Artificial Neural Networks (ANNs) have been successfully employed for predicting and forecasting groundwater levels up to some time steps ahead. In this paper, we present an application of feed forward neural networks (FFNs) for long period simulations of hourly groundwater levels in a coastal unconfined aquifer sited in the Lagoon of Venice, Italy. After initialising the model with groundwater elevations observed at a given time, the developed FNN should able to reproduce water level variations using only the external input variables, which have been identified as rainfall and evapotranspiration. To achieve this purpose, the models are first calibrated on a training dataset to perform 1-h ahead predictions of future groundwater levels using past observed groundwater levels and external inputs. Simulations are then produced on another data set by iteratively feeding back the predicted ground- water levels, along with real external data. The results show that the developed FNN can accurately reproduce groundwater depths of the shallow aquifer for several months. The study suggests that such network can be used as a viable alternative to physical-based models to simulate the responses of the aquifer under plausible future scenarios or to reconstruct long periods of missing observations provided past data for the influencing variables is available.