thermal-comfort-omp 1.1.5

Functions to calculate thermal comfort indices

thermal-comfort-omp

This version is compiled with OpenMPI support. For single core processing use thermal-comfort instead.

You can set the environment variable OMP_NUM_CORES=<number of cores> to control how many cores are used by the package. Using too many cores can lead to degraded performance.

Installation

via pypi

pip install thermal-comfort-omp

[!NOTE] For this to work, you will have to have git and gfortran installed

For every release, pre-compiled ABI-3 wheels are provided under releases

Run the tests

On version 3.12 for example...

tox -e py312

Documentation

Docs can be found here: https://rubclim.github.io/thermal-comfort/.

Quick start

The thermal-comfort package provides a limited set of commonly used functions. Which work for scalar values, but are mainly optimized for large array calculation of hundreds of thousands of values.

scalars

from thermal_comfort import utci_approx

utci_approx(ta=20.3, tmrt=50.9, v=2.7, rh=50.5)

arrays

1-dimensional arrays

import numpy as np
from thermal_comfort import utci_approx

utci_approx(
    ta=np.array([20.3, 28.5]),
    tmrt=np.array([50.9, 70.3]),
    v=np.array([2.7, 1.9]),
    rh=np.array([50.5, 70.3]),
)

n-dimensional arrays

The functions only accept 1-dimensional arrays, multi dimensional arrays must be reshaped before and after.

import numpy as np
from thermal_comfort import utci_approx

# 2D arrays e.g. a raster
ta = np.array([[20.3, 28.5], [20.3, 28.5]])
tmrt = np.array([[50.9, 70.3], [50.9, 70.3]])
v = np.array([[2.7, 1.9], [2.7, 1.9]])
rh = np.array([[50.5, 70.3], [50.5, 70.3]])
# retrieve the initial shape
orig_shape = ta.shape

# reshape the array to be 1-dimensional
ta = np.ravel(ta)
tmrt = np.ravel(tmrt)
v = np.ravel(v)
rh = np.ravel(rh)

# calculate the UTCI along the 1-dimensional array
utci = utci_approx( ta=ta, tmrt=tmrt, v=v, rh=rh)

# restore the original shape
utci = utci.reshape(orig_shape)

API

For a complete documentation look at the docs

Mean Radiant Temperature (MRT)

Calculate the mean radiant temperature based on DIN EN ISO 7726.

mean_radiant_temp(ta, tg, v, d = 0.15, e = 0.95)

• ta: air temperature in °C
• tg: black globe temperature in °C

UTCI

Calculate the Universal Thermal Climate Index (UTCI)

utci_approx(ta, tmrt, v, rh)

• ta: Air temperature in °C
• tmrt: Mean radiant temperature in °C
• v: Wind speed in m/s
• rh: Relative humidity in %

PET

Calculate the Physiological Equivalent Temperature (PET).

pet_static(ta, tmrt, v, rh, p)

• ta: air temperature in °C
• rh: relative humidity in %
• v: wind speed in m/s
• tmrt: mean radiant temperature in °C
• p: atmospheric pressure in hPa

Heat Index

Calculate the heat index following Steadman R.G (1979) & Rothfusz L.P (1990).

heat_index(ta, rh)

• ta: air temperature in °C
• rh: relative humidity in %

Extended Heat Index

Calculate the heat index following Steadman R.G (1979) & Rothfusz L.P (1990), but extends the range following The National Weather Service Weather Predicion Center.

heat_index_extended(ta, rh)

• ta: air temperature in °C
• rh: relative humidity in %

Wet Bulb Temperature (TWB)

Calculate the wet bulb temperature following the Stull (2011) equation

wet_bulb_temp(ta, rh)

• ta: air temperature in °C
• rh: relative humidity in %

Saturation Vapor Pressure

Over water

sat_vap_press_water(ta)

• ta: air temperature in °C

Over ice

sat_vap_press_ice(ta)

• ta: air temperature in °C

Dew Point Temperature

dew_point(ta, rh)

• ta: air temperature in °C
• rh: relative humidity in %

Absolute Humidity

absolute_humidity(ta, rh)

• ta: air temperature in °C
• rh: relative humidity in %

Specific Humidity

specific_humidity(ta, rh)

• ta: air temperature in °C
• rh: relative humidity in %

Performance

The benchmark was ran using an array of 100,000 values and 4 threads (OMP_NUM_THREADS=4). See the benchmark directory for more details on the benchmarks.

The hardware used is:

• 2x AMD EPYC 7702 64-Core Processor
• ubuntu 22.04

using an array of length 100,000 the following results were found:

comparing to pythermalcomfort

Benchmark	pythermalcomfort (bf9febd)	thermal-comfort	thermal-comfort (unsafe)	thermal-comfort (Open MPI)	thermal-comfort (unsafe & Open MPI)
tmrt scalar	21.0 us	20.3 us: 1.03x faster	2.21 us: 9.50x faster	23.3 us: 1.11x slower	4.38 us: 4.79x faster
tmrt array	11.9 ms	5.77 ms: 2.06x faster	11.3 ms: 1.05x faster	3.86 ms: 3.07x faster	1.29 ms: 9.17x faster
twb scalar	11.2 us	2.43 us: 4.59x faster	1.31 us: 8.52x faster	2.44 us: 4.57x faster	1.27 us: 8.80x faster
twb array	8.86 ms	2.54 ms: 3.49x faster	2.52 ms: 3.52x faster	2.63 ms: 3.37x faster	1.73 ms: 5.13x faster
heat index scalar	34.9 us	2.38 us: 14.67x faster	1.24 us: 28.10x faster	5.06 us: 6.89x faster	3.68 us: 9.47x faster
heat index array	4.06 ms	2.27 ms: 1.79x faster	2.44 ms: 1.66x faster	1.92 ms: 2.11x faster	859 us: 4.72x faster
utci scalar	59.5 us	22.7 us: 2.62x faster	2.23 us: 26.66x faster	26.0 us: 2.29x faster	4.45 us: 13.36x faster
utci array	37.1 ms	5.93 ms: 6.25x faster	11.4 ms: 3.25x faster	3.82 ms: 9.71x faster	1.24 ms: 29.84x faster
pet scalar	6.04 ms	5.86 us: 1031.44x faster	6.73 us: 897.57x faster	8.68 us: 696.14x faster	6.56 us: 922.14x faster
pet array	189 sec	284 ms: 665.14x faster	804 ms: 234.82x faster	117 ms: 1611.19x faster	113 ms: 1669.13x faster
Geometric mean	(ref)	10.00x faster	13.83x faster	10.45x faster	23.64x faster

comparing to umep

Benchmark	umep	umep (njit)	pythermalcomfort (bf9febd)	thermal-comfort	thermal-comfort (unsafe)	thermal-comfort (Open MPI)	thermal-comfort (unsafe & Open MPI)
utci scalar	44.5 us	6.56 us: 6.79x faster	59.5 us: 1.34x slower	22.7 us: 1.96x faster	2.23 us: 19.97x faster	26.0 us: 1.72x faster	4.45 us: 10.00x faster
utci array	6.87 sec	313 ms: 21.95x faster	37.1 ms: 185.23x faster	5.93 ms: 1157.28x faster	11.4 ms: 602.16x faster	3.82 ms: 1798.79x faster	1.24 ms: 5526.68x faster
pet scalar	388 us	3.17 us: 122.51x faster	6.04 ms: 15.56x slower	5.86 us: 66.29x faster	6.73 us: 57.69x faster	8.68 us: 44.74x faster	6.56 us: 59.26x faster
pet array	62.6 sec	1.33 sec: 48.25x faster	189 sec: 3.02x slower	284 ms: 220.43x faster	804 ms: 77.82x faster	117 ms: 533.95x faster	113 ms: 553.15x faster
Geometric mean	(ref)	35.61x faster	2.07x faster	53.17x faster	88.52x faster	51.68x faster	148.51x faster

While njit already gives a huge performance boost, the difference between umep (njit) and thermal-comfort increases for larger arrays e.g. 1,000,000 values as shown here:

Benchmark	umep (njit)	thermal-comfort
utci scalar	6.53 us	22.9 us: 3.51x slower
utci array	3.09 sec	48.5 ms: 63.77x faster
pet scalar	3.15 us	5.84 us: 1.85x slower
pet array	13.3 sec	2.80 sec: 4.75x faster
Geometric mean	(ref)	2.61x faster

[!CAUTION] If you're after the last bit of performance and don't care about input validation, you may use the underscored functions e.g. _utci_approx or _pet_static which fully avoid any computations in python. However, you will have to guarantee that all your arrays have the same length otherwise undefined behavior may happen. For performance reasons this package is not compiled using -fcheck=bounds compiler-flag.

Compilation

You can set the cmake flag -DUSE_OPENMP=1 to compile the package with OpenMP support