Smart cities are instrumented with several types of sensors, which allow to transmit, elaborate and exploit the collected data for different services. In this paper we focus on the urban traffic forecasting application. In such context, centralized learning (i.e., training a model in a central unit with data sent from the sensors) or having one model per sensor are the state-of-the art solutions. However, the transmission of such big amount of data, as those from a massive deployment of traffic intensity sensors, implies dense network architectures, long transmission delay, higher network congestion probability and significant energy consumption. On the other hand, training a model only with local data from each sensor lacks in generalization. In this paper we advocate Edge Intelligence and propose a federated peer-to-peer Continual Learning strategy, which applies two variants of Continual Learning principles on data from traffic intensity sensors deployed in a city with the aim to create collaboratively a single general model for all. The analysis of results, performed with real data from a district in Madrid, demonstrates that urban traffic forecasting can be successfully performed in a peer-to-peer fashion. Moreover, we prove that the proposed approaches have lower energy footprint (up to 87% less) and comparable accuracy with respect to state-of-the-art benchmarks.