Personalized Federated Learning Framework for Load Forecasting

This repository implements a federated learning framework for time series forecasting and incorporates various FL personalization mechanisms to enhance performance for non-IID clients.

Methods Implemented

• FedAvg based Federated Learning: A decentralized approach where multiple clients collaboratively train a model without sharing their raw data.
• Adaptive Local Aggregation (FedALA): Optimizes the initialization of the client model in each FL round.
• Model-Mix Client Personalization: Achieves personalization by building for each client a private local model and a public one that participates in FL, and mixing their weights.
• Client Clustering: An optional algorithm to divide clients into clusters based on load patterns similarity without sharing their data.
• The implementation of the Fedora PFL algorithm for load forecasting is available here.

Datasets

In these experiments, we use the IRISE and REFIT datasets. While IRISE is not publicly available, the REFIT dataset can be obtained here. It should be stored in data/raw/CLEAN_REFIT_081116.

File Overview

project_root/ 
├── src/                             # Directory containing the source code.  
│   ├── simulation.py                # Main script to start the FL simulation.  
│   ├── config.py                    # Contains configuration parameters for the simulation's setup.  
│   ├── server.py                    # Defines the server-side strategy for FL, including aggregation of client updates.  
│   ├── communication.py             # Manages communication protocols between clients and the server.  
│   ├── clients/                     # Contains client-related files for FL.  
│   │   ├── client.py                # Entry point for creating client objects and defining interaction with the server.  
│   │   ├── base_model.py            # Defines client logic for standard FL without personalization.  
│   │   ├── ALA.py                   # Implements Adaptive Local Aggregation for initializing the client model in each FL round.  
│   │   └── mixed_model.py           # Implements a forecasting model that combines local and federated models for personalization.  
│   ├── forecasting/                 # Contains different forecasting model implementations.  
│   │   ├── SCINet/                  # Directory for SCINet model implementation.
│   │   │   ├── SCINet.py            # Defines the model.
│   │   │   └── wrapper.py           # Handles the model training and evaluation logic.
│   │   ├── seq2seq/                 # Directory for Seq2Seq model implementation.
│   │   │   ├── model.py             # Defines the model.
│   │   │   └── wrapper.py           # Handles the model training and evaluation logic.
│   │   └── early_stop.py            # Implements early stopping logic to prevent overfitting.  
│   ├── isolated_client.py           # Implements logic for a client operating in isolation without FL.
│   ├── preprocess_dataset.py        # Preprocesses raw data.
│   ├── dataset.py                   # Handles data loading for the clients.  
│   ├── metrics.py                   # Defines custom evaluation metrics for the FL process.  
│   ├── utils.py                     # Utility functions for logging, setting seeds, and preparing folders.  
│   └── notebooks/                   # Contains Jupyter notebooks.
│   │   ├── load_clustering.ipynb    # Visualizes the client clustering process.
│   │   └── results_analysis.ipynb   # Aanalyzes and visualizes the results.
├── .gitignore                       # Specifies files and directories to be ignored by Git.  
├── LICENSE                          # License file for the project.  
├── requirements.txt                 # Lists project dependencies for easy installation.  
└── README.md                        # Overview of the project, including methods implemented and how to run the simulation.  

Running the Simulation

1. Set up a virtual environment:

   python -m venv env
   source env/bin/activate  # On Windows, use `env\Scripts\activate`

2. Install Dependencies:

   pip install -r requirements.txt

   Note: this repository was developed with the 1.8.0 version of Flower and might not work out of the box for later versions.

3. Download and preprocess the data:

   Place the raw dataset files in the data/raw directory. Each file represents a different client. These files will be preprocessed (normalized, resampled to hourly frequency and NaN values handled) using the following command:

   python src/preprocess_dataset.py

4. Configure parameters:

   Modify src/config.py to set desired parameters. Refer to the comments in that file for a descriptions of each parameter.

5. Run the simulation:

   python src/simulation.py

   Results will be saved to simulations/<simulation_id>/ and can be analyzed with src/notebooks/results_analysis.ipynb.

Citation

TBA