Prediction of Blast-induced Air Over-pressure in Open-Pit Mine: Assessment of Different Artificial Intelligence Techniques

Air over-pressure (AOp) is one of the products of blasting operations for rock fragmentation in open-pit mines. It can cause structural vibration, smash glass doors, adversely affect the surrounding environment, and even be fatal to humans. To assess its dangerous effects, seven artificial intelligence (AI) methods for predicting specific blast-induced AOp have been applied and compared in this study. The seven methods include random forest, support vector regression, Gaussian process, Bayesian additive regression trees, boosted regression trees, k-nearest neighbors, and artificial neural network (ANN). An empirical technique was also used to compare with AI models. The degree of complexity and the performance of the models were compared with each other to find the optimal model for predicting blast-induced AOp. The Deo Nai open-pit coal mine (Vietnam) was selected as a case study where 113 blasting events have been recorded. Indicators used for evaluating model performances include the root-mean-square error (RMSE), determination coefficient (R²), and mean absolute error (MAE). The results indicate that AI techniques provide better performance than the empirical method. Although the relevance of the empirical approach was acceptable (R²?=?0.930) in this study, its error (RMSE?=?7.514) is highly significant to guarantee the safety of the surrounding environment. In contrast, the AI models offer much higher accuracies. Of the seven AI models, ANN was the most dominant model based on RMSE, R², and MAE. This study demonstrated that AI techniques are excellent for predicting blast-induced AOp in open-pit mines. These techniques are useful for blasters and managers in controlling undesirable effects of blasting operations on the surrounding environment.

     

Intelligent Prediction of Blasting-Induced Ground Vibration Using ANFIS Optimized by GA and PSO

Ground vibration induced by rock blasting is one of the most crucial problems in surface mines and tunneling projects. Hence, accurate prediction of ground vibration is an important prerequisite in the minimization of its environmental impacts. This study proposes hybrid intelligent models to predict ground vibration using adaptive neuro-fuzzy inference system (ANFIS) optimized by particle swarm optimization (PSO) and genetic algorithms (GAs). To build prediction models using ANFIS, ANFIS–GA, and ANFIS–PSO, a database was established, consisting of 86 data samples gathered from two quarries in Iran. The input parameters of the proposed models were the burden, spacing, stemming, powder factor, maximum charge per delay (MCD), and distance from the blast points, while peak particle velocity (PPV) was considered as the output parameter. Based on the sensitivity analysis results, MCD was found as the most effective parameter of PPV. To check the applicability and efficiency of the proposed models, several traditional performance indices such as determination coefficient (R²) and root-mean-square error (RMSE) were computed. The obtained results showed that the proposed ANFIS–GA and ANFIS–PSO models were capable of statistically predicting ground vibration with excellent levels of accuracy. Compared to the ANFIS, the ANFIS–GA model showed an approximately 61% decrease in RMSE and 10% increase in R². Also, the ANFIS–PSO model showed an approximately 53% decrease in RMSE and 9% increase in R² compared to ANFIS. In other words, the ANFIS performance was optimized with the use of GA and PSO.

     

Calibrating cellular automata based on landscape metrics by using genetic algorithms

Landscape metrics have been widely used to characterize geographical patterns which are important for many geographical and ecological analyses. Cellular automata (CA) are attractive for simulating settlement development, landscape evolution, urban dynamics, and land-use changes. Although various methods have been developed to calibrate CA, landscape metrics have not been explicitly used to ensure the simulated pattern best fitted to the actual one. This article presents a pattern-calibrated method which is based on a number of landscape metrics for implementing CA by using genetic algorithms (GAs). A Pattern-calibrated GA–CA is proposed by incorporating percentage of landscape (PLAND), patch metric (LPI), and landscape division (D) into the fitness function of GA. The sensitivity analysis can allow the users to explore various combinations of weights and examine their effects. The comparison between Logistic- CA, Cell-calibrated GA–CA, and Pattern-calibrated GA–CA indicates that the last method can yield the best results for calibrating CA, according to both the training and validation data. For example, Logistic-CA has the average simulation error of 27.7%, but Pattern-calibrated GA–CA (the proposed method) can reduce this error to only 7.2% by using the training data set in 2003. The validation is further carried out by using new validation data in 2008 and additional landscape metrics (e.g., Landscape shape index, edge density, and aggregation index) which have not been incorporated for calibrating CA models. The comparison shows that this pattern-calibrated CA has better performance than the other two conventional models.     

Time series forecasting of river flow using an integrated approach of wavelet multi-resolution analysis and evolutionary data-driven models. A case study: Sebaou River (Algeria)

The complexity of hydrological processes and lack of data for modeling require the use of specific tools for non-linear natural phenomenon. In this paper, an effort has been made to develop a conjunction model – wavelet transformation, data-driven models, and genetic algorithm (GA) – for forecasting the daily flow of a river in northern Algeria using the time series of runoff. This catchment has a semi-arid climate and strong variability in runoff. The original time series was decomposed into multi-frequency time series by wavelet transform algorithm and used as inputs to artificial neural networks (ANN) and adaptive neuro-fuzzy inference system (ANFIS) models. Several factors must be optimized to determine the best model structures. Wavelet-based data-driven models using a GA are designed to optimize model structure. The performances of wavelet-based data-driven models (i.e. WANFIS and WANN) were superior to those of conventional models. WANFIS (RMSE = 12.15 m³/s, EC = 87.32%, R = .934) and WANN (RMSE = 15.73 m³/s, EC = 78.83%, R = .888) models improved the performances of ANFIS (RMSE = 23.13 m³/s, EC = 54.11%, R = .748) and ANN (RMSE = 22.43 m³/s, EC = 56.90%, R = .755) during the test period.     

An extended cellular automaton using case‐based reasoning for simulating urban development in a large complex region

Rule‐based cellular automata (CA) have been increasingly applied to the simulation of geographical phenomena, such as urban evolution and land‐use changes. However, these models have difficulties and uncertainties in soliciting transition rules for a large complex region. This paper presents an extended cellular automaton in which transition rules are represented by using case‐based reasoning (CBR) techniques. The common k‐NN algorithm of CBR has been modified to incorporate the location factor to reflect the spatial variation of transition rules. Multi‐temporal remote‐sensing images are used to obtain the adaptation knowledge in the temporal dimension. This model has been applied to the simulation of urban development in the Pearl River Delta which has a hierarchy of cities. Comparison indicates that this model can produce more plausible results than rule‐based CA in simulating this large complex region in 1988–2002.     

A Novel Hunger Games Search Optimization-Based Artificial Neural Network for Predicting Ground Vibration Intensity Induced by Mine Blasting

Innovation efforts in developing soft computing models (SCMs) of researchers and scholars are significant in recent years, especially for problems in the mining industry. So far, many SCMs have been proposed and applied to practical engineering to predict ground vibration intensity (BIGV) induced by mine blasting with high accuracy and reliability. These models significantly contributed to mitigate the adverse effects of blasting operations in mines. Despite the fact that many SCMs have been introduced with promising results, but ambitious goals of researchers are still novel SCMs with the accuracy improved. They aim to prevent the damages caused by blasting operations to the surrounding environment. This study, therefore, proposed a novel SCM based on a robust meta-heuristic algorithm, namely Hunger Games Search (HGS) and artificial neural network (ANN), abbreviated as HGS–ANN model, for predicting BIGV. Three benchmark models based on three other meta-heuristic algorithms (i.e., particle swarm optimization (PSO), firefly algorithm (FFA), and grasshopper optimization algorithm (GOA)) and ANN, named as PSO–ANN, FFA–ANN, and GOA–ANN, were also examined to have a comprehensive evaluation of the HGS–ANN model. A set of data with 252 blasting operations was collected to evaluate the effects of BIGV through the mentioned models. The data were then preprocessed and normalized before splitting into individual parts for training and validating the models. In the training phase, the HGS algorithm with the optimal parameters was fine-tuned to train the ANN model to optimize the ANN model's weights. Based on the statistical criteria, the HGS–ANN model showed its best performance with an MAE of 1.153, RMSE of 1.761, R² of 0.922, and MAPE of 0.156, followed by the GOA–ANN, FFA–ANN and PSO–ANN models with the lower performances (i.e., MAE?=?1.186, 1.528, 1.505; RMSE?=?1.772, 2.085, 2.153; R²?=?0.921, 0.899, 0.893; MAPE?=?0.231, 0.215, 0.225, respectively). Based on the outstanding performance, the HGS–ANN model should be applied broadly and across a swath of open-pit mines to predict BIGV, aiming to optimize blast patterns and reduce the environmental effects.

     

A hybrid genetic algorithm with local optimiser improves calibration of a vegetation change cellular automata model

Cellular automata (CA) models are commonly used to model vegetation dynamics, with the genetic algorithm (GA) being one method of calibration. This article investigates different GA settings, as well as the combination of a GA with a local optimiser to improve the calibration effort. The case study is a pattern-calibrated CA to model vegetation regrowth in central Victoria, Australia. We tested 16 GA models, varying population size, mutation rate, and level of allowable mutation. We also investigated the effect of applying a local optimiser, the Nelder?Mead Downhill Simplex (NMDS) at GA convergence. We found that using a decreasing mutation rate can reduce computational cost while avoiding premature GA convergence, while increasing population size does not make the GA more efficient. The hybrid GA-NMDS can also reduce computational cost compared to a GA alone, while also improving the calibration metric. We conclude that careful consideration of GA settings, including population size and mutation rate, and in particular the addition of a local optimiser, can positively impact the efficiency and success of the GA algorithm, which can in turn lead to improved simulations using a well-calibrated CA model.     

A partitioned and asynchronous cellular automata model for urban growth simulation

Cellular automata (CA) models are used to analyze and simulate the global phenomenon of urban growth. However, these models are characterized by ignoring spatially heterogeneous transition rules and asynchronous evolving rates, which make it difficult to improve urban growth simulations. In this paper, a partitioned and asynchronous cellular automata (PACA) model was developed by implementing the spatial heterogeneity of both transition rules and evolving rates in urban growth simulations. After dividing the study area into several subregions by k-means and knn-cluster algorithms, a C5.0 decision tree algorithm was employed to identify the transition rules in each subregion. The evolving rates for cells in each regularly divided grid were calculated by the rate of changed cells. The proposed PACA model was implemented to simulate urban growth in Wuhan, a large city in central China. The results showed that PACA performed better than traditional CA models in both a cell-to-cell accuracy assessment and a shape dimension accuracy assessment. Figure of merit of PACA is 0.368 in this research, which is significantly higher than that of partitioned CA (0.327) and traditional CA (0.247). As for the shape dimension accuracy, PACA has a fractal dimension of 1.542, which is the closest to that of the actual land use (1.535). However, fractal dimension of traditional CA (1.548) is closer to that of the actual land use than that of partitioned CA (1.285). It indicates that partitioned transition rules play an important role in the cell-to-cell accuracy of CA models, whereas the combination of partitioned transition rules and asynchronous evolving rates results in improved cell-to-cell accuracy and shape dimension accuracy. Thus, implementing partitioned transition rules and asynchronous evolving rates yields better CA model performance in urban growth simulations due to its accordance with actual urban growth processes.     

Knowledge transfer and adaptation for land-use simulation with a logistic cellular automaton

Few studies have been conducted into the use of knowledge transfer for tackling geo-simulation problems. Cellular automata (CA) have proven to be an effective and convenient means of simulating urban dynamics and land-use changes. Gathering the knowledge required to build the CA may be difficult when these models are applied to large areas or long periods. In this paper, we will explore the possibility that the knowledge from previously collected data can be transferred spatially (a different region) and/or temporally (a different period) for implementing urban CA. The domain adaptation of CA is demonstrated by integrating logistic-CA with a knowledge-transfer technique, the TrAdaBoost algorithm. A modification has been made to the TrAdaBoost algorithm by incorporating a dynamicweight-trimming technique. This proposed model, CA_trans, is tested by choosing different periods and study areas in the Pearl River Delta. The ‘Figure of Merit’ measurements in the experiments indicate that CA_trans can yield better simulation results. The variance of traditional logistic-CA is about 2–5 times the variance of CA_trans until the number of new data reaches 30. The experiments have demonstrated that the proposed method can alleviate the sparse data problem using knowledge transfer.     

Diatom-inference models for surface-water temperature and salinity developed from a 57-lake calibration set from the Sierra Nevada, California, USA

Physical, chemical, and biological data were collected from a suite of 57 lakes that span an elevational gradient of 1360 m (2115 to 3475 m a.s.l.) in the eastern Sierra Nevada, California, USA as part of a multiproxy study aimed at developing transfer functions from which to infer past drought events. Multivariate statistical techniques, including canonical correspondence analysis (CCA), were used to determine the main environmental variables influencing diatom distributions in the study lakes. Lakewater depth, surface-water temperature, salinity, total Kjeldahl nitrogen, and total phosphorus were important variables in explaining variance in the diatom distributions. Weighted-averaging (WA) and weighted-averaging partial least squares (WA-PLS) were used to develop diatom-based surface-water temperature and salinity inference models. The two best diatom-inference models for surface-water temperature were developed using simple WA and inverse deshrinking. One model covered a larger surface-water temperature gradient (13.7 °C) and performed slightly poorer (r² = 0.72, RMSE = 1.4 °C, RMSEP_jack = 2.1 °C) than a second model, which covered a smaller gradient (9.5 °C) and performed slightly better (r² = 0.89, RMSE = 0.7 °C, RMSEP_jack = 1.5 °C). The best diatom-inference model for salinity was developed using WA-PLS with three components (r² = 0.96, RMSE = 4.06 mg L^–1, RMSEP_jack = 11.13 mg L^–1). These are presently the only diatom-based inference models for surface-water temperature and salinity developed for the southwestern United States. Application of these models to fossil-diatom assemblages preserved in Sierra Nevada lake sediments offers great potential for reconstructing a high-resolution time-series of Holocene and late Pleistocene climate and drought for California.     

A dimension-independent extrusion algorithm using generalised maps

One solution to the integration of additional characteristics, for example, time and scale, into GIS data sets is to model them as extra geometric dimensions perpendicular to the spatial ones, creating a higher-dimensional model. While this approach has been previously described and advocated, it is scarcely used in practice because of a lack of high-level construction algorithms and accompanying implementations. We present in this paper a dimension-independent extrusion algorithm permitting us to construct from any (n–1)-dimensional linear cell complex represented as a generalised map, an n-dimensional one by assigning to each (n–1)-cell one or more intervals where it exists along the nth dimension. We have implemented the algorithm in C++11 using CGAL, made the source code publicly available and tested it in experiments using real-world 2D GIS data sets which were extruded to construct up to 5D models.     

Across‐axis variations in petrophysical properties of the North Porcupine Basin,offshore Ireland: New insights from long‐streamer traveltime tomography

The Porcupine Basin is a Mesozoic failed rift located in the North Atlantic margin, SW of Ireland, in which a postrift phase of extensional faulting and reactivation of synrift faults occurred during the Mid–Late Eocene. Fault zones are known to act as either conduits or barriers for fluid flow and to contribute to overpressure. Yet, little is known about the distribution of fluids and their relation to the tectono‐stratigraphic architecture of the Porcupine Basin. One way to tackle this aspect is by assessing seismic (V_p) and petrophysical (e.g., porosity) properties of the basin stratigraphy. Here, we use for the first time in the Porcupine Basin 10‐km‐long‐streamer data to perform traveltime tomography of first arrivals and retrieve the 2D V_p structure of the postrift sequence along a ~130‐km‐long EW profile across the northern Porcupine Basin. A new V_p–density relationship is derived from the exploration wells tied to the seismic line to estimate density and bulk porosity of the Cenozoic postrift sequence from the tomographic result. The V_p model covers the shallowest 4 km of the basin and reveals a steeper vertical velocity gradient in the centre of the basin than in the flanks. This variation together with a relatively thick Neogene and Quaternary sediment accumulation in the centre of the basin suggests higher overburden pressure and compaction compared to the margins, implying fluid flow towards the edges of the basin driven by differential compaction. The V_p model also reveals two prominent subvertical low‐velocity bodies on the western margin of the basin. The tomographic model in combination with the time‐migrated seismic section shows that whereas the first anomaly spatially coincides with the western basin‐bounding fault, the second body occurs within the hangingwall of the fault, where no major faulting is observed. Porosity estimates suggest that this latter anomaly indicates pore overpressure of sandier Early–Mid Eocene units. Lithological well control together with fault displacement analysis suggests that the western basin‐bounding fault can act as a hydraulic barrier for fluids migrating from the centre of the basin towards its flanks, favouring fluid compartmentalization and overpressure of sandier units of its hangingwall.     

Modeling Short-Term Spatial and Temporal Variability of Groundwater Level Using Geostatistics and GIS

Continuous depletion of groundwater levels from deliberate and uncontrolled exploitation of groundwater resources lead to the severe problems in arid and semi-arid hard-rock regions of the world. Geostatistics and geographic information system (GIS) have been proved as successful tools for efficient planning and management of the groundwater resources. The present study demonstrated applicability of geostatistics and GIS to understand spatial and temporal behavior of groundwater levels in a semi-arid hard-rock aquifer of Western India. Monthly groundwater levels of 50 sites in the study area for 36-month period (May 2006 to June 2009; excluding 3 months) were analyzed to find spatial autocorrelation and variances in the groundwater levels. Experimental variogram of the observed groundwater levels was computed at 750-m lag distance interval and the four most-widely used geostatistical models were fitted to the experimental variogram. The best-fit geostatistical model was selected by using two goodness-of-fit criteria, i.e., root mean square error (RMSE) and correlation coefficient (r). Then spatial maps of the groundwater levels were prepared through kriging technique by means of the best-fit geostatistical model. Results of two spatial statistics (Geary’s C and Moran’s I) indicated a strong positive autocorrelation in the groundwater levels within 3-km lag distance. It is emphasized that the spatial statistics are promising tools for geostatistical modeling, which help choose appropriate values of model parameters. Nugget-sill ratio (<0.25) revealed that the groundwater levels have strong spatial dependence in the area. The statistical indicators (RMSE and r) suggested that any of the three geostatistical models, i.e., spherical, circular, and exponential, can be selected as the best-fit model for reliable and accurate spatial interpolation. However, exponential model is used as the best-fit model in the present study. Selection of the exponential model as the best-fit was further supported by very high values of coefficient of determination (r ² ranging from 0.927 to 0.994). Spatial distribution maps of groundwater levels indicated that the groundwater levels are strongly affected by surface topography and the presence of surface water bodies in the study area. Temporal pattern of the groundwater levels is mainly controlled by the rainy-season recharge and amount of groundwater extraction. Furthermore, it was found that the kriging technique is helpful in identifying critical locations over the study area where water saving and groundwater augmentation techniques need to be implemented to protect depleting groundwater resources.     

Testing local spatial autocorrelation using k-order neighbours

The analysis of local spatial autocorrelation for spatial attributes has been an important concern in geographical inquiry. In this paper, we propose a concept and algorithm of k-order neighbours based on Delaunay's triangulated irregular networks and redefine Getis and Ord's (1992) local spatial autocorrelation statistic as G_i(k) with weight coefficient w_ij(k) based on k-order neighbours for the study of local patterns in spatial attributes. To test the validity of these statistics, an experiment is performed using spatial data of the elderly population in Ichikawa City, Chiba Prefecture, Japan. The difference between the weight coefficients of the k-order neighbours and distance parameter to measure the spatial proximity of districts located in the city centre and near the city limits is found by Monte-Carlo simulation.     

The potential of site-specific and local chironomid-based inference models for reconstructing past lake levels

We examined the relationship between three key environmental variables (water depth, loss-on-ignition, and bottom-water temperature) and fossil chironomid distributions sampled from within-lake gradients in three small, moderately deep (18–35 m), maar lakes on St Michael Island, western Alaska. Site-specific (one lake, 29 samples) and local (three lakes, 87 samples) inference models for reconstructing water depth were developed using partial least squares regression and calibration. These models and a previously published regional model (136 lakes, one central-lake sample from each) are used to infer water depths from 78 fossil samples spanning the last ~30,000 ¹⁴C years B.P. at Zagoskin Lake. Although the site-specific [r ² _boot = 0.90, root mean square error of prediction (RMSEP) = 1.76] and local (r _boot² = 0.68, RMSEP = 4.36) inference models have better performance statistics than the regional model, few clear trends among all three models exist in the lake-level reconstruction. We propose that multiple, within-lake sampling of gradients can be used to improve the performance statistics of water-depth transfer functions and ultimately reconstruct paleohydrology in regions known to exhibit large fluctuations in moisture balance through time given that: (1) adequate analogs are established and (2) taphonomic processes important to benthic invertebrate remains are more fully understood.     

Paleohydrology inferred from diatoms in northern latitude regions

Several recent studies have successfully applied diatom-based paleolimnological techniques to infer past hydrological changes in arctic and subarctic regions. For example, we summarize arctic studies that attempt to determine changes in peat water content, flood frequency, river discharge, effective moisture and ice cover in northern regions. Some of the investigations are still in preliminary stages, but represent innovative approaches to study arctic and subarctic paleohydrology. New data demonstrate that lake depth, which may be related to changing hydrological conditions, is a significant variable influencing the distributions of diatom taxa in lake surface sediment calibration sets from Wood Buffalo National Park (WBNP), on the border of Alberta and the Northwest Territories, Canada, and from Fennoscandia (mainly northwest Finland). Weighted averaging regression and calibration methods were used to develop quantitative inference models for lake depth using diatom assemblages preserved in surface sediments. The predictive abilities of the transfer functions were relatively high (for WBNP r² = 0.70 and RMSE = 2.6 m, and for Fennoscandia r² = 0.88 and RMSE = 1.8 m). However, evaluating the transfer functions using jack-knifing procedures indicated lower predictive abilities, possibly reflecting the relatively small sample size and/or short gradients used in these calibration sets. Such transfer functions can be used to track overall trends in lake levels, and provide an objective assessment as to directions of changing lake levels. Any interpretations of inferred lake levels, especially those related to climate change, must be made cautiously and must include some understanding of the local, present-day hydrological system.     

基于遗传算法自动获取CA模型的参数

本文提出了基于遗传算法来寻找CA模型最佳参数的方法。CA被越来越多地应用于城市和土地利用等复杂系统的动态模拟。CA模型中变量的参数值对模拟结果有非常重要的影响。如何获取理想的参数值是模型的关键。传统的逻辑回归模型运算简单,常常用来获取模型的参数值,要求解释变量间线性无关,所以获取的城市CA模型参数具有一定的局限性。遗传算法在参数优化组合、快速搜索参数值方面有很大的优势。本文利用遗传算法来自动获取优化的CA模型参数值,并获得了纠正后的CA模型。将该模型应用于东莞1988~2004年的城市发展的模拟中,得到了较好的效果。研究结果表明,遗传算法可以有效地自动获取CA模型的参数,其模拟的结果要比传统的逻辑回归校正的CA模型模拟精度高。     

How much past to see the future: a computational study in calibrating urban cellular automata

The objective of this computational study was to investigate to which extent the availability and the way of use of historical maps may affect the quality of the calibration process of cellular automata (CA) urban models. The numerical experiments are based on a constrained CA applied to a case study. Since the model depends on a large number of parameters, we optimize the CA using cooperative coevolutionary particle swarms, which is an approach known for its ability to operate effectively in search spaces with a high number of dimensions. To cope with the relevant computational cost related to the high number of CA simulations required by our study, we use a parallelized CA model that takes advantage of the computing power of graphics processing units. The study has shown that the accuracy of simulations can be significantly influenced by both the number and position in time of the historical maps involved in the calibration.     

Multi-label class assignment in land-use modelling

During the last two decades, a variety of models have been applied to understand and predict changes in land use. These models assign a single-attribute label to each spatial unit at any particular time of the simulation. This is not realistic because mixed use of land is quite common. A more detailed classification allowing the modelling of mixed land use would be desirable for better understanding and interpreting the evolution of the use of land. A possible solution is the multi-label (ML) concept where each spatial unit can belong to multiple classes simultaneously. For example, a cluster of summer houses at a lake in a forested area should be classified as water, forest and residential (built-up). The ML concept was introduced recently, and it belongs to the machine learning field. In this article, the ML concept is introduced and applied in land-use modelling. As a novelty, we present a land-use change model that allows ML class assignment using the k nearest neighbour (kNN) method that derives a functional relationship between land use and a set of explanatory variables. A case study with a rich data-set from Luxembourg using biophysical data from aerial photography is described. The model achieves promising results based on the well-known ML evaluation criteria. The application described in this article highlights the value of the multi-label k nearest neighbour method (MLkNN) for land-use modelling.     

Strength of Stacking Technique of Ensemble Learning in Rockburst Prediction with Imbalanced Data: Comparison of Eight Single and Ensemble Models

Rockburst is a common dynamic geological hazard, severely restricting the development and utilization of underground space and resources. As the depth of excavation and mining increases, rockburst tends to occur frequently. Hence, it is necessary to carry out a study on rockburst prediction. Due to the nonlinear relationship between rockburst and its influencing factors, artificial intelligence was introduced. However, the collected data were typically imbalanced. Single algorithms trained by such data have low recognition for minority classes. In order to handle the problem, this paper employed stacking technique of ensemble learning to establish rockburst prediction models. In total, 246 sets of data were collected. In the preprocessing stage, three data mining techniques including principal component analysis, local outlier factor and expectation maximization algorithm were used for dimension reduction, outlier detection and outlier substitution, respectively. Then, the pre-processed data were split into a training set (75%) and a test set (25%) with stratified sampling. Based on the four classical single intelligent algorithms, namely k-nearest neighbors (KNN), support vector machine (SVM), deep neural network (DNN) and recurrent neural network (RNN), four ensemble models (KNN–RNN, SVM–RNN, DNN–RNN and KNN–SVM–DNN–RNN) were built by stacking technique of ensemble learning. The prediction performance of eight models was evaluated, and the differences between single models and ensemble models were analyzed. Additionally, a sensitivity analysis was conducted, revealing the importance of input variables on the models. Finally, the impact of class imbalance on the prediction accuracy and fitting effect of models was quantitatively discussed. The results showed that stacking technique of ensemble learning provides a new and promising way for rockburst prediction, which exhibits unique advantages especially when using imbalanced data.