Predicting solar and wind energy generation for Germany and the United Kingdom using state-of-the-art deep learning — one hour at a time.

Energy grids are fragile. Too much or too little renewable generation at the wrong moment destabilizes the entire network. PowerSight tackles this by forecasting hourly electricity load using a Temporal Fusion Transformer (TFT) — a model that doesn't just predict, it explains which variables drove each forecast.

Trained on 5+ years of real ENTSO-E grid data (2015–2020) fused with live weather observations, PowerSight achieves a validation SMAPE of 3.1% across two of Europe's largest energy markets.

Raw OPSD Grid Data (300 columns, 50K rows)
        ↓
   Filter → DE + GB only
        ↓
   Merge with Open-Meteo Weather API
        ↓
   Clean + Impute missing values
        ↓
   StandardScale + Sliding Window sequences
        ↓
   Train TFT (1.4M params) via PyTorch Lightning
        ↓
   Evaluate + Interpret attention weights

Countries: Germany (DE) · United Kingdom (GB)
Period: January 2015 – September 2020
Granularity: 1 hour

TFT advantages over vanilla LSTM:

• Multi-horizon probabilistic forecasts (10th / 50th / 90th percentile)
• Learns per-country static context
• Interpretable attention over past time steps
• Variable importance scores via gating mechanisms

• Validation SMAPE: 3.1% on 24-step ahead load forecasting
• Attention heatmaps reveal model focuses on recent 6–12 hours for load prediction
• Variable selection gates confirm windspeed_10m and shortwave_radiation as top drivers

git clone https://github.com/shivansh052k/PowerSight.git
cd PowerSight

conda create -n powersight python=3.12
conda activate powersight

pip install -r requirements.txt

Download the OPSD dataset and place it at:

Dataset/time_series_60min_singleindex.csv

Launch dashboard:

streamlit run app.py

Python · PyTorch · PyTorch Lightning · PyTorch Forecasting · pandas · scikit-learn · XGBoost · Optuna · SHAP · Streamlit · Plotly · Open-Meteo API

Shivansh Gupta , Sada Kakarla