterized in that... The process of step S1 is as follows: 1.1 Obtaining material source characteristic parameters: According to the "Code for Construction of Rolled Earth-Rock Dams" (DL / T5129—2013), the moisture content of the rockfill was obtained through moisture content test, and the P5 content, non-uniformity coefficient and curvature coefficient were obtained through particle size analysis test. 1.2 Obtaining compaction characteristic parameters: Relying on the intelligent monitoring system for earth-rock dam compaction, and based on GPS positioning and vibration force monitoring equipment, the spatial location, vibration status and time information of the compaction machine group are collected in real time, and compaction characteristic parameters such as total number of compaction passes, number of vibratory compaction passes, compaction thickness and compaction speed are calculated. 1.3 Output Parameter Acquisition: The compacted dry density at the corresponding location is obtained through in-situ density testing; 1.4 Data aggregation: The material source characteristic parameters, compaction characteristic parameters and compaction dry density are aggregated and stored in the same database through the PDA real-time information acquisition system, forming the original data source required for subsequent modeling.
3. The method for predicting the compaction quality of dam rockfill based on THO-XGBoost according to claim 1, characterized in that... The process of step S2 is as follows: 2.1 Time Synchronization: Based on the construction reference timeline, the material source characteristic parameters, compaction characteristic parameters and compaction dry density are matched according to the collection time, and the data from different sources are unified to the same time series; 2.2 Spatial alignment: The coordinates of the compaction trajectory, the grid coordinates of the dam body surface, and the coordinates of the test pit location are unified into a reference coordinate system through coordinate transformation; 2.3 Neighborhood compaction information integration: Taking the grid of the dam body as the spatial unit, the compaction characteristic parameters within the grid are weighted and integrated based on the compaction rate and weight of the neighboring points of the compaction trajectory, so as to obtain the representative compaction characteristic index of each grid. 2.4 Spatiotemporal Registration: Based on the completion of time synchronization and spatial alignment, the material source characteristic parameters, compaction characteristic parameters and corresponding compaction dry density of the same grid and the same time period are matched one by one to generate a spatiotemporal registration data sample containing multi-dimensional input features and target output. 2.5 Dataset Construction and Partitioning: The spatiotemporally registered samples are integrated to generate a compaction quality dataset, which is then divided into a training set and a test set according to a preset ratio.
4. The method for predicting the compaction quality of dam rockfill based on THO-XGBoost according to claim 1, characterized in that... The process of step S3 is as follows: 3.1 XGBoost Model Construction: Using material source characteristic parameters and rolling characteristic parameters as input features, and rolling dry density as the prediction target, a compaction quality prediction model based on XGBoost is constructed. 3.2 Fitness Function Design: The prediction error of the XGBoost model on the training set is used as the fitness function to characterize the quality of the hyperparameter combination; 3.3 Hippo Optimization Algorithm Modeling: A swarm intelligence optimization algorithm simulating hippo group behavior is used to search for the hyperparameters of XGBoost. The hippo optimization algorithm includes: a first-stage position update behavior, which is used to perform local search in the vicinity of the current solution and global exploration based on the group center; The second stage of defensive behavior involves introducing predator positions and Lévy perturbations to achieve large-scale jumps, enhancing the ability to escape local optima; the third stage of escape behavior involves a detailed search in the vicinity of the current optimum, improving local convergence accuracy. 3.4 Tent Chaotic Map Initialization: Tent chaotic mapping sequence is introduced into the population initialization process of the Hippo optimization algorithm to generate and optimize the distribution of the initial individuals in the population, so as to improve the population diversity and global search capability. 3.5 Hyperparameter optimization and model determination: During the iterative update process, the new and old solutions are compared according to the fitness function, the optimal hyperparameter combination is retained, and finally the optimized THO-XGBoost compaction quality prediction model is obtained.