Uncertainty Quantification of Machine Learning Models to Improve Streamflow Prediction in Changing Climate and Environmental Conditions
Abstract
Machine learning (ML) models, Long Short-Term Memory (LSTM) networks in particular, have demonstrated remarkable performance in streamflow prediction and show an increasing application in hydrological community. However, most of these applications do not provide uncertainty quantification (UQ). ML models are data driven which may suffer from large extrapolation errors when applied in changing climate/environmental conditions and new geographic regions. UQ is required to ensure models' trustworthiness, improve predictive understanding of data limits and model deficiencies, and avoid overconfident predictions in extrapolation. In this study, we propose a novel PI3NN method to quantify the prediction uncertainty of ML models and integrate it with LSTM networks for streamflow predictions. PI3NN calculates Prediction Intervals by training 3 Neural Networks and uses root-finding methods to determine the interval bounds precisely. Additionally, PI3NN can accurately identify the out-of-distribution (OOD) data in a nonstationary condition to avoid the overconfident prediction. We apply the proposed PI3NN-LSTM method in both the snow-dominant East River Watershed in west region and the rain-driven Walker Branch Watershed in southeast region of US. Results indicate that for the prediction data which have similar features as the training, PI3NN precisely quantifies the prediction uncertainty with the desired confidence level; and for the OOD data where the LSTM network fails to make accurate predictions, PI3NN can produce a reasonably large uncertainty bound indicating the untrustable result to avoid overconfidence. PI3NN is computationally efficient, reliable in training and generalizable to various network structures and data with no distributional assumptions. It can be broadly applied in ML-based hydrological simulation for credible predictions.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H32A..03L