ABSTRACT This study evaluates annual country‐level renewable‐energy production forecasting using statistical, machine‐learning, deep‐learning, and explainability methods under chronological validation. The analysis uses 2500 observations from the CC0‐licensed Global Renewable Energy and Indicators Dataset, covering 10 countries, five renewable‐energy categories, and 2000–2023. The primary target is renewable‐energy production in GWh. Observations are separated into training (2000–2015), validation (2016–2018), and test (2019–2023) periods before synthetic processing. Gaussian augmentation is restricted to training partitions; validation and test observations remain unchanged. The principal models are XGBoost, long short‐term memory (LSTM), ARIMA, and VAR. Expanding‐window validation is used during development, with Optuna tuning repeated within each XGBoost fold. SHAP and LIME explain the fitted XGBoost model on held‐out observations. LSTM obtained test R2=−0.0289, RMSE = 9792.55 GWh, MAE = 7625.36 GWh, and MAPE = 17.03%; XGBoost obtained R2=−0.0835, RMSE = 10,049.20 GWh, MAE = 7798.12 GWh, and MAPE = 17.56%. The available annual predictors therefore provide limited future‐period forecasting skill. The study documents the complete evaluation protocol, statistical baselines, sequence model, and feature‐attribution results.
ABSTRACT This paper briefly introduces the enterprise environmental, social, and governance (ESG) rating and the random forest (RF) model for ESG rating prediction. The China Securities Index (CSI) ESG rating dataset is selected for simulation experiments. During the experiments, the effects of the number of decision trees, maximum tree depth, maximum number of features, and minimum number of samples per node on the RF model are first examined. The RF model is then compared with the linear logistic regression (LLR), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost) algorithms. Finally, the importance of the feature indicators for predicting the ESG rating is analyzed. Furthermore, ablation experiments and cross‐data generalization control experiments are conducted. The RF model exhibits the best predictive performance, with a precision of 0.8414, a recall of 0.8463, an F‐score of 0.8442, and an area under the curve (AUC) of 0.8971, when the number of decision trees is 300, the maximum tree depth is 10, the maximum number of features is 6, and the minimum number of samples per node is 3. The RF model has better predictive performance than the LLR, SVM, and XGBoost algorithms. The proportion of independent directors is the most important (218.74), followed by the number of green patents (196.32), information disclosure violations (182.59), and return on net assets (167.81). The results of the ablation experiment show that the RF model incorporating financial features performs significantly superior to the RF model that only uses non‐financial features; the cross‐data generalization control experiment verifies the generalization performance of the RF model.
ABSTRACT Censoring strategies in reliability testing are often adopted to control testing time, sample size, and cost. To improve inferential efficiency under such settings, this study considers an adaptive Type‐II hybrid progressive censoring scheme. The recently proposed one‐parameter Weighted Exponential–Lindley distribution is considered for modeling censored lifetime data. Both classical and Bayesian estimation approaches are developed for the model parameters, as well as for reliability and hazard rate functions. The classical inference is carried out using the maximum likelihood estimation, and approximate confidence intervals are constructed for the unknown parameter and reliability characteristics. Bayesian estimation is performed assuming a gamma prior under symmetric and asymmetric loss functions. Since the joint posterior distribution is analytically intractable, the Markov chain Monte Carlo technique is employed to obtain Bayes estimates and construct the highest posterior density credible intervals. The performance of the proposed estimators is examined through extensive simulation studies under different progressive censoring schemes. A real data application demonstrates the practical relevance and effectiveness of the proposed model.
ABSTRACT This study conducts a preconstruction assessment of the Dongyang Tunnel through three linked steps: multi‐source hydrogeological investigation, comparison of three water‐inflow prediction methods, and section‐based water–mud inrush risk assessment and zoning. The investigation integrated geological mapping, boreholes, an EM3D high‐density electrical‐resistivity survey, Lugeon tests, hydrochemistry, and regional hydrogeological records. All nine investigated boreholes intersected karst cavities; the maximum cavity height was 8.0 m, and CSZK2‐1 intersected five vertically distributed bead‐like cavity layers. Hydraulic conductivity was 0.01304–0.01357 m/day. The precipitation‐infiltration, groundwater‐runoff‐modulus, and groundwater‐dynamics‐based empirical methods yielded normal inflows of 621.00, 386.12, and 532.56 m3/day and maximum inflows of 1242.00, 772.24, and 1557.61 m3/day, respectively. The initial maximum of 1557.61 m3/day was adopted as a design control reference, not an observed construction inflow. The resulting qualitative zoning identifies higher‐risk and high‐risk sections for targeted prediction, drilling, grouting, and controlled drainage. The predictions and zoning remain design‐stage results without systematic construction monitoring or back‐analysis.
ABSTRACT This computational study introduces a novel hybrid framework that combines advanced numerical simulations with Artificial Neural Networks (ANN) for prediction and classification. This approach addresses a critical research gap by optimizing the application of hybrid nanofluids to enhance cooling performance in horizontal wavy channels. The proposed problem is formulated as a set of partial differential equations for momentum and energy, with specific boundary conditions. The Galerkin finite element method is applied to obtain accurate results by varying critical parameters, including the Richardson number ( Ri = 0.1–10), Hartmann number ( Ha = 0–60), and volume fraction of nanoparticles ( ϕ = 0.01–0.06), at Re = 100 and Pr = 6.837. The effect of key parameters on the average Nusselt number datasets is examined and compared with the predictive and classification performance of ANN. The numerical model is validated against published benchmark solutions, showing excellent agreement. The results reveal that the maximum heat transfer performance is achieved at Ri = 10, Ha = 0, and ϕ = 0.06 under the default parameter settings for different boundary conditions. The proposed ANN accurately predicts the average Nusselt number, achieving R 2 values of 0.9938 and 0.9922, MAPE values of 1.62% and 1.51%, and classification accuracies of 97.8% and 97.2% for the two heating configurations, respectively. These findings demonstrate that the proposed FEM–ML framework enables accurate, real‐time prediction of heat‐transfer performance without requiring repeated numerical simulations, thereby providing an efficient computational tool for the design, optimization, and control of MHD‐assisted hybrid nanofluid thermal systems used in HVAC systems, electronic cooling devices, solar thermal systems, and other energy‐efficient cooling applications.
