a b s t r a c tUrmia lake basin located in northwestern Iran is the second largest saline lake in the world. Due to many reasons i.e. climate changes, several dam constructions, building a bridge across the Lake, extra agricultural consumption and improper management of water resources, the water level of the lake has been decreased since 1997 and thousand hectares of emerged salty land has made numerous ecological and environmental problems. Therefore, an accurate forecast of the entrance runoff to the lake is important in managing the river flow and water transfer within basins. There are various methods for time-series based forecasting; in the presented study Feed-forward Neural Network and Autocorrelation Regressive Integrated Moving Average (ARIMA) models were applied to forecast the monthly rainfall in Urmia lake basin. The results showed that the estimated values of monthly rainfall through Feed-forward NN were close to ARIMA model with coefficient of correlation 0.62 and the root mean square error of 12.43 mm over the 6 years test period; then rainfall amount were predicted for a 6-year period starting from

The present study aimed to develop a hybrid model to predict the rainfall time series of Urmia Lake watershed. For this purpose, a model based on discrete wavelet transform, ARIMAX and least squares support vector machine (LSSVM) (W-S-LSSVM) was developed. The proposed model was designed to handle linear, nonlinear and seasonality of rainfall time series. In the proposed model, time series were decomposed into sub-series (approximation (a) and details (d)). Next, the sub-series were predicted separately. In the proposed model, sub-series were fed into SARIMAX to be predicted. The residual of predicted sub-series (error) of the rainfall time series was then fed into LSSVM to predict the residual components. Then, all predicted values were aggregated to rebuild the predicted time series. In order to compare results, first a classic modeling was performed by LSSVM. Later, wavelet-based LSSVM was used to capture the peak values of rainfall. Results revealed that Daubechies 4 and decomposition level 4 (db(4,4)) led to the best outcome. Due to the performance of db(4,4), it was selected to be applied in the proposed model. Based on results, it was observed that the W-S-LSSVM's performance was improved in comparison with other models.