Hybrid neural network models for rainfall runoffs: Comparative study

Ojugo AA^1*, Abere R¹, Eboka AO², Yerokun MO², Yoro RE³, Onochie CC² and Oyemade D²

Research Article | Published October 2013
Advancement in Scientific and Engineering Research, Vol. 1(2), pp. 22-34


¹Department of Mathematics and Computer Science, Federal University of Petroleum Resources Effurun, Delta State, Nigeria

²Department of Computer Science, Federal College of Education Technical, Asaba, Nigeria.

³Delta State Polytechnic Ogwashi Uku, Delta State, Nigeria.

*Corresponding author. E-mail: ojugo_arnold@yahoo.com.

……………………………………………………………………………………………......................................

View Full Text - PDF

The study aims to compare hybrid Artificial Neural Networks that have been employed as appropriate models in the prediction of rainfall with historical data as explanatory variables. The selected area is the Chad River Basin in Nigeria with collected data set over the period from 1996 to 2007. A structured analysis of the existing system (SIMHYD) shows bottleneck such as excessive data requirement, large computational demand, and the fact that validation is still an on-going process. The 12-years historical data each on rainfall, relative humidity, cloud cover, temperature difference, and sunshine from the National Metrological Center Oshodi in Nigeria, with lag plot of the cross correlation values of rainfall with each of the other variables to choose explanatory variables that exerts significant influence on rainfall. The dataset was split into three dataset for the developed ANN model into these percentages: training (50%), Cross Validation (25%) and Testing (25%). This was used to validate results obtained, which shows significant correlation of rainfall as established for relative humidity, cloud cover and temperature difference (that were used as explanatory variables in this study). The Tansig activation function was adopted for the three explanatory variables and two parameter weights for each variable identified as the most appropriate model for modeling and predicting rainfall in Chad. Various ANN hybrids have been successful in their implementation, showing high degree of accuracy with many practical implications to water resource operations as well as provide lead time warning in flood management. Results show computed COE as 58, 24, 56 and 42% respectively for the various stations. Observed annual rainfall variations from long-term runoff, is an effect of variation cycle with significant correlation between rainfall and runoff (as indicative in the dataset used). The study implementation will create a synergy between Artificial Intelligence and other fields, which in this case, hydrology via the hybrid ANN models, so that the trained system will help simulate future flood and provide, lead time warning in flood management.

Keywords: Catchment, stochastic, algorithms, evolutionary, fitness function.

Citation: Ojugo AA, Abere R, Eboka AO, Yerokun MO, Yoro RE, Onochie CC and Oyemade D (2013). Hybrid neural network models for rainfall runoffs: Comparative study. Adv. Sci. Eng. Res. 1(2): 22-34
……………………………………………........................................................................................................…..