Similarity-based error prediction approach for real-time inflow forecasting Mahmood Akbari and Abbas Afshar ABSTRACT Regardless of extensive researches on hydrologic forecasting models, the issue of updating the outputs from forecasting models has remained a main challenge. Most of the existing output updating methods are mainly based on the presence of persistence in the errors. This paper presents an alternative approach to updating the outputs from forecasting models in order to produce more accurate forecast results. The approach uses the concept of the similarity in errors for error prediction. The K nearest neighbor (KNN) algorithm is employed as a similarity-based error prediction model and improvements are made by new data, and two other forms of the KNN are developed in this study. The KNN models are applied for the error prediction of flow forecasting models in two catchments and the updated flows are compared to those of persistence-based methods such as autoregressive (AR) and artificial neural network (ANN) models. The results show that the similarity- based error prediction models can be recognized as an efficient alternative for real-time inflow forecasting, especially where the persistence in the error series of flow forecasting model is relatively low. Mahmood Akbari (corresponding author) Department of Civil Engineering, University of Kashan, Kashan, Iran E-mail: makbari@kashanu.ac.ir Abbas Afshar Department of Civil Engineering, Iran University of Science and Technology, Tehran, Iran Key words | K nearest neighbor (KNN), new data, real-time inflow forecasting, similarity-based error prediction INTRODUCTION Inflow forecasts are a fundamental requirement for mana- ging water resources systems and the successful operation of river-reservoir systems. Increased computer capacity has led to the increased use of hydrological models in forecast- ing. Along with sophisticated models and longer time series, also the data acquisition systems that are used to col- lect real-time hydrologic data have improved and made it possible to use up-to-date information from basins in real- time forecasting. The forecasting models that operate in real-time are often supported by observed inflow or water level at the time of forecasting. This feedback process of assimilating the measured data into the forecasting procedure to improve the performance of a real-time forecasting system is referred to as updating (Refsgaard ). The updating procedures may concentrate on input variables, state variables, model parameters, and output variables depending on the variables modified during the feedback process (World Meteorologi- cal Organization (WMO) ; Refsgaard ). If forecasting interest is limited to only a few variables at some specific locations with a high degree of accuracy and for a considerably long forecast lead-time, a data assim- ilation scheme based on the updating of output variables may be the most suitable approach (Babovic et al. ). The key advantage of an output updating procedure is the simplicity of its application in a totally automated way to any flow forecasting model (hereafter called primary model), without any need to alter its structure and physical meaning, or its operational implementation (Brath et al. ). This method, which is often called error prediction, has been widely used in different fields of real-time fore- casting as well as in inflow forecasting. In this method, an error prediction model is provided in the updating mode to estimate the errors likely to occur in the next 589 © IWA Publishing 2014 Hydrology Research | 45.4-5 | 2014 doi: 10.2166/nh.2013.098 Downloaded from https://iwaponline.com/hr/article-pdf/45/4-5/589/372624/589.pdf by guest on 22 May 2020