The liberalization of the electricity market and the expansion of new forms of electricity production and consumption are paving the way for new smart digital services. These new services will certainly be relying on a new generation of smart meter (SM) that will offer, among other things, prediction of consumption and production at both household and microgrid levels.These predictions can be obtained either through a generic model trained on data collected from all SMs, or through specific models developed for each SM based on its individual data. The benefit of the generic model is that it guarantees an optimal solution. However, its implementation is not possible for security reasons. The use of specific models requires managing a large number of SMs, which poses a significant challenge.This paper presents a Federated Learning (FL) approach, a decentralized privacy-preserving paradigm that achieves com-parable performance to the generic model, considering both consumption and production scenarios.The dataset, gathered from 1153 SMs over a period of 18 months, is provided by a Swiss Distribution System Operators (DSO). Although, data are individually collected per device, the generic model is trained to holistically process this data.Our experimental results demonstrate that our FL based Long Short-Term Memory (LSTM) model performs as well as the generic model and outperforms the specific models, while preserving data privacy and security.