Intelligent Prediction and Prevention of Coal Mine Water Inrush: Integrating Hybrid Data Augmentation, HO-SVR, and RAG-LLM Technologies

Abstract

This study proposes a novel integrated framework that combines a Hippopotamus-Optimized Support Vector Regression (HO-SVR) prediction model with a Retrieval-Augmented Generation-enhanced Large Language Model (RAG-LLM)-based intelligent decision module, addressing the core challenge of bridging prediction and prevention in coal mine water inrush disasters. It represents the first application of the combined HO-SVR and RAG-LLM approach in this field. Methodologically, a hybrid data augmentation technique (SMOTE–GN–Bootstrap) alleviates data scarcity and imbalance, while feature selection and dimensionality reduction optimize the input features. The developed HO-SVR model demonstrates superior prediction accuracy over benchmark models. The key innovation lies in the RAG-LLM module which automatically generates interpretable reports and actionable prevention strategies based on the prediction results and key influencing factors, thereby establishing a closed-loop intelligent system from accurate prediction to informed prevention. Practically, this framework enables proactive risk management through data-driven predictions, significantly reduces water inrush incidents, and provides intelligent decision support for field operations, substantially enhancing mine safety. Furthermore, the study discusses the model’s potential and challenges across different geological settings, charting a course for developing more generalized models

1. Introduction

Coal mine safety production is crucial for sustainable development. Floor water inrushes are destructive due to their complexity and suddenness, threatening safety and mine viability [1]. These incidents not only cause significant economic losses but also hinder the sustainable and safe development of the coal mining industry [2,3]. Furthermore, the significant diversity in hydrogeological conditions across China’s major coalfields poses a fundamental challenge to developing a universally applicable prediction model. Consequently, achieving high-precision prediction of water inrush volume and, more importantly, intelligently formulating executable risk mitigation strategies based on these predictions are essential for transitioning from passive response to proactive, precise prevention and control [4].

However, the extreme complexity of the underground environment and limitations in data acquisition present significant challenges [5]. These challenges are not only technical but also relate to the knowledge gap in effectively translating predictive insights into actionable decisions for mine engineers. This translation bottleneck currently impedes the intelligent and precise management of water inrush risks. Therefore, developing an integrated framework that seamlessly connects high-precision prediction with intelligent decision-support, is of paramount importance for advancing sustainable safety management practices in coal mining. This study aims to address these challenges by proposing an integrated framework that bridges high-precision prediction with intelligent decision-support, thereby contributing to the sustainable management of water hazards in coal mining.

2. Literature Review

2.1. Research on Coal Mine Water Inrush Prediction

The pursuit of accurate water inrush prediction has led to the development of various methodologies, which can be broadly categorized into traditional techniques and data-driven approaches. Conventional forecasting techniques primarily include statistical methods and numerical simulations. While statistical approaches offer computational efficiency, their ability to characterize the intricate, nonlinear dynamics of water inrush systems is often limited [6,7,8]. Numerical simulations, on the other hand, can model the physical processes but require extensive precise geological parameters, leading to high computational costs and potential inaccuracies due to parameter uncertainty.

The advancement of machine learning (ML) has provided new paradigms for enhancing mining safety [9,10]. Models such as Support Vector Regression (SVR) [11,12,13,14], Random Forest (RF) [15], various neural networks [16,17,18], and LightGBM have been extensively applied to predict water inrush volume, demonstrating superior capability in handling nonlinear relationships from historical data. To further improve model performance, researchers have actively addressed inherent challenges. To mitigate issues of data scarcity and imbalance, data augmentation strategies are employed. To enhance model efficiency and generalization, feature selection or dimensionality reduction techniques are used to eliminate redundant or irrelevant features [19,20]. Furthermore, to overcome the limitations of manual parameter tuning, metaheuristic optimization algorithms such as Gray Wolf Optimization (GWO) [21], Bald Eagle Search (BES) [22], and Particle Swarm Optimization (PSO) [23] have been successfully integrated to automate and optimize model hyperparameters.

Recent research trends over the past two years have further expanded the boundaries of this field. Advances in physics-informed machine learning, explainable artificial intelligence (XAI), and metaheuristic optimization algorithms provide valuable insights for coal mine floor water inrush prediction.

(1) Integration of Physical Mechanisms and XAI

To enhance model generalizability and trustworthiness, research trends have shifted towards integrating domain knowledge into machine learning models and improving model explainability [24]. Wang et al. [25] highlighted the challenge of balancing accuracy and interpretability in traditional machine learning models for coal and gas outburst prediction, particularly with limited data or under complex geological conditions. Physics-Informed Neural Networks (PINNs) integrate physical monotonicity constraints with data-driven learning, effectively improving prediction accuracy and generalization capability [26]. The PINN framework embeds physical laws and constraints into the neural network’s training process, enabling predictions consistent with physical principles even without extensive data, which is particularly crucial for water inrush scenarios lacking sufficient historical cases. Deep learning models are often considered “black boxes” due to their opaque decision-making processes, a significant drawback in safety-critical applications like coal mine water inrush prediction [27]. XAI techniques aim to improve the understandability and transparency of ML models by interpreting the factors behind model decisions, thereby increasing user trust in predictions. For instance, in predicting tunnel crown convergence in TBM projects, researchers proposed an interpretable ML method combining Bayesian Optimization (BO) and SHapley Additive exPlanations (SHAP) values to quantify each input feature’s contribution to the prediction, thereby enhancing model interpretability [28]. Applying XAI techniques to water inrush prediction can help engineers understand which geo-hydrogeological parameters contribute most to the risk, enabling more precise preventive measures.

(2) Application and Innovation of Metaheuristic Optimization Algorithms

Metaheuristic optimization algorithms play an increasingly important role in optimizing parameters for water inrush prediction models due to their global search capability and efficiency in handling complex optimization problems [29]. The recently developed Hippopotamus Optimization Algorithm (HO), which mimics the unique behaviors of hippopotamuses, shows considerable promise in convergence efficiency and global exploration abilities, offering a novel method for parameter optimization [30]. Recent research [31,32,33,34] reveals emerging application trends of HO in engineering optimization, highlighting its competitiveness in handling complex problems and indirectly demonstrating its potential. The Aquila Optimizer (AO) is a novel metaheuristic algorithm based on the hunting behavior of Aquila birds. To address its limitations in narrow exploration capability and susceptibility to local optima when solving high-dimensional problems, an improved Aquila Optimizer (IAO) was proposed, incorporating mechanisms like dynamic random walk and dynamic opposition-based learning to enhance its global optimization capability and efficiency [35,36]. Furthermore, many other novel metaheuristic algorithms have been introduced and applied in engineering optimization, such as the Slime Mould Algorithm (SMA) [37], Election Optimizer Algorithm (EOA) [38], Groupers and Moray Eels (GME) optimization algorithm [39], and African Vultures Optimization Algorithm (AVOA) [40]. By simulating various natural behaviors or physical phenomena, these algorithms provide diverse options and potential for solving complex optimization problems in coal mine water inrush prediction models.

2.2. Research Gaps

Despite these advancements, current research exhibits several critical gaps that hinder the transition from high-precision prediction to effective prevention, which is essential for sustainable risk management:

(1) Prediction-to-Action Disconnect

Current research primarily focuses on enhancing predictive accuracy through increasingly sophisticated models and optimization algorithms. However, a critical and often overlooked issue is the translational bottleneck between high-precision numerical forecasts and executable, context-aware prevention and control strategies for field engineers. Although emerging Explainable AI techniques like SHAP begin to address model interpretability, there is still a lack of an integrated framework capable of automatically synthesizing model predictions, mechanistic interpretations, and domain knowledge into directly actionable, intelligible reports and preventive measures. This disconnect significantly limits the practical utility of predictive models in achieving proactive and precise risk management.

(2) Data Limitations

The performance of machine learning models is highly dependent on the quality and quantity of training data. Although recent advances have introduced techniques such as hybrid data augmentation and Physics-Informed Neural Networks to mitigate data scarcity, a fundamental challenge persists. Acquiring high-quality, annotated data for coal mine water inrushes remains exceptionally difficult due to the rarity of major events and the complexity of the underground environment, often resulting in datasets that are small, geographically concentrated, and imbalanced. This data scarcity constrains the model’s ability to accurately detect and predict crisis conditions.

(3) Feature Redundancy and Model Optimization

Input data for water inrush prediction may contain redundant or irrelevant features, which can adversely affect model efficiency and generalization capability. Furthermore, machine learning models are typically highly sensitive to their parameter configurations, necessitating comprehensive optimization to achieve reliable prediction outcomes. This optimization process is inherently complex and computationally intensive.

2.3. Research Objectives

To address the identified gaps, this paper proposes an integrated framework for the prevention and control of coal mine floor water outbursts. The specific objectives of this study are threefold:

(1) To establish an intelligent decision-support system that effectively bridges predictive analytics and operational prevention, this study proposes a novel framework that leverages a Retrieval-Augmented Generation (RAG)-powered Large Language Model (LLM) framework, termed RAG-LLM. This system converts prediction results and feature importance rankings into interpretable mechanism analyses and executable prevention strategies, enabling seamless data-driven decision-making.

(2) To develop a robust prediction model by addressing data scarcity through a hybrid augmentation approach and reducing feature redundancy via advanced selection techniques, thereby enhancing model interpretability and performance.

(3) To develop a high-precision water inrush prediction model by integrating a hybrid feature selection method to eliminate redundant and irrelevant features, and subsequently applying the HO to optimize the SVR hyperparameters. This dual optimization aims to enhance model interpretability, efficiency, and generalization capability.

3. Methodology

3.1. Research Framework

The accurate prevention and management of coal mine water inrush disasters from the floor aquifer rely on the integration of high-precision prediction models with the effective implementation of preventive and control measures. This paper proposes a closed-loop research framework that integrates hybrid data augmentation, HO-SVR prediction, and RAG-LLM intelligent decision-making to address the limitations and bottlenecks in current research. As illustrated in Figure 1, the proposed methodology consists of six key steps:

Step 1: Data preparation. Execute missing value imputation, categorical variable encoding, and more actions on the raw dataset. Following thorough preparation, precisely partition the dataset into a training set and an independent test set in a 7:3 ratio to establish a standardized data foundation for following procedures.

Step 2: Hybrid data augmentation. This paper offers the new SMOTE–GN–Bootstrap mixed augmentation technique to effectively mitigate the challenges posed by limited sample sizes and data imbalance, thereby addressing the issues of insufficient training samples and the prevalence of outlier samples.

Step 3: Data standardization. Standardization is conducted independently on the training and test sets to systematically scale the data to a uniform range, thereby mitigating the potential impact of dimensional discrepancies on distance-sensitive models, such as SVR, and ensuring the uniformity of the model input data.

Step 4: Feature engineering. Feature sets are generated independently by mutual information feature selection, polynomial feature generation, and PCA dimensionality reduction. The efficacy of various feature sets in the model is assessed using R² and RMSE metrics, and the most effective feature set is chosen to ensure superior feature input for model training.

Step 5: Model optimization and comparison. A high-precision HO-SVR prediction model was developed based on the optimization outcomes of feature engineering. To thoroughly assess the efficacy and superiority of HO-SVR, systematic comparative experiments were executed using HO-RF and HO-LightGBM models, in addition to baseline models SVR, RF, and LightGBM.

Step 6: RAG-LLM intelligent decision-making. The predicted water inrush volume from HO-SVR and the critical feature importance data obtained via perturbation analysis are fed into the RAG-LLM decision-making module to generate a report on the water inrush risk mechanism and associated prevention and control measures, thereby providing a robust foundation and intelligent support for management personnel’s scientific decision-making.

3.2. Data Preparation

This study collected 134 valid water inrush volume sample data from multiple mining areas in Shandong Province, covering water inrush events at different mining stages and geological structural locations. There are 24 initial features ($X_n$), and the target variable is the water inrush volume ($Y$) (m³/h), which represents the volume of water suddenly gushing out from the coal seam floor per unit time. The names and units of the 24 feature variables are shown in

Table 1. Description of the 24 initial feature variables for water inrush prediction.

No.	Feature Variable Names	Unit	No.	Feature Variable Names	Unit
X1	Development degree of Water-Conducting Structures	None	X13	Structural Water Retention	None
X2	Fracture Permeability Characteristics	None	X14	Collapse Pillar	None
X3	Inclined Length of Working Face	m	X15	Fault	None
X4	Aquifer Water Pressure	MPa	X16	Fracture Zone	None
X5	Monthly Advancement Distance of Working Face	m	X17	Fault Displacement	m
X6	Aquitard Thickness	m	X18	Mining Depth of Working Face	m
X7	Depth of Floor Damage	m	X19	Water Source	None
X8	Effective Thickness of Floor Aquitard	m	X20	Water Quality	None
X9	Percentage of Sandstone in the Aquitard	None	X21	Water Temperature	°C
X10	Percentage of Mudstone in the Aquitard	None	X22	Strike Length of Working Face	m
X11	Percentage of Limestone in the Aquitard	None	X23	Coal Seam Dip Angle	°
X12	Mining Height of Working Face	m	X24	Coal Seam Thickness	m

. Among these, X₁₄, X₁₅, X₁₆, X₁₉, and X₂₀ are discrete data, while the remaining features are continuous data.

The State Administration of Work Safety of China stipulates in its “Regulations on Water Prevention and Control in Coal Mines” that a water inrush exceeding 300 m³/h for the first time constitutes a major water inrush accident. Among the 60 samples collected, non-water inrush samples $\left(Y=0\right)$ accounted for 24% of the total samples, small water inrush samples ($0\le Y\le 300$) accounted for 58%, and large water inrush samples $\left(Y>300\right)$ accounted for only 18%. This imbalanced distribution with scarce large-water-inrush samples can weaken the predictive model’s ability to identify major water inrush risks, making it difficult to meet regulatory requirements and practical needs for accurate early warning of serious accidents.

To improve data quality and meet modeling requirements, preprocessing steps such as missing value imputation and categorical variable encoding were performed on the raw dataset. After thorough data preparation, the dataset was accurately split into a training set and an independent test set in a 7:3 ratio, establishing a standardized data basis for subsequent modeling procedures.

3.3. Hybrid Data Augmentation

To address the issues of limited sample size and imbalanced data distribution, a hybrid data augmentation technique, integrating Synthetic Minority Over-sampling Technique (SMOTE), Gaussian Noise (GN), and Bootstrap Sampling (hence referred to as SMOTE–GN–Bootstrap), was developed. SMOTE effectively balances data distribution [41], while the addition of GN further enhances data diversity [42]. In contrast, Bootstrap sampling generates multiple varied training subsets through sampling with replacement [43]. Together, these methods collectively enlarge the dataset and improve the robustness and generalization ability of the model.

This study effectively increased the number of training samples from 94 to 200 using this technique (

Table 2. Sample of the enhanced training dataset after SMOTE–GN–Bootstrap augmentation.

NO.	X1	X2	X3	X4	X5	X6	…	Y
1	0.5	0.5	185	1.5	90	32	…	82
2	0.8	0.8	82	3.6	90	33	…	156
3	0.5	0.5	120	4.25	84	49	…	195
…	…	…	…	…	…	…	…
198	0.3	0.5	110	1.48	30	29	…	12
199	0.5	0.5	80	2.18	30	24.5	…	60
200	0.3	0.1	105	1.3	60	74.55	…	585

). Figure 2 illustrates the distribution of the original coal mine floor water inrush volume data (left figure) and the augmented training data (right figure) after processing with the linear axis scaling technique. This technique retains statistical information by implementing scaling restrictions on the coordinate axes, proportionally compressing data point locations while preserving their relative positions, therefore resolving readability challenges in small-scale charts for extensive datasets. Comparative analysis reveals that the original data exhibits sparse sampling, particularly in regions characterized by high water inrush volumes where samples are limited. Conversely, the improved data distribution is more consistent and uniform, exhibiting a substantial increase in the quantity of high-water inrush samples. This hybrid improvement technique effectively addresses the scarcity of high-water inrush samples, providing a more comprehensive and representative database for model training.

3.4. Data Standardization

To mitigate the influence of varying units, types, and significant absolute value disparities among the feature variables in the dataset on the model’s predictive efficacy regarding bottom plate water ingress, the original data features are normalized to the interval [0, 1]. This work employs the Min-Max standardization method, as delineated in Formula (1).

$$X^{\prime }={\displaystyle \frac{X-X_{min}}{X_{max}-X_{min}}}$$

(1)

In the formula, $X$ is the original value, $X_{max}$ is the maximum value of the sample data, and $X_{min}$ is the minimum value of the sample data. Standardized data helps improve the training efficiency and prediction performance of models such as SVR.

3.5. Feature Engineering

In high-dimensional data, feature redundancy and irrelevance can significantly impair model performance. Feature engineering can efficiently reduce feature dimensionality by selecting the most discriminative subset of features, thereby mitigating noise interference [44] and improving the model’s predicted accuracy and generalization. This study evaluates two complementary feature engineering methodologies to identify optimal feature subsets while determining their comparative effectiveness. First, feature selection via mutual information analysis quantifies feature-goal variable associations to isolate maximally informative subsets. Second, feature extraction expands the original feature space through polynomial feature construction, subsequently applying principal component analysis (PCA) for dimensionality reduction to derive essential low-dimensional representations. This study utilized the original features, feature subsets selected using mutual information, and feature subsets generated through polynomial feature engineering and PCA dimensionality reduction in three models—RF, LightGBM, and SVR—for training and prediction. The models’ performance was assessed and contrasted utilizing the coefficient of determination (R²) and root mean square error (RMSE) as primary evaluation criteria. Ultimately, based on the comprehensive evaluation results from 5-fold cross-validation, the feature subset generation method exhibiting the optimal overall performance and its corresponding feature subset were selected from various feature subsets to ensure that the final feature set demonstrates robust generalization ability and predictive performance across multiple models. The process of feature engineering strategy is illustrated in Figure 3.

3.5.1. Feature Selection: Mutual Information Method

Predicting water inrush volume in coal mines is a typical nonlinear regression problem. Mutual information, as a non-parametric information-theoretic method, can efficiently measure the nonlinear statistical dependence between features and target variables, making it more suitable for intricate feature interaction patterns in practical situations [45]. It elucidates the interdependence between characteristics and goal variables from an informational standpoint by quantifying the shared information between them. Mutual information quantifies the shared knowledge between two random variables, $X$ (features) and $Y$ (target), as specified in Formula (2):

$$I\left(X;Y\right)={\displaystyle \sum }_{xϵX}{\displaystyle \sum }_{yϵY}p\left(x,y\right)log{\displaystyle \frac{p\left(x,y\right)}{p\left(x\right)p\left(y\right)}}$$

(2)

Among them, $p\left(x,y\right)$ is the joint probability distribution of $X$ and $Y$, and $p\left(x\right)$ and $p\left(y\right)$ are the marginal probability distributions of X and Y, respectively. $I\left(X;Y\right) \ge$0, and the larger the value, the stronger the correlation between the feature and the target.

Features are ranked in descending order based on their mutual information scores, and the cumulative importance curve method is employed to determine the number of features to retain. Initially, all features are arranged according to their mutual information values. Subsequently, the cumulative mutual information percentage of the top N features is computed. Ultimately, the N value at which the cumulative percentage attains 80% is identified, and the top N features are preserved.

3.5.2. Feature Extraction: Polynomial Feature Construction and PCA

This research develops a two-stage processing system that combines polynomial feature creation with PCA. Higher-order interaction terms are constructed to expand the feature space, followed by a linear transformation that extracts the most information-dense low-dimensional representation, thereby mitigating redundancy in high-dimensional data and improving the model’s capacity to identify complex patterns. The essence of polynomial feature generation involves generating new features by augmenting the dimensionality of the original features, thereby amplifying the model’s complexity and its capacity for nonlinear fitting [46]. Assuming that the original feature space contains a feature x, polynomial features can be constructed by expanding the original features through linear terms ($x_i)$, quadratic terms (${x_i}^2$), interaction terms $\left(x_i\ast x_i\right)$, and higher-order terms $\left({x_i}^3,{x_i}^4,\dots ,etc.\right)$. Through this expansion, the original feature space is mapped to a higher-dimensional feature space, allowing linear models to better fit nonlinear data relationships.

The dimensionality of the feature space increases substantially following the construction of polynomial features. PCA is employed for feature dimension reduction to minimize dimensionality and eliminate redundant information. PCA is a method for feature extraction and dimensionality reduction commonly employed in high-dimensional datasets [47,48], which can map the high-dimensional polynomial feature space to a low-dimensional feature space while retaining the most important information in the data.

3.6. Model Construction and Optimization

3.6.1. Fundamentals of the SVR Algorithm

SVR is a machine learning method founded on the notion of structural risk minimization, aiming to design an appropriate regression hyperplane in the feature space to reduce the discrepancy between predicted and actual values [49]. This study adopts the widely used $\epsilon -SVR$ form, which introduces a $\epsilon -insensitive loss function$ to find a function $f\left(x\right)$ such that, for a given training dataset ${\left(x_i,y_i\right)}_{i=1}^n$ satisfies $|y_i-f\left(x_i\right)|\le \epsilon$, where $x_i$ is the input feature, $y_i$ is the target value, and $\epsilon$ is the predefined tolerance error. This ensures that the prediction values $f\left(x_i\right)$ of all training samples fall within an ε-wide “tube” relative to the true values $y_i$ The SVR schematic diagram is shown in Figure 4. To allow some errors to exceed $\epsilon$ slack variables ${\xi }_i$ and ${\xi }_i^{*}$ are introduced. In this diagram, the blue and yellow points represent training data samples. The two black dashed lines form the $\epsilon -insensitive$ band, within which errors are not penalized. The blue solid line is the fitted regression function $wx+b=0$. Thus, the optimization problem can be represented as shown in Formula (3):

$$min{\displaystyle \frac{1}{2}}\left|\left| w^2\right|\right|+C{\displaystyle \sum }_{i=1}^n({\xi }_i+{\xi }_i^{*})$$

$$s.t\left\{\begin{array}{c}{y}_i-wx-b\le \epsilon +{\xi }_i\\ wx+b-y_i\le \epsilon +{\xi }_i^{*}\\ {\xi }_i,{\xi }_i^{*}\ge 0\left(i=1,2\dots \dots ,m\right)\end{array} \right.$$

(3)

In the formula, $w$ represents the weight vector, $b$ denotes the bias, and $C$ serves as the penalty factor, which balances model complexity and training error.

The performance of SVR is highly dependent on the settings of its core hyperparameters. $C$ is the penalty parameter, which controls the model’s tolerance for training error. When $C$ is large, the model imposes stricter penalties on training error, leading to increased model complexity and potentially causing overfitting; $\epsilon$ determines the tolerance range for error. When $\epsilon$ is large, the model becomes more tolerant of small errors, reducing model complexity and potentially leading to underfitting; The $\mathrm{kernel}$ represents the kernel function, with common options including the linear kernel, polynomial kernel, and Radial Basis Function Kernel (RBF). This study selected the RBF, which has strong representational capabilities for nonlinear relationships and is highly compatible with the characteristics of geological data. $\gamma$ is a parameter used with the RBF, determining the width of the RBF. When $\gamma$ is large, the RBF has a narrower width, resulting in a more complex decision boundary for the model, which may lead to overfitting. SVR combines high-precision modeling capabilities for small-sample nonlinear systems with strong generalization performance, offering unique application advantages in the field of coal mine water inflow prediction.

3.6.2. Fundamentals of the HO

HO is an innovative meta-heuristic algorithm derived from the foraging and social behaviors of hippopotamuses. The HO replicates three behaviors of hippopotamuses in aquatic environments: exploration, predator protection, and predator evasion. It develops a three-stage model to facilitate the optimization process, adeptly balancing global exploration and local exploitation capabilities, thereby enhancing convergence time and solution accuracy in multidimensional, multimodal optimization challenges. In this study, the HO was selected as the optimizer to explore the applicability of novel metaheuristic methods in coal mine water inrush prediction, providing a new perspective for algorithmic diversity in this field.

The HO encodes the parameters for optimization, creating Hippo individuals, each of which is regarded as a potential solution. A specific quantity of Hippo individuals is randomly generated within the search space to constitute the initial population. The caliber of the initial population directly influences the convergence rate of the algorithm and the quality of the ultimate answer. The population of hippos can be denoted as a matrix $X$:

$$X=\left[\begin{array}{c}{X}_1\\ \dots \\ X_2\\ \dots \\ X_3\end{array}\right]=\left[\begin{array}{ccccc}{x}_{11}& \dots & x_{1j}& \dots & x_{1m}\\ ⋮& \ddots & ⋮& ⋰& ⋮\\ x_{v1}& \dots & x_{vj}& \dots & x_{vm}\\ ⋮& ⋰& ⋮& \ddots & ⋮\\ x_{n1}& \dots & x_{nj}& \dots & x_{nm}\end{array}\right]$$

(4)

In the matrix, $x$ denotes the position of the $v$th candidate solution, where $v=1,2,\dots ,n$ and $j=1,2,\dots ,m$; n signifies the size of the hippopotamus population in the herd, while $m$ indicates the number of decision variables in the problem.

Phase 1: Hippopotamus location updates in rivers or ponds. Hippopotamus groups comprise many adult females, juvenile hippos, multiple adult males, and a dominant male. Hippos generally congregate together, with the dominant male safeguarding the herd and territory from potential dangers, while several female hippos remain near the male. Upon reaching maturity, male hippos are expelled from the herd by the alpha male. Occasionally, juvenile hippos may grow estranged from the herd or their maternal figure. Male hippos navigate unpredictably in the water when foraging for food, hence broadening the search area and improving the algorithm’s global exploration efficacy. Their habitats in lakes or ponds are as follows:

$$X_i^M:x_{ij}^M=x_{ij}+y_1\left(D_h-I_1{x}_{ij}\right) i=1,2\dots ,{\displaystyle \frac{n}{2}} ,j=1,2,\dots \dots ,m$$

(5)

In the formula, $X_i^M$ denotes the position of the male hippopotamus; $x_{ij}^M$ indicates the position value of the male hippopotamus on the $j$th decision variable; $x_{ij}$ signifies the position value of the $i$th hippopotamus on the $j$th decision variable; $D_h$ signifies the position of the dominant male hippopotamus; $I_1$ is an integer value ranging from 1 to 2.

$$T=\exp \left({\displaystyle \frac{-t}{M}}\right)$$

(6)

In the formula, $T$ is employed to ascertain whether a juvenile hippopotamus has departed from its mother; $t$ represents the number of iterations, and when $T>0.6$, it signifies that the juvenile hippopotamus has left its mother; $M$ denotes the maximum number of iterations.

Phase 2: Hippopotamus defense against predators. In the presence of predators, hippopotamuses display defensive behavior by orienting towards the threat and potentially advancing to deter it. This phase seeks to conduct an exhaustive exploration of the local solution space to identify an improved solution. The predator’s location is:

$$P_j=l_i+r_1\left(u_i-l_j\right)$$

(7)

In the formula, $P_j$ denotes the position of the predator; $l_i$ represents the lower limit of the variable; $r_1$ is a stochastic vector ranging from 0 to 1; $u_i$ signifies the upper limit of the variable.

Phase 3: Hippos flee from predators. When faced with predators they cannot resist, hippos flee quickly. The HO simulates the escape behavior of hippos, enabling individuals to jump out of local optima and thereby enhancing the algorithm’s global search capabilities. The nearest safe location found by the hippos is:

$$X_i^s:x_{ij}^s=x_{ij}+r_2\left[l_j^{loc}+s_1\left(u_j^{loc}-l_j^{loc}\right)\right]$$

(8)

$$l_j^{loc}={\displaystyle \frac{l_i}{t}};u_j^{loc}={\displaystyle \frac{u_j}{t}};s_1=\left\{\begin{array}{c}2r_3-1\\ randn\left( \right)\\ rand\left( \right)\end{array}\right.$$

(9)

In the formula, $X_i^s$ represents the closest secure point identified; $t=1,2\dots ,M_{it}$; $r_2$ denotes a random number within the interval [0, 1]; $s_1$ signifies a random number or vector; $r_3$ is a random vector within the interval [0, 1]; $randn\left( \right)$ generates a normally distributed random number; $rand\left( \right)$ produces a random number within the interval [0, 1].

HO identifies the optimal solution by continuously adjusting the placements of the hippos. To evaluate the efficacy of the solution, it is necessary to establish a fitness function. In machine learning, RMSE is a prevalent evaluation metric. Upon reaching the maximum number of iterations, HO ceases operations and presents the set of potential solutions exhibiting the optimal fitness values as the refined model parameters [50].

3.6.3. HO-SVR Model

The HO-SVR model markedly enhances predictive performance by including HO for the adaptive optimization of SVR hyperparameters. The construction process is illustrated in Figure 5, with the specific processes outlined as follows:

Step 1: Initially, preprocess the training sample data to guarantee data quality and relevance.

Step 2: Establish the population size and the maximum number of iterations, while delineating the range of hyperparameters for the SVR model to be optimized, which includes the penalty coefficient $C\in \left[C_{min},C_{max}\right]$, the kernel function parameter $\gamma \in \left[{\gamma }_{min},{\gamma }_{max}\right]$, and the insensitivity bandwidth $\epsilon ϵ \left[{\epsilon }_{min},{\epsilon }_{max}\right]$. Simultaneously, establish the fitness function as Fitness = RMSE to assess the model’s predictive performance.

Step 3: Initialize the hippo population. Each hippo individual represents a set of random hyperparameter combinations $(C,\gamma ,\epsilon$) of the SVR model. The population size controls the breadth of the search. A bigger population may cover a wider search space but also raises the computing cost.

Step 4: Iterative optimization. For each individual hippo, utilize the augmented training set to conduct K-fold cross-validation and derive the Fitness value. Utilizing the Fitness value, implement the three behaviors of the hippo optimization algorithm to revise the population positions, create a new population, and update the current global optimum.

Step 5: Ascertain whether the maximum iteration count has been attained. If not, revert to Step 4 to persist with the iteration; if it has been reached, advance to the subsequent step.

Step 6: After reaching the maximum iteration count, identify the hippo individual with the best Fitness value. The corresponding $(\mathrm{C},\mathrm{\gamma },\mathrm{\epsilon }$) represents the optimal parameter combination, which is output as the determined optimal parameter combination.

Step 7: Retrain the SVR prediction model using the optimal parameters on the training set to obtain the HO-SVR model, which is used for prediction tasks in practical applications.

3.6.4. Establishment of Comparison Models

To evaluate the effectiveness of the HO in parameter optimization, five comparison models were established in this study, including HO-RF, HO-LightGBM, and SVR, as well as RF and LightGBM, all based on standard parameter settings. The construction processes of HO-RF and HO-LightGBM are similar to those of HO-SVR, both of which use the HO to optimize the key parameters of the model. The target optimization parameters for HO-RF and HO-LightGBM are shown in

Table 3. Target optimization parameters for HO-RF and HO-LightGBM.

Model	Key Parameters to Be Optimized	Meaning
HO-RF	n_estimators	Number of decision trees
	max_depth	Maximum depth of a single tree
	min_samples_split	Minimum number of samples for node splitting
	max_features	Proportion of features considered when splitting
HO-LightGBM	learning_rate	Learning rate
	num_leaves	Number of leaf nodes
	max_depth	Maximum depth of the tree
	feature_fraction	Feature sampling ratio
	lambda_l1	L1 regularization coefficient
	lambda_l2	L2 regularization coefficient

, with both models using RMSE as their fitness function. Ultimately, these optimized models were trained on the enhanced training set, following feature engineering, to obtain the optimal random forest and LightGBM models. The baseline models for SVR, RF, and LightGBM were trained using the enhanced training set with commonly used parameters based on experience.

3.6.5. Prediction Model Evaluation Metrics

This study utilizes the Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Coefficient of Determination (R²) as evaluation metrics to objectively assess the performance of various prediction models. MAE, MAPE, and RMSE quantify the deviation between predicted and actual values, with lower values indicating superior model accuracy. R², ranging from 0 to 1, indicates the model’s ability to explain variations in the target variable; values closer to 1 signify enhanced fitting performance. The specific formulas are detailed in Formulas (10)–(13).

To ensure a robust evaluation, a 10-fold cross-validation procedure was rigorously adopted. The final performance for each model is reported as the mean value and standard deviation calculated across the ten folds for all four metrics. The mean value reflects the central tendency of the model’s performance, while the standard deviation quantifies its stability and variability, together providing a comprehensive assessment of the model’s generalization capability.

$$MAE={\displaystyle \frac{1}{n}}{\displaystyle \sum }_{i=1}^n\left|{\widehat{y}}_i-y_i\right|$$

(10)

$$APE={\displaystyle \frac{100\%}{n}}{\displaystyle \sum }_{i=1}^n|{\displaystyle \frac{{\widehat{y}}_i-y_i}{yi}}|$$

(11)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum }_{i=1}^n{({\widehat{y}}_i-y_i)}^2}$$

(12)

$$R^2=1-{\displaystyle \frac{{{\displaystyle \sum }}_{i=1}^n{({\widehat{y}}_i-y_i)}^2}{{{\displaystyle \sum }}_i^n{(y_i-\overline{y})}^2}}$$

(13)

In the formula, ${\widehat{y}}_i$ denotes the predicted value, $y_i$ denotes the actual value of the data, $\overline{y}$ denotes the average of the actual values, and n denotes the number of prediction points.

3.7. RAG- LLM Intelligent Decision-Making

LLM have demonstrated powerful semantic understanding and text generation capabilities in the field of natural language processing, particularly excelling at multidimensional analysis and knowledge condensation of complex data [51,52]. In this study, the core value of LLM lies in converting the numerical results predicted by HO-SVR and the importance of features within the model into interpretable reports on water inrush mechanisms and structured prevention strategies, significantly enhancing the operability of prediction results. However, LLM has limitations in terms of insufficient domain knowledge depth and hallucination risks, which may severely constrain its reliability in safety-critical scenarios [53,54]. This study innovatively introduces the RAG architecture (Figure 6) to overcome these limitations through a domain knowledge anchoring mechanism [55,56]. The core process of RAG is divided into two stages: retrieval and enhancement. In the retrieval stage, after the user inputs a query, the query text is first converted into an embedding vector using an embedding model. The query’s embedding vector is then compared with a pre-built vector store index. The vector store index contains the embedding vectors of various documents or information fragments from a large corpus. To identify the most relevant documents, metrics such as cosine similarity are typically used to calculate the similarity between the query embedding vector and the document embedding vectors. In the generation phase, the retrieved relevant document fragments are used as contextual information and input into the LLM along with the original query. The model understands the query intent based on the provided context and synthesizes the final response.

The RAG-LLM intelligent decision-making module developed in this work employs predictive outcomes and essential attributes as query vectors to align pertinent regulatory provisions and historical cases from the established knowledge base. The LLM produces outputs derived from authorized knowledge, guaranteeing that the suggested countermeasures adhere exactly to industry norms. The “retrieval-generation” collaborative mechanism separates the LLM’s semantic generating abilities from domain-specific knowledge, maintaining the adaptability of language generation while essentially mitigating the possibility of technical parameter hallucinations. The choice of the LLM is paramount. DeepSeek excels at comprehending Chinese engineering texts, significantly outperforming general-purpose models, and accommodates an ultra-long context of 128K, facilitating the efficient processing of extensive technical documents like the “Regulations on Water Prevention and Control in Coal Mines.” Moreover, its open-source, configurable framework facilitates future local deployment to guarantee data security for the sensitive data management needs of coal mines. This study used DeepSeek-R1 as the primary LLM engine. The precise application procedure is as follows:

3.7.1. Input Layer

To drive the RAG intelligent decision-making module, it is necessary to analyze the key features in the HO-SVR prediction results. Since the standard SVR model does not directly output feature importance, this study uses permutation importance analysis to indirectly evaluate feature contribution. First, the initial RMSE of HO-SVR is calculated on the test set (denoted as $RMS{E}_{base}$). Next, the sample values of the target feature $X_j$ are randomly permuted, and the permuted RMSE is calculated (denoted as $RMS{E}_{permuted}$). Finally, the feature importance $Imp{X}_j$ is calculated according to Formula (14), and features with $Imp{X}_j$ > 5% are selected as key features. Ultimately, the predicted water flow values, the top 5 important features, and their feature importance values are extracted to form the input query vector for the RAG module.

$$Imp{X}_j={\displaystyle \frac{RMS{E}_{permuted}-RMS{E}_{base}}{RMS{E}_{base}}}\times 100\%$$

(14)

3.7.2. Retrieval Enhancement Layer

This study primarily gathers authoritative knowledge in coal mine water prevention and control to establish a domain-specific knowledge base (

Table 4. Composition of the domain-specific knowledge base for RAG-LLM decision support.

Category	Source Example	Quantity
National regulations	Detailed Rules for Water Prevention and Control in Coal Mines, Coal Mine Safety Regulations, Management Measures for the “Three Zones” of Water Prevention and Control in Coal Mines, etc.	12
Local standards and regulations	Shandong Province’s “Implementation Rules for the Management Measures for the Three Zones of Coal Mine Water Prevention and Control”, etc.	8
Engineering cases	Shandong Huafeng Coal Mine “6·1” Major Water Hazard Accident, etc.	63
Research literature	Papers on the mechanism of water inrush in coal mine floors, factors causing disasters, and water hazard prevention and control technologies.	364

), which facilitates RAG decision-making through semantics-oriented chunk optimization and vectorized retrieval. Initially, documents are input into the system for preliminary examination and categorization. based on the attributes and structure of the document content, they are segmented into text blocks that convey reasonably comprehensive meanings. Subsequently, semantic analysis is conducted on the text segments to identify significant entities, concepts, and semantic links, thereby consolidating semantically related content into more comprehensive semantic blocks. The mxbai-embed-large model generates embedding vectors for the semantic blocks, transforming textual semantic information into high-dimensional vector space representations. In the retrieval phase, the user-inputted sudden water flow prediction values and feature importance are vectorized. Utilizing cosine similarity, the system conducts real-time matching of the top-K highly relevant semantic fragments in the vector database, directly outputting them to the generation layer.

3.7.3. Generation Layer

The Top-K highly relevant semantic fragments output by the retrieval layer are fused with the original query and input into the LLM. Through a triple constraint mechanism, a decision report is generated. First, an authoritative context is constructed based on the retrieval content to correct the LLM’s free generation bias forcefully. Second, feature importance is analyzed to inform the model’s quantification of the contribution of disaster-causing factors. Finally, using structured prompt engineering, a prompt template (

Table 5. RAG-LLM Structured Prompt Design.

As a Coal Mine Water Prevention Expert, Please Generate an Implementable Water Inrush Risk Prevention Plan Based on The Following Forecast Data and Professional Knowledge, and Cite Content from the Knowledge Base as Core Support:
<input>
1. predicted water discharge: {y_hat} m3/h
2. top 5 key features and their importance: {feat.name_1}:{x_1_value} {unit} (importance: {imp_score_1}%) {feat.name_2}:{x_2_value} {unit} (importance: {imp_score_2}%) …… {feat.name_5}:{x_2_value} {unit} (importance: {imp_score_5}%)
<output>
generation requirements: (1) please explain the mechanism of water inrush based on the above information and professional knowledge, and describe the risk mechanism. (2) propose specific prevention and control measures, strictly ranked in order of priority, and cite relevant knowledge base provisions.

) is designed to drive DeepSeek-R1 to generate a report on the analysis and prevention measures for water inrush mechanisms.

3.7.4. Decision-Making Model Evaluation Metrics

This study utilizes a 5-point Likert scale (1 = extremely poor, 2 = poor, 3 = average, 4 = good, 5 = excellent) to perform a multi-dimensional evaluation of the RAG-LLM decision report, thereby validating the scientific rigor and engineering feasibility of the intelligent decision-making module’s output. Three specialists in coal mine water prevention and control were requested to assess 15 randomly selected generated reports. The assessment dimensions and scoring criteria are presented in

Table 6. Expert Assessment Scale for Evaluating RAG-LLM Output Quality.

Evaluation Dimensions	Scoring Criteria (Scale: 1–5)
Scientific explanation of the mechanism	Assessing the scientific accuracy of the water eruption mechanism explanation and evaluating the rationale of the analysis of the significance of its characteristics.
Feasibility of prevention and control measures	Determining if the measures are quantifiable and operational.
Compliance with regulations	Assess the accuracy of the citation of standard clauses and the compliance of the parameters with the standards.
Reasonableness of priority	Assessing the consistency of the measure sequence with the significance of attributes and evaluating the rationality of the logic.

, with each dimension rated on a scale of 1 to 5, resulting in a total of 20 points.

4. Results

4.1. Feature Engineering Results

The intricate nonlinear properties of coal mine floor water inrush data reveal that the mutual information feature selection technique exhibits considerable advantages across all three model types and dual-indicator assessments. Figure 7 illustrates that the mutual information feature selection method attains maximum R² values for SVR, RF, and LightGBM, with enhancements of 19.16%, 3.64%, and 15.25% relative to the original features, and 18.63%, 2.78%, and 18.45% compared to the polynomial construction with PCA strategy; additionally, its RMAE value is the lowest among the three methodologies. The mutual information feature selection method directly identifies critical disaster-inducing factors that are strongly correlated with water inrush volume, circumventing the noise interference that may arise from polynomial construction and the loss of essential nonlinear information associated with PCA dimensionality reduction. The consistency verification of the three heterogeneous models corroborates the universal analytical capability of the mutual information feature selection method for the nonlinear mechanisms of water inrush. Therefore, this study adopts the mutual information feature selection method as the final feature engineering strategy to maximize the model’s identification accuracy and generalization performance for disaster-driving factors. Based on mutual information feature analysis, this study employs the cumulative importance curve method to determine the optimal dimension of the feature subset. Features were sorted from high to low based on their mutual information scores with water inrush volume, as shown in Figure 8. By sequentially calculating the cumulative mutual information contribution of each feature, an engineering experience threshold of at least 80% cumulative mutual information score was established to identify the minimum number of features that meet the criteria. Empirical results show that the cumulative contribution rate of the first 14 features is 78.23%. After including the 15th feature, the cumulative contribution rate increases to 81.76%, which is significantly higher than the empirical threshold. Therefore, the top 15 feature subset is ultimately retained, achieving efficient feature space reduction while ensuring the integrity of core disaster-causing information.

The feature ranking derived from mutual information analysis provides a data-driven revelation of the core mechanisms behind coal mine floor water inrush. The highest-ranked features show remarkable consistency with established hydrogeological theory: Aquifer Water Pressure (X₄) acts as the fundamental driving force, Aquitard Thickness (X₆) constitutes the key barrier, and Fault Displacement (X₁₇) determines the development of water-conducting pathways, together forming the essential “source-barrier-conduit” system of inrush risk. The high ranking of mining engineering features such as Strike Length of Working Face (X₂₂), Depth of Floor Damage (X₇), Effective Thickness of Floor Aquitard (X₈), and Mining Height of Working Face (X₁₂) underscores the decisive impact of mining-induced disturbances on floor stability. The underlying mechanism is that the comparison between the mining-induced failure depth and the effective aquitard thickness directly controls the occurrence of water inrush. Furthermore, Coal Seam Dip Angle (X₂₃) influences stress distribution, while the Percentage of Mudstone in the Aquitard (X₁₀) reflects the plasticity of the rock mass, each contributing to the risk from different aspects. This analysis not only validates the effectiveness of the feature selection method but, more importantly, strongly demonstrates the consistency between the patterns captured by the data-driven model and the physical mechanisms, significantly enhancing the model’s interpretability and its value in guiding engineering practice.

4.2. Optimization Process Analysis

This study established the value ranges for the hyperparameters to thoroughly identify the ideal combination. The penalty parameter $C$ has a value range of $\left[0.1, 100.0\right]$, the RBF parameter $\gamma$ has a value range of $\left[0.0001, 10.0\right]$, and the insensitivity loss function parameter $\epsilon$ has a value range of $\left[0.001, 1.0\right]$. Figure 9 shows the SVR hyperparameter optimization process using RMSE as the fitness value. As shown in the figure, the RMSE value exhibits a clear decreasing trend as the number of iterations increases. The initial RMSE was 0.8500, and after 13 iterations, the final RMSE decreased to 0.6780, representing a 20.24% reduction compared to the initial value, indicating that the optimization process effectively improved the model’s predictive performance. By automatically searching for hyperparameters within the aforementioned range using the Hippo optimization algorithm, an optimal set of hyperparameter combinations was determined, including $C=45.67, \gamma =0.0123, \epsilon =0.0056$. Additionally, to compare the effectiveness of HO in improving the performance of the SVR model, HO was further applied to the core hyperparameter optimization of the RF and LightGBM models. The hyperparameter search space for RF model optimization is set as follows: $n_{estimators}\in \left[10, 200\right], ma{x}_{depth}\in \left[2, 20\right], mi{n}_{sample{s}_{split}}\in \left[2, 10\right], max\_features \in \left[0.1, 1.0\right]$. The hyperparameter search space for optimizing the LightGBM model was set as follows: $learnin{g}_{rate}\in \left[0.001, 0.3\right], nu{m}_{leaves}\in \left[10, 100\right], ma{x}_{depth}\in \left[2, 20\right], featur{e}_{fraction}\in \left[0.5, 1.0\right], lambd{a}_{l1}\in \left[0, 5\right], lambda\_l2 \in \left[0, 5\right]$. HO performed efficient searches in the RF and LightGBM parameter spaces, converging successfully after 21 and 33 iterations, respectively, and yielding a set of optimal hyperparameter combinations for each.

4.3. Comparison of Model Prediction Performance

4.3.1. Multi-Metric Comparative Evaluation of Model Performance

In the study of predicting water inrush volume in coal mine floor strata, to validate the effectiveness and superiority of the constructed HO-SVR model, five comparison models were established. To provide a more robust statistical evaluation, the model performance was re-assessed using 10-fold cross-validation, with the results shown in

Table 7. Comparison of the performance of coal mine floor water inrush volume prediction models.

Prediction Models	MAE (Mean ± Std)	MAPE (Mean ± Std)	RMSE (Mean ± Std)	R2 (Mean ± Std)
RF	2.5482 ± 0.2575	36.94% ± 3.65%	3.9075 ± 0.3920	0.8781 ± 0.0495
LightGBM	1.7639 ± 0.1634	22.77% ± 2.15%	2.9046 ± 0.2709	0.8958 ± 0.0384
SVR	4.8121 ± 0.5220	53.68% ± 5.36%	5.0213 ± 0.5443	0.7215 ± 0.1110
HO-RF	1.5375 ± 0.1026	19.36% ± 1.33%	1.9682 ± 0.1316	0.9325 ± 0.0301
HO-LightGBM	1.1971 ± 0.1240	15.22% ± 1.57%	1.5758 ± 0.1634	0.9169 ± 0.0415
HO-SVR	0.4239 ± 0.0464	5.38% ± 0.59%	0.7298 ± 0.0795	0.9539 ± 0.0285

and Figure 10. Table 7 clearly presents the mean value and standard deviation of each model across the four evaluation metrics. As displayed, the HO-SVR model achieved the lowest mean values for MAE, MAPE, and RMSE, at 0.4239, 5.38%, and 0.7298, respectively, while achieving the highest mean R² value of 0.9539. Figure 10 visualizes these results using bar charts with error bars, where the height of the bars represents the mean value of the performance metrics, and the error bars (vertical line segments) illustrate the standard deviation (±Std) of the data. Critically, the HO-SVR model also demonstrated the smallest standard deviations across all metrics (e.g., R² Std = 0.0285), indicating that it possesses not only superior predictive accuracy but also exceptional stability across different data subsets, yielding highly reliable performance estimates. These high-precision and high-stability metrics are of critical importance in practical safety engineering, as they allow for more reliable early warning of water inrush disasters, helping to reduce the probability of false alarms and missed detections, thereby providing a solid decision-making basis for the development of underground engineering safety plans.

The effectiveness of the HO is evident by comparing the metrics of HO-SVR with SVR, HO-RF with RF, and HO-LightGBM with LightGBM. Taking SVR as an example, its mean MAE is 4.8121 (±0.5220). In contrast, HO-SVR reduces this error by approximately 91.2%, and its significantly smaller standard deviation (±0.0464) conclusively demonstrates that the HO optimization effectively overcomes the limitations of the initial SVR parameter settings, drastically enhancing both its precision and stability. Similarly, compared to RF and LightGBM, HO-RF and HO-LightGBM showed reductions in the mean values of error metrics and an increase in the mean R², with their standard deviations suggesting improved stability post-optimization. In the horizontal comparison among HO-optimized models and the overall comparison of all models, HO-SVR maintains a definitive performance advantage. Its error levels remain substantially lower than other optimized models: the mean MAE of HO-RF (1.5375) is approximately 3.6 times that of HO-SVR, while the mean MAE of HO-LightGBM (1.1971) is about 2.8 times higher. Furthermore, the error standard deviations of HO-RF and HO-LightGBM (e.g., RMSE Std of 0.1316 and 0.1634, respectively) are notably higher than that of HO-SVR (0.0795), underscoring HO-SVR’s dual advantage in both accuracy and robustness. These advantages enable more accurate risk assessment and allow for the targeted allocation of safety measures, thereby enhancing both safety and cost-effectiveness.

HO-SVR integrates the robust parameter optimization features of the HO with the benefits of the SVR model for high-dimensional space modeling. The 10-fold cross-validation results confirm that its predictive accuracy, error control, and model stability are significantly superior to those of other comparative models, thereby highlighting its substantial potential and practical relevance in addressing complex nonlinear engineering prediction challenges. The application of this model provides a reliable technical means for achieving proactive and accurate prevention and control of water disasters in coal mines, effectively enhancing the intrinsic safety level of mining operations.

4.3.2. Prediction Error Distribution and Model Diagnostics

Figure 11a–f illustrate the prediction error distributions and normality test outcomes for the six models, facilitating a deeper comprehension of the models’ prediction mistakes. The HO-SVR model exhibits superior error convergence properties, characterized by tightly clustered error distributions and a core interval of [−0.6, 0.8], the narrowest range on the horizontal axis, indicating minimal divergence between predicted and actual values. The error ranges for HO-RF and HO-LightGBM are [−6, 2] and [−1, 1.5], respectively; LingtGBM has a range of [−1.5, 1.5]; RF ranges from [−6, 4]; and SVR ranges from [−15, 20], with the latter demonstrating the greatest variability. Furthermore, the HO-SVR model demonstrates the highest percentage of samples with an error close to zero since the majority of samples display near-zero errors. In the Quantile-Quantile plot (Q-Q plot) of HO-SVR, the blue points are predominantly aligned with the red theoretical reference line, suggesting that the error adheres to a normal distribution $(e~N(0,{\sigma }^2$)). Conversely, HO-RF and HO-LightGBM demonstrate discrepancies at the distribution’s extremes, whereas the baseline model has a consistent bias.

A comparative analysis of error distribution characteristics reveals that the error range of the HO-SVR model is narrower than that of alternative models, rendering it highly dependable for predicting water inrush volume near the critical threshold and substantially mitigating safety decision-making risks in coal mines. In comparison to the error range of the RF model, the HO-SVR model displays the least sensitivity to outlier samples and exhibits markedly enhanced robustness. In comparison to the original SVR model, the error bandwidth of the HO-SVR model is reduced by a factor of 25, and the Q-Q plot shifts from a significant deviation to an impeccable alignment with the normal distribution, thereby substantially enhancing the predictive capacity of SVR. The HO-SVR, characterized by its compact error distribution, stringent normal distribution assumption, and markedly enhanced performance relative to other HO optimization models, serves as a high-precision, quantifiable predictive instrument for early warning of water inrush in coal mines, rendering it essential for engineering safety standards.

Despite the outstanding overall performance of the HO-SVR model, an in-depth analysis of prediction errors is crucial for uncovering its performance boundaries and charting directions for improvement. By analyzing the error distribution in Figure 11, we focused on samples with high prediction errors. A case-by-case study revealed that these high-error samples typically correspond to extreme or complex geological conditions. For instance, several high-error samples simultaneously exhibited feature combinations such as “extremely high aquifer water pressure (X₄ > 12 MPa)” and “significant fault displacement (X₁₇ > 8 m)”. The strong nonlinear coupling effects potentially arising from such multiple extreme geological factors, coupled with the relative scarcity of similarly complex cases in the training dataset, are likely the primary reasons for the model’s challenged predictive accuracy under these boundary conditions. This analysis does not undermine the model’s reliability; on the contrary, it enhances the credibility of the conclusions by clearly defining its applicable boundaries. Furthermore, it provides a clear path for future model enhancement: the primary task is to systematically collect cases involving complex geological conditions to enrich the diversity of the training data; secondly, exploring the explicit embedding of physical mechanism constraints into the model, or employing advanced techniques like quantile regression, could improve the model’s capability to characterize complex nonlinear relationships and prediction uncertainties.

4.4. RAG-LLM Intelligent Decision-Making

4.4.1. Typical Application Case Testing

To assess the usability and efficacy of the RAG-LLM intelligent decision-support module developed in this study, a representative test set #38 of water inrush samples was chosen for thorough validation. The sample forecasted a water inrush volume of up to 585 m³/h, indicating a significant risk level. Key features and their significance were derived using perturbation importance analysis utilizing HO-SVR, as illustrated in Figure 12.

Table 8. Key features and their importance (Sample #38).

Key Features	Value	Importance
$X_4$ (Aquifer Water Pressure)	12 MPa	63.21%
$X_{17}$ (Fault Displacement)	5 m	57.42%
$X_{10}$ (Percentage of Mudstone in the Aquitard)	0.11	54.64%
$X_{21}$ (Water Temperature)	55.72 °C	32.83%
$X_6$ (Aquitard Thickness)	167.5 m	25.65%

lists the feature names along with their corresponding values. Among them, X₄ (aquifer water pressure), X₁₇ (fault displacement), and X₁₀ (percentage of mudstone in the aquitard) all have importance values exceeding 50%, making them the dominant factors in this water inrush event. This study developed a structured prompt template to guide the DeepSeek-R1 model, which is based on the RAG architecture, resulting in a report that analyzes the water inrush mechanism and proposes feasible prevention and control measures (

Table 9. Prompt Design Template of the RAG-LLM and the Generated Prevention and Control Measures Report.

The Following Is the Prompt Entered Into the LLM and the Content Output by the LLM.
Prompt: As a coal mine water prevention expert, please generate an implementable water inrush risk prevention plan based on the following forecast data and professional knowledge, and cite content from the knowledge base as core support: <Input>
1. Predicted water inflow rate: 585 m3/h
2. Top 5 key features and their importance: Aquifer Water Pressure: 12 MPa (importance: 63.21%) Fault Displacement: 5 m (importance: 57.42%) Percentage of Mudstone in the Aquitard: 0.11 (importance: 54.64%) Water Temperature: 55.72 °C (importance: 32.83%) Aquitard Thickness: 167.5 m (importance: 25.65%)
<Output>
Generation requirements: (1) Kindly elucidate the method of water eruptions using the aforementioned information and your professional expertise, and delineate the risk mechanism. (2) Recommend precise prevention and control methods, meticulously prioritized, and reference pertinent sections from the knowledge base.
Mechanism of water inrush and prevention measures:
This document outlines a risk prevention and management strategy for coal mine floor water inrushes, anticipating a water influx of 585 m3/h. Feasible technical measures are offered in order of priority based on projected characteristics, the water inrush mechanism, and industry norms.
I. Analysis of the Mechanism of Water Inrush Risk Aquifer Water Pressure (12 MPa): High water pressure can easily break through weak areas of the water barrier, which is the main factor controlling water breakthrough. Fault Displacement (5 m): Fault formation creates water channels connecting aquifers and mining spaces. Percentage of Mudstone in the Aquitard (11%): The low proportion of mudstone results in insufficient waterproofing and permeability resistance, making it susceptible to high-pressure water penetration. Water Temperature (55.7 °C): High temperatures exacerbate rock fissure expansion and reduce the stability of the water barrier layer. Aquitard Thickness (167.5 m): The mining depth is shallow, the overlying rock layer pressure is insufficient, and the water-resistant layer has weak compressive strength. Water inrush path: High water pressure (12 MPa) → Penetration of brittle aquifer (11% mudstone) → Activation of fault zone (5 m drop) → Formation of water conduit → High-temperature water (55.7 °C) accelerates rock destruction → Water inrush volume reaches 585 m3/h. II. Decision Report on Mine Floor Water Inrush Prevention and Control Generated by the RAG-LLM (in order of priority) 1. Water Pressure Control and Regional Management Measures: Ground-directed drilling and grouting: For aquifer-bearing strata in the base plate, ground-directed drilling holes are created, and a cement-aggregate mixed grout (water-cement ratio 0.8:1) is injected to convert the aquifer-bearing strata into weak aquifer-bearing strata. Underground dewatering and pressure reduction: Bore cross-layer holes (94 mm diameter) in the tunnel, install pressure control valves, and incrementally reduce water pressure to a safe level (≤0.1 MPa/m water inrush coefficient). References: Article 74 of the Detailed Rules for Water Prevention and Control in Coal Mines, Article 311 of the Coal Mine Safety Regulations, and a case study of a mine in the Huanghebei Coalfield, Shandong Province. 2. Fault Sealing and Aquifer Reinforcement Measures: Fault Grouting Curtain: Bore holes along the fault line at intervals of 20 m and inject a dual-liquid grout composed of ultra-fine cement and water glass (setting time 30 s) to create a water-impermeable curtain. Base plate reinforcement: Underground radial grouting (depth of 30 m) is employed to enhance the water resistance of compromised mudstone regions. References: “Grouting Diffusion Mechanism in Non-Homogeneous Fault Zones”, “Coal Mine Safety Regulations” Article 305. 3. Real-time monitoring and early warning response Measures: A microseismic monitoring system comprises 12-channel sensors positioned at the working face to continuously monitor bedrock fracture signals, triggering an alarm when energy exceeds 103 J. Hydrological dynamic monitoring involves assessing water pressure, water temperature (initiating the cooling plan when the temperature above 55 °C), and water volume (evacuating when the change rate reaches 10%) on a per-shift basis. References: Article 47 of the Detailed Rules for the Prevention and Control of Coal Mine Impact Pressure, Articles 285 and 235 of the Coal Mine Safety Regulations, “A Method for Real-Time Spatial-Temporal Early Warning of Floor Water Inrush in Working Faces Based on Microseismic Monitoring”, and the case study of the 1604 working face at a particular mine in the Jiaozuo Mining Area. 4. Emergency Response and Management Support Measures: Water-resistant coal pillars must be preserved on either side of the fault, with dimensions determined in accordance with the “Detailed Rules for Water Prevention and Control in Coal Mines,” and shall not be less than 20 m. Quarterly flood evacuation drills will be executed, targeting evacuation to the +50 m elevation safety zone, with a reaction time of under 8 min. References: Articles 91 to 94 of the Detailed Rules for Water Prevention and Control in Coal Mines, “Research on the Activation Characteristics of Faults and the Rational Retention of Water-Resistant Coal Pillars”, and the Key Points for Supervision, Inspection, and Enforcement of Water Hazard Prevention and Control in Coal Mines (2022 Edition).

). This effectively illustrates the robust capabilities and practical utility of the RAG-LLM intelligent decision support module in facilitating decision-making for the prevention and control of water inrush risks in coal mine bottom plates.

4.4.2. Expert Evaluation

This study engaged three experts in coal mine water prevention and control to perform a multi-dimensional quantitative evaluation of 15 randomly selected RAG-LLM decision-making reports, ensuring scientific rigor, adherence to standards, and practical engineering value. The assessment results are presented in

Table 10. Expert Evaluation Results on the Decision Report Generated by the RAG-LLM.

Expert	Average Score for Scientific Explanation of the Mechanism	Average Score for Feasibility of Prevention and Control Measures	Average Score for Compliance with Regulations	Average Score for Reasonableness of Priority	Average Total Score
Expert 1	4.6	4.5	4.6	4.7	18.4
Expert 2	4.2	4.3	3.8	4.4	16.7
Expert 3	4.4	4.0	4.2	4.5	17.1

, where each dimension’s score reflects the average of the 15 reports. According to the evaluation results, RAG-LLM demonstrated exceptional performance in the scientific validity of explanations regarding sudden water inrush mechanisms and the practicality of recommendations for prevention and control measures, highlighting the model’s fundamental ability to accurately link hydrogeological characteristics with engineering decision-making processes. In the realm of decision-making priority rationality, the model attained the highest average score of 4.53, confirming the efficacy of the RAG mechanism in dynamically filtering essential information and enhancing emergency resource allocation strategies, thus substantially bolstering the engineering viability of decisions. The average score for conformance with requirements was a modest 4.2, suggesting that the model requires enhancement in this area. The mean expert score was 17.4, equivalent to an average of 4.35 on a 5-point scale. The score findings confirm the sophisticated characteristics and practical utility of the RAG-LLM in intelligent decision-making contexts related to coal mine water inrush concerns. Subsequent research priorities should target standardization challenges through enhanced knowledge base integration and optimized constraint mechanisms, thereby advancing computational model efficacy and practical deployment capabilities.

5. Discussion

5.1. Theoretical Contributions

This study makes several significant theoretical contributions to the field of intelligent mine safety management. Primarily, it moves beyond the traditional “siloed” approach of prediction modeling by proposing a novel closed-loop intelligent prediction–prevention framework that seamlessly bridges the gap from data to decision. By integrating hybrid data augmentation, advanced feature engineering, HO-SVR, and RAG-LLM intelligent decision support, this framework systematically addresses the critical challenge of translating high-precision numerical predictions into interpretable mechanistic reports and executable strategies, thereby providing a methodological paradigm for the intelligent prevention and control of complex engineering problems. At the algorithmic level, the introduction of the HO into the domain of water inrush prediction demonstrates its exceptional efficacy in optimizing parameters for complex, non-linear engineering models, offering a new algorithmic tool to enhance the reliability and accuracy of predictive models. Most notably, the successful application of the RAG and LLM paradigm establishes a new paradigm for human–AI collaborative governance in geological hazards. This paradigm achieves a safe and controlled integration of domain-specific knowledge with powerful generative AI capabilities, providing a replicable methodological framework for effectively synergizing machine intelligence with human expertise in geological risk management. It promotes a theoretical paradigm shift from traditional model-driven approaches to a deep integration of knowledge-guided and data-driven methodologies.

5.2. Practical Contributions

The practical implications of the integrated intelligent framework proposed in this research are substantial for the coal mining industry’s pursuit of safety and sustainability. At the level of prediction and early warning, the developed HO-SVR model, with its high prediction accuracy and concentrated error distribution, provides a reliable and robust tool for the early warning of water inrush disasters. Its exceptional stability significantly enhances the reliability of risk identification, enabling mine operators to proactively and accurately identify high-risk conditions, thereby shifting safety management from a passive response to a proactive early warning paradigm. Regarding decision support, the core value of the RAG-LLM intelligent decision-support module lies in its ability to translate complex numerical predictions into directly understandable and actionable prevention and control measures for field engineers. As validated by expert evaluation, the generated reports not only offer scientifically sound mechanistic analysis but also deliver structured, prioritized action plans. This dramatically shortens the decision-making cycle from obtaining predictions to implementing strategies, providing crucial technological support for near-real-time decision-making in dynamic mining operations. In terms of systematic management, the framework effectively demystifies the “black box” nature of complex ML models and bridges the critical knowledge gap between data scientists and field engineers. The entire system offers a scalable blueprint for the intelligent upgrade of safety management systems. Its closed-loop paradigm of “precision prediction–intelligent decision-making” holds potential applicability beyond water inrush prevention, extending to the precise prevention and control of other geological hazards such as rockburst and gas outburst, thereby promoting a fundamental shift within the industry towards proactive, precise prevention and control.

5.3. Data Limitations and Generalization Ability: A Mechanistic Analysis and Framework

This study acknowledges that the HO-SVR model was trained on a dataset of 134 samples collected primarily from mining areas in Shandong Province. While the hybrid data augmentation technique (SMOTE–GN–Bootstrap) mitigated the issue of sample imbalance to some extent, the fundamental challenges associated with the relatively small sample size and its confined geographical origin for the model’s generalization ability require in-depth analysis.

Furthermore, while the hybrid data augmentation technique (SMOTE–GN–Bootstrap) effectively expanded the training set, it is crucial to address the potential introduction of artificial correlations and the lack of physical plausibility validation. For instance, the SMOTE method, which relies on nearest-neighbor interpolation, may generate samples that deviate from real geological mechanisms. To mitigate this, future implementations should incorporate physical constraints directly into the augmentation process, ensuring synthetic samples align with known hydrogeological laws. This approach would enhance the physical plausibility of augmented data and reduce the risk of learning spurious relationships.

5.3.1. Impact of Limited Sample Size and Geographical Specificity

From a machine learning perspective, models trained on relatively small datasets from a specific geographical region are susceptible to learning localized patterns that may not be universally applicable. The high prediction accuracy achieved by the HO-SVR model demonstrates its excellent performance within the geological context of the Shandong mining areas. However, this high performance may partly reflect the model’s adaptation to regional-specific “noise,” increasing the risk of overfitting to local characteristics rather than capturing the fundamental, transferable principles of water inrush mechanisms. Consequently, the model’s predictive performance could potentially degrade when applied to mining regions with significantly different geological genesis types. Secondly, the identified feature importance ranking, where factors are dominant, is intrinsically linked to the data context from which it was derived. The stability of this ranking across different geological settings cannot be guaranteed, as the primary controlling factors for water inrush may vary. Furthermore, the concentrated error distribution of the HO-SVR model, as shown in Figure 11, indicates robust performance on the available test set. However, the model’s sensitivity to samples that fall outside the dominant patterns within the training data underscores the potential challenge of generalizing to the broader and more diverse conditions found across different coalfields in China.

5.3.2. Applicability Analysis Based on Geological Mechanisms

Given the current impossibility of conducting large-scale cross-regional validation, a knowledge-driven, mechanistic deduction is employed to discuss the model’s potential applicability limitations.

(1) Application in North China-type Coalfields (Ordovician Limestone Aquifers): Water inrushes in North China-type coalfields (e.g., Shanxi, Hebei) are predominantly sourced from highly water-rich Ordovician limestone karst aquifers. The core mechanism hinges on the connectivity of karst conduits. While the key input features of our model, such as Aquifer Water Pressure (X₄), remain relevant, the decisive factor—the degree of karst development—is not explicitly captured in our feature set. Therefore, directly applying this model to North China-type coalfields might fail to accurately capture the core controlling elements of inrush risk, potentially leading to systematic prediction bias.

(2) Application in Northwestern Jurassic Coalfields (Shallow Burial, Thin Bedrock): Northwestern coalfields are characterized by shallow burial depths, thin bedrock, and overlying thick loose layers. Their inrush types often involve roof water or surface water seepage, which is fundamentally different from the “confined water breakout from the floor” mechanism typical in Shandong. The importance of key features relied upon by our model, such as X₇ (Depth of Floor Damage) and X₈ (Effective Thickness of Floor Aquitard), is expected to diminish significantly in the Northwest. In contrast, features like the height of the water-conducting fracture zone and the water abundance of loose aquifers, which are not included in the current model, become paramount. Thus, the direct applicability of the model in Northwestern regions is anticipated to be low, necessitating fundamental feature engineering reconstruction.

5.3.3. A Framework for Enhancing Model Generality

To address the identified limitations and pave the way for creating a more universally applicable model, we propose a clear framework for future research:

(1) Data Layer: Initiatives should be taken to construct a cross-regional collaborative database or knowledge graph for coal mine water inrush cases, systematically incorporating data from diverse geological units across China.

(2) Algorithm Layer: Explore Transfer Learning techniques. The HO-SVR model from this study can serve as a pre-trained base model. When applied to a new mining area, it can be fine-tuned using a small amount of local data from that area, enabling it to adapt quickly to the new geological environment and effectively address the data scarcity problem in new regions.

(3) Mechanism Layer: Develop Physics-Informed Machine Learning models. By embedding constraints based on hydrogeological and rock mechanics principles into the data-driven model, its reasoning capability and interpretability in data-scarce regions can be enhanced, reducing reliance on purely data-driven approaches.

6. Conclusions

This study successfully addressed key challenges in coal mine floor water inrush prediction and control, including data scarcity, feature redundancy, model parameter optimization, and the translation of predictions into actionable strategies. The main findings indicate that the HO-SVR model demonstrates superior and highly stable predictive performance, achieving an R² value of 0.9539 ± 0.0285 on the test set, with a highly concentrated prediction error distribution that significantly outperforms other benchmark models. Furthermore, the RAG-LLM intelligent decision-making module effectively bridged the gap between numerical predictions and actionable strategies by generating interpretable reports and prevention measures that were validated by domain experts. To overcome data limitations, a SMOTE–GN–Bootstrap hybrid data augmentation technique was developed, which effectively mitigated sample imbalance by expanding the training set from 94 to 200 samples and improving data representativeness. Finally, mutual information feature selection was identified as the optimal feature engineering strategy, as it directly selected critical disaster-inducing factors and outperformed polynomial feature construction combined with PCA. These components are integrated into a comprehensive framework that provides a practical technical solution for mine water hazard control. The framework’s high-precision prediction capability enables advanced risk early warning while the RAG-LLM module generates executable strategies for on-site guidance, collectively enhancing the accuracy and responsiveness of safety management. This research establishes a replicable methodological paradigm which is directly applicable in regions with analogous hydrogeological conditions and adaptable to other areas after validation, offering significant value for the precise prevention and control of water hazards in coal mines nationwide.

This study acknowledges certain limitations, primarily concerning the geographical specificity and scale of the dataset, as extensively analyzed in Section 5.3. To address these constraints and enhance the framework’s broader applicability, several pathways for future research are outlined. Future efforts will prioritize enhancing system adaptability through the integration of real-time monitoring data with physical models of rock mechanics and groundwater flow, thereby developing a dynamic hybrid system that combines data-driven predictions with physical principles. Concurrently, validating and improving the physical plausibility of data augmentation will be pursued through methods such as Physics-Informed Generative Adversarial Networks, ensuring synthetic samples adhere to geological and mechanical constraints. Furthermore, future work will focus on validating and enhancing the model’s generality by collecting multi-regional data and employing transfer learning techniques to adapt it across China’s major coalfields. Substantial advancements are also envisioned in the spatiotemporal visualization of prediction results and decision support, aiming to develop a digital twin system for mine water hazards that enables virtual–real mapping, real-time updates, and intelligent simulation. Additionally, exploring the lightweight deployment of the RAG-LLM module for on-site use and extending this framework to other geological hazards represent critical future directions. Finally, a systematic comparative validation of the HO optimizer with other established optimizers will be undertaken to comprehensively evaluate its performance boundaries and relative advantages.