Asiedu13/windenergy_ml_pipeline

A pipeline for training and producing wind resource prediction models for Ghana # Ghana Wind Power Forecasting Pipeline Two-stage wind power forecasting pipeline for Ghanaian sites. - **Stage 1 (ML):** predicts hub-height wind speed some hours ahead from reanalysis weather. - **Stage 2 (Simulator):** converts predicted wind speed to power using a manufacturer power curve via `windpowerlib`. Ghana has no operating utility wind farm publishing SCADA data, so this pipeline forecasts *potential* output from reanalysis weather (ERA5 or NASA POWER) plus a manufacturer power curve. See `outputs/reports/report.md` after a run for feature rankings, three-model comparison, and error-propagation analysis. ## Quick start ```bash pip install -r requirements.txt python -m src.pipeline --config config.yaml ``` Outputs land in `outputs/models/`, `outputs/figures/`, `outputs/reports/`. ## Data sources The pipeline supports two reanalysis sources (choose via `data.source` in `config.yaml`): ### ERA5 (primary, higher quality) Requires a free Copernicus Climate Data Store (CDS) account. 1. Register at cds.climate.copernicus.eu 2. Copy your API key from your CDS profile. 3. Create `~/.cdsapirc` with: ``` url: cds.climate.copernicus.eu key: ``` 4. Accept the ERA5 licence once via the CDS web UI. First run downloads ~decade of hourly variables for a small bbox around the site — expect a slow first run (minutes to hours depending on date range) and cached NetCDF in `data/raw/`. ### NASA POWER (no auth, coarser) Set `data.source: nasa_power` in `config.yaml`. Uses the public POWER hourly API, no credentials needed. ## Changing site or turbine Edit `config.yaml`. Retraining on a new site is a one-line change: ```yaml site: name: takoradi # any short slug latitude: 4.90 longitude: -1.75 hub_height_m: 100 ``` Turbine name must exist in the `oedb` library (see `windpowerlib.get_turbine_types()`). ## Retraining note Wind-speed models are **site-specific** — a model trained at Anloga will not generalize to Ada Foah. Rerun the …