Short Duration Rainfall Forecasting Modeling through ANNS

Abstract

Present research paper is articulated the application of Artificial Neural Networks (ANNs) in the field of rainfall forecasting. Research shows ability of ANNs for daily rainfall forecasting. Two Different combinations of weather parameters, one day lag and previous moving average week as case I and Case II respectively has been prepared to generate nonlinear relationship. There are single and multi-hidden layer ANNs generated by increasing and decreasing of hidden layer(s) and Processing Element (PE) by trial and error method. Developed models are selected based on mainly two basics criteria least Mean Square Error along with higher Correlation Coefficient and low Value of Akaike Information Criteria(AIC). Different models were developed and tested by using two input dataset. Models were trained and tested using last 30 (1979-2008) years and 5 (2009-13) year of weather parameter respectively. Result showed that multi hidden layer model ANN Model (7-4-1-1) of case II has good Correlation Coefficient (0.93) and least Mean Square Error (0.001) which was selected as best among four models. It clearly revealed that monsoon depends on long term of weather parameter. It unveils that it does not necessary have more number of Processing Element (PE) and more number of hidden layer(s) which always give good result. Sensitivity analysis revealed that wet bulb temperature is most sensible parameter followed by mean temperature, dry bulb temperature, relative humidity, evaporation, rainfall, and wind velocity.

U radu je predstavljena aplikacija Veštačke Neuronske Mreže (ANNs) u oblasti prognoze padavina. Istraživanje pokazuje mogućnost ANNs za dnevnu prognozu padavina. Dve različite kombinacije vremenskih parametara, jednodnevni pomak i prethodna pokretna srednja nedelja kao Slučaj I i Slučaj II, redom, bile su pripremljene da se generiše nelinearna zavisnost. Postoje jedan i više skrivenih slojeva ANNs generisanih povećanjem i smanjenjem skrivenih slojeva i Elementom za obradu (PE) metodom probe i greške. Razvijeni modeli su izabrani na osnovu dva kriterijuma, najmanja srednja kvadratna greška zajedno sa koeficijentom korelacije i kriterijumom donje vrednosti Akaike informacije (AIC). Različiti modeli su razvijeni i testirani upotrebom dva ulazna seta podataka. Modeli su isprobani i testirani upotrebom poslednjih 30 (1979-2008) godina i 5 (2009-13) godina vremenskih parametara, redom. Rezultat je pokazao da je model više skrivenih slojeva ANN Model (7-4-1-1) slučaja II imao dobar koeficijent korelacije (0.93) i srednju kvadratnu grešku (0.001) i bio je izabran kao najbolji od svih modela. Jasno je pokazao da monsun dugoročno zavisi od vremenskih parametara. On otkriva da ne mora da ima veći broj elemenata obrade (PE) i veći broj skrivenih slojeva koji uvek daje dobar rezultat. Analiza osetljivosti pokazuje da je temperature vlažnog termometra najosetljiviji parameter, a za njim sledi srednja temperatura, temperatura suvog termometra, relativna vlažnost, evaporacija, padavine i brzina vetra.