Neuroforecasting of daily streamflows in the UK for short- and medium-term horizons: A novel insight

Predicting streamflows, which is crucial for flood defence and optimal management of water resources for drinking, irrigation, hydropower generation and ecosystem conservation, is a challenging task in most practical cases. The limitations of physically based models and the increasing availability of time series data on flow rates and other weather and climate variables of interest are increasingly driving the use of models based on Machine Learning algorithms. Of these, neural networks have proven to be among the best-performing prediction tools.In this study, four types of neural networks, Multilayer Perceptron (MLP), Radial Basis Function Neural Network (RBF-NN), Long Short-Term Memory Network (LSTM), and Bi-directional Long Short-Term Memory Network (Bi-LSTM), were compared in the prediction of short-term (1-3 days ahead) and medium-term (7-15 days ahead) daily flow rates of six different rivers in the United Kingdom. The predictors consisted only of the lagged values of flow rates and daily cumulative precipitation. The optimal number of these and the hyper-parameters of the different algorithms were selected according to a Bayesian optimization procedure.The various algorithms demonstrated comparable and strong short-term forecasting abilities, with a slight inclination to underestimate the maximum flood flows. In particular, the coefficient of determination (R2) for 1-day ahead forecasts ranged from 0.909 to 0.986, and the Mean Absolute Percentage Error (MAPE) ranged from 3.36% to 13.94%. However, as the forecast horizon extended, a reduction in forecasting accuracy was identified, despite all models being able to predict the overall flow pattern, even up to 7-15 days ahead. Compared to LSTM-and Bi-LSTM-based models, RBF-NN-based models showed less of a tendency to underestimate flood peaks and overestimate low flows and could predict both with good accuracy. Additionally, the relative error distribution exhibited a general skew in all models. The findings of this study suggest that RBF-NNs are a powerful tool for obtaining accurate forecasts in both the short-and medium-term while requiring a limited number of parameters to be optimized, thus reducing the calculation time required.

Automated summary

Predicting streamflows is crucial for flood defence and optimal management of water resources. Machine Learning algorithms, particularly neural networks, have shown promise in accurately predicting short-term and medium-term flow rates. In this study, four types of neural networks were compared, and RBF-NN models demonstrated better accuracy in predicting flood peaks and low flows.