Vol.9, No 3, 2011 pp. 473 - 480
UDC 551.573:004.032.26:551=111
DOI: 10.2298/FUACE1103473T
ESTIMATING HOURLY REFERENCE EVAPOTRANSPIRATION FROM LIMITED WEATHER DATA BY SEQUENTIALLY ADAPTIVE RBF NETWORK
Slaviša Trajković
Faculty of Civil Engineering and Architecture, University of Niš, Serbia
E-mail: slavisa.trajkovic@gaf.ni.ac.rs
This study investigates the utility of adaptive Radial Basis Function (RBF) networks for estimating hourly grass reference evapotranspiration (ET0) from limited weather data. Nineteen days of micrometeorological and lysimeter data collected at half-hour intervals during 1962-63 and 1966-67 in the Campbell Tract research site in Davis, California were used in this study. Ten randomly chosen days (234 patterns) were selected for the RBF networks training. Two sequentially adaptive RBF networks with different number of inputs (ANNTR and ANNTHR) and two Penman-Monteith equations with different canopy resistance values (PM42 and PM70) were tested against hourly lysimeter data from remaining nine days (200 patterns). The ANNTR requires only two parameters (air temperature and net radiation) as inputs. Air temperature, humidity, net radiation and soil heat flux were used as inputs in the ANNTHR. PM equations use air temperature, humidity, wind speed, net radiation and soil heat flux density as inputs. The results reveal that ANNTR and PM42 were generally the best in estimating hourly ET0. The ANNTHR performed less well, but the results were acceptable for estimating ET0. These results are of significant practical use because the RBF network with air temperature and net radiation as inputs could be used to estimate hourly ET0 at Davis, California.
Key words: evapotranspiration, neural networks, temperature, net radiation, lysimeters.
PRORAČUN ČASOVNE REFERENTNE EVAPOTRANSPIRACIJE IZ MINIMALNOG BROJA KLIMATSKIH PODATAKA KORIŠĆENJEM SEKVENCIJALNE ADAPTIVNE RBF MREŽE
Ova studija istražuje pogodnost korišćenja adaptivnih RBF mreža za proračun časovnih referentnih evapotranspiracija iz minimalnog broja klimatskih podataka. Polučasovni podaci iz devetnaest dana sakupljanih 1962-63 and 1966-67 u Campbell Tract research site u Davis-u, California su korišćeni u ovoj studiji. Deset slučajno izabranih dana (234 podataka) su korišćeni za trening RBF mreža. Dve sekvencijalne adaptivne RBF mreže sa različitim brojem ulaza (ANNTR and ANNTHR) I dve Penman-Monteith jednačine sa različitim vrednostima površinskog otpora (PM42 and PM70) su testirane uporedjivanjem sa lizimetarskim podacima sa preostalih devet dana (200 podataka). ANNTR zahteva samo dva parametra (temperatura vazduha i neto radijacija) kao ulaze. Temperatura vazduha, vlažnost vazduha, neto radiacija i zemljišni toplotni fluks su korišćeni kao ulazi za ANNTHR. Rezultati pokazuju da su ANNTR i PM42 najbolji u proračunu časovnih ET0. Ovi rezultati su od velikog praktičnog značajazato što RBF mreža sa temperaturom vazduha i neto radijacijom može da se koristi za proračun časovne ET0.
Ključne reči: evapotranspiracija, neuronske mreže, temperatura vazduha, neto radijacija, lizimetri