This study presents two innovations. First, an artificial neural network (ANN) is trained to generate a two-dimensional wind-farm power curve, which is a function of not only wind speed, like the conventional power curves provided by manufacturers for a single wind turbine, but also wind direction, thus effectively taking into account wake losses. The resulting two-dimensional power curve predicts with high accuracy (error ~2%) the power of the entire Lillgrund wind farm. Second, the ANN is trained a second time using two geometric (GM) properties of wind farms (blockage ratio and blockage distance) to replace wind direction. The resulting GM-trained ANN has transfer-learning ability because it is no longer depending on the local wind conditions, but on general geometric properties, and therefore it can predict the power generated by any wind farm, not just the one it was trained on (Lillgrund). A validation at an onshore wind farm (Nørrekær in Denmark) demonstrates the high accuracy (error ~6%) and transfer-learning ability of the GM-ANN.