Redevelopment of LobithNN : on working methods, model choices and model quality

This project aims to improve LobithNN, the machine learning model for short-term (2 days ahead) discharge forecasting at Lobith with an hourly timescale and improved quality over the existing multi linear regression model. Through experimentation with various test cases, we identified the optimal setup for the new LobithNN model. This model leverages water levels from six reliable upstream stations and historical discharge data at Lobith, trained on BfG data and operationalized with Matroos data. The new LobithNN model exhibits improved performance over LobithW and Old LobithNN in normal discharge ranges across nearly all lead times. Future works are recommended to improve operational reliability, enhance model accuracy, and standardize and generalize machine learning workflows and tools.