Real-time forecasting of streamflow and water loss/gain in a river system by using a robust multivariate Bayesian regression model

This research presents a model that simultaneously forecasts streamflow one and two days ahead, and water loss/gain in a river reach between two reservoirs one day ahead and for the next two days. The reservoir operator can take into account these real-time predictions and decide whether to increase/decrease the releases from the upstream reservoir in order to compensate the water loss/gain and manage the streamflow entering the downstream reservoir efficiently. The model inputs are the past daily data of climate (maximum and minimum temperature), streamflow, reservoir releases, water loss/gain in the river, and irrigation canal diversions. The model is developed in the form of a multivariate relevance vector machine (MVRVM) that is based on a multivariate Bayesian regression approach. Based on this Bayesian approach, a predictive confidence interval is obtained from the model that captures the uncertainty of both the model and the data. The model is applied to the river system located in the Lower Sevier River Basin near Delta, Utah. The results show that the model learns the input-output patterns with good accuracy. A bootstrap analysis is used to guarantee robustness of the estimated model parameters. Test results demonstrate good performance of predictions and statistics that indicate robust model generalization abilities