Transformer-Based Energy Demand Forecasting in Smart Grids and Inference Task Offloading
Hadish, Siem
Hadish, Siem
Author
Supervisor
Department
Machine Learning
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Smart grids are transforming energy management and are at the forefront of many green initiatives and smart city projects. Their integration with multi-access edge computing (MEC) makes them even more attractive tools to tackle many energy and compute-related challenges. Enabling them with tools to forecast energy consumption and allocate tasks more efficiently makes them even better management of resources. This thesis leverages transformer-based models to address these challenges, with a focus on edge deployment to enhance smart grid performance. The first part of the thesis includes an extensive review of the current state of edge-deployed large language models (EdgeLLMs), exploring the challenges, enablers, and current state of the solutions. Common LLM compression techniques like pruning, knowledge distillation, and quantization are discussed along with efficient training methods. A holistic framework is also presented integrating the main components of a successful EdgeLLM ecosystem. Building on those foundations, the second part proposes a novel transformer-based model for energy consumption forecasting in smart grids, designed for edge deployment. The model utilizes resource-conscious attention mechanisms such as FlashAttenttion and ProbAttention for efficient robust energy demand forecasting in microgrid settings. This model is trained and validated using the Pecan Street dataset and predicts the energy consumption of a grid, over 24 and 48 hour horizons, outperforming traditional models enabling realtime demand supply optimization. The application of this model could extend to improving grid reliability and reducing energy waste, demonstrating a tangible application of edge LLMs in sustainable energy management. The third part focuses on operational efficiency, proposing an LLM-enabled distributed framework for task offloading in smart grids. With smart grids that power multi-access edge computing (MEC) environments, a matching gamebased scheme is utilized to optimize the allocation of computational tasks. These servers could host a range of tasks such as smart meter analytics and user query processing, across edge nodes. LLMs are deployed to reason about the task priorities and resource availability providing a scalable solution to the resource contention challenges inherent in MEC and smart grid settings. Overall, this thesis aims to bridge theoretical insights and practical innovations in advancing the deployment of edge LLMs in smart grids.
Citation
Siem Hadish, “Transformer-Based Energy Demand Forecasting in Smart Grids and Inference Task Offloading,” Master of Science thesis, Machine Learning, MBZUAI, 2025.
Source
Conference
Keywords
Machine Learning, Transformers, Smart Cities, Smart Grids