Grid-connected photovoltaic systems with battery energy storage require tight direct-current-link voltage regulation to preserve converter stability and grid compliance, while practical operation also favors high efficiency and strong self-consumption. This paper compares super-twisting sliding mode control, model-free adaptive control, extreme learning machine control, and online recursive least squares with feedforward for a 100 kW photovoltaic plant coupled to a 200 kWh lithium-ion battery energy storage system. Year-long simulations over 8,784 h are driven by irradiance and ambient-temperature data from the National Aeronautics and Space Administration Prediction of Worldwide Energy Resources dataset for Hanoi, Vietnam (21.03°N, 105.85°E). The controllers are evaluated using direct-current-link voltage tracking accuracy, frequency quality, total harmonic distortion, power factor, system efficiency, self-consumption ratio, and a weighted multi-criteria score. Quantitatively, model-free adaptive control achieves the lowest voltage root-mean-square error of 4.18 V, online recursive least squares with feedforward obtains the highest efficiency of 98.9 %, and extreme learning machine control gives the highest self-consumption ratio of 83.6 % at the expense of larger voltage deviations. All controllers maintain the direct-current link within 630–770 V, with total harmonic distortion not exceeding 1.52 % and power factor above 0.991. The TOPSIS-based composite ranking is MFAC (0.829) > ORLS-FF (0.722) > ELM-Control (0.470) > ST-SMC (0.445), providing quantitative guidance for selecting control laws according to voltage-quality, efficiency, and energy-autonomy priorities. Under out-of-distribution grid disturbances (voltage sag −15 %, frequency step + 0.3 Hz), MFAC achieves 30 × lower DC-link RMSE than ST-SMC (0.47 V vs. 14.19 V), confirming superior out-of-distribution robustness without any retraining.
Accurate classification of respiratory sounds is critical for early assessment and continuous monitoring of lung diseases. Abnormal auditory patterns such as wheezes and crackles serve as key indicators of underlying respiratory disorders. However, challenges including signal variability, environmental noise, and class imbalance limit reliable automated analysis. This study proposes a deep learning–based, non-invasive framework for multi-class respiratory disease classification using lung sound recordings. Two publicly available datasets, SPRSound 2024 and HLS-CMDS, consisting exclusively of lung sound recordings, are employed. A comprehensive preprocessing pipeline involving resampling, silence removal, noise reduction, normalization, segmentation, and data augmentation is applied to improve signal quality and robustness. Two complementary acoustic representations—Improved Mel Spectrograms and Constant-Q Transform (CQT)—are extracted to capture temporal and harmonic characteristics. Feature fusion is achieved using the proposed Global–Local Acoustic Fusion Attention Module (GLAFAM), integrating local feature selection and global channel attention. The fused features are classified using RespiratorySoundNet, an attention-based residual deep learning architecture with hybrid channel–spatial attention. The proposed framework achieves 98.74 % accuracy on SPRSound 2024 and 98.54 % accuracy on HLS-CMDS, demonstrating promising performance on these benchmark datasets. Overall, RespiratorySoundNet with GLAFAM shows potential as an automated approach for multi-class respiratory sound classification. Further external validation and real-time deployment studies are required to establish its generalizability and clinical applicability.
Turkish abstractive text summarization remains underexplored despite the language’s agglutinative morphology, which poses distinct challenges for subword tokenization. WordPiece, Byte-Pair Encoding (BPE), and SentencePiece tokenizers decompose Turkish words in fundamentally different ways, each capturing distinct morphological patterns. We adapt and extend a multi-encoder transformer originally proposed for Korean to Turkish, investigating whether combining pre-trained language models with diverse tokenizers improves summarization quality. Starting from a pool of ten candidate pre-trained language models spanning three tokenizer families, of which eight could be successfully trained within our framework (two ModernBERT-based models were excluded due to architectural incompatibilities), we conduct a systematic study of 29 trained configurations (eight single-encoder baselines, five same-encoder ablations, twelve diverse pairs, and four triple-encoder variants) together with three Lead-n extractive baselines on the MLSum Turkish dataset. We propose two fusion strategies: Gated Parallel, which learns layer-wise encoder importance through softmax-normalized gates, and Encoder-Aware Flat, which augments concatenated encoder outputs with learnable identity embeddings. Our best configuration, BERTurk paired with multilingual BERT using parallel fusion, achieves a ROUGE-1 F1 of 42.53, outperforming the single BERTurk baseline of 41.99 (p < 0.001 on all four metrics: ROUGE-1, ROUGE-2, ROUGE-L, and BERTScore). A controlled same-encoder ablation across all four fusion strategies shows that the gains are driven primarily by the fusion mechanism itself rather than by encoder heterogeneity: both proposed mechanisms significantly outperform their counterparts even on duplicated encoders, and the best same-encoder configuration closes most of the gap to the best diverse pair. Gate-weight analysis reveals layer-wise encoder specialization: BPE encoders contribute more at lower decoder layers while WordPiece encoders dominate at higher semantic layers. Finally, we identify an anchor encoder pattern: pairs containing the decoder initialization source achieve a mean ROUGE-1 that is 1.34 points higher than pairs without it, a finding that is consistent with a representational alignment mechanism and offers practical guidance for multi-encoder design.
The increasing penetration of renewable energy sources and the growing complexity of modern power systems have made efficient reactive power management essential for ensuring system stability, reducing transmission losses, and maintaining voltage quality. However, the ORPD problem remains challenging due to its nonlinear, non-convex nature, as well as the intermittent behaviour of wind power and the limited coordination of reactive compensation devices under geo-physical constraints. To address these challenges, this paper proposes a comprehensive ORPD framework that integrates SVCs and WPS while explicitly considering geo-physical limitations. The problem is formulated under both mono-objective and multi-objective perspectives, targeting the minimization of active power losses, total voltage deviation, voltage stability index, and capital cost. A hybrid SVC placement strategy is developed based on sensitivity analysis and minimum voltage criteria, further refined using Multi Fuzzy Linear Programming. In addition, the COA is applied for the first time to the ORPD problem to enhance global search capability and convergence performance. The proposed approach is validated on the IEEE 30-bus and Algerian DZ-114 systems. For the IEEE 30-bus system, COA achieved a minimum active power loss of 2.406 MW, reduced total voltage deviation to 0.0776p.u., and improved the voltage stability index to 0.0825. Furthermore, under combined SVC and WPS integration, transmission losses remained low across different loading conditions, demonstrating strong scalability and robustness. Comparative results confirm that the hybrid integration of SVCs and WPS provides the best trade-off between technical performance and economic efficiency. Overall, the findings demonstrate that the proposed COA-based framework is a robust and effective solution for ORPD in renewable-integrated power systems.
As a critical strategy for enhancing the safety and efficiency of hazardous materials (hazmat) logistics, the adoption of multi-modal transportation remains constrained by conventional modeling approaches that fail to account for spatial heterogeneity in regional emergency response capacity. To address this gap, this study proposes an integrated risk-cost optimization framework that incorporates emergency response accessibility into a Conditional Value-at-Risk (CVaR) model under fuzzy temporal uncertainty. A depth-first search, Dirichlet initialization, and grid-sparsity-enhanced NSGA-II (DGS-NSGA-II) algorithm is developed to solve the multi-objective optimization problem. Compared with the model assuming fixed consequences, the proposed model achieves a 32.74% reduction in transportation risk in terms of CVaR, and an 8.53% reduction in transportation cost under the safety-oriented scenario. By integrating route optimization with emergency response preparedness, this research provides a decision-support approach for balancing economic efficiency, risk equity, and disaster resilience.
