PyWeatherEnriched

Real geocoding + real historical weather, wired into a Rust core with pandas/numpy climate feature engineering on top.

What it is

Given a location name and a timestamp, PyWeatherEnriched geocodes the location (OpenStreetMap Nominatim) and looks up the real, genuinely observed historical weather for that place and hour (Open-Meteo's free Archive API) — no formula-generated or fabricated numbers. The lookup/geocoding core is Rust (via PyO3) for speed and a small memory footprint; a Python layer on top adds pandas/numpy feature engineering (rolling aggregates, heating/cooling degree-days, cyclical time encoding, anomaly z-scores) useful for feeding weather into an ML pipeline.

Features

• Real historical weather. Open-Meteo Archive API — observed data, not forecasts or formulas.
• Real forward geocoding. OpenStreetMap Nominatim turns a location name into coordinates for you.
• Historical backfill. enrich_range(location, start_date, end_date) geocodes once and fetches every hourly observation across the whole range in a single Open-Meteo request.
• pandas/numpy feature engineering via standalone functions or the build_features convenience pipeline: degree-days (hdd/cdd), cyclical hour/day-of-week/day-of-year encodings, trailing rolling aggregates, and anomaly z-scores.
• Two-tier caching. A built-in in-memory LRU cache, plus an opt-in SQLite-backed EnhancedCache with geospatial-proximity matching, TTL expiration, batch deduplication, and date-range queries.
• Honest failure handling. Rows whose lookup fails get NaN rather than raising or fabricating a value; hours with no observation are omitted, not filled.

Who it's for

• ML/data engineers enriching a training dataset with real historical weather for a location/timestamp pair, instead of a synthetic or averaged fill-in.
• Teams building weather-sensitive models that need ready-made climate features (rolling aggregates, degree-days, cyclical encodings, anomaly z-scores).
• Anyone who wants Open-Meteo data without writing the geocoding and feature math themselves.

What it is not

• Not real-time or forecast weather. This is historical, observed data only.
• Not a source of hyperlocal data beyond what Open-Meteo provides. The package layers geocoding and feature engineering on top of the same public archive; it does not produce new weather measurements.
• Some Rust building blocks are not exposed to Python. Elevation/lapse-rate adjustment (real SRTM GeoTIFF parsing), urban-heat-island modeling (real OSM building-density analysis), and OSM-based reverse geocoding are independently unit-tested but not yet wired into the Python API. They live under src/geospatial/.
• Some layers are deliberately unimplemented. Vegetation/NDVI, soil, and flood-risk layers, plus additional commercial reverse-geocoding provider backends, are framework stubs (src/geospatial/optional.rs) that return a clear "not yet implemented" error rather than fake data.
• Not fully covered by cargo test. Plain cargo test (without --lib) currently fails to compile because a stale integration test references source files not declared as mods in src/lib.rs. Use cargo test --lib.

Project status

Beta. Version 0.7.0. The Python-facing geocoding, historical weather, caching, and feature-engineering paths are real and tested against live network data; the extra Rust building blocks above are implemented but not yet exposed. Changes are tracked in CHANGELOG.md; see also ROADMAP_HONEST.md.

Installation

pip install pyweatherenriched

Requires Python 3.10+. CI builds prebuilt wheels for Linux (x86_64/ARM64), macOS (Intel/Apple Silicon), and Windows (x86_64) on every release, but the CI→PyPI publish step is currently broken (PyPI rejects the upload with 403 Invalid or non-existent authentication information — a stale/invalid API token), so only a macOS ARM64 abi3 wheel plus an sdist is actually published on PyPI today. On any other platform, pip install falls back to building the sdist from source, which does need a Rust toolchain (see Requirements below).

Quick start

import pyweatherenriched as pwe

enricher = pwe.WeatherEnricher()
result = enricher.enrich_row("New York", "2024-06-15T12:00:00")

print(result)
# {'location': 'New York', 'latitude': 40.7127281, 'longitude': -74.0060152,
#  'temperature': 19.4, 'humidity': 74.0, 'condition': 'Clear',
#  'timestamp': '2024-06-15T12:00:00'}

enrich_row makes two real network calls the first time it sees a location/timestamp pair (one to Nominatim to geocode location, one to Open-Meteo for the historical weather at those coordinates); the in-process cache means repeating the same lookup doesn't hit the network again:

fresh = pwe.WeatherEnricher()
fresh.enrich_row("New York", "2024-06-15T12:00:00")  # network
fresh.enrich_row("New York", "2024-06-15T12:00:00")  # cache hit
print(fresh.cache_stats())  # {'hits': 1, 'misses': 1, 'size': 1}

Usage

Historical backfill

enrich_row/enrich_batch fetch one day of data per call. To backfill a date range for a location, enrich_range geocodes once and fetches every hourly observation across the whole range in a single Open-Meteo request:

enricher = pwe.WeatherEnricher()
rows = enricher.enrich_range("Chicago", "2024-06-01", "2024-06-07")
print(len(rows))  # ~168 (7 days x 24 hours, minus any hours Open-Meteo has no data for)

Hours Open-Meteo has no observation for (e.g. the tail of a range that runs past the latest available data) are omitted from the result rather than filled with a fabricated value.

Feature engineering on a DataFrame

enrich_dataframe bridges the Rust core to pandas — fetch real weather for every row of a DataFrame in one call — and the features functions build standard climate ML features on top of the result:

import pandas as pd
import pyweatherenriched as pwe

orders = pd.DataFrame({
    "order_id": ["A1", "A2", "A3"],
    "location": ["Chicago", "Miami", "Denver"],
    "timestamp": ["2024-01-15T12:00:00", "2024-06-15T12:00:00", "2024-03-15T12:00:00"],
})

enricher = pwe.WeatherEnricher()
enriched = pwe.enrich_dataframe(enricher, orders)     # + latitude/longitude/temperature/humidity/condition
features = pwe.build_features(enriched)                # + hdd/cdd, cyclical time, rolling stats, anomaly z-score

print(features[["order_id", "temperature", "hdd", "cdd", "temperature_zscore"]])

Rows whose lookup fails (unresolvable location, upstream error) get NaN in the new columns instead of raising or fabricating a value, so one bad row doesn't lose the batch.

build_features is a convenience pipeline; each step is also a standalone function:

Caching

WeatherEnricher has a small built-in LRU cache. For larger workloads, EnhancedCache adds a second, SQLite-backed persistent tier with geospatial-proximity matching, TTL expiration, batch deduplication, and date-range queries:

from pyweatherenriched import EnhancedCache

cache = EnhancedCache(cache_size=5000, db_path="weather_cache.db")
cache.set_proximity_radius(10.0)  # treat lookups within 10km as cache hits
cache.set_ttl(72)                 # hours before an entry expires

cache.put("New York", 40.7128, -74.0060, 15.2, 65.0, "Partly Cloudy", "2024-01-15T12:00:00Z")
result = cache.get("New York", 40.7128, -74.0060, "2024-01-15T12:00:00Z")

# Deduplicate a batch before making any API calls.
batch = [
    ("New York", 40.7128, -74.0060, "2024-01-15T12:00:00Z"),
    ("New York", 40.7128, -74.0060, "2024-01-15T12:00:00Z"),  # duplicate
]
missing_indices, cache_hits = cache.deduplicate_batch(batch)

# Every observation in a date range near a point (requires db_path).
rows = cache.get_range(40.7128, -74.0060, "2024-01-01T00:00:00Z", "2024-01-31T23:59:59Z")

stats = cache.stats()
print(f"hit ratio: {stats['hit_ratio']:.1%}")

get() checks the memory tier, then the persistent tier (if db_path was given), then falls back to a proximity search; get_range only queries the persistent tier, so it always returns [] without db_path.

vs raw Open-Meteo

PyWeatherEnriched is built directly on Open-Meteo's free Archive API, so the honest comparison isn't "who has better weather data" (same real source) — it's "what do you get for the extra network round-trip and processing time." Tested live against real data for 3 real location/timestamp pairs (Chicago/Jan, Miami/Jun, Denver/Mar 2024, noon local):

Correctness was verified by hand against real output: hdd/cdd matched max(0, 18-t)/max(0, t-18) exactly for all 3 real temperatures, and hour_sin/hour_cos matched sin/cos(2π·hour/24) exactly. Returned temperature/humidity also matched the raw Open-Meteo baseline for the same coordinates/timestamps (rounding only).

Requirements

• Python 3.10+
• Rust 1.75+ (only if building from source — not needed to pip install on a platform with a published wheel)
• Network access to nominatim.openstreetmap.org and archive-api.open-meteo.com at call time (no API key needed for either)

Documentation

• BUILD.md — wheel-building/release details.
• ROADMAP_HONEST.md — what's real, what's deferred, and why.
• CHANGELOG.md — release history.
• docs/ — additional documentation.

Contributing

git clone https://github.com/Mullassery/PyWeatherEnriched.git
cd PyWeatherEnriched

maturin develop --release   # build the Rust extension + install editable
cargo test --lib            # Rust unit tests
pip install -e ".[dev]"
pytest tests/ -v            # Python tests, incl. live-network ones

cargo run --bin pyweatherenriched-validate is a small smoke-test CLI that exercises real geocoding + weather fetch end to end — useful to confirm a build/environment can actually reach Nominatim and Open-Meteo.

Related projects

Other tools from the same family, for adjacent parts of the data/ML stack:

• PySynthData — synthetic data generation for ML training: realistic datasets for model development and testing.
• StatGuardian — fast, declarative data quality framework (Rust): schema validation, drift detection, anomaly detection.
• PyReverseETL — reverse ETL runtime: move data into CRMs and marketing tools with lineage and compliance enforcement.
• PyDBTGuard — analyzes a dbt project's manifest.json to flag risky tests, map blast radius, and estimate test cost.
• PyAirflowTester — dependency intelligence and quality assurance for Airflow and dbt pipelines.

Browse all projects →

License

This project is licensed under the Apache License 2.0.