Pavement temperature is a critical factor in winter road maintenance as it directly affects operational decisions related to de-icing, antiicing, and other safety measures. Accurate forecasting of pavement temperature enables road agencies to optimize maintenance strategies, reduce operational costs, and improve roadway safety outcomes. This study proposes a novel machine-learning algorithm, termed LSTMCNN, which integrates convolutional neural networks (CNNs) with long short-term memory (LSTM) networks for pavement temperature prediction. The proposed model enables the LSTM component to capture sequential dependencies, whereas the CNN component extracts local and spatial features embedded in time-series temperature records. Therefore, the proposed model can effectively identify long-range temporal relationships while uncovering localized or spatial features within the dataset. The input data—comprising pavement, atmospheric, and soil temperatures—were obtained at the entrance of a tunnel where a multivehicle pile-up due to black ice had occurred previously. The proposed LSTM-CNN model achieved an average prediction error of 0.61 ℃ and was benchmarked against other well-established machine-learning models, including Transformer and standalone LSTM architectures. The results show that the proposed algorithm delivers statistically superior predictive performance. The LSTM-CNN approach offers significant potential for enhancing the efficiency and effectiveness of winter road maintenance operations.

This study optimizes three machine learning models—Decision Tree, Random Forest (RF), and Gradient Boosting—to classify concrete structure types (C2, C3, C4, and C5) using information from a building register. Although the initial models achieved high overall accuracy, the minority class C5 exhibited relatively low performance due to class imbalance and inherent complexity. To address this, an exhaustive grid search over discrete parameter candidates was performed, and a class-weighting strategy was integrated into the RF model to prioritize accurate classification of the minority class. The optimized RF model preserved a high overall accuracy of 94% while markedly improving C5 recall from 0.81 to 0.86 and its F1-score from 0.85 to 0.87. These results demonstrate that strategic hyperparameter tuning with class weights can effectively enhance classification reliability for rare structural types. Future research should include feature importance analysis to refine data configurations and the expansion of minority class samples to further improve model robustness in practical applications.

본 연구는 다양한 천연염재로 염색조건을 달리하여 염색한 견직물로 준비된 동일색조의 2-배색 100종에 대하여 주관적 색채감성 요인구조를 규명하고, 동일색조 유형과 유/무채색 색조, 물리적 색채특성 및 배색변인의 객관적 변 인들이 색채감성요인에 미치는 영향을 분석하였으며, 인공지능 기계학습 기반의 Random Forest를 이용하여 색채감 성요인 예측모델을 제안하였다. 연구 결과로서 천연염색 견직물의 동일색조 2-배색에 대한 색채감성요인으로 ‘유쾌 함’, ‘클래식’, ‘소프트’, ‘모던’의 4개 감성이 추출되었는데, 각 요인은 단색의 물리적 색채특성, 동일배색 유형, 유채 색/무채색, 정량적 배색 변인을 포함한 객관적 색채 변인으로부터 유의한 영향을 받음이 확인되었다. Random Forest 를 이용하여 동일색조 2-배색의 색채감성요인 별로 수립한 예측 모델에서 요인 ‘유쾌함’과 ‘소프트’ 예측모델의 예측 성능이 가장 우수하였으며, 색채감성요인 예측 모델에서 변수 중요도와 대체선형모델의 구조를 통해 요인 ‘유쾌함’ 은 색채 밝기 관련 변인, 요인 ‘소프트’는 색채 진하기 관련 변인의 영향력이 가장 큰 것으로 파악되었다. 또한 실험 값과 예측값 간 높은 상관성을 확인함으로써, 인공지능 기계학습 알고리즘 Random Forest를 천연염색직물의 색채감 성예측에 활용할 수 있을 것으로 기대되었다.

AI-driven automation for structural design has been actively studied in structural engineering. In particular, reinforcement learning (RL) has attracted attention as a framework in which an agent interacts with an environment to autonomously search for optimal design solutions in complex design spaces. This study proposes an automated design model for rectangular reinforced-concrete (RC) columns based on a multi-agent Double Deep Q-Network (Double DQN). Extending prior RL-based automation developed for RC beam design to column members, the proposed environment explicitly incorporates key column-specific behaviors, including axial force–bending moment (P–M) interaction and moment magnification due to column buckling. Four agents independently determine the section width (b), section depth (h), number of longitudinal bars (n), and bar size. The reward function combines (i) penalty terms for violations of ACI 318-19 design constraints and (ii) an economic reward defined relative to an approximate optimal cost predicted by a quadratic regression model. After training for approximately 10,000 episodes, the proposed multi-agent Double DQN consistently generated ACI-compliant column designs across all test load cases and produced solutions with improved cost efficiency compared with the approximate optimal baseline. These results demonstrate the feasibility and practical potential of multi-agent RL for automated RC column section design.

The application of machine learning in concrete technology has expanded rapidly, yet its reliability is often constrained by limited experimental data, heterogeneous testing conditions, and inconsistencies across published studies. This study investigates the integration of machine learning and synthetic data augmentation to predict the compressive strength of concrete incorporating biochar as a partial replacement for cement. An experimental dataset was compiled from peer-reviewed journal articles indexed in Web of Science, focusing on biochar-modified concrete mixtures. Input variables included cement content, fine and coarse aggregates, biochar dosage, water to binder ratio, superplasticizer content, and curing age, with compressive strength as the target variable. Extreme Gradient Boosting was adopted due to its strong performance on nonlinear tabular data. Model performance was evaluated using the mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R²), alongside five-fold cross-validation. Hyperparameter optimization was performed using Optuna. To address data scarcity, a synthetic dataset of 1000 samples was generated using ChatGPT. the large language model approach relied solely on natural language prompts. Only feature definitions and the target variable were provided, without exposing the original data or implementing data generation algorithms. Three modeling strategies were examined. First, model trained and tested solely on experimental data achieved a testing R² of approximately 0.91. Second, model trained on synthetic data and evaluated exclusively on experimental data showed reduced generalization, achieving a testing R² of about 0.42, indicating pronounced domain shift effects. Third, synthetic and experimental data were combined through data augmentation and jointly modeled, a testing R² of 0.93 was achieved. The result showed that the use of LLMs for augmentation improved the performance of the model.

본 연구는 항만물류 분야의 산업 설비에 대한 머신러닝 기반 예지정비 시스템 개발을 목적으로 수행되었다. UCI Repository의 Dataset을 활용하여 10,000개의 데이터 포인트를 분석하였으며, 설비 고장 발생 여부를 예측하는 이진 분류와 고장 유형을 분류하는 다중 클래스 분류 과제를 수행하였다. 데이터 전처리 과정에서 클래스 불균형 문제 해결을 위해 SMOTE 기법을 적용하였고, StandardScaler를 이 용한 정규화를 수행하였다. 주성분 분석을 통해 온도 변수, 기계 출력, 공구 마모가 주요 예측 변수로 확인되었다. 로지스틱 회귀, K-최근 접 이웃, 서포트 벡터 머신, 랜덤 포레스트, XGBoost 등 다섯 가지 머신러닝 알고리즘을 적용하여 성능을 비교하였다. 분석 결과, KNN은 상대적으로 낮은 성능을 보였으나 빠른 응답속도를 제공하였고, XGBoost가 모두에서 최고 성능을 보였으며, 이진 분류에서 F1 점수 0.958, 다중 클래스 분류에서 0.989를 달성하였다.

The loss of soil available nutrients may affect soil quality and crop growth. Biochar can form a multi-level fixed network because of its rich pore structure and surface functional groups, which can effectively fix available nutrients in soil and maintain nutrient utilization rate. Because it is difficult to directly prepare biochar materials with good adsorption characteristics through experimental results. This study employed an XGBoost machine learning prediction model to determine the optimal nutrient-rich biochar preparation conditions. The R2 value ranged from 0.97 to 0.99. The results indicated that specific surface area was the primary factor influencing ammonium nitrogen adsorption, with a feature importance of 56.13%. Production conditions (hydrothermal temperature and time) significantly affected the adsorption of nitrate nitrogen and available phosphorus, with feature importances of 75.91% and 81.54%, respectively. Mean pore diameter was negatively correlated with potassium ion adsorption characteristics. Biochar prepared under hydrothermal conditions at 202.50–251.25 °C for 3 h exhibited favorable adsorption characteristics for multiple soil available nutrients. This study provides new insights into biochar’s application in the field of soil nutrient adsorption through data analysis. It is helpful to avoid the waste in the process of energy utilization from biomass to biochar.

Based on a carbon emission inventory of China’s cement industry, this study evaluates the performance of six machine learning models—ridge regression (RR), polynomial regression (PR), random forest (RF), support vector machine (SVR), gradient boosted regression tree (GBRT), and feed-forward neural network (FNN)—in predicting carbon emissions. Model accuracy, feature importance, and residual distributions were analyzed. Results show that clinker production and coal consumption are the dominant factors, contributing 83.7% and 11.95% to emissions, respectively. PR and FNN achieved the best performance with R2 values up to 0.99 and lowest mean square errors (0.11 and 1.82). Their mechanisms were further adapted to improve the generalization of other models. Spatial analysis revealed that North, South, and Southwest China are major emission regions. Using the optimal model, emissions in 2035 are projected to reach 519.14 million tonnes. This study offers technical insights for model optimization and supports low-carbon policymaking in the cement industry.

This study compares the shear behavior of anisotropic magnetorheological elastomers (MREs) using natural rubber (NR) and silicone rubber (Si) as matrices. The effects of magnetic flux density and compressive pre-stress on the shear modulus were experimentally investigated. Results showed that silicone-based MREs exhibited a 10–20% higher magnetorheological effect than NR-based ones due to stronger particle–matrix bonding and stable chain alignment under magnetic fields. In contrast, NR-based MREs showed greater stiffness variation under compressive stress, attributed to strain-hardening and volumetric constraint effects. These findings indicate that matrix selection significantly governs the magneto-mechanical response: silicone MREs are suitable for precision control and sensing, while NR MREs perform better in high-stress damping systems. This study provides fundamental insight for tailoring MREs according to design requirements.

With the rapid expansion of renewable energy deployment, power systems are increasingly exposed to issues such as higher output variability. Photovoltaic generation, as the most widely installed variable renewable energy source both domestically and internationally, exhibits significant fluctuations due to weather conditions. These characteristics lead to operational challenges including increased curtailment, higher reserve requirements, and even risks of large-scale outages. This study aimed to improve the accuracy of photovoltaic power generation forecasting by developing a data quality control procedure for meteorological data collected at a PV plant. The quality-controlled data were used as inputs to SVM and XGBoost, resulting in improved forecasting accuracy, with MAPE decreasing from 7–10% to 6.32% and 6.08%, respectively. The results demonstrate that meteorological data quality control significantly enhances PV forecasting performance and can contribute to distributed energy resource operation and curtailment mitigation strategies.

우리나라 농촌 인구의 고령화는 빠르게 진행되고 있으며, 2020년 기준 농업경영주의 평균 연령은 66.1세, 65세 이상 경영주 비율은 56%에 이른다. 2022년 기준 밭작물 전체의 기계화율은 63.3%인 반면, 정식 작업의 기계화율은 12.2%에 불과하며, 특히 고추 정식의 기계화율은 거의 0% 수준이다. 이러한 문제를 해결하기 위해 본 연구에서는 1회전 2식부 방식의 식부 메커니즘을 적용한 고추 정식기를 설계하였다. 식부 장치는 식부 프레임, 암(arm), 그리고 호퍼로 구성되며, 호퍼는 상사점에서 모종을 공급받아 하사점에서 식부한 후 모종과의 충돌 없이 복귀하도록 캠 메커니즘과 스윙 구조의 보조를 받는다. 호퍼는 하강 시 시계 방향의 반원 궤적을 따라 이동하고, 상승 시에는 타원 궤적으로 이동하며, 이때 상승 궤적은 사이클로이드 곡선과 높은 유사성을 보였고 주행속도가 증가할수록 그 유사성이 더욱 증가하였다. 재배 환경에 따른 주행속도는 하우스 재배에서 2.0 km/h(55.6 rpm), 노지 일반 재배에서 2.5 km/h(52.1 rpm), 노지 터널 재배에서 3.0 km/h(50.0 rpm)로 설정하였다. 식부 암의 회전속도를 60 rpm으로 고정한 조건에서, 주간거리별 최대 주행속도를 산출하였다. 주행속도와 주간거리가 증가할수록 식부 호퍼의 후퇴 면적은 감소하는 경향을 보였으며, 노지 터널 재배 조건(주간거리 500 mm, 주행속도 3.6 km/h)에서 가장 작은 후퇴 면적(10,560.0 mm²)이 나타났다. 본 연구는 1구 2식부 방식의 식부 메커니즘이 우수한 작동 성능을 가짐을 입증하였으며, 고추 정식기의 구조 및 구동 시스템 최적화를 위한 궤적 및 운전 조건 설정에 기초 자료를 제공한다.

본 연구는 무인기로 촬영한 다중분광 영상으로부터 취득한 반사값을 통해 산출된 식생지수로 콩(Glycine max (L.) Merr.)의 경태를 추정하는 머신러닝 회귀모델 개발을 목표로 한다. 연구 대상은 경상남도 밀양시 국립식량과학원 남부작물부 실험포장에서 2022년 6월 20일과 2023년 6월 24일에 파종한 선풍 품종의 콩이며, 관행구와 처리구로 나누어 재배하였다. 생육조사는 2022년 8월 20일과 9월 20일, 2023년 8월 21일과 9월 25일에 수행하였고, 영상은 2022년 8월 22일과 9월 21일, 2023년 8월 22일과 9월 20일에 촬영하였다. 촬영된 영상으로부터 5가지 반사값을 추출하여 9가지 식생지수를 산출하였다. 모델 구축에는 Ridge Regression (RR)과 LASSO Regression (LR), Random Forest Regression (RFR)과 K-Nearest Neighbor Regression (KNR)을 사용하였고, 단계적 변수 선택법을 사용하였다. 훈련과 검증의 비율은 8:2, 7:3, 6:4로 설정하였고, 모델은 R2, RMSE, MAPE로 평가하였다. 단년차 월별 모델의 경우, 8월과 9월 모두 2023년의 모델이 좋은 모델로 선정되었다. 다년차 월별 모델의 경우, 환경적 조건에 편중되어 군집화 현상이 나타나는 경우(8월)와 통계적으로 유의한 차이가 있음에도 불구하고 군집화 현상이 나타나지 않는 경우(9월)가 확인되었다. 따라서 월별 모델에 비해 성능은 낮지만 군집화가 발생하지 않고, 더 많은 샘플 수를 가진 전체 통합 모델을 최적 모델로 선택하였고, Calibration에서 R2=0.916, RMSE=0.683mm, MAPE=5.644%, Validation에서 R2=0.708, RMSE=1.002mm, MAPE=8.957% 의 성능을 나타내었다.

This study estimated whole crop maize (WCM; Zea mays L.) yield damage under abnormal climate conditions using a machine-learning approach based on Representative Concentration Pathway (RCP) 8.5 and visualized the results as spatial maps. A total of 3,232 WCM observations were compiled, and climate data were obtained from the Korea Meteorological Administration (KMA) Open Data Portal. The machine learning model used DeepCrossing. Dry matter yield (DMY) was predicted using the DeepCrossing model and climate data from the Automated Synoptic Observing System (ASOS; 95 stations). The calculation of damage was the difference between the DMYnormal and DMYabnormal. The normal climate was set as the 40-year of climate data according to the year of WCM data (1978-2017). The level of abnormal climate by temperature and precipitation was set as RCP 8.5 standard. The predicted DMYnormal ranged from 13,845-19,347 kg/ha. The damage from WCM varied by region and the severity of abnormal climate, including abnormal temperature and precipitation. Under abnormal temperature conditions, damage in 2050 and 2100 ranged from –243 to –133 and –1,797 to –245 kg/ha, respectively. Under abnormal precipitation conditions, damage in 2050 and 2100 ranged from –2,998 to 1,447 and –11,308 to 29 kg/ha, respectively. Overall, DMY of WCM tended to increase with higher mean monthly temperature. The damage calculated through the RCP 8.5 standard was presented as a spatial distribution using QGIS. Although this study used an RCP scenario based on greenhouse gas concentrations, further research is needed to apply an integrated Shared Socioeconomic Pathway (SSP) that accounts for socioeconomic factors.

Crash risk in metropolitan areas arises from the interaction among driver behavior, enforcement infrastructure, and urban environmental conditions; however, microspatial evidence remains scarce. This study examines the effects of automated speed-enforcement cameras on the crash risk in Seoul at the legal-dong level using the accident risk index, which accounts for both crash frequency and injury severity. The dataset combines crash records, enforcement infrastructure, school-zone information, demographic indicators, and land-use characteristics. Methodologically, a Bayesian negative binomial regression model was employed to address overdispersed crash data, whereas gradient-boosting machine and eXtreme Gradient Boosting models with SHAP interpretations were applied to capture nonlinear effects, heterogeneity, and variable interactions. The results reveal that the crash risk is spatially concentrated, with a small proportion of districts contributing disproportionately to the overall exposure. Regression analysis highlights enforcement and behavioral factors as significant predictors, whereas machine-learning models emphasize the added importance of structural and demographic characteristics, such as road area and floating population. This divergence reflects the sensitivity of the regression to collinearity and linearity assumptions in contrast to the flexibility of tree-based methods in uncovering nonlinear and context-dependent influences. In general, the findings reflect the value of integrating statistical and machine-learning approaches for a more comprehensive understanding of crash risk at the microspatial scale. This study advances the methodological diversity in traffic-safety research and provides practical evidence that cameradeployment strategies should be context sensitive while targeting areas with concentrated risks and distinct structural and demographic profiles.

As renewable energy penetration continues to increase, the output variability and forecasting uncertainty of photovoltaic generation have emerged as major operational risks in power systems. This study establishes a sensor-based data quality control procedure to ensure the reliability of meteorological data collected at a PV plant. For temperature, humidity, and wind speed, a four stage QC process physical range check, persistence check, step change check, and median filtering was applied. Solar radiation, which exhibits strong temporal and distributional characteristics, was processed using a three-stage QC procedure consisting of physical range, step change, and frequency distribution checks. Using the quality-controlled meteorological data, PV generation forecasting was performed with SVM and XGBoost models. As a result, the MAPE values improved to 6.32% for SVM and 6.08% for XGBoost after QC application. The findings confirm that meteorological data quality control significantly enhances PV forecasting accuracy and can support future strategies for distributed energy resource management, curtailment mitigation, and power system risk reduction.