An ANN Data-driven Model for Flood Level Forecast in Non-regulated Catchments: A Case Study from Norway

In the present study an Artificial Neural Network (ANN) model is developed for forecasting the peak flood level in a river based on weather forecasts. The development steps of the ANN model (input and time lag selection, data division and resampling, data scaling, model structure selection and optimization, and performance evaluation) will be presented, taking into consideration physical based knowledge and the stochastic character of these type of machine learning models. The input selection will be discussed comprising the daily precipitation, average daily temperature and the maximum daily water level in several stations available for previous days and forecasted for the upcoming days. The influence of the length of the data set will be discussed by testing different data set lengths, allowing to check the possible influence of climate change on past hydrological datasets. The data is divided in three subsets (training, test-validation and test) and the influence of the size of the training subset will be assessed. Namely, the effect of including new data in the training will be studied, aiming at getting a model that can be update throughout time. The ANN model is then use has a flood forecasting tool for an extreme event occurred in October 2017, for which the model has not be trained. For that purpose, data from the weather forecast system of the Norwegian Meteorological Institute will be used. The forecasting tool is then developed for 4 days ahead, considering all the information available and the permanent updates of the weather forecast.

The results show that the length of the data set can have some influence on the final result, but a data set from the last 10 to 20 years is a good choice. The length of the training data subset has also some influence on the final results, being length around 50% of the full dataset an adequate choice. The inclusion of new data, including one severe extreme, has a marginal effect on the overall results of the model. Finally, the application of the developed ANN model to forecast an extreme event, higher than the highest extreme on the training data subset, showed very goods results for a 4 day ahead forecast. This points out that this type of data-driven models can be useful forecasting tools and complementary to hydrologic and hydraulic physical-based models. The advantages and limitations of the methodology will be discussed in a practical point of view.