Machine learning (ML) is a promising methodology for generating prediction and forecasting models on multiple application fields. With regards to Photovoltaic (PV) energy production, the prediction of the next-day energy outputs is an extremely important to electricity grid maintainers to guarantee the stability of the supply networks. Generating quality prediction models highly depends on the volume and quality of the data used during training. Such data are not always available in a central repository due to legal or technical reasons so other approaches previously used in big data applications can be used to overcome such limitations. In this work, we investigate how to engineer ML system specifically designed for edge-enabled deployments using a real-world deployment. We evaluate 3 different reference architectures and investigate their capacity, data-related limitations and robustness of the computed prediction models for PV installations.