Ensemble quantile-based deep learning framework for streamflow and flood prediction in Australian catchments

TL;DR

This study tackles flood forecasting in Australia by marrying ensemble learning with quantile regression in a deep learning framework. It uses the CAMELS dataset to build multi-step, multi-catchment streamflow predictions, incorporating both temporal and static catchment features to enable regional modelling. The proposed Ensemble Quantile-LSTM delivers uncertainty-aware predictions by employing $q=0.05$ and $q=0.95$ quantiles alongside a mean estimator, and floods are assessed via a frequency-based threshold $\gamma$ derived from exceedance probabilities $\alpha = m/(n+1)$. Findings indicate that per-catchment models yield the best extreme-value forecasts, with the ensemble approach providing robust flood probability estimates suitable for early warning systems; code and data areOpen Source to extend to other regions.

Abstract

In recent years, climate extremes such as floods have created significant environmental and economic hazards for Australia. Deep learning methods have been promising for predicting extreme climate events; however, large flooding events present a critical challenge due to factors such as model calibration and missing data. We present an ensemble quantile-based deep learning framework that addresses large-scale streamflow forecasts using quantile regression for uncertainty projections in prediction. We evaluate selected univariate and multivariate deep learning models and catchment strategies. Furthermore, we implement a multistep time-series prediction model using the CAMELS dataset for selected catchments across Australia. The ensemble model employs a set of quantile deep learning models for streamflow determined by historical streamflow data. We utilise the streamflow prediction and obtain flood probability using flood frequency analysis and compare it with historical flooding events for selected catchments. Our results demonstrate notable efficacy and uncertainties in streamflow forecasts with varied catchment properties. Our flood probability estimates show good accuracy in capturing the historical floods from the selected catchments. This underscores the potential for our deep learning framework to revolutionise flood forecasting across diverse regions and be implemented as an early warning system.

Figures (9)

• Figure 1: Location of the Camels catchments across Australia across the different states.
• Figure 2: Framework for comparing the different models for the different architectures (Stage 2 and 3) after acquiring and processing data (Stage 1). Stage 4 features our ensemble quantile deep learning model that predicts 2 extreme quantiles ($5^{th}$ and $95^{th}$ quantiles) along with regular streamflow prediction. We then utilise the model from Stage 4 to selected catchments across Australian states in Stage 5 and show results using the respective metrics (RMSE, NSE and SER). Finally, in Stage 6, we provide a hydrograph visualisation of flood probability and streamflow prediction for selected catchments.
• Figure 3: Embedding: the transformation of a two-dimensional temporal dataset of features (daily time interval) into a three-dimensional format through a sliding window technique (windowing).
• Figure 4: Configurations
• Figure 5: Ensemble model utilising Quantile-LSTM for prediction of flood probability and streamflow with uncertainty bounds using quantile regression.