Research Areas

Exploring the frontiers of artificial intelligence and machine learning through innovative research and cutting-edge applications.

Solar Energy & Photovoltaic AI

AI for the full solar pipeline: segmenting PV panels across image resolutions with multi-agent, uncertainty-aware frameworks, mapping rooftop installations from satellite and aerial imagery with minimal labelled data, and tracking the true global power peak in module-level solar optimizers.

Applications:

  • PV Panel Segmentation & Mapping
  • Rooftop Solar Detection from Satellite Imagery
  • Behind-the-Meter Installation Discovery
  • Generation Capacity Estimation
  • Global MPPT for Module-Level Optimizers

Active Projects:3

MAPVNet-Q: Multi-Agent PV Segmentation
A multi-agent framework that uses uncertainty-based quality features and a gain-prediction router to pick the best segmentation model, improving PV mapping reliability across imagery sources and acquisition conditions
Active
SolarMap AI: Rooftop Solar Mapping
Segments PV panels address by address across Jeju Island, estimates generation capacity, and flags unregistered installations, with a human-in-the-loop loop that retrains the model on corrected masks
Active
ML-Seeded Global MPPT
A learned seed predicts the global peak region from about six voltage probes, then hands over to Perturb & Observe and re-seeds when shading changes, tracking the true peak at a fraction of the measurement cost
Active

Probabilistic Energy & Grid Forecasting

Reliable, uncertainty-aware forecasting for power systems: physics-guided and physics-informed deep learning, meta-learning for unseen operating regimes, conformal calibration of prediction intervals, and explainable multivariate time-series models for generation, demand and grid state.

Applications:

  • Grid-State & Power-Flow Forecasting
  • Solar Power Forecasting
  • Electricity Demand & Consumption Prediction
  • Renewable Generation Forecasting
  • Uncertainty-Aware Grid Operations

Active Projects:3

Physics-Guided Meta-Learning for Grid-State Forecasting
Combines graph-based learning, uncertainty quantification and differentiable power-flow modeling to forecast nodal power, bus voltages and line loadings that stay physically consistent in unseen operating regimes
Active
Regime-Aware Conformal Solar Forecasting
Calibrates prediction intervals of physics-informed deep learning models per solar regime, for reliable coverage under changing irradiance and weather conditions
Active
Multivariate Time-Series Forecasting for Energy
Neural architectures that model temporal dynamics and cross-variable dependencies for multi-horizon forecasting of renewable generation, electricity demand and consumption
Active

Efficient Vision & Edge AI

Real-time, edge-deployable computer vision: confidence-guided hybrid architectures for language-guided video segmentation, and lightweight, distilled models that judge the analytical value of satellite imagery onboard to cut unnecessary data transmission.

Applications:

  • Real-Time Language-Guided Video Segmentation
  • Onboard Satellite Image Selection
  • Cloud-Aware Land-Cover Analysis
  • Bandwidth-Efficient Satellite Downlink
  • Edge-Deployable Vision Systems

Active Projects:2

ConfTrackNet: Real-Time Video Instance Segmentation
A confidence-guided hybrid framework pairing a heavy vision-language detector on key frames with lightweight mask propagation, re-detecting only when a multi-signal confidence estimator sees quality drop
Active
Task-Retention Utility for Satellite Downlink
Predicts how much land-cover segmentation performance survives cloud cover using clear-to-cloud knowledge distillation and a lightweight model, so satellites prioritize imagery by analytical value rather than cloud coverage alone
Active

Trustworthy LLMs & Adaptive RAG

Making generative AI reliable: adaptive retrieval-augmented generation that chooses retrieval channels per query, answers or abstains based on visual evidence and confidence, and selectively revises responses as video and sensor observations change.

Applications:

  • Smart Farming Decision Support
  • Evidence-Grounded Question Answering
  • Video & Sensor-Based Reasoning
  • Answer-or-Abstain Response Control
  • Context-Aware Response Revision

Active Projects:1

Adaptive RAG for Smart Farming
Selects retrieval channels by query and candidates, learns to answer or abstain from visual evidence, and revises only the affected parts of a response when sensor states change, delivering evidence-grounded farm decision support
Active

Interested in Collaboration?

We welcome partnerships with industry, academia, and research institutions to advance the frontiers of AI and machine learning.