Segment Based Water Temperature Prediction
Supervised by: Benjamin Fankhauser
If you are interested in this topic or have further questions, do not hesitate to contact me.
Context/Background/Current State
In weather forecasting one often uses rather primitive functions to model the change of state of neighboring grid points. In the present project we aim to build a graph neural network model with the same spirit: each grid point (or river segment) is dependent of the state of his neighbors. To do so, we plan to create a specifically designed graph, i.e. a grid with edges based on watersheds. On this novel graph a message passing neural network will be trained. As we do only have a sparse amount of river segments with actual measurement stations this adds another difficulty to the project (moreover, we also have to leave some out for assessing the performance on scgeisse test set).
Goal(s)
- Model the Swiss river graph based on watersheds with appropriate segments or grid points
- Predict the water temperature at any node of the graph based on sparse measurements
Approach
- Get inspired of current weather forecast models.
- Model the Swiss river network as grid points or river segments in appropriate length.
- Create a message passing network with a sparse loss function.
- Train the network and assess its performance on held out test stations.
Required Skills
The student should be familiar with deep learning (LSTMs, Neural Networks, Graph Neural Networks) and graph structures in general. A strong background in programming and project management skills is advantageous. The large scope allows the student to do further research in a topic of his choice.
Remarks
The project might be tailored in scope.