Get Weather Data

Historical daily weather for any US ZIP code or latitude/longitude. NOAA station data (GHCN Daily and GSOD), automatic nearest-station selection, consistent metric or imperial output.

No-code app

Prefer clicking to coding? A Streamlit app gives the same data through a web UI — enter a ZIP or lat/lon, pick dates, get a table, chart, and CSV download. Run it with streamlit run app/streamlit_app.py or deploy it free on Streamlit Community Cloud.

Installation

pip install get-weather-data

Or with uv:

uv pip install get-weather-data

Quick Start

Python API

from get_weather_data import Weather

# Initialize and set up database (downloads ~60MB first time)
weather = Weather()
weather.setup()

# By ZIP code...
result = weather.get("10001", "2024-01-15")
print(f"Max temp: {result.tmax} °C")  # e.g. -1.6
print(f"Precip:  {result.prcp} mm")  # 0.0 means zero, None means no data
print(f"Station: {result.station_name} ({result.station_distance_meters} m away)")

# ...or by coordinates
result = weather.get((40.7484, -73.9967), "2024-01-15")

# Imperial units if you want them
weather_f = Weather(units="imperial")
result = weather_f.get("10001", "2024-01-15")
print(f"Max temp: {result.tmax} °F")

Command Line

# Set up database (first time only)
get-weather setup

# By ZIP or by coordinates; metric by default
get-weather get 10001 2024-01-15
get-weather get "40.75,-73.99" 2024-01-15 --units imperial

# Process a CSV file
get-weather process input.csv output.csv

Online Mode (no setup)

Skip the station-database build by querying NOAA's Climate Data Online API directly (only a small ZIP-coordinates file, a few MB, is cached on first use). Get a free token at https://www.ncdc.noaa.gov/cdo-web/token and set NCDC_TOKEN:

from get_weather_data import Weather

weather = Weather(online=True)  # requires NCDC_TOKEN
result = weather.get("10001", "2024-01-15")

NCDC_TOKEN=your-token get-weather get 10001 2024-01-15 --online

Notes on online mode:

• Tokens are limited to 5 requests/second and 10,000 requests/day, so process_csv (batch jobs) requires the local database.
• Same result contract as the local path: nearest reporting station first, same units, real station distances. Online covers GHCN stations only (no GSOD fallback).

Features

• ZIP or lat/lon everywhere: get(), get_range(), CSV batch, and the CLI all take either a ZIP code or coordinates
• Consistent units: real metric values (°C, mm, m/s) across the API, CLI, and CSV output — or units="imperial" (°F, in, mph)
• pandas-ready: Weather.get_frame(...) returns a tidy DataFrame (pip install get-weather-data[pandas])
• 13 variables: temperatures, precipitation, snow, wind + gust, dew point, pressures, and visibility
• Quality-controlled: values that failed NOAA's QC are dropped by default; include_flags=True surfaces the per-field QC flag
• Present weather: include_weather_types=True reports the day's phenomena (fog, thunder, hail, freezing rain, ...) from GHCN WT** codes and GSOD's FRSHTT indicator
• Explainable gaps: explain=True tells you why a value is missing (stations_considered + a per-field missing reason) — a blank cell becomes diagnosable, not silent
• Automatic station selection: nearest station first, farther stations fill in missing variables; coverage(...) reports how well a point is served
• Interpolation: Weather(interpolate=True) inverse-distance-weights the nearest stations instead of taking the single closest one
• Universal CONUS coverage: source="grid" (or "auto") pulls NOAA's authoritative 5-km nClimGrid daily grid, so any Lower-48 lat/long returns real temperature + precipitation — no station gaps (pip install get-weather-data[grid])
• Three data sources: GHCN Daily (~93K US/CA/MX stations), GSOD (~9K airport stations), and the nClimGrid gridded product
• Robust batch processing: streams CSVs in chunks, one bad row gets a weather_error note instead of killing the job
• Parquet + SQL: stream batch output to columnar Parquet and query it (or a glob of many files) with query_weather(...) via DuckDB (pip install get-weather-data[parquet])
• Cache management: TTL-based refresh of station lists, get-weather cache info / cache clear

Usage Examples

Get Weather for a Date Range

from datetime import date
from get_weather_data import Weather

weather = Weather()

results = weather.get_range(
    "90210",
    start_date=date(2024, 7, 1),
    end_date=date(2024, 7, 7),
)

for r in results:
    if r.tmax is not None:
        print(f"{r.date}: High {r.tmax:.0f}°C")

Get a pandas DataFrame

from get_weather_data import Weather

weather = Weather()  # needs: pip install get-weather-data[pandas]

df = weather.get_frame("90210", "2024-07-01", "2024-07-07")
# one row per day; columns = metadata + tmax/tmin/prcp/... in your units
df[["date", "tmax", "prcp"]].head()

Process a CSV File

from get_weather_data import Weather

weather = Weather()

# ZIP-based input (zip, year, month, day columns)
weather.process_csv("locations.csv", "with_weather.csv")

# Coordinate-based input
weather.process_csv(
    "points.csv",
    "with_weather.csv",
    lat_column="lat",
    lon_column="lon",
    date_column="date",
)

Output rows carry the weather columns below (already in your chosen units), plus weather_error explaining any row that could not be resolved. More examples in the examples/ directory.

Parquet output and SQL queries

For the analytical "millions of rows" workflow, write results as Parquet (typed, compressed columns — far smaller than CSV and directly queryable) and run SQL over them with DuckDB. Needs the parquet extra (pip install get-weather-data[parquet]):

from get_weather_data import Weather, query_weather

# Streaming Parquet output (inferred from the .parquet suffix; no pandas
# needed, memory stays bounded over millions of rows):
Weather().process_csv("locations.csv", "weather.parquet", date_column="date")

# Query one file or a glob of many with SQL:
rows = query_weather(
    "SELECT station_id, avg(tmax) AS mean_high "
    "FROM t WHERE prcp > 0 GROUP BY station_id ORDER BY mean_high DESC",
    tables={"t": "weather.parquet"},  # or "runs/*.parquet"
)

get-weather process in.csv out.parquet (or --format parquet) does the same on the CLI. The daily and hourly DataFrames also write Parquet directly: get_frame(...).to_parquet(path) / get_hourly_frame(...).to_parquet(path).

Interpolate between stations

For a point between stations, interpolate=True blends the nearest stations by inverse-distance weighting rather than returning the single closest station's raw value:

weather = Weather(interpolate=True)
result = weather.get((40.75, -73.99), "2024-01-15")
# result.station_type == "interpolated"

Check coverage before trusting the numbers

cov = weather.coverage("59718", "2024-01-01", "2024-12-31")
print(cov.station_name, cov.station_distance_meters, "m away")
print(f"tmax present on {cov.fraction('tmax'):.0%} of days")

Hourly data

For sub-daily detail, get_hourly(...) returns one HourlyResult per hour from NOAA's ISD-Lite (nearest airport/USAF-WBAN station). Timestamps are UTC and the local station database is required (setup()); there is no online equivalent.

weather = Weather(units="imperial")

hours = weather.get_hourly("11371", "2023-07-16")  # LaGuardia, one UTC day
for h in hours[:3]:
    print(h.observed_at, h.temp, "F", h.wind_speed, "mph", h.wind_direction, "deg")

# pandas: one row per hour (needs the [pandas] extra)
df = weather.get_hourly_frame("11371", "2023-07-16", "2023-07-17")

CLI: get-weather hourly 11371 2023-07-16 [--end 2023-07-17] [--units imperial]. Fields: temp, dewpoint, sea_level_pressure, wind_direction (degrees), wind_speed, sky_condition, precip_1h, precip_6h.

Weather Variables

All values are floats in the units below (or their imperial equivalents); None/empty means the station network had no reading — a genuine zero is reported as 0.0.

With include_flags=True, result.flags maps each field to its GHCN quality-control flag (blank = passed all checks; GHCN stations only).

Present-weather types

The numeric variables above say how much rain fell but not whether it was thunder, freezing rain, or fog. With include_weather_types=True, result.weather_types is a set of the day's phenomena, drawn from GHCN WT** occurrence codes (fog, thunder, hail, glaze, blowing snow, ...) and, for GSOD airport stations, the FRSHTT indicator (fog / rain / snow / hail / thunder / tornado):

w = Weather(online=True, include_weather_types=True)
r = w.get("11430", "2023-07-16")  # JFK on a stormy afternoon
print(sorted(r.weather_types))  # ['fog', 'heavy_fog', 'thunder']

get-weather get <loc> <date> --weather-types shows them on the CLI, and process --weather-types adds a comma-joined weather_types column to the batch CSV.

Why is a value missing?

A blank tmax could mean many things: no station nearby, a station that measures only precipitation, or a gap on that date. With explain=True, the result carries stations_considered and a missing map giving a reason for each requested field that came back empty:

w = Weather(online=True, explain=True)
r = w.get("59221", "2023-01-15")  # remote NE Montana
print(r.stations_considered)  # e.g. 20
print(r.missing.get("tmax"))
# 'none of the 20 CDO stations near this point reported tmax on 2023-01-15'

get --explain prints the reasons on the CLI, process --explain adds stations_considered and missing columns to the batch CSV, and get_frame(...) includes them when present. (For a data consumer, this turns a silent NaN into a diagnosable gap.)

Which backend?

from get_weather_data import Weather  # pip install get-weather-data[grid]

# Any Lower-48 coordinate, guaranteed — no local database needed
weather = Weather(source="grid")
result = weather.get((44.06, -121.31), "2024-01-15")  # remote Oregon
print(result.tmax, result.station_type)  # -> value, "gridded"

nClimGrid caveats: contiguous US only (no Alaska/Hawaii/PR), maximum, minimum, and average temperature plus precipitation only, and a ~2-3 day latency for the most recent days. Its daily values follow NOAA's "24-hour period ending in the early morning of the specified day" convention, which can differ slightly from a single station's calendar-day value.

Data Sources

This package uses data from NOAA's National Centers for Environmental Information:

• GHCN Daily: Global Historical Climatology Network daily summaries (~93K stations across the US, Canada, and Mexico — border ZIPs get the truly nearest station)
• GSOD: Global Summary of the Day from USAF/WBAN airport stations (~9K); GSOD reports no snowfall, so snow comes from GHCN stations
• nClimGrid-Daily: NOAA's authoritative 5-km gridded daily temperature and precipitation for the contiguous US, 1951-present, interpolated from GHCN-Daily (used by source="grid"/"auto")
• GeoNames: ZIP code to coordinates mapping

Database Setup and Disk Use

setup() downloads the station lists and ZIP coordinates (~60MB) and builds a nearest-stations index; it takes a few minutes, once. Station lists refresh automatically when older than 30 days.

Weather data itself is fetched lazily per year: each GHCN year you touch builds a local SQLite file (roughly 1–3 GB for recent years); GSOD adds one small CSV per station-year. Historical years never re-download; the current and previous year refresh monthly. Inspect or reclaim space anytime:

get-weather cache info
get-weather cache clear --ghcn        # or --gsod / --stations / --all

weather = Weather()
weather.setup()

info = weather.info()
print(f"GHCN stations: {info['ghcn_stations']:,}")
print(f"USAF stations: {info['usaf_stations']:,}")
print(f"ZIP codes: {info['zipcodes']:,}")

Configuration

Custom database location:

weather = Weather(database_path="/path/to/my.db")

Or via CLI:

get-weather --database /path/to/my.db setup

Upgrading from v3

v4 is a breaking release:

• Values are now real metric floats (°C, mm, m/s) everywhere — previously the Python API returned raw GHCN tenths. Divide-by-10 code should be removed.
• Run get-weather setup --force (or weather.setup(force=True)) after upgrading: v4 fixes a nearest-station ranking bug, so indexes built by v3 contain wrong distances.
• WeatherResult gained tobs, latitude, longitude, and units; it is now importable from the package root.

License

MIT License. See LICENSE for details.

Authors

• Suriyan Laohaprapanon
• Gaurav Sood