direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments



Neural Networks approaches to optimize water level predictions for inland navigation
Zitatschlüssel MattaE.;MaY;MeißnerD.;RichterJ.;SchellenbergH.;Hinkelmann
Autor Matta, E.; Ma, Y; Meißner, D.; Richter, J.; Schellenberg, H.; Hinkelmann, R.
Buchtitel EUROCONTROL TIM/ART Workshop on Machine Learning
Jahr 2018
Adresse Brétigny-sur-Orge, France
Zusammenfassung The German Federal Ministry of Transport and Digital Infrastructure (BMVI) expects a significant increase of inland navigation traffic by the year 2030; since the national waterways are envisaged to remain nearly unaltered, a higher efficiency of the entire shipping logistics and management will be required. In this context, the mFund research project Digital Skipper Assistant (DSA) has the main scope to provide a multi-device demonstrator adapted to the requirements of inland navigation, available as web browser and mobile application, in order to support skippers, carriers and administration agencies, such as the Federal Waterways and Shipping Administration (WSV). The DSA is able to calculate and optimize routes, calculate Estimated Time of Arrival (ETA), provide traffic information and water level forecasts. In the DSA project framework, the research conducted here has the objective to explore the capabilities of artificial neural networks (ANNs) in predicting water levels up to ten-days ahead at some crucial gauges along the German part of the Rhine River Basin (Fig. 1). The basic idea consists in predicting downstream gauges using the significant upstream ones as inputs, focusing exclusively on water levels. Different open-source tools characterized by different ANN architectures have been explored and sensitivity analyses (e.g. varying the number of hidden neurons or gradient descent methods) have been conducted to define the optimized model for this application. OpenNN (Lopez 2017) and Keras (Keras, n.d.; Brownlee, 2016) were used as open-source software to set up respectively feed-forward back-propagation and long short term memory neural networks (FFBP, LSTM). Subsequently to the implementation and testing of single- and multiple-outputs models for water levels predictions, only using the water level measurements provided by the Federal Institute of Hydrology (i.e. BfG, partner in the project), the BfG water level forecasts have been additionally integrated in the ANNs as additional input data (predictors), which are obtained using the BfG physically-based hydrological model. The most relevant results have been compared regarding two different input scenarios for the water level forecasting: I) daily measured water levels; II) daily measured water levels integrated with the BfG forecasts. The results of the latter scenario delivered the best results for the longer-term predictions (leading time of about 2-10 days), with a higher accuracy than the ones of the BfG hydrological model chain. On the other hand, the BfG physically based model series or the single-output ANNs provide superior performances for predictions up to 2 days ahead. E.g. the comparison over time between some of the most relevant results at the gauge Oestrich is shown in Fig. 2. In future work, the developed networks will be further refined in gauges which have been not yet investigated or in reaches where water is highly regulated e.g. by locks or dams. Further ANN architectures such as echo state networks (ESN) or support vector machine (SVM) can be investigated for longer-term water levels predictions (longer than 10 days) and compared systematically with the current approaches developed at BfG.
Download Bibtex Eintrag

Zusatzinformationen / Extras


Schnellnavigation zur Seite über Nummerneingabe