Modelling of downward longwave radiation using artificial neural networks

Abstract

The assessment of the energy balance at Earth's surface is often complicated by uncertainties in evaluating downward longwave radiation (Rld), which is essential for estimating evapotranspiration, weather forecasting, and taking advantage of radiative cooling. Empirical models based on air temperature (Ta) and relative humidity (RH) measurements represent a common alternative to observation, even if their accuracy is inherently site-specific and often degrades when applied to locations different from their original development site. Artificial neural networks could represent an alternative tool for evaluating R ld . For this purpose, an ANN was developed and trained to predict Rld under clear and cloudy conditions, using Ta, RH, global horizontal irradiance (GHI), direct normal irradiance (DNI) and fractional cloud cover (CFC) as input variables. The R ld predicted by the ANN was then compared with experimental observations and R ld calculated using empirical models. The ANN achieved good accuracy, with R 2 above 0.8 under clear-sky conditions, outperforming classical empirical approaches. In addition, the ANN was trained to predict CFC, achieving R2 values close to 0.7. One practical application of the ANN model is calculating sky temperature, a key variable in exploiting radiative passive cooling of building roofs and façades without the need for thermal or compression chillers.