Yildirimhan Aydin

🌬️ Wind Forecast ML Pipeline

This project presents a modular machine learning pipeline designed to forecast wind speed and wind direction for multiple wind farms using weather forecast model data.

Built as part of a broader energy optimization effort, this system helps improve the forecast accuracy of wind resources by training per-site ML models, automating predictions, and logging performance daily.

🧠 Key Goals

• Increase short-term forecast accuracy for wind assets.
• Automate daily forecasting (24 hours ahead).
• Track per-site model performance over time.
• Support easy retraining and feature testing.

🏗️ System Architecture

The pipeline is built using modular notebooks (or scripts), each responsible for a distinct stage:

1. Feature Selection

• Evaluate and compare feature sets per wind site.
• Models tested: XGBoost, LightGBM, Random Forest.
• Metrics calculated: MAE, RMSE, R², MAPE, Bias.
• Best features and results stored in a tracking table.

2. Model Training

• Trains a dedicated model per site using selected features.
• Stores trained models in cloud storage (e.g., object store).
• Logs metadata: hyperparameters, metrics, model paths.

3. Daily Prediction

• Runs once per day (e.g., at 10:00 AM) for next-day forecasts.
• Loads weather model inputs + trained models.
• Predicts 24 hourly values for wind speed and direction.
• Skips sites with missing data and logs the issue.

📥 Data Sources (Generalized)

• Weather Forecast Models:
• Global Forecast System (GFS)
• ECMWF

• Other regional weather models

• Data Format: tabular format with fields like:

• timestamp_to
• wind_speed, wind_direction
• facility_id

🗃️ Output Tables (Generalized Schema)

🔹 model_features

Tracks selected features and performance per model.

🔹 model_registry

Stores paths, parameters, and metrics for trained models.

| Facility ID | Model Path | Best Params | MAE | R² | Timestamp | ... |

🔹 daily_predictions

Daily predictions stored per facility, per hour.

| Facility ID | DateTime | Wind Speed | Wind Direction | Model Version | Prediction Time |

🔹 skipped_sites

Logs skipped predictions due to missing features.

| Facility ID | Date | Missing Inputs | Logged At |

🛠️ Tools & Tech

• Python (pandas, scikit-learn, xgboost, etc.)
• Notebook-based pipeline (Jupyter or Databricks style)
• Object storage or Lakehouse for data + model storage
• Cloud scheduling or job orchestration
• Versioned model tracking tables

📈 Outcomes

• Trained models per wind farm with optimized feature sets.
• Forecasts stored daily for operational use.
• Fail-safe logic to skip and log incomplete predictions.
• Easily extendable to power forecasting or anomaly detection.

🚀 Possible Enhancements

• Auto-retraining if accuracy drops.
• Dashboards for forecast vs. actual (e.g., Power BI).
• Alerts for persistent data gaps or model drift.
• Integration with power curve models for generation estimation.

📝 Notes

• This is an anonymized version of a real production pipeline used in renewable energy forecasting.
• All data, identifiers, and systems have been generalized for public sharing.