Bochum students explore deep learning techniques for rainfall runoff predictions
In recent years, the applicability of Machine Learning approaches for hydrological use cases has reached great attendance. Especially long short-term memory (LSTM) networks show comparable results to conceptual models when being applied to hydrological time series forecasting problems. Now, Interdisciplinary students from Bochum explored innovative deep learning techniques for rainfall-runoff modeling.
Several large sample studies found that LSTM networks are capable of modeling hydro-meteorological processes in catchment areas (Hu et al., 2018; Kratzert et al., 2018; Lees et al., 2021). Their special recurrent neural network architecture enables the learning of long term dependencies, which make them appropriate for rainfall-runoff forecastings.
As part of a research course at the Bochum University of Applied Sciences, several students from the interdisciplinary fields of Geoinformatics and Technical Informatics used the current state of research to explore different deep learning techniques for rainfall-runoff modeling on their own. The students achieved practical data science skills in Python, developed innovative neural network architectures and learned methods for bringing reproducible research into practice. Three groups worked on final practical exercises to complete the course.
- Johannes Kopka explored techniques to take into account spatial distributed meteorological datasets for rainfall-runoff forecasting by combining LSTM layers with convolutional neural network (CNN) layers.
- Thomas Blindert & Tim Kurowski developed a custom mass-conserving LSTM cell for the deep learning framework Tensorflow, which is capable of preserving physical laws.
- Kadir Polat & Hamza Salhi compared the performance of CNN LSTM and pure LSTM networks for hydrological time series prediction.
The results provide great contributions to deep learning based rainfall-runoff modeling and show the wide range of applications for neural networks in the field of hydrology. Future studies will focus on techniques that support model interpretation and help explore what physical implications neural networks have learned.