Monthly and seasonal streamflow forecasting in the Rio Grande basin

Improving the quality of streamflow forecasting has always been an important task for researchers and water resources managers. In this research, the seasonal and monthly streamflow forecasting using various data-driven statistical models was investigated for naturalized streamflow at Del Norte Gaging Station, Rio Grande, Colorado and observed Elephant Butte Reservoir net inflow, Rio Grande, New Mexico. The application of partial least squares regression (PLSR) and hybrid models in seasonal streamflow forecasting, the inclusion of snowpack and El Nino Southern Oscillation (ENSO) information in the monthly and seasonal streamflow forecasting were investigated. The modeling methods included autoregressive integrated moving average (ARIMA) models, transfer function-noise (TFN) models, artificial neural networks (ANN) models, principal components regression (PCR), and PLSR. Two hybrid modeling approaches, including TFN forecast modification using ANN (TFN+ANN) and a combination of principal components analysis (PCA) and ANN (PCA+ANN), were also applied in seasonal streamflow forecasting. The ARIMA models were used as a benchmark for the comparison of the performance of the models. Additionally, the forecasting results were compared to the Natural Resources Conservation Service (NRCS) official forecasts to evaluate the performance of the proposed models.;The results of seasonal flow modeling indicated that using a composite precipitation index is a relatively effective method in both improving forecast accuracy and developing parsimonious regression models with fewer and readily-available input variables. In comparison of PLSR and PCR, similar forecast accuracies were obtained for both methods in jackknife cross validation and test period (2003-2007) although PLSR has higher calibration coefficient of determination (R2) and can reach its minimum prediction error with a smaller number of components than PCR. The comparison with NRCS official forecasts showed that the application of PLSR in seasonal streamflow forecasting is promising. The application of hybrid modeling approaches showed potential capability of hybrid models to improve forecast accuracy in seasonal streamflow modeling as compared to single models. For Elephant Butte net inflow modeling, the normalized root mean square errors (NRMSE) of forecasted and observed net inflow for April-July decreased from 0.36 to 0.19 from single TFN model to the TFN+ANN hybrid approach. The performance of PCA+ANN approach was also comparable to the TFN+ANN.;The results of monthly flow modeling suggested that the forecast modification using a combination of TFN and ANN methods (TFN+ANN) displayed better performance than the ANN models that were specifically calibrated for each month of the snowmelt season and was able to improve forecast accuracy significantly compared to other models. The normalized root mean square errors (NRMSE) for one-month-ahead forecasts for Del Norte Gaging Station were 0.46, 0.41, 0.24 and 0.21 for simple ARIMA, TFN model, ANN models and TFN+ANN approach respectively. These findings suggested that the TFN+ANN method is an advantageous approach in improving forecast accuracy and the ANN is a useful tool in forecasting monthly streamflow, whether it is used for direct modeling or used as a forecast modification technique.;The findings of this study may provide an impetus for streamflow forecasting by using hybrid modeling approach and PLSR method with various operationally available climatic variables. PLSR approach can be combined into NRCS's operational forecasting environment for possible forecast improvement.