An application of neural network technique to correct the dome temperature effects on pyrgeometer measurements

Abstract

This work describes an application of a multilayer perceptron neural network technique to correct dome emission effects on longwave atmospheric radiation measurements carried out using an Eppley Precision Infrared Radiometer (PIR) pyrgeometer. It is shown that approximately 7-month-long measurements of dome and case temperatures and meteorological variables available in regular surface stations (global solar radiation, air temperature, and air relative humidity) are enough to train the neural network algorithm and correct the observed longwave radiation for dome temperature effects in surface stations with climates similar to that of the city of São Paulo, Brazil. The network was trained using data from 15 October 2003 to 7 January 2004 and verified using data, not present during the network-training period, from 8 January to 30 April 2004. The longwave radiation values generated by the neural network technique were very similar to the values obtained by Fairall et al., assumed here as the reference approach to correct dome emission effects in PIR pyrgeometers. Compared to the empirical approach the neural network technique is less limited to sensor type and time of day (allows nighttime corrections).