Evaluation of deep-learning and tree-boosting machine learning models in automatic error correction of forecasts from a physics-based model: A case study on Storå river, Denmark

Accurate real-time flood predictions play a vital role in flood early warning systems, which further helps in mitigating the damage and saving lives. Error correction using machine learning (ML) in physics-based models (alternatively known as physicallybased models) has been widely considered and recommended in the literature to improve forecast accuracy. This study mainly focuses on evaluating the ability of novel tree-based ML methods and Bidirectional LSTM (BLSTM) at different lead times and high flow conditions. Also, the performance of these methods is compared with the traditionally used autoregression (AR), Multilayer perceptron (MLP), and naïve models. So overall, we evaluated six data-driven models and one naïve model on Storå river to correct the errors in the physics-based model: Two tree-boosting ML models (XGBoost, Gradient boosting), two deep learning-based models (MLP, BLSTM), and then simple models like autoregression (AR) & persistence (or naïve). Then, a stacked model combining XGBoost, and AR is developed and tested. Hyperparameter tuning is performed using Bayesian optimization. Results on the independent test set show that all the methods can improve the discharge simulations from a physics-based model. However, the Bidirectional LSTM and stacked model are consistently performed slightly better than other models in all lead times. At shorter lead times, tree-boosting approaches marginally underperformed. While gradient boosting performed better at longer lead times and produced results comparable to BLSTM and stacked models, XGBoost continues to underperform but gave better results than AR and PERS & MLP. The BLSTM and stacked models performed well under high flow conditions as well. Even though the difference is minor, they consistently outperformed all the other models. Furthermore, while tree-based methods (XGBoost & gradient boosting) fared somewhat worse than BLSTM & stacked model, they outperformed basic methods (AR/Pers) and MLP at high flow conditions. One additional key finding in this study is that even when the stacked model was built using less computationally intensive methods (XGBoost & AR), it produced equivalent results to BLSTM.