The increasing consumption of surface and groundwater resources in Iran coincided with global climate change has led to increasing pressure on the existing water resources. Therefore, without comprehensive attention to the existing limitations and challenges, sustainable planning and management of water resources will not be possible in Iran, which is located in the arid and semi-arid regions of the world. Lack of appropriate frameworks and laws, inefficient operation of the available water resources, excessive consumption of groundwater resources, and the intensified pollution are some of serious challenges with respect to water resources management in Iran. Therefore, it is important to regenerate and update the existing frameworks in order to provide the necessary conditions for optimal operation of the limited resources.
Generally, any permanent change in the climate patterns, whether regionally or globally, is called "climate change". Throughout the life of the planet, climate has been changing continuously. However, in recent decades, there have been far greater and more significant changes than what have ever been recorded before. The recent research works show the crucial impact of climate change on trends in climatic and hydrological variables (Ehsanzadeh and Adamowski, 2010). In general, extensive changes in precipitation and temperature patterns are known to be the driving forces of these changes. In other words, increased temperature leads to intensified evaporation from moist surfaces and this in turn leads to more humidity in the atmosphere. Increased humidity in the atmosphere leads to intensified hydrologic cycle which causes massive floods in some areas and extensive droughts in other regions. In addition, reviewing previous numerous research works indicate that changes in the melting history of snow and its inclination in winter lead to a considerable change in flow patterns of rivers in different parts of the world. One of the most important impacts of this phenomenon on water resources in Iran and some other regions of the world is the reduction of precipitation in the form of snow and the shift of snowmelt timing towards the cold seasons. This leads to reduction of river flow in the spring and summer and therefore water inflow into the dam reservoirs are inconsistent with the values considered at the time of design. Efficient and reliable planning of water resources is impossible without considering the effects and consequences of climate change phenomenon. In other words, the occurrence of such changes in the future leads to the problematic management of water resources which is based on the assumption of stationary hydrology (Middle Cope et al., 2001). In near future, this issue is one of the most important challenges for planning and management of water resources, and therefore operation policies should be updated and modified according to new conditions (Steele-Dunne et al., 2008).
In recent years, numerous studies have been performed in the field of climate change and its effects on water resources systems and hydrological variables. The results of these researches indicate the gradual changes of climatic conditions for different regions of the world, including Iran (Dudangeh et al., 2010; Pirnia et al., 2015). The climatic instabilities lead to uncertain determination of design parameters, which is one of the main challenges in water resources engineering (Xu et al., 1997).
For solving reservoir operation problems, different uncertain and unstable variables should be considered including water inflow into the reservoir and downstream water demand under future conditions. Using realistic values or at least appropriate estimated values of these variables taking into account climate instability impact on these variables can lead to a significant improvement of the reservoir operation performance (Loucks et al., 1981). For optimal operation of reservoirs, therefore, it is necessary to consider the impact of the mentioned uncertainties and possible instabilities (Soleimani et al., 2016).
One of the simplest and most practical methods to study climate variations and changes in hydrological data is to study the trends of hydrological and climatic variables. From a statistical and climatic point of view, the values of decreasing or increasing trends for variables such as precipitation, runoff, and evaporation can be attributed to increasing trend of average temperature at the Earth's surface.
Statistical tests are commonly used to determine hydrological trends. In general, statistical tests are divided into two categories of parametric and non-parametric tests. Non-parametric tests are superior to parametric tests because no information on the probability distribution of data is required (Rood et al., 2005). The most common non-parametric tests for trend analysis are Mann-Kendall and Spearman tests, which have been widely used in recent years to study trend in climatic variables (Burn & Elnur, 2002; ZarehAbyaneh et al., 2011; Pirnia et al., 2015; Azarakhshi et al., 2013; Asakereh and Doostkamian, 2014).
One of the limitations of classical statistical trend tests is the assumption of data independence and lack of storage capacity (Ehsanzadeh and Adamowski, 2010). If a short-term or long-term correlation is observed among the data, the mentioned assumptions are problematic. Since serial dependencies or short-term persistence (STP) are observed in most hydrological and climatic time series, statistical tests should be revised to account for this phenomenon (Ehsanzadeh and Adamowski, 2010). In 1951, Hurst observed long-term serial correlation among some time series, known as Hurst phenomenon (Hurst, 1951). This phenomenon which shows long term correlation among time series can affect the results of trend detection studies (Fathian et al., 2014). The intensity of the Hurst phenomenon or long-term persistence (LTP) for time series is measured using the Hurst exponent (H). If long-term persistence is detected in climate and hydrological data, its impact on statistical analysis of data should be considered.
For most hydrological designs, the homogeneity of climate data is very important. However, it is often observed that climate data are heterogeneous (Buishaund, 1982). In many studies, in addition to analysis of trends, the homogeneity of time series have also been investigated (Modaresi et al., 2010). Generally, heterogeneity in hydrological data can be caused by various factors including climate change. The heterogeneity of time series affects the results of a meaningful trend detection of a specific variable. Various tests such as statistical tests of Cumulative Deviations, Worsley's Likelihood Ratio Test, and Bayesian inference have been proposed and used to investigate the existence of heterogeneity in the time series.
A wide range of different factors affects river flows at different times of the year including climatic variables, physical, and geometric characteristics of the upstream basin condition, and the amount of upstream water allocation. The degree of extensiveness and variability of these factors leads to the uncertain discharges of rivers. Climate change due to increased greenhouse gases, on the other hand, enhances the mentioned uncertainty
Reviewing past researches in the field of reservoir operation shows that the amount of water allocated to a user or released from a reservoir is determined using the measured inflow to the reservoir in previous years. This management practice leads to unfavorable results due to changes in inflows over time. In other words, one of the most important parameters for water resources management is predicting the true amounts of inflows and outflows in the future. For this purpose, various methods have been proposed that can be generally classified into two categories: data-driven methods and conceptual methods. In data-driven methods, the modeler relies only on runoff time series and maybe precipitation and temperature to predict inflow. Conceptual methods, however, are based on the accurate comprehension of the physical mechanisms governing the hydrological processes of the basin. Therefore, conceptual methods require a wide range of hydrological and meteorological data of the basin for modeling and predicting the flow. Less dependence of data-driven methods on the volume of input data, as well as less complexity of these methods, has led to greater popularity of these models (Lima et al., 2014).
Data-driven models are more commonly used in recent years due to some specific features of these models (Takeuchi and Sivaarthitkul, 1995; Faber and Stedinger, 2001; Chiew et al., 2003; Maurer and Letnmeier, 2004; Dong et al., 2005; George Jacques and Graham, 2008; Madsen et al., 2009; Tang et al., 2010; Ahmadi et al. 2014; Zhao et al., 2015 and Arena et al., 2015). One of the most usual data-driven models is the artificial neural network (ANN) model. The ANN models are applied for solving various engineering problems including water resources engineering problems. So far, a wide range of different combinations of input variables, network structures and frameworks, and various training algorithms have been used in ANN modeling. ANN models are powerful computational methods that are capable of identifying and finding complex relationships between hydrological and climatic variables, leading to relatively accurate predictions. It is worth noting that modeling hydrological processes is difficult but extremely important. The complex and dynamic nature of hydrological processes makes the use of intelligent and nonlinear models such as ANN models inevitable. Generally, ANN models are divided into two general categories: static and dynamic or recursive networks (Gómez-Ramos and Wongas-Martinez, 2013). Reviewing the researches shows that ANN models, especially dynamic ANNs are widely used in different fields of water engineering problems (Daniell, 1991; French et al., 1992; Hall and Minns, 1993; Hsu et al., 1995; Hamlet et al., 2002; Lin and Chen, 2005; Cigizoglu, 2008; Hung et al., 2009; Shen and Chang, 2013; Ruslanet al., 2013; Rani and Parekh, 2014; Banihabib et al., 2014; Yin et al., 2014; Zhe et al., 2015).
Although accurate flow prediction improves the efficiency of water resources management policies, so far no comprehensive approach has been proposed for this purpose. One of the most important challenges of finding a comprehensive approach is the low efficiency of current predicting methods and the complexity of finding connections between uncertain predictions and decision variables (Georgakakos and Graham, 2008). However, the results of numerous studies have shown that simultaneous use of predicting water inflow into the reservoir and adaptive reservoir management can not only counteract the potential effects of climate change, but it also leads to a significant improvement of the efficiency of current reservoir operation methods (Yao and Georgakakos, 2001; Georgakakos and Graham, 2008; Raje and Mujumdar, 2010;Zhao et al., 2015).
Pishgah Hadian et al. (2016) investigated the trends of various climatic parameters such as runoff, precipitation, temperature, and evaporation for Sefidruod Dam reservoir located in northern Iran, in order to investigate the effects of climate change and emphasize on the updating operation policies of the Sefidruod Dam reservoir. The results showed decreasing trends for precipitation and minimum temperature and also increasing trend for maximum temperature. Therefore, predicting the water inflow values of Sefidruod River in the future is necessary for effective management and optimal operation of the Sefidruod Dam reservoir. The purpose of the current study is to investigate trends of water inflow into the Sefidruod Dam reservoir for the last decades. Prediction of water inflow into the Sefidruod Dam reservoir using dynamic neural networks for the coming years is another objective of this study. The parametric regression and non-parametric Mann-Kendall tests with the assumptions of independence, short-term and long-term persistence are used to determine trends. The homogeneity of the data used is studied using the statistical tests such as cumulative deviations, Worsley's Likelihood Ratio Test, and Bayesian inference. Two dynamic ANN models of NAR and NARX will be used to predict water inflow into the reservoir.