Rainfall-runoff modeling for real-time ecological reserve implementation.

Abstract

Real-time ecological reserve implementation systems are essential for preventing deterioration, or facilitating the restoration of riverine ecosystems. These conditions are a result of modified flow regimes that alter essential facets of flow, and which are in turn due to water resource development projects and other activities undertaken along river courses. The essential facets of flow for ecological reserve implementation are presently seen as the low or base flows, the small increases in flow, referred to as freshes, and the small to medium floods. Large floods, which may not be impounded, are not considered because they are unmanageable. The objective therefore is to bring back certain facets of the original natural flow regime that are essential for the proper functioning of the riverine ecosystem. These facets rely strongly on the natural variability of river flow. To accomplish this objective, there is need of effecting optimized reservoir releases that mimic natural variability of flow, in this way to satisfy both ecosystem needs as well as human needs for water. Therefore defining an approach that will provide trigger information to release the required flow becomes imperative. The scope of this project is confined to the study of high flows or flood events in a sub-catchment of the Thukela River in an attempt to build a model that can be used to forecast an impending flood event and hence trigger appropriate releases. The models are for making a 1 day forecast of flow from rainfall at two ii i rain gauge stations i.e. Heartsease and Monks Cowl stations. The study employs 3 methods: regression analysis, a simple empirical model and Artificial Neural Networks (ANN) to build the models. The flood events were selected from stream flow records of a stream gauge at Instream Flow Requirement (IFR) station V1H010 in the little Thukela River. The results obtained from the study show that the ANN performed better than the other two models which yielded unsatisfactory results for prediction of flood events. ANN analysis produced a coefficient of determination of 0.60 with a correlation coefficient of 0.78. Results obtained from regression analysis were 0.49 as coefficient of determination and 0.70 for correlation coefficient. Analysis results from the empirical model showed the worst performance of the three models with a coefficient of determination of 0.36 and a correlation coefficient of 0.60. The results bring forth the need to further analyze data using more powerful models in order to achieve better results than those obtained from ANN. The analysis also indicates that the recommended ecological reserve implementation from a reserve determination study of the selected catchment cannot even be met by the natural series and are therefore most likely invalid.