Offshore Wind Project Assessment

Python toolkit for offshore wind project analysis: budget normalization, data quality validation, CAPEX/OPEX/LCOE calculation, ML-based CAPEX benchmarking, mapping, and new-project classification.

The project is designed as an end-to-end offshore wind decision-support prototype. It keeps the domain formulas transparent while adding source verification, confidence scoring, model validation, and reproducible reports.

What It Does

• normalizes project budgets into EUR 2026;
• validates budget plausibility with EUR/MW bounds;
• ranks budget source quality with A/B/C/D verification levels;
• applies manual curated overrides from input/budget_overrides.csv;
• handles combined budgets by allocating them proportionally by capacity;
• computes CAPEX, OPEX, CRF, annual production, and LCOE;
• classifies projects against validated historic projects;
• estimates a new project's budget with validated nearest neighbours;
• reports budget range, confidence score, confidence reasons, and outlier type;
• compares ML models against a median baseline.

Project Structure

.
├── input/
│   ├── european_offshore_wind_capex.csv
│   ├── wind_dataset.csv
│   └── budget_overrides.csv
├── examples/
│   └── new_project_500mw.json
├── output/
├── src/
│   ├── budget/
│   └── lcoe/
├── tests/
├── wind_budget.py
├── wind_farm_classifier.py
├── model_training.py
├── windAlgorithm.ipynb
├── METHODOLOGY.md
├── pyproject.toml
└── uv.lock

Setup

uv sync --group dev
cp .env.example .env

Set a Serper API key in .env if you want to run web budget enrichment:

SERPER_API_KEY=your_serper_api_key_here

The analysis and classifier can run from the existing cleaned dataset without rerunning web search.

Data

The main dataset is:

input/wind_dataset.csv

Current budget quality fields include:

• total_budget_raw - raw selected budget text;
• original_budget_raw - original dataset budget before enrichment/selection;
• total_budget_EUR_2026 - normalized budget in EUR 2026;
• budget_EUR_per_MW - normalized budget per installed MW;
• budget_quality_status - valid, missing, or rejected;
• budget_source_type - declared_csv, internet_lookup, manual_override, or none;
• budget_verification_level - A, B, C, or D;
• budget_confidence - numeric confidence score.

Verification levels:

• A: official developer, regulator, government, EIB, or equivalent primary source;
• B: specialist industry source or curated declared dataset value;
• C: generic web lookup that passed plausibility checks but requires manual review;
• D: rejected or missing.

LCOE, budget estimation, and ML training use only A/B budget rows.

Manual budget corrections are stored in:

input/budget_overrides.csv

Schema:

wind_farm_name,budget_raw,budget_year,source_url,verification_level,notes

Budget Enrichment

Run web budget enrichment:

uv run python wind_budget.py --input input/wind_dataset.csv --output-dir output --workers 5

The script writes:

output/wind_dataset.csv
output/wind_dataset_budget_lookup_2026EUR.json

Do not blindly replace the input dataset with the output. Web lookup can find sector-wide or unrelated values. The quality layer prefers manual overrides and declared project-level budgets over weaker web lookup values.

Classify A New Project

Run the included 500 MW example:

uv run python wind_farm_classifier.py \
  --dataset input/wind_dataset.csv \
  --json examples/new_project_500mw.json

The classifier estimates budget from validated similar projects and writes:

output/classification_result.json
output/classification_preview.html
output/classification_history/
output/history/

The report includes:

• point budget estimate;
• budget range;
• confidence label and reasons;
• CAPEX/OPEX/LCOE metrics;
• outlier type;
• nearest historical project;
• ML model comparison and prediction.

ML Training

Run model comparison:

uv run python model_training.py --dataset input/wind_dataset.csv

The training report compares:

• median baseline;
• Ridge regression;
• Random Forest;
• Histogram Gradient Boosting.

Target:

log(budget_EUR_per_MW)

Metrics:

• holdout MAE in EUR/MW;
• MAPE;
• R2 on log target;
• cross-validation MAE on log target;
• errors by country and foundation;
• feature importance where available.

The report is saved to:

output/model_training_report.json

Notebook

uv run --with jupyter jupyter lab

Open:

windAlgorithm.ipynb

Tests

uv run pytest

The test suite covers:

• budget parsing and currency conversion;
• blocked social sources;
• source verification ranking;
• combined budget allocation;
• protection against bad web lookup replacing good declared budgets;
• validated-reference budget estimation;
• outlier classification;
• classifier CLI smoke test.

Methodology

See:

METHODOLOGY.md

It documents data sources, budget normalization, source verification, combined-budget handling, CAPEX/OPEX/LCOE formulas, ML validation, confidence scoring, and limitations.

Current Status

The cleaned dataset currently contains:

182 projects
127 A/B verified projects usable for LCOE and ML

The project is suitable as a strong educational decision-support prototype. It is not a bankable investment model; project-level financial decisions still require source-by-source manual review.