Stream-flow Prediction in Ergene River Basin via Kalman Filter Emrecan Ozdogan * , Mohsen M. Vanolya † , Levent Ucun * , Seref Naci Engin * * Department of Control and Automation Engineering Yildiz Technical University, Davutpasa, Istanbul Emails: emrecanozdogan95@gmail.com, lucun@yildiz.edu.tr, nengin@yildiz.edu.tr † Department of Civil Engineering Yildiz Technical University, Davutpasa, Istanbul Email: mvanolya1976@gmail.com Abstract—This study deals with the implementation of Kalman filter for the prediction of stream-flow in Ergene River Basin. In the study, stream-flow, precipitation and wastewater are chosen as the state variables during the prediction process since these parameters are highly effective on the stream-flow. Effects of precipitation and wastewater are calculated via Soil and Water Assessment Tool (SWAT) model in the study. Covariance matrices are calculated by using real-time data with 5 year length and model performance is tested with short and long-term predictions based on measurements and the accuracy of the proposed method is evaluated with Nash-Sutcliffe efficiency coefficient(NS) and root mean squared error (RMSE). I. I NTRODUCTION A large number of empirical and analytical models are available for streamflow forecasting that can be classified as short, medium and long-term forecasting models [1], [2]. Linear quadratic estimation (LQE) and Kalman filtering are considered as empirical stochastic models, which combine the dynamics and probability distribution of the measured variables in current state for forecasting future ones [3]. Jens et al. (1985) used Kalman filter for real time operation of surface water flow by forecasting in stochastic space in rainfall-runoff model of Mike 11 hydrodynamic model [4]. They discussed the source of uncertainty and stated that it came from the precipitation that is the input to rainfall-runoff. Ngan (1986) compared autoregressive models with Kalman filter based flow forecasting in his PhD thesis [5]. He showed that Kalman filter had better reliability in flow prediction compared to ARMAX. Jean (2004) used it for groundwater level forecasting as well as rainfall-runoff prediction in Danish Hydraulic Institution (DHI) [6]. Moradkhani et al. (2005), forecasted one-day ahead streamflow of the Leaf River watershed by using a dual state parameter estimation approach based on the Ensemble Kalman Filter (EnKF) and showed that the results are very consistent with the observations [7]. Clark et al. (2008) described an application of the EnKF in which streamflow observations are used to update the states in a distributed hydrological model for extracting the source of uncertainty [8]. In another study similar to their work, Noh et al. (2013) assessed EnKF and particle filter (PF) with another distributed hydrologic model and showed that the Kalman filter model is sensitive for the length of lag time [9]. Rasmussen et al. (2015) assessed the assimilation of groundwater and streamflow data in integrated hydrologic model in the size of ensemble and localization of Kalman filter [10]. They concluded that the required ensemble size depends heavily on the assimilation of discharge observations and estimation of parameters as well as on the number of observed variables. Deng et al. (2016) used ensemble Kalman filter for identification of temporal variation of hydrologic parameters in a monthly water balance model [11]. They used the filter for Wudinghe basin in China and showed the effectiveness of its detection on storage capacity. Mathematical models involved in streamflow prediction to provide more simplistic solutions considering physical ones require comprehensive geographic and measured data. They chose a few of hundreds of variables that affect the streamflow most and dealt with the error caused by linearization and vari- able ignorance. For this purpose, Kalman Filters are used [12]. They achieved promising results. Later, regression models and Artificial Neural Networks are added to the methods with their own approach to the problem and successful predictions [13], [14], [15]. Today, numerous different methods are used to predict streamflow or enhance the ones that are already being used such as Chaos Theory to improve prediction length of Kalman Filter [16]. Another recent addition to this study area is wavelets, by adding periodic knowledge to the model, they increase the accuracy of it [17], [18]. Kalman filter is first proposed by R.E. Kalman [19]. This method takes observation errors and disturbances into account, minimizes the modelling errors and its convergence is guar- anteed. Because of these features, Kalman filter is commonly used in, but not limited with, aircraft position estimation and control systems [20], [21]. Chemical processes are other study areas that prediction accuracy of Kalman filter is frequently exploited [22]. Also, increasing awareness of global warming is attracting more attention every year to prediction and management of water resources [12]. In some cases, Kalman Filter’s accuracy outperforms other prediction methods [23]. SWAT is used in many studies with the help of its wide access to environmental data such as soil moisture, snow cover fraction, streamflow and many more. In its cooperation with Kalman Filter, generally SWAT is the predicting part and Kalman Filter is a tool that prepares inputs to the model by