Smart Grids aim at leveraging communication and computation capabilities to better operate and control electrical networks. The idea of controlling the output of decentralized photovoltaic (PV) generators so as to avoid current and/or voltage issues in the distribution grid is, in particular, thoroughly investigated by the research community. The simplest and most commonly considered control methods consist in using a feedback controller based on a local voltage measurement at PV generator level. The main advantages of such methods is that they can be implemented at low cost and require no specific information about the network on which they are deployed. In this paper, we consider an alternative, that uses the advanced metering infrastructure (AMI) as the basis for PV generation control. Our contributions are as follows. First, we cast the problem of PV generation curtail-ment in the general framework of control theory, which sheds some light on the structure of previously proposed controllers. Second, we show how the advanced metering infrastructure may be used on one hand to infer some knowledge about the underlying network (learning phase), and on the other hand, to observe some of the disturbances that apply to our system so that they can be taken into account in the control phase. While doing so, we take great care of accounting for the inherent limitations of the advanced metering infrastructure in terms of measurements and communication, which is a major difference between this paper and previous work on the topic. Third, by means of numerical simulations, we compare our proposed feed-forward controller with two other controller structures: open-loop, and feedback. We demonstrate that our feed-forward controller — that requires no prior knowledge of the underlying electrical network — brings significant performance improvements as it can effectively suppress over-voltage and over-current while requiring low power curtailment.