Chakkrit Termritthikun

The advancement of intelligent multimedia computing, particularly through deep neural networks (DNNs), has gained remarkable attention across a wide array of research domains and practical applications, such as audio-visual joint processing, image classification, video computing, computer vision, etc. However, the concern for sustainability and inherent black-box characteristics of DNNs have emerged as significant challenges, especially when dealing with multiple data sources. To tackle these challenges, alternative learning models have been actively explored in recent years, with green learning and interpretable learning standing out as two key directions. In this paper, a discriminative correlation perceptron-style neural network (DC-PNN) is proposed for feature representation learning with application to multimodal information computing. By leveraging the joint strength of statistics machine learning (SML) principles and the cascade architecture of a perceptron-style neural network (PNN), the DC-PNN model is green in nature with interpretable properties. To evaluate its effectiveness and generalizability, we conduct experiments on six benchmark databases with different scales and then apply to a real world problem, Glaucoma image classification. The results demonstrate that the DC-PNN model outperforms state-of-the-art (SOTA) methods. In addition, the parameters and FLOP analysis are implemented, validating the advantage of the model in terms of efficiency and resource requirements.

Automating ICD-10 coding from discharge summaries remains demanding because coders analyze clinical narratives while justifying decisions. This study compares three automation patterns: PLM-ICD as a standalone deep learning system emitting 15 codes per case, LLM-only generation with full autonomy, and a hybrid approach where PLM-ICD drafts candidates for an agentic LLM audit to accept or reject. All strategies were evaluated on 19,801 MIMIC-IV summaries using four LLMs spanning compact (Qwen2.5-3B-Instruct, Llama-3.2-3B-Instruct, Phi-4-mini-instruct) to large-scale (Sonnet-4.5). Precision guided evaluation because coders still supply any missing diagnoses. PLM-ICD alone reached 55.8% precision while always surfacing 15 suggestions. LLM-only generation lagged severely (1.5–34.6% precision) and produced inconsistent output sizes. The agentic audit delivered the best trade-off: compact LLMs reviewed the 15 candidates, discarded weak evidence, and returned 2–8 high-confidence codes. Llama-3.2-3B-Instruct, for example, improved from 1.5% as a generator to 55.1% as a verifier while trimming false positives by 73%. These results show that positioning LLMs as quality controllers, rather than primary generators, yields reliable support for clinical coding teams, while formal recall/F1 reporting remains future work for fully autonomous implementations.

Grid-Scale Battery Energy Storage Systems (GS-BESS) play a crucial role in modern power grids, addressing challenges related to integrating renewable energy sources (RESs), load balancing, peak shaving, voltage support, load shifting, frequency regulation, emergency response, and enhancing system stability. However, harnessing their full potential and lifetime requires intelligent operational strategies that balance technological performance, economic viability, and environmental sustainability. This systematic review examines how artificial intelligence (AI)-based intelligent optimization enhances GS-BESS performance, focusing on its techno-economic, environmental impacts, and policy and regulatory implications. Following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, we review the evolution of GS-BESS, analyze its advancements, and assess state-of-the-art applications and emerging AI techniques for GS-BESS optimization. AI techniques, including machine learning (ML), predictive modeling, optimization algorithms, deep learning (DL), and reinforcement learning (RL), are examined for their ability to improve operational efficiency and control precision in GS-BESSs. Furthermore, the review discusses the benefits of advanced dispatch strategies, including economic efficiency, emissions reduction, and improved grid resilience. Despite significant progress, challenges persist in data availability, model generalization, high computational requirements, scalability, and regulatory gaps. We conclude by identifying emerging opportunities to guide the next generation of intelligent energy storage systems. This work serves as a foundational resource for researchers, engineers, and policymakers seeking to advance the deployment of AI-enhanced GS-BESS for sustainable, resilient power systems. By analyzing the latest developments in AI applications and BESS technologies, this review provides a comprehensive perspective on their synergistic potential to drive sustainability, cost-effectiveness, and energy systems reliability.

Automating ICD-10 coding from discharge summaries remains demanding because coders analyze clinical narratives while justifying decisions. This study compares three automation patterns: PLM-ICD as a standalone deep learning system emitting 15 codes per case, LLM-only generation with full autonomy, and a hybrid approach where PLM-ICD drafts candidates for an agentic LLM filter to accept or reject. All strategies were evaluated on 19,801 MIMIC-IV summaries using four LLMs spanning compact (Qwen2.5-3B, Llama-3.2-3B, Phi-4-mini) through large scale (Sonnet-4.5). Precision guided evaluation because coders still supply any missing diagnoses. PLM-ICD alone reached 55.8% precision while always surfacing 15 suggestions. LLM-only generation lagged severely (1.5--34.6% precision) and produced inconsistent output sizes. The agentic filter delivered the best trade-off: compact LLMs reviewed the 15 candidates, discarded weak evidence, and returned 2--8 high-confidence codes. Llama-3.2-3B, for example, improved from 1.5% as a generator to 55.1% as a verifier while trimming false positives by 73%. These results show that positioning LLMs as quality controllers, rather than primary generators, yields reliable support for clinical coding teams, while formal recall/F1 reporting remains future work for fully autonomous implementations.

Short-term electricity demand forecasting is a critical task for real-time decision-making. This research investigates the impact of solar irradiance components on electricity demand forecasting. The study utilized two years of hourly electricity data in addition to three key solar irradiance components: Global Horizontal Irradiance (GHI), Direct Normal Irradiance (DNI), and Diffuse Horizontal Irradiance (DHI), as clear sky indices and exogenous features. Extensive feature selection was performed to discover the best feature that enhances the forecasting accuracy of the model. Temporal Fusion Transformer (TFT) was employed as the primary model to study the influence of these exogenous variables on the electricity demand forecasting. The results demonstrated that use of solar irradiance data in the electricity demand forecasting improves the forecasting accuracy of the model. Particularly the combination of DHI and DNI clear sky indices produced the most accurate predictions. However, the addition of GHI introduced redundancy and reduced performance. The performance of the TFT model was also compared with various benchmark models. TFT outperformed the other compared models in various forecast horizons.

Electricity load forecasting is an important aspect of power system management. Improving forecasting accuracy ensures reliable electricity supply, grid operations, and cost savings. Often, collected data consist of Missing Values (MVs), anomalies, outliers, or other inconsistencies caused by power failures, metering errors, data collection errors, hardware failures, network failures, or other unexpected events. This study uses real-world data to investigate the possibility of using synthetically generated data as an alternative to filling in MVs. Three datasets were created from an original one based on different imputation methods. The imputation methods employed were linear interpolation, imputation using synthetic data, and a proposed hybrid method based on linear interpolation and synthetic data. The performance of the three datasets was compared using deep learning, machine learning, and statistical models and verified based on forecasting accuracy improvements. The findings demonstrate that the hybrid dataset outperformed the other interpolation methods based on the forecasting accuracy of the models.

The growing global demand for electrical energy, together with the significant problem of increasing carbon dioxide emissions, have become urgent issues. The inefficient use of electricity exacerbates these challenges. Smart grid technology is emerging as a solution, employing innovative approaches to solve these problems. Within this context, this research emphasizes the identification of the most effective forecasting model for demand prediction. Utilizing the Transformer-based model, Temporal Fusion Transformer (TFT), together with the Naresuan University, School of Renewable Energy and Smart Grid Technology (SGtech) net metering dataset, we explored the influence of additional features on forecasting models, categorizing them into net metering data, weather-related attributes including temperature, dew point, weather conditions, and wind direction, and supplementary features related to the operational behavior of SGtech, specifically workday and time-of-day. Our experimentation shows that integrating workday and time-of-day data alongside net metering data significantly enhances prediction precision compared to other combinations. The TFT model outperforms popular time series forecasting models, including Neural Basis Expansion Analysis for Time Series (N-BEATS) and Neural Hierarchical Interpolation for Time Series (N-HiTS), in accuracy and parameter efficiency while maintaining inference times.

Automated multi-label chest X-rays (CXR) image classification has achieved substantial progress in clinical diagnosis via utilizing sophisticated deep learning approaches. However, most deep models have high computational demands, which makes them less feasible for compact devices with low computational requirements. To overcome this problem, we propose a knowledge distillation (KD) strategy to create the compact deep learning model for the real-time multi-label CXR image classification. We study different alternatives of CNNs and Transforms as the teacher to distill the knowledge to a smaller student. Then, we employed explainable artificial intelligence (XAI) to provide the visual explanation for the model decision improved by the KD. Our results on three benchmark CXR datasets show that our KD strategy provides the improved performance on the compact student model, thus being the feasible choice for many limited hardware platforms. For instance, when using DenseNet161 as the teacher network, EEEA-Net-C2 achieved an AUC of 83.7%, 87.1%, and 88.7% on the ChestX-ray14, CheXpert, and PadChest datasets, respectively, with fewer parameters of 4.7 million and computational cost of 0.3 billion FLOPS.