Wavelet and cuckoo search-support vector machine conjugation for drought forecasting using Standardized Precipitation Index (case study: Urmia Lake, Iran) Mehdi Komasi, Soroush Sharghi and Hamid R. Safavi ABSTRACT In this study, wavelet-support vector machine (WSVM) is proposed for drought forecasting using the Standardized Precipitation Index (SPI). In this way, the SPI time series of Urmia Lake watershed is decomposed to multiple frequency time series by wavelet transform. Then, these time sub-series are applied as input data to the support vector machine (SVM) model to forecast drought. Also, a cuckoo search (CS)-based approach is proposed for parameter optimization of SVM, finding the best initial constant parameters of the SVM algorithm. The obtained results indicate that the radial basis function (RBF)-kernel function of the SVM algorithm has high efficiency in the SPI modeling, resulting in a determination coefficient (DC) of 0.865 in verification step. In the WSVM model, the Coif1, which is considered as a mother wavelet function with decomposition level of five, shows a better performance with DC of 0.954 in verification step, revealing that the proposed hybrid WSVM model outperforms the single SVM model in forecasting SPI time series. Also, DC of cuckoo search-support vector machine (CS-SVM) is calculated to be 0.912 in verification step, indicating the fact that the proposed CS-SVM model shows better efficiency than single SVM model. Mehdi Komasi (corresponding author) Faculty of Engineering, University of Ayatollah Ozma Borujerdi, Borujerd, Iran E-mail: komasi@abru.ac.ir Soroush Sharghi Water Resources Management Engineering, University of Tehran, Tehran, Iran Hamid R. Safavi Department of Civil Engineering, Isfahan University of Technology, Isfahan, Iran Key words | cuckoo search, drought forecasting, SPI, SVM, Urmia Lake watershed, wavelet transform INTRODUCTION Unlike other natural menaces, droughts have a slow evol- ution time such that the outcomes of droughts take a considerable amount of time to come into effect. As a result, the ability to forecast and model the characteristics of drought, especially their initiation, frequency, and sever- ity is important in order to manage water resources for agricultural and industrial uses. The conventional method to monitor drought conditions is a drought index. Some drought indices, such as the Palmer Drought Severity Index (PDSI) and the Standardized Precipitation Index (SPI), are more commonly used than others. A major advantage of the SPI is that it makes the description of droughts on multiple time scales possible (Cacciamani et al. ). One of the differences between the PDSI and SPI is that the PDSI index has a complex structure with a very long memory whereas the SPI is an easily inter- preted and simple-moving average process (Tsakiris & Vangelis ). Furthermore, unlike the PDSI, the charac- teristics of SPI are constant site to site and the calculations of which only take precipitation data (Belay- neh & Adamowski ; Blagojevic et al. ). As a result, SPI is used as an appropriate drought index in this study. The SPI needs machine learning tools to be forecasted. Today, artificial intelligence (AI) models such as the artifi- cial neural network (ANN) and adaptive network-based fuzzy inference system (ANFIS) have been used in several studies to forecast the hydrological, geological, and 975 © IWA Publishing 2018 Journal of Hydroinformatics | 20.4 | 2018 doi: 10.2166/hydro.2018.115 Downloaded from https://iwaponline.com/jh/article-pdf/20/4/975/245342/jh0200975.pdf by guest on 16 June 2020