Forecasting river discharge using machine learning methods

Abstract

The objective of this study is find out whether maximum daily discharge of the Geul and Rur catchments can be forecast using machine learning (ML) methods, and if so, to what extent. In addition, these ML models are compared to a conceptual model to see which performs better. A second objective is to test whether soil moisture content (SMC) and NDVI increase performance of the two ML models. The Geul and Rur catchments are both partly situated in the administrative area of Waterschap Limburg, a water authority in the Netherlands. They use discharge forecasts in order to prepare flood defenses and to monitor high water levels more closely. Currently, discharge is forecast using the conceptual HBV model for the Geul. Forecasting is done based on experience for the Rur and only in case of high water levels. Conceptual and physical models are based on physical laws, i.e. conservation of mass and energy. However, some relations are not yet fully understood, or are hard to translate to equations, and assumptions have to be made. This is why a data-driven approach is used, as no explicit relationship between variables has to be specified. In this study a gradient boosted decision tree framework (XGBoost) and a recurrent deep learning model (LSTM) are used to map input to output. XGBoost is a relatively new framework that has shown promising results in other water resources related studies. Long Short-Term Memory is a type of recurrent neural network and chosen for its ability to handle long-term dependencies and for its ability to model non-linearities. In order to see whether the machine learning methods outperform conceptual models, they are compared to the GR4J model. GR4J is a simple yet effective soil moisture accounting model. Beside SMC and NDVI, meteorological variables are used as input. Results show that the deep learning model performs best for simulating today’s discharge and when forecasting up to three days ahead. The GBDT model has a slightly higher Nash-Sutcliffe Efficiency (NSE) for the daily simulation of the Geul, but also a higher mean absolute error (MAE) compared to the deep learning model. The same holds for the three-day-ahead forecast for the Geul and the Rur. Peak timing is accurate for most models but peak discharge is often underestimated. When comparing the ML models to the conceptual model for the daily simulation, deep learning performs best in terms of MAE, but GBDT is better in terms of NSE. When looking at the one-day-ahead forecast, deep learning outperforms the GBDT and conceptual model in both NSE and MAE. In any case, when looking at the metric they outperform the conceptual model. However, the conceptual model has only a couple of parameters to calibrate, is transparent and has only two input variables. The ML models, on the other hand, have my parameters to train, are difficult to physically interpret and have four to five input variables. Besides comparing the two types of models, it is tested whether adding soil moisture content and NDVI as input improve performance of the machine learning models. The former undoubtedly improves performance, whereas NDVI at best improves performance as much as some other meteorological variables. Overall, this study finds that a conceptual model still outperforms the two ML models from a holistic point of view. However, machine learning is not yet fully exploited in water resources management. It already gives promising results and is likely to perform even better in the future.