Skip to content

README.md

Latest commit

 

History

History
135 lines (98 loc) · 4.23 KB

README.md

File metadata and controls

135 lines (98 loc) · 4.23 KB

🌞🌬️ Renewable Power Generation Forecast – Day-Ahead

This project aims to forecast day-ahead wind and solar power generation to reconstruct the global power stack and support a fundamental-based price forecast for electricity markets.

🧠 Objective

The primary goal is to model and forecast the renewable power output for the next day based on weather forecasts. This includes:

  • Forecasting solar and wind generation using machine learning.
  • Combining this with consumption forecasts to estimate the power stack.
  • Enabling a fundamental pricing model for day-ahead markets.

📦 1. Data Sources

🌤️ Weather Data

  • Source: Open-Meteo API
  • Storage: Supabase (access via API)
  • Key features: [ 'Solar_Radiation', 'Direct_Radiation', 'Diffuse_Radiation', 'Direct_Normal_Irradiance', 'Global_Tilted_Irradiance', 'Cloud_Cover', 'Cloud_Cover_Low', 'Cloud_Cover_Mid', 'Cloud_Cover_High', 'Temperature_2m', 'Relative_Humidity_2m', 'Dew_Point_2m', 'Precipitation', 'Wind_Speed_100m', 'Wind_Direction_100m', 'Wind_Gusts_10m', 'Surface_Pressure', 'SR_capa', 'WIND_capa' ]

⚡ Generation Data

  • Source: ENTSO-E API
  • Labels: Historical solar and wind power generation
  • Storage: Supabase (separate table)

🧹 2. Data Cleaning

☀️ Solar Generation

Challenge: Solar generation data can include physical inconsistencies (e.g., production at night).

Hybrid Outlier Detection Strategy:

  1. Binning by Solar Radiation:
  • Bins reflect different irradiation regimes: night, dawn/dusk, midday.
  1. Low Radiation (< 5 W/m²):
  • Any positive generation is flagged and removed (physically implausible).
  1. Moderate Radiation:
  • Interquartile Range (IQR) filtering used: 
    IQR = Q3 - Q1
Outlier if x < Q1 - 1.5*IQR or x > Q3 + 1.5*IQR
  1. High Radiation:
  • Median Absolute Deviation (MAD) method used: 
    z = |x - median| / MAD
Outlier if z > 4

image

🌬️ Wind Generation

Challenge: Affected by curtailment, noise, and inconsistencies.

Two-Stage Outlier Detection:

  1. Bin-wise IQR Filtering by Wind Speed:
  • Low Speeds (< 7 m/s): Any positive generation flagged.
  • 7–16 m/s: Apply IQR method.
  1. Model-Based Residual Filtering:
  • Fit a 3rd-degree polynomial regression: 
    Y = a₀ + a₁x + a₂x² + a₃x³
r = y - Y
z = |r - median(r)| / MAD
Outlier if z > 5

image

🤖 3. Model Training

  • Algorithms Used: LightGBM / XGBoost
  • No normalization required due to tree-based models
  • Data cleaning is applied prior to training

✅ Model Pipeline

  • Cleaned data is piped into training
  • Model and pipeline saved in .pkl format for deployment

📈 4. Model Performance

Best results after tuning and cleaning:

Model Target MAE RMSE R²
LGBMR_cleaned_tuned WIND 639.42 836.27 0.875
LGBMR_cleaned_tuned SOLAR 128.53 229.83 0.997

📊 5. Dashboard

Track daily forecasts against actuals in near-real time:

🔗 Forecast vs Realized Power Generation – France

Compare ResPy's morning forecast, RTE’s evening forecast, and the final realized RES generation.

🛠️ Future Improvements

  • Introduce normalization and feature scaling for advanced models (e.g., neural nets)
  • Extend location coverage beyond central France
  • Add dynamic capacity estimation from asset metadata or satellite imagery
  • Integrate with electricity price forecasting module

👤 Author

Hugo Lambert – Energy Forecasting & Market Modeling
Feel free to reach out hugo.lambert.perso@gmail.com