Addressing effective real-time forecasting inflows to dams through predictive uncertainty estimate

Abstract

The effect of flood waves entering into artificial reservoirs has to be carefully considered for properly addressing dam management to warrant water resources utilization and to provide flood control within a context of safety for the upstream and downstream territories. The optimization of the dam management has to be mainly based on the forecast of the water volume entering the reservoir and of the planning releases from the bottom outlets and the spillways. The flood forecasting models are able to provide discharge predictions at sections of particular interest, i.e. upstream of artificial reservoirs, with forecast horizons appropriate to support the civil protection activities for the mitigation of flooding effects. In 2019, Nanda et al. developed five deterministic forecasts of inflows to the Indian dam of Hirakud, in the Mahanadi River basin, by using the semi-distributed Variable Infiltration Capacity (VIC) model and four error correction models for updating the streamflow predictions by VIC. These models only provide deterministic forecasts for the future events and do not deal with the uncertainty on decisions. With the aim to shed light on the benefits for appropriately using probabilistic forecasts and uncertainty estimate, the deterministic forecasts provided in Nanda et al. (2019) are here used to feed the Model Conditional Processor (MCP) that allows to estimate the predictive uncertainty through the analytical treatment of multivariate probability densities, enabling a decision based on multiple forecasts of different models at the same time.The study is performed considering the monsoon periods of the years 2008–2011: the results achieved for the calibration and validation analysis show that MCP application involving a larger number of deterministic models, up to five, improves the probabilistic forecast accuracy in terms of the expected value accuracy and 90% uncertainty band width. Moreover, when only the two best deterministic models are considered as inputs for MCP, the results are similar to the ones achieved by using all the available deterministic forecasts.The results indicate that MCP is able to provide effective probabilistic real-time inflow forecasting into the reservoir to be used as an appropriate support for the artificial reservoir management during significant flood events.