Input Dataset Research Articles

Wind Power Forecasts and Network Learning Process Optimization through Input Data Set Selection

Energy policies of the European Union, the United States, China, and many other countries are focused on the growth in the number of and output from renewable energy sources (RES). That is because RES has become increasingly more competitive when compared to conventional sources, such as coal, nuclear energy, oil, or gas. In addition, there is still a lot of untapped wind energy potential in Europe and worldwide. That is bound to result in continuous growth in the share of sources that demonstrate significant production variability in the overall energy mix, as they depend on the weather. To ensure efficient energy management, both its production and grid flow, it is necessary to employ forecasting models for renewable energy source-based power plants. That will allow us to estimate the production volume well in advance and take the necessary remedial actions. The article discusses in detail the development of forecasting models for RES, dedicated, among others, to wind power plants. Describes also the forecasting accuracy improvement process through the selection of the network structure and input data set, as well as presents the impact of weather factors and how much they affect the energy generated by the wind power plant. As a result of the research, the best structures of neural networks and data for individual objects were selected. Their diversity is due to the differences between the power plants in terms of location, installed capacity, energy conversion technology, land orography, the distance between turbines, and the available data set. The method proposed in the article, using data from several points and from different meteorological forecast providers, allowed us to reduce the forecast error of the NMAPE generation to 3.3%.

A Study to Predict Ignition Delay of an Engine Using Diesel and Biodiesel Fuel Based on the ANN and SVM Machine Learning Methods.

Over time, machine learning methods have developed, but there have not been many studies comparing how well they predict ignition delays. In this study, a model that forecasts the ignition delay of a diesel engine utilizing diesel fuel and biodiesel fuel was developed using Artificial Neural Network (ANN) and Support Vector Machine (SVM) machine learning techniques. This work has clarified the problems in designing and training the model. The effectiveness of the ANN and SVM machine learning methods' ignition delay prediction models has been evaluated under various input variable conditions. The authors employed a data set of over 700 input data sets from diesel fuel and biodiesel in the B0 to B60 range for this purpose. To evaluate the accuracy of the models, the authors compared the average accuracy of the overall classification as well as the standard deviation. The results after training and verifying the accuracy of the models show that the SVM model has a better ability to predict the fire ignition delay than the ANN model. Specifically, with the test data set and the SVM model at compression ratio (ε) = 15, RMSE = 34.45 μs, MAPE = 1.30%, MAE = 28.33 μs, MAE = 28.33 μs, and R 2 = 0.967, respectively, and the SVM model can predict well. At compression ratio ε = 17, RMSE = 30.18 μs, MAPE = 1.30%, MAE = 23.48 μs, and R 2 = 0.908, respectively. With an ANN neural network model, the prediction error value at compression ratio ε = 15 is RMSE = 41.29 μs, MAPE = 1.35%, MAE = 29.68 μs, and R 2 = 0.952, respectively; at compression ratio ε = 17, it is RMSE = 30.28 μs, MAPE = 1.25%, MAE = 23.00 μs, and R 2 = 0.975, respectively. With this accuracy, the SVM model is fully capable of forecasting the ignition delay combustion time of diesel/biodiesel engines.

Transforming Türkiye's power system: An assessment of economic, social, and external impacts of an energy transition by 2030

Türkiye has the long-term goal of transforming its power system to one that is cleaner, more secure and more affordable. According to this paper's scenario analyses, low-cost renewables can supply 55% of Türkiye's total electricity demand. Coupled with the electrification of end-use sectors, energy efficiency can reduce total power demand by 10% compared to a business as usual scenario by 2030. The paper assesses the social, economic, and environmental impacts of this transformation by soft linking a power system model with an applied computable general equilibrium model, using an updated input and output dataset, and employing a novel analysis of job creation and fossil fuel externalities. The power system transformation significantly improves social welfare with net socioeconomic benefits estimated at 1% of GDP by 2030. Positive impacts include a reduction in human health and climate change externalities by a third, which are further enhanced by wage income growth that is driven by higher skilled and better paid jobs. A carbon tax emerges as a critical instrument to realize these benefits whilst reducing the power sector's emissions to 2030. The assessment should be expanded with more ambitious clean energy technology deployment for the entire energy system to operationalize Türkiye's Paris-aligned 2053 net-zero emission target and just transition policies.

Ant colony optimisation and fuzzy system for prediction of computational data of fluid flow in a bubble column reactor

ABSTRACT Computational data of a complex problem could be mapped and solved by artificial intelligence (AI) for postprocessing purposes and numerical time-saving. A fuzzy adaptive network as a hybrid AI system has already been used for CFD data machine learning. However, no investigation has been done for the ant colony optimisation (ACO) performance evaluation compared to the adaptive network. As a state-of-the-art method, the ACO and fuzzy system are combined and tuned to estimate the flow in a reactor. The liquid-phase velocity in the y-direction is predicted using a combination of fuzzy system and ACO in a large domain. The findings further indicate that, in addition to the CFD dataset and Navier-Stokes’s equations, employing an ant colony optimisation structure as a training technique may give an accurate approximation of two-phase flow. Furthermore, for numerical methods, a continuous link between input and output datasets may be a significant data source for the AI mechanism to train the physical process, incredibly bubbly flow. The ant colony optimisation shows a more accurate prediction than the adaptive network. As far as authors know, there are no such investigations to develop a new AI algorithm and performance analysis for fluid dynamics problems.

P2P Investment Data Analytics: A Case Study of Lending Club

In this study, a forecasting model is constructed for default risk by applying several machine learning models, including logistic regression, random forest, gaussian naive bayes, artificial neural networks, support vector machine, and XGBoost. Cross validation and voting ensemble learning are also utilized for best suitable hyperparameters as well. Dataset comes from Lending Club and then cleaned, normalized and standardized it into an ideal input dataset which are splitting into training and testing ones later. Among the methods, XGBoost achieved the best performance.

Sensor Data Reconstruction for Dynamic Responses of Structures Using External Feedback of Recurrent Neural Network.

An event of sensor faults in sensor networks deployed in structures might result in the degradation of the structural health monitoring system and lead to difficulties in structural condition assessment. Reconstruction techniques of the data for missing sensor channels were widely adopted to restore a dataset from all sensor channels. In this study, a recurrent neural network (RNN) model combined with external feedback is proposed to enhance the accuracy and effectiveness of sensor data reconstruction for measuring the dynamic responses of structures. The model utilizes spatial correlation rather than spatiotemporal correlation by explicitly feeding the previously reconstructed time series of defective sensor channels back to the input dataset. Because of the nature of spatial correlation, the proposed method generates robust and precise results regardless of the hyperparameters set in the RNN model. To verify the performance of the proposed method, simple RNN, long short-term memory, and gated recurrent unit models were trained using the acceleration datasets obtained from laboratory-scaled three- and six-story shear building frames.

Data-driven reliability assessment with scarce samples considering multidimensional dependence

This study proposes a data-driven method for assessing reliability, based on the scarce input dataset with multidimensional correlation. Since considering the distribution parameters estimated from the scarce dataset as those of the population may lead to epistemic uncertainty, the bootstrap resampling algorithm is adopted to infer the distribution parameters as interval parameters. To account for the variable dependence, vine copula theory is utilized to construct the joint probability density function (PDF) of input variables, and maximum likelihood estimation (MLE) and Akaike information criterion (AIC) analysis are employed to select optimal copulas based on the samples for the vine structure. Subsequently, the failure probability bounds of a response function are calculated based on the constructed joint PDF with interval distribution parameters by the active learning Kriging (AK) method combining the sparse grid integration (SGI) method. Finally, several examples are provided to demonstrate the feasibility and efficiency of the proposed method.

A Novel Two‐Step Method for Enforcing Water Budget Closure and an Intercomparison of Budget Closure Correction Methods Based on Satellite Hydrological Products

AbstractEnforcing terrestrial water budget closure is critical for obtaining consistent data sets of budget components to understand the changes and availability of water resources over time. However, most existing budget closure correction methods (BCCs) are significantly affected by errors in the budget‐component products. Moreover, these existing BCCs do not fully account for the preselection of high‐precision input data sets before enforcing the water budget closure, resulting in uncertainties in the budget‐corrected data sets. In this study, a two‐step method was proposed to enforce the water budget closure of satellite‐based hydrological products. First, high‐precision budget‐component data sets were selected and second, the water budget closure of the selected high‐precision data sets was then enforced by proposing an improved BCC strategy, that is, the Minimized Series Deviation method (MSD). The performance of the proposed two‐step method was verified in 24 global basins by comparing it to three existing BCCs of varying complexity, that is, Proportional Redistribution (PR), Constrained Kalman Filter (CKF), and Multiple Collocation (MCL). The results showed that compared to the existing BCCs, the proposed two‐step method significantly improved the accuracy of budget‐corrected data sets in the range of 2%–19% (statistical analysis was based root mean square error (RMSE) and mean absolute error (MAE) with estimated observation (EO) as a reference). This study also summarized the main factors influencing the performance of the existing BCCs and their further development prospects based on the results. This provides insight into the expansion of theories and methods related to closing the terrestrial water budget.

FPGA Implementation of Batch-Mode Depth-Pipelined Two Means Decision Tree

Decision tree for classification tasks are learned from the input dataset and consist of split nodes and leaf nodes. This letter presents the hardware implementation of learning of two means decision tree (TMDT). To accommodate large-size datasets and hence, to increase accuracy, the training data is divided into small batches and one batch at a time is loaded into chip memory. The hardware is divided into two pipelines to optimize timing and resource consumption. The critical path of the architecture enables the field-programmable gate array (FPGA) to operate with maximum frequency of 62 MHz. Simulation results show that the proposed FPGA runs at least <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$27\times $ </tex-math></inline-formula> and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$26\times $ </tex-math></inline-formula> faster than the C implementation and existing hardware, respectively.

O3–precursor relationship over multiple patterns of timescale: a case study in Zibo, Shandong Province, China

Abstract. In this study, we developed an approach that integrated multiple patterns of timescale for box modeling (MCMv3.3.1) to better understand the O3–precursor relationship at multiple sites and through continuous observations. A 5-month field campaign was conducted in the summer of 2019 to investigate the ozone formation chemistry at three sites in a major prefecture-level city (Zibo) in Shandong Province of northern China. It was found that the relative incremental reactivity (RIR) of major precursor groups (e.g., anthropogenic volatile organic compounds (AVOCs), NOx) was overall consistent in terms of timescales changed from wider to narrower (four patterns: 5-month, monthly, weekly, and daily) at each site, though the magnitudes of RIR varied at different sites. The time series of the photochemical regime (using RIRNOx / RIRAVOC as an indicator) in weekly or daily patterns further showed a synchronous temporal trend among the three sites, while the magnitude of RIRNOx / RIRAVOC was site-to-site dependent. The derived RIR ranking (top 10) of individual AVOC species showed consistency between three patterns (i.e., 5-month, monthly, and weekly). It was further found that the campaign-averaging photochemical regimes showed overall consistency in the sign but non-negligible variability among the four patterns of timescale, which was mainly due to the embedded uncertainty in the model input dataset when averaging individual daily patterns into different timescales. This implies that utilizing narrower timescales (i.e., daily pattern) is useful for deriving reliable and robust O3–precursor relationships. Our results highlight the importance of quantifying the impact of different timescales to constrain the photochemical regime, which can formulate more accurate policy-relevant guidance for O3 pollution control.

Modelling PM2.5 for Data-Scarce Zone of Northwestern India using Multi Linear Regression and Random Forest Approaches

ABSTRACT PM2.5 (Particulate matter with aerodynamic diameter <2.5 m) concentrations above permissible limit causes air quality deterioration and hampers human health. Due to the lack of a good spatial network of ground-based PM monitoring sites and systematic checking, the availability of continuous data of PM2.5 concentrations at macro and meso scales is restricted. Present research estimated PM2.5 concentrations at high (1 km) resolution over Faridabad, Ghaziabad, Gurugram and Gautam Buddha Nagar, a data-scarce zone of the highly urbanized area of northwestern India for the year 2019 using Random Forest (RF), Multi-Linear Regression (MLR) models and Hybrid Model combining RF and MLR. It included Aerosol Optical Depth (AOD), meteorological data and limited in-situ data of PM2.5. For validation, the correlation coefficient (R), Root-Mean-Square Error (RMSE), Mean Absolute Error (MAE) and Relative Prediction Error (RPE) have been utilized. The hybrid model estimated PM2.5 with a greater correlation (R = 0.865) and smaller RPE (22.41%) compared to standalone MLR/RF models. Despite the inadequate in-situ data, Greater Noida has been found to have a high correlation (R = 0.933) and low RPE (32.13%) in the hybrid model. The most polluted seasons of the year are winter (137.28 µgm−3) and post-monsoon (112.93 µgm−3), whereas the wet monsoon (44.56 µgm−3) season is the cleanest. The highest PM2.5 level was recorded in Noida followed by Ghaziabad, Greater Noida and Faridabad. The findings of the present research will provide an input dataset for air pollution exposure risk research in parts of northwestern India with sparse monitoring data.

A comprehensive remote sensing-based Agriculture Drought Condition Indicator (CADCI) using machine learning

Agriculture drought is a decrease in soil moisture during a growing season. In this study, a comprehensive remote sensing-based Agriculture Drought Condition Indicator (CADCI) was developed to monitor the agriculture drought in semi-arid environments and assess its effectiveness in rainfed agriculture regions in (A) Jordan and (B) Syria. First, remote sensed-based drought-condition spectral indices [i.e., Vegetation Condition Index (VCI), Temperature Condition Index (TCI), Evapotranspiration Condition Index (ETCI), Precipitation Condition Index (PCI), Soil Moisture Condition Index (SMCI), and Vegetation Health Index (VHI)] were calculated using data from Moderate Resolution Imaging Spectroradiometer satellite (MODIS) [Land Surface Temperature (LST), Normalized Difference Vegetation Index (NDVI), and evapotranspiration (ET)]; the Global Precipitation Measurement (GPMs); the Soil Moisture Active Passive (SMAPs); and Sentinel-1A. Second, Random Forest (RF) was used to estimate and determine the relative importance of these indices based on Standardized Precipitation Index (SPI) values to select the three spectral indices that have the most monthly short-term relative importance in identifying the agriculture drought for semi-arid environments, which were PCI, TCI, and VCI. Third, these indices were integrated to identify the drought severity based on specific thresholds that compare the pixel-specific value with the study area average value. For instance, a severe drought condition is identified if all three indices indicate a drought condition, a moderate drought or mild drought conditions are identified if any two or any one of the indices indicate drought conditions, respectively. Lastly, a none drought condition is identified if none of the indices indicate a drought condition. Finally, the SPI sets for 1 and 3-months (SPI-1 and SPI-3) were used to evaluate the performance of the CADCI. The results showed the CADCI has a high agreement with SPI-1 classes in the study areas, with overall accuracy and Kappa-values of 85% and 0.80, for study area A and 83% and 0.76 for study area B, respectively. Consequently, the CADCI shows its ability to monitor agricultural drought in semi-arid environments. Perhaps, it could be applicable for larger areas due to the spatial resolution of the input dataset.

A stepwise regression to identify relevant variables affecting the environmental impacts of clinker production

The construction sector contributes to 8% of carbon dioxide emissions, most of them caused by clinker production: in recent years, the efforts to reduce environmental impacts are growing. This paper investigates the environmental performances of clinker production and the factors driving emissions. For this purpose, the environmental impacts of 41 different grey clinkers produced in Italy from 2016 to 2021 certified by a third party according to the European standard EN 15804 have been assessed. Three input datasets (i.e., energy consumption, raw materials, and transport) have been investigated to identify the most significant variables that affected the environmental impact categories in the period 2016–2020: a stepwise regression allowed the identification of the optimal predictors to be included in the multiple regression models. The most significant variables are pet coke, fossil fuels, natural raw materials, and lorry transport. The identified models have been validated having regard to 2021 data. Therefore, for each impact category, multiple positive and negative dependencies which bear an effect on the environmental results have been quantified. The output provides proven quantitative results that can support cement producers to develop low-impacting clinker recipes.

An Entropy-Based Directed Random Walk for Cancer Classification Using Gene Expression Data Based on Bi-Random Walk on Two Separated Networks.

The integration of microarray technologies and machine learning methods has become popular in predicting the pathological condition of diseases and discovering risk genes. Traditional microarray analysis considers pathways as a simple gene set, treating all genes in the pathway identically while ignoring the pathway network's structure information. This study proposed an entropy-based directed random walk (e-DRW) method to infer pathway activities. Two enhancements from the conventional DRW were conducted, which are (1) to increase the coverage of human pathway information by constructing two inputting networks for pathway activity inference, and (2) to enhance the gene-weighting method in DRW by incorporating correlation coefficient values and t-test statistic scores. To test the objectives, gene expression datasets were used as input datasets while the pathway datasets were used as reference datasets to build two directed graphs. The within-dataset experiments indicated that e-DRW method demonstrated robust and superior performance in terms of classification accuracy and robustness of the predicted risk-active pathways compared to the other methods. In conclusion, the results revealed that e-DRW not only improved the prediction performance, but also effectively extracted topologically important pathways and genes that were specifically related to the corresponding cancer types.

Smaller is better? Unduly nice accuracy assessments in roof detection using remote sensing data with machine learning and k-fold cross-validation

Deriving the thematic accuracy of models is a fundamental part of image classification analyses. K-fold cross-validation (KCV), as an accuracy assessment technique, can be biased because existing built-in algorithms of software solutions do not handle the high autocorrelation of remotely sensed images, leading to overestimation of accuracies. We aimed to quantify the magnitude of the overestimation of KCV-based accuracies and propose a method to overcome this problem with the example of rooftops using a WorldView-2 (WV2) satellite image, and two orthophotos. Random split to training/testing subsets, independent testing and different types of repeated KCV sampling strategies were used to generate input datasets for classification. Results revealed that applying the random splitting of reference data to training/testing subsets and KCV methods had significantly biased the accuracies by up to 17%; overall accuracies (OAs) can incorrectly reach >99%. We found that repeated KCV can provide similar results to independent testing when spatial sampling is applied with a sufficiently large distance threshold (in our case 10 m). Coarser resolution of WV2 ensured more reliable results (up to a 5–9% increase in OA) than orthophotos. Object-based pixel purity of buildings showed that when using a majority filter for at least of 50% of objects the final accuracy approached 100% with each sampling method. The final conclusion is that KCV-based modelling ensures better accuracy than single models (with better pixel purity on the object level), but the accuracy metrics without spatially filtered sampling are not reliable.

Machine learning-based ozone and PM2.5 forecasting: Application to multiple AQS sites in the Pacific Northwest.

Air quality in the Pacific Northwest (PNW) of the U.S has generally been good in recent years, but unhealthy events were observed due to wildfires in summer or wood burning in winter. The current air quality forecasting system, which uses chemical transport models (CTMs), has had difficulty forecasting these unhealthy air quality events in the PNW. We developed a machine learning (ML) based forecasting system, which consists of two components, ML1 (random forecast classifiers and multiple linear regression models) and ML2 (two-phase random forest regression model). Our previous study showed that the ML system provides reliable forecasts of O3 at a single monitoring site in Kennewick, WA. In this paper, we expand the ML forecasting system to predict both O3 in the wildfire season and PM2.5 in wildfire and cold seasons at all available monitoring sites in the PNW during 2017-2020, and evaluate our ML forecasts against the existing operational CTM-based forecasts. For O3, both ML1 and ML2 are used to achieve the best forecasts, which was the case in our previous study: ML2 performs better overall (R2 = 0.79), especially for low-O3 events, while ML1 correctly captures more high-O3 events. Compared to the CTM-based forecast, our O3 ML forecasts reduce the normalized mean bias (NMB) from 7.6 to 2.6% and normalized mean error (NME) from 18 to 12% when evaluating against the observation. For PM2.5, ML2 performs the best and thus is used for the final forecasts. Compared to the CTM-based PM2.5, ML2 clearly improves PM2.5 forecasts for both wildfire season (May to September) and cold season (November to February): ML2 reduces NMB (-27 to 7.9% for wildfire season; 3.4 to 2.2% for cold season) and NME (59 to 41% for wildfires season; 67 to 28% for cold season) significantly and captures more high-PM2.5 events correctly. Our ML air quality forecast system requires fewer computing resources and fewer input datasets, yet it provides more reliable forecasts than (if not, comparable to) the CTM-based forecast. It demonstrates that our ML system is a low-cost, reliable air quality forecasting system that can support regional/local air quality management.

Neural Networks for Hyperspectral Imaging of Historical Paintings: A Practical Review.

Hyperspectral imaging (HSI) has become widely used in cultural heritage (CH). This very efficient method for artwork analysis is connected with the generation of large amounts of spectral data. The effective processing of such heavy spectral datasets remains an active research area. Along with the firmly established statistical and multivariate analysis methods, neural networks (NNs) represent a promising alternative in the field of CH. Over the last five years, the application of NNs for pigment identification and classification based on HSI datasets has drastically expanded due to the flexibility of the types of data they can process, and their superior ability to extract structures contained in the raw spectral data. This review provides an exhaustive analysis of the literature related to NNs applied for HSI data in the CH field. We outline the existing data processing workflows and propose a comprehensive comparison of the applications and limitations of the various input dataset preparation methods and NN architectures. By leveraging NN strategies in CH, the paper contributes to a wider and more systematic application of this novel data analysis method.

Authentication of Polish Red Wines Produced from Zweigelt and Rondo Grape Varieties Based on Volatile Compounds Analysis in Combination with Machine Learning Algorithms: Hotrienol as a Marker of the Zweigelt Variety.

The aim of this study was to determine volatile compounds in red wines of Zweigelt and Rondo varieties using HS-SPME/GC-MS and to find a marker and/or a classification model for the assessment of varietal authenticity. The wines were produced by using five commercial yeast strains and two types of malolactic fermentation. Sixty-seven volatile compounds were tentatively identified in the test wines; they represented several classes: 9 acids, 24 alcohols, 2 aldehydes, 19 esters, 2 furan compounds, 2 ketones, 1 sulfur compound and 8 terpenes. 3,7-dimethyl-1,5,7-octatrien-3-ol (hotrienol) was found to be a variety marker for Zweigelt wines, since it was detected in all the Zweigelt wines, but was not present in the Rondo wines at all. The relative concentrations of volatiles were used as an input data set, divided into two subsets (training and testing), to the support vector machine (SVM) and k-nearest neighbor (kNN) algorithms. Both machine learning methods yielded models with the highest possible classification accuracy (100%) when the relative concentrations of all the test compounds or alcohols alone were used as input data. An evaluation of the importance value of subsets consisting of six volatile compounds with the highest potential to distinguish between the Zweigelt and Rondo varieties revealed that SVM and kNN yielded the best classification models (F-score of 1, accuracy of 100%) when 3-ethyl-4-methylpentan-1-ol or 3,7-dimethyl-1,5,7-octatrien-3-ol (hotrienol) or subsets containing one or both of them were used. Moreover, the best SVM model (F-score of 1) was built with a subset containing 2-phenylethyl acetate and 3-(methylsulfanyl)propan-1-ol.

Machine learning light hypernuclei

We employ a feed-forward artificial neural network to extrapolate at large model spaces the results of ab-initio hypernuclear No-Core Shell Model calculations for the Λ separation energy BΛ of the lightest hypernuclei, Λ3H, Λ4H and Λ4He, obtained in computationally accessible harmonic oscillator basis spaces using chiral nucleon-nucleon, nucleon-nucleon-nucleon and hyperon-nucleon interactions. The overfitting problem is avoided by enlarging the size of the input dataset and by introducing a Gaussian noise during the training process of the neural network. We find that a network with a single hidden layer of eight neurons is sufficient to extrapolate correctly the value of the Λ separation energy to model spaces of size Nmax=100. The results obtained are in agreement with the experimental data in the case of Λ3H and the 0+ state of Λ4He, although they are off of the experiment by about 0.3 MeV for both the 0+ and 1+states of Λ4H and the 1+ state of Λ4He. We find that our results are in excellent agreement with those obtained using other extrapolation schemes of the No-Core Shell Model calculations, showing this that an ANN is a reliable method to extrapolate the results of hypernuclear No-Core Shell Model calculations to large model spaces.

Reconstruction of the meso-scale concrete model using a deep convolutional generative adversarial network (DCGAN)

The deep learning has been widely used in engineering mechanics, composite materials and other fields due to its powerful information analysis ability. A Deep Convolutional Generative Adversarial Network (DCGAN) is trained in this paper and used to generate a large number of meso-scale concrete models. The CT scanning technology is employed to obtain the 2D slice images of concrete, which are transformed into binary images through a series of preprocessing operations and used as the training dataset of neural network to train the DCGAN. The pre-processing of input dataset, such as the CT image preprocessing and the data enhancement, and some training strategies of the DCGAN are provided in detail. The meso-scale concrete model generated by the trained DCGAN is similar with the training samples and cannot be identified with the naked eye. The reliability of the generated concrete samples is further verified by two indexes (i.e. the aggregate area fraction and the aggregate fractal dimension), and the results show that the generated samples have high consistency with the training samples on the boxplot and box-counting D (the errors are within around 10% and 2%, respectively). In addition, the similar stress–strain curves and the damage evolution processes of the training and generated samples are obtained by the uniaxial compression tests. Compared with the traditional modeling method of meso-scale concrete, the proposed method is more convenient on generating the mass of meso-scale concrete samples and resolves the conflict between computational efficiency and sample quantity. It can be used to study the relationship between the macroscopic and mesoscopic behaviors of concrete from the statistical perspective.