maxamud123/federated-learning-smart-grid

Optimizing Federated Learning for Resource-Constrained Edge Devices in Smart Grids - Master's Thesis Research (ULK Rwanda) # Optimizing Federated Learning for Resource-Constrained Edge Devices in Smart Grids ## A Case Study in Somalia **Thesis:** Optimizing Federated Learning for Resource-Constrained Edge Devices in Smart Grids **Author:** Mohamoud Abukar | Reg No: 202413001 **Supervisor:** Dr. KAMUHANDA Danny **University:** Kigali Independent University (ULK) | MSc Internet Systems | 2024-2025 --- ## Project Structure ```text federated-learning-smart-grid/ ├── data/ │ ├── raw/ ← Place UCI dataset here │ ├── processed/ ← Cleaned hourly data │ └── splits/ ← Per-client numpy arrays ├── src/ │ ├── models/ │ │ └── lstm_model.py ← LSTM + INT8 quantization │ ├── preprocessing/ │ │ └── data_loader.py ← UCI loading, windowing, normalization │ ├── client/ │ │ └── fl_client.py ← Flower client + Top-K compression │ ├── server/ │ │ └── fl_server.py ← Flower server + FedAvg │ └── evaluation/ │ └── metrics.py ← RMSE, MAE, memory, communication metrics ├── experiments/ │ ├── results/ ← JSON result files │ └── logs/ ← Per-round training logs ├── DEVICE_SETUP.md ← Hardware setup guide (Pi 4 emulated on laptop, Raspberry Pi Zero 2 W, ESP32) ├── main.py ← Entry point for all modes ├── requirements.txt └── README.md ``` --- ## Three Optimizations | Optimization | Technique | Target Reduction | | ------------- | -------------------------- | --------------------- | | Memory | INT8 Quantization | ~50% RAM reduction | | Communication | Top-K Compression (k=0.1) | ~60% bandwidth saving | | Training Time | Adaptive Local Epochs | ~40% time reduction | --- ## Physical Setup (No Router Needed) The Raspberry Pi 4 client runs on the laptop/host machine using the `pi4` device profile — it is not run on physical Pi 4 hardware. Physical hardware validation is performed on the Raspberry Pi Zero 2 W and …