Forecasting models represent a paramount opportunity to support decision-making in modern real-world applications, especially in fields characterized by high variability due to geophysical phenomena. In this context, major challenges are the effective combination of temporal and spatial information from multiple geo-distributed nodes, and supporting scalability for larger and growing sensor networks. Current deep learning methods are generally unable to properly model node relationships throughout the time sequences, as they either treat all nodes independently, or exclusively focus on capturing high-level global network information. Additionally, their complexity usually grows rapidly as more nodes are added, making their applicability in large sensor networks costly. In this paper, we propose a novel geo-distributed forecasting method that simultaneously deals with these two challenges. In particular, we adopt a neural architecture combining recurrent and graph neural networks to jointly analyze sensor network time series data at two levels of granularity: while the GCN sub-network analyzes global network information, the LSTM sub-network is specific to a single node under consideration and extracts temporal autocorrelation from its time series. The model is designed so that multiple sub-models can be trained independently, one for each node of the sensor network, enabling high parallelization capabilities. Quantitative experiments with real-world energy datasets show that our method is highly competitive with respect to state-of-the-art forecasting methods.