Abstract

The presentation begins with a comprehensive overview of Vietnam’s power system, encompassing the structure and current status of electricity generation, high-voltage transmission networks, and the distribution and retail sector. The next section will discuss key development directions for Vietnam’s power system based on the Power Development Plan 8. Following this, the presentation addresses some pressing research issues for Vietnam’s power system. In the final part, the presentation highlights Vietnam’s latest policies to promote innovation in the energy sector, including new regulations, R&D incentives, and increased collaboration among industry, academia, and policymakers. These efforts aim to foster technological advancement and support the sustainable growth of Vietnam’s power sector.

Abstract

This presentation summarizes our group's research on applying artificial intelligence models to forecast electricity loads and solar power generation systems. In Vietnam, electricity demand is growing rapidly, at approximately 10% per year. Our research team has proposed AI methods that utilize time series data for these forecasts. The results indicate that enhanced techniques derived from long short-term memory models yield promising outcomes. Furthermore, Vietnam leads among ASEAN countries in terms of installed solar power capacity. Based on the data collected from weather patterns, our research demonstrates that using artificial intelligence models produces highly accurate forecast results.

Abstract

Direct Load Control(DLC) program is an advanced agreement between a power utility and customers to control some customers’ appliances (e.g., air conditioners and water heaters). It is an incentive-based demand response system in which the utility provides the affected customer with financial incentives. Under the existing DLC program, the utility turn on/off customers’ appliances without considering the customer's inconvenience. that causes a lot of customers are reluctant to participate in the DLC program. In this paper, To solve this problem, the author present weather based intelligent demand controller to analyze correlation between weather condition and customers energy consumption and to predict customers energy consumption. To demonstrate the validity of the proposed demand controller, porto type demand controller are developed and applied to real customer load.

Abstract

In the VPN (with data encryption) applied to protect the distributed network over the public internet, the small computing resource of low-spec devices limits TCP/IP-based VPN use. Although lightweight IoT communication protocols such as LoWPAN are used, TCP/IP-based VPNs such as IPsec and OpenVPN require bandwidth, CPU/memory, and electric power.

To provide a safe and lightweight network, we propose a secure and scalable IoT platform (SSI) that can prevent security threats while minimizing the usage of computing resources. SSI, which has a lower load than TCP/IP-based VPN, is a layer 2 VPN and supplies data link frame encryption. L2TP and VXLAN are provided for a scalable layer 2 VPN, and the MACsec algorithm encrypts layer 2 frames. Furthermore, SSI shows 30% network speed improvement and 31.6% CPU usage reduction compared to the network applied OpenVPN.

Abstract

Modern challenges—such as the growing volume of freight and passenger transportation, the need to reduce greenhouse gas emissions, and the global shift toward sustainable technologies—demand a rethinking of energy strategies in Kazakhstan's railway sector. This keynote will explore key development directions including electrification, the integration of renewable energy sources, improvements in rolling stock energy efficiency, and the potential use of alternative fuels. Special attention will be given to the role of rail transport energy in achieving the national decarbonization strategy and building a sustainable transportation infrastructure.

Abstract

Adversarial modifications of input data can degrade the stability of deep neural networks in medical image segmentation. This study evaluates the robustness of UNet and Att-UNet++ architectures using the NuInsSeg dataset with annotated nuclear regions from various tissue sources. Both models were trained and tested under eight perturbation types, including gradient-based, iterative, and stochastic methods, with identical parameter settings. In the absence of perturbations, Att-UNet++ produced higher segmentation results with a Dice of 0.7160 and a mean IoU of 0.6190 compared to 0.6424 and 0.4732 for UNet. Under NI-FGSM and Gaussian noise, Att-UNet++ experienced a greater reduction in mean IoU, reaching 0.1215 and 0.0658, while UNet maintained 0.1968 and 0.2329. Loss landscape analysis showed smoother surfaces for Att-UNet++, yet revealed increased responsiveness to directional gradients. The findings suggest that improvements in segmentation accuracy through architectural modifications may be accompanied by increased vulnerability to input perturbations, highlighting the necessity of robustness evaluation in model development for medical image analysis.

Abstract

The increasing demand for sustainable and reliable railway transportation has spurred innovations in traction power supply systems, particularly through the integration of microgrids. This talk presents a comprehensive overview of resilient traction power supply systems in modern railways, emphasizing the role of interconnected microgrids in enhancing operational reliability, energy efficiency, and system flexibility.

We explore key architectural advancements that enable bi-directional power flow, dynamic load management, and fault tolerance. Particular attention is given to the coordination between railway substations, renewable energy sources, energy storage systems, and intelligent control frameworks that form the backbone of interconnected microgrids. Case studies and simulation insights highlight how these systems can effectively mitigate power outages, reduce dependency on centralized grids, and support the transition toward greener railway networks. The speech concludes with a discussion on current challenges, future research directions, and policy implications for large-scale deployment.

Abstract

The global transition toward renewable energy and intelligent computing is driving the development of multifunctional, environmentally integrated technologies. This study presents innovations in transparent photovoltaics and neuromorphic devices designed for both terrestrial and underwater applications. A highly efficient transparent photovoltaic (TPV) module was developed, comprising an FTO/n-ZnO/p-NiO/silver nanowires (AgNWs)/ZnO structure with a transparent p–n junction. By employing metal sputtering techniques to assemble module arrays, we significantly increased the voltage output per unit area, offering new opportunities for building-integrated transparent photovoltaics (BITPVs) that generate solar energy without obstructing visibility.

In parallel, we introduce a water-enhanced transparent photovoltaic system for underwater power generation, addressing critical needs in sensing, networking, and communication for ocean exploration. This wide-bandgap heterojunction device achieves an average visible transmittance of 41% and an underwater power conversion efficiency of 10.25% under monochromatic 470 nm light. The water medium not only focuses incident light but also improves light collection angles, while impedance spectroscopy reveals that a water-induced electric double layer enhances electron transport and photogeneration. Demonstrating its practicality, the system successfully powered a 10 mW load underwater, establishing a pathway toward secure, energy-autonomous underwater environments.

Additionally, we report on a photonic neuromorphic device based on 2D layer (SnS) into the transparent heterojunction device, capable of performing basic logic operations (OR, NOT, AND) using light as a stimulus. This artificial synapse emulates neurotransmitter-like behavior, modulating signal processing for precise control of logic functions. Such a system promises significant advancements in neuromorphic computing, with potential applications in AI, humanoid robotics, and adaptive intelligent electronics.

Together, these developments represent a significant step toward seamlessly integrated energy-harvesting, intelligent computing, and environmentally adaptive systems for the future.

Abstract

The building sector plays a crucial role in global decarbonization efforts, accounting for a significant share of energy consumption and carbon emissions. Decarbonization in this sector spans both pre-construction and post-construction phases. While pre-construction focuses on sustainable materials, design, and construction practices, the post-construction phase—the operational lifetime of buildings—presents the greatest opportunity for impactful and immediate carbon reduction through energy efficiency.

Buildings consume a substantial portion of total global energy, with electricity being the dominant form of energy used, especially in urban environments. The primary systems responsible for electricity consumption in buildings include air conditioning (approximately 40–50%), lighting (15–20%), and other electrical appliances and systems (30–40%), though this can vary by region and building type.

Improving energy efficiency in each of these systems can lead to significant reductions in energy use and associated emissions. For air conditioning, strategies such as high-efficiency HVAC equipment, smart thermostats, and building envelope improvements can cut energy use by up to 30–40%. In lighting, transitioning to LED technology and using daylighting strategies can reduce consumption by over 50%. Efficient appliances, load management, and the integration of smart control systems can also optimize overall building energy use.

When applied holistically, energy efficiency measures can reduce total building energy consumption by 20–40%, making it a cornerstone of decarbonization strategies. As buildings are expected to remain operational for decades, prioritizing energy efficiency in the post-construction phase is not only cost-effective but also essential for achieving long-term carbon neutrality goals in the built environment.

Abstract

This presentation discusses system design and modeling aspects of VRF systems when integrated into net-zero energy building (NZEB) designs. Various system configurations of VRF systems that can be used in NZEB designs and their influence on the NZEB energy performance will be illustrated. Modeling strategies of VRF systems for NZEB buildings will also be discussed in this presentation. Simulation-based case studies of an office prototype building model with a VRF based NZEB model will be presented to demonstrate potential economic and energetic benefits in various US climate locations.

Abstract

My current line of research, entitled Sub One Shot Learning (SOSL), tackles the hitherto neglected situation in which the expected number of labelled examples per class (α) is strictly between zero and one. This problem formulation naturally occurs in ultra rare disease imaging, endangered language processing, military anomaly detection, and privacy sensitive on device personalisation—domains where even a single labelled instance can be prohibitively expensive or ethically constrained.

A rigorous PAC Bayesian analysis demonstrates that, as α → 0, empirical evidence vanishes while the KL divergence to a prior becomes the dominant term of the generalisation bound. Consequently, our first contribution is the construction of high fidelity priors distilled from state of the art vision language and medical foundation models. Building upon this theoretical footing, we propose a three layer SOSL pipeline:

1. Variational Bayesian Core — produces a calibrated posterior by jointly minimising task loss and KL divergence;
2. Ultra light Adaptation — adapter modules and prompt tuning inject task specific knowledge with only a few trainable parameters, thereby constraining the divergence term and enabling low power deployment;
3. Single Step Meta Learning — a degenerate MAML loop that adapts in one gradient step, or in the literal zero shot case, defers entirely to the prior.

To prevent synthetic bias when data are nearly absent, we restrict augmentation to group theoretic transformations that respect known symmetries without inducing distributional drift. Empirical studies confirm the merit of this design: on α constrained versions of mini ImageNet and Omniglot the method secures up to 18 % relative accuracy gains, while on a 14 slice MRI set for an ultra rare neoplasm it achieves a clinically actionable AUC 0.87 with tightly calibrated uncertainty.

Ongoing extensions include (i) Continual SOSL, which allows the model to transition smoothly from vacuum level data to moderate abundance without catastrophic forgetting, and (ii) Trustworthy SOSL, which emphasises Bayesian calibration for safety critical decisions. To foster community engagement, our group is releasing an open benchmark and leaderboard devoted to SOSL research.

I warmly invite collaborators who share the ambition of enabling reliable artificial intelligence precisely where data are scarcest.

Abstract

This presentation addresses the digitalization of overhead contact line maintenance for railway electrification through a novel management process utilizing 3D scanning and printing. The initiative aims to overcome the limitations of current 2D-based management and manual inspections by creating a 3D database of overhead line components. This approach seeks to enhance maintenance efficiency, improve component development, and support a digital twin framework for the railway industry. The process involves reverse engineering of components, creating precise 3D models with integrated metadata, and utilizing 3D printing for on-demand production and rapid prototyping. This methodology is positioned as a key technology for advancing railway infrastructure management and ensuring a more reliable and safer railway service.