suncalc

A fast, vectorized Python port of suncalc.js


Keywords
suncalc, sun, numpy, sunrise, sunset
License
MIT
Install
pip install suncalc==0.1.0

Documentation

suncalc-py

Test Package version Downloads

A fast, vectorized Python implementation of suncalc.js for calculating sun position and sunlight phases (times for sunrise, sunset, dusk, etc.) for the given location and time.

While other similar libraries exist, I didn't originally encounter any that met my requirements of being openly-licensed, vectorized, and simple to use 1.

Install

pip install suncalc

Using

Example

suncalc is designed to work both with single values and with arrays of values.

First, import the module:

from suncalc import get_position, get_times
from datetime import datetime

There are currently two methods: get_position, to get the sun azimuth and altitude for a given date and position, and get_times, to get sunlight phases for a given date and position.

date = datetime.now()
lon = 20
lat = 45
get_position(date, lon, lat)
# {'azimuth': -0.8619668996997687, 'altitude': 0.5586446727994595}

get_times(date, lon, lat)
# {'solar_noon': Timestamp('2020-11-20 08:47:08.410863770'),
#  'nadir': Timestamp('2020-11-19 20:47:08.410863770'),
#  'sunrise': Timestamp('2020-11-20 03:13:22.645455322'),
#  'sunset': Timestamp('2020-11-20 14:20:54.176272461'),
#  'sunrise_end': Timestamp('2020-11-20 03:15:48.318936035'),
#  'sunset_start': Timestamp('2020-11-20 14:18:28.502791748'),
#  'dawn': Timestamp('2020-11-20 02:50:00.045539551'),
#  'dusk': Timestamp('2020-11-20 14:44:16.776188232'),
#  'nautical_dawn': Timestamp('2020-11-20 02:23:10.019832520'),
#  'nautical_dusk': Timestamp('2020-11-20 15:11:06.801895264'),
#  'night_end': Timestamp('2020-11-20 01:56:36.144269287'),
#  'night': Timestamp('2020-11-20 15:37:40.677458252'),
#  'golden_hour_end': Timestamp('2020-11-20 03:44:46.795967773'),
#  'golden_hour': Timestamp('2020-11-20 13:49:30.025760010')}

These methods also work for arrays of data, and since the implementation is vectorized it's much faster than a for loop in Python.

import pandas as pd

df = pd.DataFrame({
    'date': [date] * 10,
    'lon': [lon] * 10,
    'lat': [lat] * 10
})
pd.DataFrame(get_position(df['date'], df['lon'], df['lat']))
# azimuth	altitude
# 0	-1.485509	-1.048223
# 1	-1.485509	-1.048223
# ...

pd.DataFrame(get_times(df['date'], df['lon'], df['lat']))['solar_noon']
# 0   2020-11-20 08:47:08.410863872+00:00
# 1   2020-11-20 08:47:08.410863872+00:00
# ...
# Name: solar_noon, dtype: datetime64[ns, UTC]

If you want to join this data back to your DataFrame, you can use pd.concat:

times = pd.DataFrame(get_times(df['date'], df['lon'], df['lat']))
pd.concat([df, times], axis=1)

API

get_position

Calculate sun position (azimuth and altitude) for a given date and latitude/longitude

  • date (datetime or a pandas series of datetimes): date and time to find sun position of. Datetime must be in UTC.
  • lng (float or numpy array of float): longitude to find sun position of
  • lat (float or numpy array of float): latitude to find sun position of

Returns a dict with two keys: azimuth and altitude. If the input values were singletons, the dict's values will be floats. Otherwise they'll be numpy arrays of floats.

get_times

  • date (datetime or a pandas series of datetimes): date and time to find sunlight phases of. Datetime must be in UTC.

  • lng (float or numpy array of float): longitude to find sunlight phases of

  • lat (float or numpy array of float): latitude to find sunlight phases of

  • height (float or numpy array of float, default 0): observer height in meters

  • times (Iterable[Tuple[float, str, str]]): an iterable defining the angle above the horizon and strings for custom sunlight phases. The default is:

    # (angle, morning name, evening name)
    DEFAULT_TIMES = [
        (-0.833, 'sunrise', 'sunset'),
        (-0.3, 'sunrise_end', 'sunset_start'),
        (-6, 'dawn', 'dusk'),
        (-12, 'nautical_dawn', 'nautical_dusk'),
        (-18, 'night_end', 'night'),
        (6, 'golden_hour_end', 'golden_hour')
    ]

Returns a dict where the keys are solar_noon, nadir, plus any keys passed in the times argument. If the input values were singletons, the dict's values will be of type datetime.datetime (or pd.Timestamp if you have pandas installed, which is a subclass of and therefore compatible with datetime.datetime). Otherwise they'll be pandas DateTime series. The returned times will be in UTC.

Benchmark

This benchmark is to show that the vectorized implementation is nearly 100x faster than a for loop in Python.

First set up a DataFrame with random data. Here I create 100,000 rows.

from suncalc import get_position, get_times
import pandas as pd

def random_dates(start, end, n=10):
    """Create an array of random dates"""
    start_u = start.value//10**9
    end_u = end.value//10**9
    return pd.to_datetime(np.random.randint(start_u, end_u, n), unit='s')

start = pd.to_datetime('2015-01-01')
end = pd.to_datetime('2018-01-01')
dates = random_dates(start, end, n=100_000)

lons = np.random.uniform(low=-179, high=179, size=(100_000,))
lats = np.random.uniform(low=-89, high=89, size=(100_000,))

df = pd.DataFrame({'date': dates, 'lat': lats, 'lon': lons})

Then compute SunCalc.get_position two ways: the first using the vectorized implementation and the second using df.apply, which is equivalent to a for loop. The first is more than 100x faster than the second.

%timeit get_position(df['date'], df['lon'], df['lat'])
# 41.4 ms ± 437 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

%timeit df.apply(lambda row: get_position(row['date'], row['lon'], row['lat']), axis=1)
# 4.89 s ± 184 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Likewise, compute SunCalc.get_times the same two ways: first using the vectorized implementation and the second using df.apply. The first is 2800x faster than the second! Some of the difference here is that under the hood the non-vectorized approach uses pd.to_datetime while the vectorized implementation uses np.astype('datetime64[ns, UTC]'). pd.to_datetime is really slow!!

%timeit get_times(df['date'], df['lon'], df['lat'])
# 55.3 ms ± 1.91 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

%time df.apply(lambda row: get_times(row['date'], row['lon'], row['lat']), axis=1)
# CPU times: user 2min 33s, sys: 288 ms, total: 2min 34s
# Wall time: 2min 34s

1: pyorbital looks great but is GPL3-licensed; pysolar is also GPL3-licensed; pyEphem is LGPL3-licensed. suncalcPy is another port of suncalc.js, and is MIT-licensed, but doesn't use Numpy and thus isn't vectorized. I recently discovered sunpy and astropy, both of which probably would've worked but I didn't see them at first and they look quite complex for this simple task...