Flood Stage Forecasting Using Machine-Learning Methods: A Case Study on the Parma River (Italy)

Real-time river flood forecasting models can be useful for issuing flood alerts and reducing or preventing inundations. To this end, machine-learning (ML) methods are becoming increasingly popular thanks to their low computational requirements and to their reliance on observed data only. This work aimed to evaluate the ML models' capability of predicting flood stages at a critical gauge station, using mainly upstream stage observations, though downstream levels should also be included to consider backwater, if present. The case study selected for this analysis was the lower stretch of the Parma River (Italy), and the forecast horizon was extended up to 9 h. The performances of three ML algorithms, namely Support Vector Regression (SVR), MultiLayer Perceptron (MLP), and Long Short-term Memory (LSTM), were compared herein in terms of accuracy and computational time. Up to 6 h ahead, all models provided sufficiently accurate predictions for practical purposes (e.g., Root Mean Square Error < 15 cm, and Nash-Sutcliffe Efficiency coefficient > 0.99), while peak levels were poorly predicted for longer lead times. Moreover, the results suggest that the LSTM model, despite requiring the longest training time, is the most robust and accurate in predicting peak values, and it should be preferred for setting up an operational forecasting system.