Genetic modeling for the optimal forecasting of hydrologic time-series: Application in Nestos River

摘要

River flow forecasting consists one of the most important applications in modern hydrology, especially for the effective hydropower reservoir management. In this paper, an innovative non-linear time-series fitting and forecasting model is proposed, consisting of the following sub-modules: (a) the division of the time-series into generations, by identifying the structural change points, (b) the generation decomposition into linear trend, harmonic component and autoregressive component, consisting of several gene ARMA models, (c) the use of fuzzy methods to determine the relative weight of each gene model, and (d) the time-series expansion for the optimal forecasting. The method was applied to the mean monthly Nestos River discharge data for the 1966-2006 period, recorded at the Greek-Bulgarian border, serving as inflow to the Thissavros Hydropower Reservoir. The selected series was divided into five distinctive generations representing periods of gradual surface runoff reduction. It occurred that mean monthly discharge during the fifth generation was almost halved, compared to the corresponding value of the first generation. Harmonic decomposition produced periods in agreement with the large-scale atmospheric fluctuations, while ARIMA modeling performed on the residuals from a pre-defined gene model-base, produced satisfactory fitting to the measured flow series. Model expansion for the last 2 years (2005-2006) of the time-series illustrated reasonable good approximations. Model forecasts during 2005 followed closely the recorded variability, with MAPE and RMSE statistics indicating increased model accuracy. Overall, it is evident that the model slightly underpredicts Nestos River discharge, while slight overestimation occurs under high river flow conditions (winter 2006). Although current research proved model's distinct capability and advantages in river flow fitting and medium-scale forecast, future improvement could facilitate the use of more elaborative models (ARCH, fuzzy, ANN), in the developed gene model-base.