A third of the food consumed by humankind depends on bees’ activities. These insects have a fundamental role in pollination and they are disappearing from the planet. An understanding of their behavior, discussed here from the point of view of their activity level, can help detect adverse situations and even improve the employment of bees in crops. In this work, several Recurrent Neural Networks’ architectures, alternating topologies with GRU and LSTM structures, are evaluated in the task of forecasting bees’ activity level based on the values of past levels. We also show how RNNs can improve its accuracy by evaluating how different input time windows impact on results.