Ten-Day Streamflow Prediction of the Dry Season Integrated by Seasonal Rainfall and Hydrological Model              

                                    Chun-Chao Kuo

A combined, climate-hydrologic model with three components to predict the streamflow of two river basins of Taiwan at one season (3-month) lead time for the November-January (NDJ) and January-March (JFM) seasons was developed. The first component consists of the wavelet-based, ANN-GA model (Artificial Neural Network calibrated by Genetic Algorithm) which predicts the seasonal rainfall by using selected sea surface temperature (SST) as predictors. For the second component, three disaggregation models, Valencia and Schaake (VS), Lane, and Canonical Random Cascade model (CRCM), were tested to compare the accuracy of seasonal rainfall disaggregated by these three models to 3-day time scale rainfall data. The third component consists of the continuous rainfall-runoff model modified from HBV (called the MHBV) and calibrated by a global optimization algorithm against the observed rainfall and streamflow data of the Shihmen and Tsengwen river basins of Taiwan.

From the scale average wavelet power (SAWP) computed for the seasonal rainfall, it seems that the data exhibit interannual oscillations at 2-4-year period. On the basis of correlation fields between DCR-WPC1 of seasonal rainfall and DCR-SAWP of SST of Pacific Ocean at one-season lead time, SST of some domains of the western Pacific Ocean (July-September SST around 5°N-30°N, 120°E-150°E; October-December SST around 0°N-60°N, 125°E-160°W) were selected as predictors to predict seasonal NDJ and JFM rainfall of Taiwan at one season lead time respectively, using an Artificial Neural Network calibrated by Genetic Algorithm (ANN-GA).  The ANN-GA was first calibrated using the 1974-1998 data and independently validated using 1999-2005 data. In terms of summary statistics such as the correlation coefficient, root-mean-square error (RMSE), and Hansen-Kuipers (HK) scores, the prediction of seasonal rainfall of northern and western Taiwan using ANN-GA are generally good for both calibration and validation stages (correlation coefficient ranged from 0.3 to 0.7), but not so for southeastern Taiwan because the seasonal rainfall of the former are much more significantly correlated to the SST of selected sectors of the Pacific Ocean than the latter.

In the part of rainfall disaggregation, the VS and Lane models were tested for disaggregating seasonal rainfall to 10-day and 3-day rainfall. The results revealed no obvious differences between them. In order to be closer to the time scale of the hydrological model, 3-day disaggregated rainfall was used for later application. The disaggregated results of these two models were further compared with the one of CRCM. Overall, the performances of the VS and the Lane models are better than the one of CRCM. The Lane model is chosen for further application after evaluation.

In the aspect of streamflow simulation, the analytical results of the two-stage calibration are better than the ones of the one-stage calibration. The proposed framework was tested, first by disaggregating the predicted seasonal rainfall of ANN-GA to rainfall of 3-day time step using the Lane model; then the disaggregated rainfall data was used to drive the calibrated MHBV to predict the streamflow for both river basins at 3-day time step up to a season’s lead time.  Overall, the correlation of 10-day streamflow prediction is between 0.48 and 0.53. That will be useful for the seasonal planning and management of water resources of these two river basins of Taiwan.

Keywords: Wavelet, sea surface temperature, seasonal rainfall, disaggregation model, conceptual rainfall-runoff model, basin streamflow.