This study explores Spatial-Temporal Graph Neural Networks (STGNNs) for predicting and discovering weather patterns in South Africa. We compared the predictive performance of three recent STGNNs, i.e. the Adaptive Graph Convolutional Recurrent Network (AGCRN), Conditional Local Convolution Recurrent Network (CLCRN) and Graph WaveNet (GWN) to a Temporal Convolution Network (TCN) to predict temperature, humidity, pressure and wind speed at 45 weather stations in South Africa. We also analysed the quality and usability of the spatial-temporal dependency graph learnt by the different STGNNs. Using the TCN as a baseline we analyse the increase in performance by adding spatial information and identify the stations most likely to be influenced by changes in neighbouring stations. While AGCRN achieved the best overall predictive performance, the dependency graph learned by CLCRN captured the most plausible spatial-temporal dependencies. The findings highlight the strengths and limitations of AGCRN and CLCRN for weather prediction and for automatic discovery of prominent weather patterns in South Africa.