A hybrid of ANN and CLA to predict rainfall

Abstract

Accurate prediction of rainfall is extremely important to design of hydrological system and water resources management. In this study, two different techniques, artificial neural networks (ANNs) and Learning-Cellular Automation (CLA) were used to classify rainy days. A wide range of dataset (2249 dataset) was collected from Shiraz synoptic stations including temperature, humidity, wind speed, pressure, and daily rainfall. The results indicated that the CLA with R2 = 0.796 and RMSE = 0.431 provides better classification in comparison with ANN (R2 = 0.566 and RMSE = 0.407). Then, daily rainfall is predicted using a novel hybrid method of ANN and CLA (ANN-CLA) that has the ability to automatically remove non-rainy days. To investigate the performance of proposed method, the results of ANN-CLA were compared with one of the most common neural network method, namely Feed Forward Back Propagation (FFBP) with two learning function of LM (FFBP-LM) and BFGS (FFBP-BFGS). The result indicated that hybrid of ANN-CLA successfully predicts daily rainfall (R2 = 0.881 and RMSE = 0.202) in comparison with FFBP-LM (R2 = 0.839 and RMSE = 0.222) and FFBP-BFGS (R2 = 0.698 and RMSE = 0.246). A sensitivity analysis is performed on the data to determine the effect of input parameters on the daily rainfall. This analysis indicates that the average wind speed is the main parameter for prediction of rainfall. This research highlights that the presented techniques can be successfully employed as a robust method for prediction of problems related to water resources, hydrology, and water sciences.