In our latest ‘Performance Report’ series, we start with analysing BWI’s discharge forecast model’s performance for the La Dives river!
As climate change intensifies hydrological extremes, such as floods and droughts, the high-frequency monitoring of water resources becomes crucial for effective water resource management.
BWI employs a semi-distributed hydrological model. With the integration of machine learning techniques, it monitors and forecasts river discharges. This blog post delves into a detailed performance evaluation of BWI’s discharge prediction model, specifically focusing on the ‘La Vie’ tributary of the ‘La Dives’ river, with its outlet located in Le Bourg, Mézidon Vallée d’Auge, in the Normandy region of France.
The La Dives basin is located in the east-central part of the Calvados department and the north of the Orne department. It is encircled by the Orne basins to the west and south, and by the Touques basin to the east. ‘La Vie,’ the basin’s primary tributary, extends for approximately 66.9 kilometers. Ultimately, it merges with the river in Biéville-Quétiéville. The catchment area of ‘La Vie,’ with its outlet at ‘Le Bourg,’ spans 338.203 square kilometers. Within this catchment, there are six small lakes covering a total surface area of 0.095 square kilometers.
BWI utilizes a robust model performance evaluation framework that considers a range of metrics. Some of these include Nash-Sutcliffe model efficiency scores (NSE), normalized NSE (NNSE), Kling-Gupta Efficiency (KGE), and their respective modified and normalized versions. The preferred metric within this framework is NNSE, which ranges between 0 and 1. A higher NNSE score indicates superior model performance. This metric is particularly useful for tracking the model’s performance over time and effectively communicating the results.
At the La Vie’s Le Bourg outlet, the model achieves an NNSE of 0.8. This is a score that scientific publications classify as indicative of good model performance (NNSE > 0.65). During the training phase, following an initial warm-up and calibration period, the model’s output demonstrates strong convergence with measured discharges. The optimal parameter sets and the final state of the basin at the end of the training are recorded and utilized to initialize the subsequent training batch.
