Permanent reservoir surveillance is an invaluable monitoring tool for CO2 storage projects, as it tracks spatial-temporal evolution of the gas plume. The frequent images of CO2 plumes will facilitate history-matching of the reservoir simulations and increase confidence of early leakage detection. However, continuous data acquisition and real-time interpretation require a new approach to data analysis. Here we propose a data-driven approach to forecasting future time-lapse seismic images based on the observed past images and test this approach on the Otway Stage 2C data. The core component of the predictor is a convolutional neural network, which considers subsequent plume maps as colour layers, similarly to standard red-green-blue blending. Based on the spatial distribution of these ‘colours’ we may predict the future contour of the seismically visible part of the plume. The neural networks absorb the physics of CO2 migration through training on reservoir simulations for a wide range of injection scenarios and subsurface models. Extensive testing shows that realistic plumes for Stage 2C are too complicated and the neural network should be pre-trained on simpler reservoir simulations that include only one or two geological features, such as: faults, spill-points. Such staged training enables a gradual descent of the neural network optimization to a global minimum. In an upshot, the proposed algorithms are proven accurate. The approach is practical, because each CO2 storage project requires extensive pre-injection reservoir simulations. Once the predictor has been trained, it forecasts plume evolution almost instantly and quickly adapts to changing dynamics of the CO2 migration.