Impact of Precipitation and Model Uncertainty on Change in Design-Flood Prediction Under Climate Change Using Process-Based and Deep Learning Models

Over the last five years, the adoption and development of machine learning (ML) based hydrologic models has advanced rapidly, motivated by a step change in predictive performance over process-based models when measured on benchmark datasets. However, such advances have become increasingly incremental, are often discussed outside the context of whether models are fit-for-purpose, and commonly ignore major sources of predictive uncertainty (e.g., precipitation error). This study forwards a novel, model-as-truth experimental design to isolate and quantify the effects of meteorological uncertainty on ML-based hydrologic model error, with a focus on extreme flood events. We first calibrate a process-based HBV model using grid-based estimates of mean areal precipitation (MAP) and observed streamflow for 80 basins in Massachusetts. The calibrated HBV models are then considered the true hydrologic system and are forced with ‘true’ MAP (estimated from the N precipitation grid cells that overlap each basin) to generate ‘true’ streamflow series at each site. We then randomly sample M<N precipitation grid cells and develop noisy estimates of MAP, mimicking the situation of having a limited number of point measurements to estimate areal precipitation. These noisy MAPs are developed for various values of M (i.e., various point measurement densities), and for each sample, we calibrate long-short term memory (LSTM) networks and a HYMOD model to the ‘true’ HBV streamflow. We then simulate the calibrated LSTM and HYMOD models, as well as the true HBV model, with additional samples of noisy MAP. This experimental design results in 1) HBV-based observations of streamflow free from all uncertainties; 2) simulations of HBV streamflow tarnished with precipitation uncertainty only; and 3) LSTM and HYMOD streamflow predictions tarnished with precipitation, structural, and parametric uncertainties. We compare the various estimates of streamflow to the HBV-observed streamflow and contrast the magnitude of error arising from precipitation uncertainty to the performance improvements of LSTM over HYMOD, focusing on the largest flood events. This study will help establish an upper bound for the limits of predictability in ML-based hydrologic models due to precipitation uncertainty and contextualize ML-based improvements against the residual uncertainty in predicted hydrologic extremes.