FedEDF

Data and source code for the paper "On Electric Vehicle Energy Demand Forecasting and the Effect of Federated learning"

Table of Contents

• Overview
• Prerequisites & installation
• Data preparation
  • EVSE transaction processing
  • Metadata processing
  • Feature engineering
• Model training
  • Centralized learning (baseline; ARIMA, SARIMA, SARIMAX)
  • Centralized learning (ML-based; XGBoost)
  • Centralized learning (ML-based; [Bi-]LSTM/GRU)
  • Federated learning (ML-based; XGBoost)
  • Federated learning (ML-based; [Bi-]LSTM/GRU)
• Inference & visualization
• Contributors
• Acknowledgement

Overview

This repository provides a complete end‑to‑end pipeline for Electric Vehicle Supply Equipment (EVSE) demand forecasting. The pipeline covers:

• Raw data ingestion & cleaning – outlier filtering, timeseries creation.
• Metadata enrichment – location (geocoding), charger type, nominal power output.
• Feature engineering – resampling, lag variables, cyclical encodings, rolling statistics, extrapolation.
• Centralised models – statistical baselines (ARIMA, SARIMA, SARIMAX), XGBoost, RNN ([Bi-]LSTM/GRU).
• Federated learning – privacy‑aware training of RNNs and XGBoost across multiple "EVSE hubs" using Flower.
• Evaluation & visualisation – performance tables (MASE, SMAPE, RMSE, R²) and geographic maps of EVSE locations.

All scripts are written in Python 3.10. For the sake of reproducibility, the /data directory includes the preprocessed datasets, along with pretrained models.

Prerequisites & installation

1. Python 3.10
2. System libraries – git, curl, wget (for dataset download)
3. (Optional) GPU – A CUDA‑enabled GPU speeds up model training; otherwise the CPU fallback works out‑of‑the‑box.
4. Python packages (cf., requirements.txt)

pip install -r requirements.txt

Note – The code uses the Nominatim OpenStreetMap API for geocoding. No API key is required, but respect the usage policy (https://operations.osmfoundation.org/policies/nominatim/). The original EVSE datasets can be found at: https://github.com/yvenn-amara/ev-load-open-data

Data preparation

EVSE transaction processing

python 0-data-processing.py \
    --dataset {dundee,boulder,paloalto} \
    --sr_freq 12H \         # resampling frequency
    --min_pts 100           # minimum transactions per EVSE

Build Metadata

python 1-metadata-preprocessing-dundee-v3.py
python 1-metadata-preprocessing-boulder-v3.py
python 1-metadata-preprocessing-paloalto-v3.py

Feature Engineering

Open and execute all cells in the notebook 1.5-feature-engineering-v3.ipynb for each dataset.

Model training

Centralized learning (baselines; ARIMA, SARIMA, SARIMAX)

python 2-statistical-training-v4.py --dataset {dundee, boulder, paloalto} --sr_freq 12H --min_pts 100

Centralized learning (ML-based; XGBoost)

Open and execute all cells in the notebook 3-xgboost-training-v5.ipynb for each dataset.

Centralized learning (ML-based; [Bi-]LSTM/GRU)

python 4-rnn-training-v3.py \
    --data {dundee,boulder,paloalto} \
    --njobs 96 \
    --hidden_size 24 \
    --rnn_cell {lstm,gru} 
    [--bi] \
    --num_layers 1 \
    --fc_layers 24 \
    --sr_freq 12H \
    --min_pts 100 \
    --bs 32 \
    --length 48 \
    --stride 1 \
    --n_epochs 100 \
    --patience 10 \

Federated learning (ML-based; XGBoost)

# FedXGBllr (Average)
python fededf_server.py \
    --federation {dundee,boulder,paloalto} \
    --num_rounds 40 \
    --local_epochs 10 \
    --early_stop \
    --patience 3 \
    --port 28080 \
    --mu 0.0 \
    --clients 8 \
    --conv_channels 16 \
    --dropout_rate 0.13 \
    --strategy fedxgbllr

# FedXGBllr (Proximal)
python fededf_server.py \
    --federation {dundee,boulder,paloalto} \
    --num_rounds 40 \
    --local_epochs 10 \
    --early_stop \
    --patience 3 \
    --port 28080 \
    --mu 1e-1 \
    --clients 8 \
    --conv_channels 16 \
    --dropout_rate 0.13 \
    --strategy fedxgbllr

# Launch federation
bash launch_fedxgb_clients.sh \
    --federation {dundee,boulder,paloalto} \
    --n_estimators 37 \
    --sr_freq 12H \
    --min_pts 100 \
    --bs 64 \
    --port 28080

Federated learning (ML-based; [Bi-]LSTM/GRU)

# FedProx
python fededf_server.py \
    --federation {dundee,boulder,paloalto} \
    --num_rounds 50 \
    --local_epochs 5 \
    --early_stop \
    --patience 2 \
    --port 28080 \
    --mu 1e-1 \
    --clients 8 \
    --strategy fedprox

# Launch federation
bash launch_fedrnn_clients.sh \
    --federation {dundee,boulder,paloalto} \
    --rnn_cell {lstm, gru} \
    [--bi] \
    --hidden_size 24 \
    --num_layers 1 \
    --fc_layers 24 \
    --sr_freq 12H \
    --min_pts 100 \
    --bs 32 \
    --length 48 \
    --stride 1 \
    --port 28080

Note – Before running any FL-related experiments, use notebook 6-data-federation.ipynb to generate the data for each client in the federation.

Inference & visualization

After training, the server stores the global checkpoint under /data/pth/.

To obtain model forecasts, execute the code in notebooks 4.1-rnn-inference-v3.ipynb, 6.1-federated-rnn-inference.ipynb and 6.2-federated-xgb-inference.ipynb.

Use notebooks 5-centralized-edf-evaluation.ipynb and 6.3-federated-edf-evaluation.ipynb to generate the comparison tables.

Contributors

• Andreas Tritsarolis; Department of Informatics, University of Piraeus
• Gil Sampaio; Center for Power and Energy Systems, INESC TEC
• Nikos Pelekis; Department of Statistics and Insurance Science, University of Piraeus
• Yannis Theodoridis; Department of Informatics, University of Piraeus

Acknowledgement

This work was supported in part by the Horizon Europe Research and Innovation Programme of the European Union under grant agreement No. 101070416 (Green.Dat.AI; https://greendatai.eu). In this work, INESC TEC provided the requirements of the business case, as well as access to the FEUP dataset and their Energy benchmarking tool.