This paper presents the neural network model that was used by the author in the Weather4cast 2021 Challenge Stage 1, where the objective was to predict the time evolution of satellite-based weather data images. The network is based on an encoder-forecaster architecture making use of gated recurrent units (GRU), residual blocks and a contracting/expanding architecture with shortcuts similar to U-Net. A GRU variant utilizing residual blocks in place of convolutions is also introduced. Example predictions and evaluation metrics for the model are presented. These demonstrate that the model can retain sharp features of the input for the first predictions, while the later predictions become more blurred to reflect the increasing uncertainty.