The unprecedented growth of Renewable Energy Sources (RES) positions solar power as a leading contender in the global energy mix. Solar energy offers a sustainable alternative to fossil fuels, mitigating carbon emissions and promoting environmental sustainability. This study explores the crucial role of forecasting algorithms within photovoltaic (PV) systems. We aim to provide a comprehensive understanding of methodologies, datasets, and recent advancements for enhancing predictive accuracy in solar power generation forecasting. While machine learning has dominated previous research, recent studies highlight challenges in achieving optimal efficiency and accuracy. A significant obstacle lies in the deficiency of real-world application for large-scale specifically for solar power generation forecasting. To address this gap, this study defines prevalent forecasting methodologies and illuminates datasets with diverse characteristics and their relevance. This study meticulously provides and explore recent advanced methods and datasets, emphasizing their impact on forecasting performance. This study not only deepens our understanding of existing methodologies but also provides valuable insights for future advancements in solar power generation forecasting.

Solar energy, a cornerstone of renewable energy, for optimal grid integration and management, requires precise forecasting. Photovoltaic (PV) forecasting must be accurate to ensure energy stability and maximize resource utilization. This study compares Multi-Layer Perceptron (MLP) and Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) models for forecasting solar power generation. Both models were trained with 13 features using an open-source dataset from 10 PV sites in Hebei Province, China, spanning 300 days (2018-07-01 to 2019-06-13). The CNN-LSTM was configured with 50 epochs and particular hyperparameters. CNN-LSTM demonstrated superior performance, with Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE) values of 0.088, 0.051, and 0.227 versus MLP's 0.260, 0.156, and 0.395. The findings demonstrate CNN-LSTM's potential for enhancing solar power forecasting and facilitating the management of renewable energy sources.

Accurate short-term forecasting of photovoltaic (PV) power generation is essential for maintaining the stability and efficiency of modern power grids. However, conventional statistical and regression-based methods often fail to capture the nonlinear and stochastic patterns of solar generation data, leading to unreliable forecasts and operational inefficiencies. This study addresses these challenges by systematically benchmarking eight deep learning architectures for PV power forecasting, including Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), Transformer, Temporal Fusion Transformer (TFT), TiDE, N-HiTS, DLinear, and Neural Basis Expansion Analysis for Time Series (N-BEATS). Real-world 15-minute interval PV data from 2018 to 2019 were used to evaluate the models in terms of forecasting accuracy, parameter efficiency, training time, and memory usage. Results demonstrated that compact variants of the N-BEATS architecture, specifically the optimized G-NBEATS and I-NBEATS models, delivered the highest predictive performance, achieving R² scores of 98.7% and 98.5%, respectively. These models substantially reduced computational demands, lowering parameter counts by up to 98% relative to N-HiTS (3.1 M parameters), training times by over 90% compared to Transformer and TFT model (> 2300 s), and peak GPU memory usage by more than 95% relative to Transformer (997.67MB VRAM). Qualitative assessments further confirmed their capability to accurately capture both diurnal ramp patterns and seasonal transitions. This study introduces a lightweight, deployment-ready forecasting framework that combines high accuracy with low computational cost, offering practical solutions for real-time solar power prediction in resource-constrained environments.

Laily Nur Qomariyati, Nurul Jannah, Suryo Adhi Wibowo, and Thomhert Suprapto Siadari. International Journal of Fuzzy Logic and Intelligent Systems 2024;24:194-202. https://doi.org/10.5391/IJFIS.2024.24.3.194

Spinal X-rays are still often read through manual measurements, yet the patients who most need timely assessment cannot afford delay, inconsistency, or heavy computational pipelines. Motivated by this clinical tension, this study proposes an efficiency-aware deep learning framework for three-class spinal disorder classification that asks a practical question rarely centered in prior work: not only which model is most accurate, but which model is accurate enough, light enough, and fast enough to matter in real screening settings. Using a public dataset of 338 subjects, five lightweight backbones, CBAM-enhanced variants, and a dual-branch fusion model were evaluated through stratified 5-fold cross-validation under multiple balancing strategies, with performance measured by accuracy, precision, recall, F1-score, parameter count, FLOPs, model size, latency, and throughput. The results reveal an unexpected pattern: bigger models do not win. MobileNetV3Small delivers the strongest efficiency-performance balance, reaching an F1-score of 0.962 with only 1.0 million parameters, while the best overall result is achieved by the Fusion_MNv3_MNAS model under augmentation-only training, with an F1-score of 0.976. Ablation findings further show that attention and fusion are not universally beneficial, but become most effective when paired with sufficient data-driven regularization, and that fine-tuning about 30% of backbone parameters yields the most favorable adaptation. Taken together, these findings show that performance in spinal X-ray classification depends less on model size alone than on the fit between architecture and training strategy. The study therefore offers a concrete and clinically relevant message: lightweight, well-regularized models can match or surpass heavier alternatives while remaining more practical for scalable deployment. ABSTRAK: Radiograf tulang belakang masih kerap dinilai melalui pengukuran manual, sedangkan pesakit yang memerlukan rawatan awal tidak dapat menanggung kelewatan, ketidakselarasan, atau kebergantungan pada sistem pengiraan yang berat. Berpunca pada masalah klinikal ini, kajian ini mengemukakan satu rangka kerja pembelajaran mendalam berpaksikan kecekapan bagi pengelasan tiga kelas gangguan tulang belakang, dengan menumpukan persoalan praktikal yang jarang diberi perhatian dalam kajian terdahulu, iaitu bukan pada model mana yang paling tepat, tetapi model mana yang cukup tepat, cukup ringan, dan cukup pantas bagi persekitaran saringan klinikal. Dengan menggunakan satu set data awam yang melibatkan 338 subjek, lima model asas ringan, varian yang dipertingkatkan dengan CBAM, serta model gabungan dwi-cabang telah dinilai melalui pengesahan silang berstrata lima lipatan di bawah beberapa strategi pengimbangan kelas, dengan prestasi diukur menggunakan ketepatan, kejituan, keboleh ingatan, skor F1, bilangan parameter, FLOPs, saiz model, kependaman, dan kadar pemprosesan. Dapatan kajian menunjukkan satu corak yang tidak dijangka, iaitu model yang lebih besar tidak semestinya memberi prestasi terbaik. MobileNetV3Small memperlihatkan keseimbangan paling kukuh antara kecekapan dan prestasi dengan mencapai skor F1 sebanyak 0.962 hanya menggunakan 1.0 juta parameter, manakala prestasi keseluruhan terbaik dicapai oleh model Fusion_MNv3_MNAS di bawah latihan berasaskan augmentasi sahaja dengan skor F1 sebanyak 0.976. Analisis ablasi seterusnya menunjukkan bahawa mekanisme perhatian dan gabungan tidak sentiasa memberikan manfaat secara menyeluruh, sebaliknya menjadi paling berkesan apabila dipadankan dengan regularisasi berasaskan data yang mencukupi, dan penalaan halus sekitar 30% parameter rangka asas menghasilkan penyesuaian yang terbaik. Secara keseluruhan, dapatan ini menunjukkan bahawa prestasi dalam pengelasan sinar-X tulang belakang kurang bergantung pada saiz model semata-mata, sebaliknya lebih dipengaruhi oleh kesesuaian antara seni bina model dan strategi latihan. Oleh itu, kajian ini membawa mesej yang jelas dan signifikan dari sudut klinikal, iaitu model ringan yang diregularisasikan dengan baik mampu menandingi malah mengatasi model yang lebih berat, iaitu kekal lebih praktikal bagi pelaksanaan berskala dalam persekitaran saringan klinikal.

Anyone who has watched a distant scene shimmer above hot pavement has seen atmospheric turbulence destroy image detail. In long-range imaging, turbulence produces spatially varying blur, geometric warping, scintillation, and temporal instability. Recovering the underlying scene is therefore an ill-posed inverse problem, and learned priors must compensate for distortions that simplified optical models capture only partially. We use turbulence mitigation as the umbrella term for all countermeasures and turbulence restoration for its computational core, the estimation of a clean image from degraded observations. From a cognitive-computing perspective, mitigation is not merely image enhancement. It is an uncertainty-constrained visual inference problem in which an intelligent system must reconstruct, interpret, and act on observations relayed through a stochastic physical channel. This critical review examines how the deep learning era has reshaped turbulence mitigation, with physics as the foundation for understanding degradation and designing inductive biases. It traces the architectural progression from convolutional and adversarial networks to Transformers, denoising diffusion models, Mamba and other state-space architectures, and physics-informed frameworks. For each family, we ask a common question: How does it treat the aleatoric uncertainty intrinsic to a random optical channel and the epistemic uncertainty introduced by scarce and simulator-dominated training data? The review also analyzes datasets, simulation strategies, loss functions, and evaluation metrics, and it separates the small body of shared-protocol benchmark evidence from the far larger body of self-reported results that cannot be compared across studies. Persistent obstacles include the synthetic-to-real domain gap, the scarcity of paired real turbulence data, the mismatch between fidelity metrics and downstream task performance, and the computational cost that limits operational deployment. We close with an AI-centered agenda in which uncertainty quantification stands alongside domain adaptation as a first-order priority.