1. Improving Daily and Monthly River Discharge Forecasts using Geostatistical Ensemble Modeling.
- Author
-
Rezaei, Farshid, Ghorbani, Rezvane, and Mahjouri, Najmeh
- Subjects
ARTIFICIAL neural networks ,FORECASTING ,K-means clustering ,KRIGING ,EXTREME value theory ,MOVING average process ,WATERSHEDS - Abstract
In this paper, daily and monthly runoff discharge forecasts are improved by developing an ensemble model based on the Bayesian maximum entropy (BME), which integrates the outcomes of several single-source simulation models. The performance of the developed ensemble model in daily and monthly discharge forecasting is evaluated by comparing its results with those obtained from 12 other ensemble models, including simple averaging (SA), weighted averaging (WA), Bayesian model averaging (BMA), the best model in last time step (BMLS), artificial neural network (ANN), artificial neural network-based on differenced data, error analysis (EA), wavelet-artificial neural network (Wavelet-ANN), multi-model super ensemble (MMSE), modified multi-model super ensemble (MMMSE), Kriging and an ensemble model based on K-means clustering. Six models, namely artificial neural network (ANN), auto-regressive integrated moving average (ARIMA), adaptive neuro-fuzzy inference system (ANFIS), support vector regression (SVR), M5tree, and Wavelet-ANN, are used as single-source forecasting models. These models provide the inputs of the ensemble models. After analyzing different hydrologic variables, the best set of predictors is chosen for the single-source models. Then, the single-source models are calibrated and validated and the ones with the best performances are selected for being used in the ensemble phase. The performance of the models is evaluated by applying them to two different case studies of the Liqvan watershed, a sub-watershed of Lake Urmia catchment in the north-west of Iran, and the Latyan watershed, in the east of Tehran, Iran. The results show that the ensemble models, especially the BME and ANN-based ensemble models, significantly improve both monthly and daily river discharge forecasts. The values of the performance criteria of Nash–Sutcliffe efficiency (NSE) and coefficient of determination (R
2 ) in forecasting the extreme values of monthly runoff discharge, using the best single-source model, are respectively 0.787 and 0.869. The corresponding values for the best ensemble model are 0.933 and 0.962, respectively. For the daily runoff discharge forecast, the values of the mentioned performance criteria for the best single-source model are respectively 0.874 and 0.877, while corresponding values for the best ensemble model are 0.927 and 0.935, respectively. [ABSTRACT FROM AUTHOR]- Published
- 2022
- Full Text
- View/download PDF