Machine learning models for river flow forecasting in small catchments.

In consideration of ongoing climate changes, it has been necessary to provide new tools capable of mitigating hydrogeological risks. These effects will be more marked in small catchments, where the geological and environmental contexts do not require long warning times to implement risk mitigation measures. In this context, deep learning models can be an effective tool for local authorities to have solid forecasts of outflows and to make correct choices during the alarm phase. In this study, we investigate the use of deep learning models able to forecast hydrometric height in very fast hydrographic basins. The errors of the models are very small and about a few centimetres, with several forecasting hours. The models allow a prediction of extreme events with also 4-6 h (RMSE of about 10-30 cm, with a forecasting time of 6 h) in hydrographic basins characterized by rapid changes in the river flow rates. However, to reduce the uncertainties of the predictions with the increase in forecasting time, the system performs better when using a machine learning model able to provide a confidence interval of the prediction based on the last observed river flow rate. By testing models based on different input datasets, the results indicate that a combination of models can provide a set of predictions allowing for a more comprehensive description of the possible future evolutions of river flows. Once the deep learning models have been trained, their application is purely objective and very rapid, permitting the development of simple software that can be used even by lower skilled individuals.