Weather data often contains missing measurements due to sensor failures, communication
errors, or maintenance interventions, which complicates forecasting
and monitoring. The purpose of this thesis is to replace missing temperature measurements
by exploiting the spatial-temporal dependencies between meteorological
stations.
The goal of our thesis is to develop a machine learning model using graph
neural networks to impute temperature measurements in incomplete data sets.
We structured the data into a spatio-temporal graph and developed two main
approaches. The GNN-GRU model combines graph attention and recurrent GRU
units to predict temperature for the next time step and is used when historical
measurements at the target station are available. Meanwhile the GAT-Spatial
model uses multiple consecutive GAT layers to replace current temperature values
without historical data at the target station are not used.
Our best results were achieved with the GNN-GRU model with an average absolute
error of 0.371 °C, while the GAT-Spatial model achieved an average absolute
error of 1.220 °C.
|
