Forecasting daily streamflow using hybrid ANN models

In this paper, the classic ‘divide and conquer (DAC)’ paradigm is applied as a top-down black-box technique for the forecasting of daily streamflows from the streamflow records alone, i.e. without employing exogenous variables of the runoff generating process such as rainfall. To this end, three forms of hybrid artificial neural networks (ANNs) are used as univariate time series models, namely, the threshold-based ANN (TANN), the cluster-based ANN (CANN), and the periodic ANN (PANN). For the purpose of comparison of forecasting efficiency, the normal multi-layer perceptron form of ANN (MLP–ANN) is selected as the baseline ANN model. Having first applied the MLP–ANN models without any data-grouping procedure, the influence of various data preprocessing procedures on the MLP–ANN model forecasting performance is then investigated. The preprocessing procedures considered are: standardization, log-transformation, rescaling, deseasonalization, and combinations of these. In the context of the single streamflow series considered, deseasonalization without rescaling was found to be the most effective preprocessing procedure. Some discussions are presented (i) on data preprocessing and (ii) on selection of the best ANN model. Overall, among the three variations of hybrid ANNs tested, the PANN model performed best. Compared with the MLP–ANN fitted to the deseasonalized data, the PANN based on the soft seasonal partitioning performed better for short lead times (≤3 days), but the advantage vanishes for longer lead times.