AI and Optimization for Resilient and Efficient Local Energy Markets

Researcher: Qiao Xiao, Eindhoven University of Technology

Meet Qiao Xiao and His Work at Megamind 

Qiao Xiao is a final-year PhD researcher at Eindhoven University of Technology and part of the MEGAMIND project. His work focuses on making artificial intelligence (AI) more efficient, scalable, and privacy-preserving, with a strong drive to apply these innovations in decentralised and resource-constrained environments such as local energy markets.

By combining AI and optimisation techniques, Qiao’s research develops intelligent solutions that:

  • Reduce the need for high-performance computing,

  • Work within system-level constraints,

  • Safeguard data privacy, and

  • Enable smooth coordination between decentralised actors.

His ultimate goal: to create AI that works for everyone, not just in large data centres, but also in households, communities, and energy grids of the future.

 

Why Efficient AI Matters for Energy Systems

The energy transition is driving a shift towards decentralised and dynamic energy networks. Households, communities, and businesses are becoming both consumers and producers of energy, which makes grid management increasingly complex.

However, conventional AI models often demand large amounts of data and computing power, raising challenges around scalability, privacy, and sustainability.

Qiao addresses these challenges by developing methods that are:

  • Lightweight → enabling AI models to run in resource-limited settings,
  • Data-efficient → focusing on the most important and informative data,
  • Privacy-friendly → ensuring collaboration without data sharing.

This research directly supports MEGAMIND’s vision of creating resilient, secure, and efficient local energy markets.

Research Highlights

🔹 Model-Level Efficiency

Qiao explores sparse neural networks that drastically reduce computational and memory requirements. His methods include sparsity-aware architectures for time series analysis and parameter-efficient fine-tuning for large language models (LLMs). These innovations bring advanced AI within reach of real-world energy applications.

🔹 Data-Level Efficiency

Through adaptive data selection and prioritisation, his models learn faster, become more robust, and require less data. For grid operators, this means AI systems that can handle dynamic conditions without unnecessary overhead.

🔹 Federated and Communication-Efficient Learning

Since energy data is both sensitive and distributed, Qiao applies federated learning. This allows multiple parties to jointly train AI models without sharing raw data. To make it scalable, he integrates lightweight communication protocols and sparsity-aware updates, cutting both costs and privacy risks.

From Research to Real-World Impact

Qiao’s work bridges advanced AI methods with practical benefits for the energy sector. His innovations help local energy actors—such as households with solar panels and community energy initiatives—coordinate efficiently, securely, and sustainably.

Selected Publications

  • Dynamic Sparse Network for Time Series Classification: Learning What to “See” (NeurIPS 2022)
    Introduces efficient sparse networks for time series analysis, cutting computation by over 50% while maintaining high accuracy.
  • Leave it to the Specialist: Repair Sparse LLMs with Sparse Fine-Tuning via Sparsity Evolution (arXiv 2025)
    Proposes a method to efficiently adapt large language models in resource-constrained scenarios with minimal parameter updates.
  • Dynamic Data Pruning for Automatic Speech Recognition (Interspeech 2024)
    Demonstrates how training with only 70% of the data can achieve the same performance as full-data training.
  • Addressing the Collaboration Dilemma in Low-Data Federated Learning via Transient Sparsity (arXiv 2025)
    Presents an efficient federated learning framework that reduces communication costs and strengthens privacy protection.

Looking Ahead

Qiao Xiao’s research shows how AI and optimisation can accelerate the energy transition by making our grids more intelligent, efficient, and resilient. His work not only advances the state of AI research but also ensures that innovations translate into real-world impact for local energy systems.

Key Takeaways

⚡ Sparse AI models lower computing and memory demands.
📊 Adaptive data strategies speed up training and improve robustness.
🔒 Federated learning enables secure collaboration without sharing sensitive data.
🌍 Efficient communication protocols make decentralised AI scalable.

References

  • Xiao, Q., Wu, B., Zhang, Y., Liu, S., Pechenizkiy, M., Mocanu, E., & Mocanu, D. C. (2022). Dynamic Sparse Network for Time Series Classification: Learning What to “See”. NeurIPS 2022. Link
  • Xiao, Q., Ansell, A., Wu, B., Yin, L., Pechenizkiy, M., Liu, S., & Mocanu, D. C. (2025). Leave it to the Specialist: Repair Sparse LLMs with Sparse Fine-Tuning via Sparsity Evolution. arXiv:2505.24037. Link
  • Xiao, Q., Ma, P., Fernandez-Lopez, A., Wu, B., Yin, L., Petridis, S., Pechenizkiy, M., Pantic, M., Mocanu, D. C., & Liu, S. (2024). Dynamic Data Pruning for Automatic Speech Recognition. Interspeech 2024. Link
  • Xiao, Q., Wu, B., Yin, L., Gadzinski, C. N., Huang, T., Pechenizkiy, M., & Mocanu, D. C. (2024). Are Sparse Neural Networks Better Hard Sample Learners?. BMVC 2024. Link
  • Wu, B., Xiao, Q., Wang, S., Strisciuglio, N., Pechenizkiy, M., van Keulen, M., Mocanu, D. C., & Mocanu, E. (2025). Dynamic Sparse Training versus Dense Training: The Unexpected Winner in Image Corruption Robustness. ICLR 2025. Link
  • Wu, B., Xiao, Q., Liu, S., Yin, L., Pechenizkiy, M., Mocanu, D. C., van Keulen, M., & Mocanu, E. (2024). E2ENet: Dynamic Sparse Feature Fusion for Accurate and Efficient 3D Medical Image Segmentation. NeurIPS 2024. Link
  • Xiao, Q., Wu, B., Poddubnyy, A., Mocanu, E., Nguyen, P. H., Pechenizkiy, M., & Mocanu, D. C. (2025). Addressing the Collaboration Dilemma in Low-Data Federated Learning via Transient Sparsity. arXiv:2506.00932. Link