Enhancement of Model Reliability by Integrating Prediction Interval Optimization into Hydrogeological Modeling

This paper presents a single-objective optimization-based perturbation analysis to quantify model prediction uncertainty. A new index named coverage width index (CWI), which combines two commonly used uncertainty indices, the percentage of coverage (POC) and the average width (AW), was proposed to facilitate the optimization. Considering the outperformance of the wavelet neural network (WNN) among various data-driven modeling approaches in hydrogeological modeling, the proposed approach was integrated into WNN (called OPWNN). A case study was conducted to demonstrate the application of OPWNN in groundwater level forecasting at two wells in the Amaravathi River Basin, India. The sensitivity analysis of the effect of initial perturbation range on CWI suggested that uncertainty is sensitive to the selected perturbation range and a small perturbation does not guarantee an acceptable prediction interval (PI). The modeling results demonstrated that the OPWNN can optimize the PI effectively with minimized AW corresponding to an expected high POC. Therefore, this approach can yield more reliable predictions/forecasts for water resources management.