daily discharge estimation in talar river using lazy learning model

introduction: river discharge as one of the most important hydrology factors has a vital role in physical, ecological, social and economic processes. so, accurate and reliable prediction and estimation of river discharge have been widely considered by many researchers in different fields such as surface water management, design of hydraulic structures, flood control and ecological studies in spetialand temporal scale. therefore, in last decades different techniques for short-term and long-term estimation of hourly, daily, monthly and annual discharge have been developed for many years. however, short-term estimation models are less sophisticated and more accurate.various global and local algorithms have been widely used to estimate hydrologic variables. the current study effort to use lazy learning approach to evaluate the adequacy of input data in order to follow the variation of discharge and also simulate next-day discharge in talar river in kasilianbasinwhere is located in north of iran with an area of 66.75 km2. lazy learning is a local linear modelling approach in which generalization beyond the training data is delayed until a query is made to the system, as opposed to in eager learning, where the system tries to generalize the training data before receiving queries materials and methods: the current study was conducted in kasilian basin, where is located in north of iran with an area of 66.75 km2. the main river of this basin joins to talar river near valicbon village and then exit from the watershed. hydrometric station located near valicbon village is equipped with parshall flume and limnogragh which can record river discharge of about 20 cubic meters per second.in this study, daily data of discharge recorded in valicbon station related to 2002 to 2012 was used to estimate the discharge of 19 september 2012. the mean annual discharge of considered river was also calculated by using available data about 0.441 cubic meters per second. to estimate the discharge of considered day, three methods of constant, linear and quadratic functionscontrollers based on the local linearization provided by the lazy learning algorithm were considered. lazy learning is a memory-based linear technique for local modeling approach which is reported as a high-efficient algorithm for simulating variables with low input data.the series of input data was categorized into previous 6, 8, 10, 15 and 20 days, 1 and 2 months, 1, 2 and 3 seasons and also 1 and 2 years to evaluate which series is appropriately enough to predict next-day discharge inthe river. then, mean absolute error and root-mean square error were calculated for all series and modelsin order to find the best estimator model and the most appropriate series of input data. results: results showed that constant and linear model had the minimum root-mean square error of 0.001 and 0.057 respectivelywith previous 60 days’ data series. whilethe quadratic model had its best estimation with previous 2 season data series with the minimum root-mean square error of 0.059. the result indicated that the more input data increase, the best quadratic model estimate until 60 days. but after 60 days, estimation error gradually increased. consequently, not more data but adequate areneeded for accurate estimation. also, rmse in linear model had less fluctuation and therefore less sensitivity compared with other models. and quadratic model had less fluctuation and sensitivity to neighborhoods. also, according to results, the more variation in each period increase, the better estimation is accrued by lazy learning algorithm. hence, it was expected that next-day discharge prediction in low-water period needs longer data series than high-water period. conclusion: regarding to thousands of prepared training models, constant model with previous 60 days’ data and minimum error of 0.0001 was selected as the most accurate estimatefor next-day river discharge in talar river. results showed that despite of some limitation and demerits, the local lazy learning algorithm has significant efficiency in time series simulating. although the accuracy of simulation increase with more input data, but this algorithm can runby at least 5 training data. however we find lazy learning to be the best performing approach on average goodness indicators (such as mean absolute error and root-mean square error). on the other hand, the lazy learning predictor can be quickly developed and easily kept up-to-date by adding new data to its database. also, it does not face with overfitting problems which are common in global modeling approaches.according to some noteworthy features of lazy learning noticed in this regards, this approach will have good performance for time-series studies.