Abstract
Identifying significant large‐scale climate indicators has the potential to improve long‐range streamflow forecasts. In this research, we develop streamflow forecasts for Lake Urmia basin, Iran, specifically for inflow into the Boukan and Mahabad reservoirs. In doing so, two types of inflow forecast models are considered: a single site univariate model ignoring the cross correlation between streamflow at different stations, and a multi‐site multivariate forecast model which takes into consideration the cross correlations among stations. Predictor selection is performed through a principal component analysis and an adaptive‐network‐based fuzzy inference system is used to forecast streamflow. Forecast performance is investigated by employing different combinations of large‐scale climatic information and hydrologic data. We found that gridded ocean‐atmospheric circulation variables, including surface precipitation rate and omega (pressure vertical velocity), have the highest correlations (about 0.7) with annual streamflow. In general, multivariate models are able to better preserve the annual cross‐correlations between streamflow at different stations, as expected, without sacrificing forecast skill as compared to the univariate forecast model approach. Additionally, as compared with the baseline feed‐forward artificial neural network‐ and traditional multiple linear regression‐forecast models, the results were approximately the same. This similarity in the forecast performance between the linear and nonlinear models is likely due to the short of data (44‐sample record).