AbstractCharacteristics at nodes in a network, such as values of demand, evolve over time. To make time-dependent decisions for a network, making time series predictions at each node in the network over time is often necessary. Typical time series prediction approaches are based on historical information. However, these fail to account for network-level factors that might affect nodal values. This paper proposes an approach for the time series prediction in nodal networks that accounts for both time history information and nodal characteristics in the prediction. The approach is based on recurrent neural networks and, in particular, gated recurrent units (GRU), creating a new GRU structure called a Pairwise-GRU to include the influence of both historical data and neighboring node information to predict values at each node in the network. The result is a more accurate and confident time series prediction. The performance of the proposed approach is tested using an electricity network in the southeastern United States. The results indicate that the proposed Pairwise-GRU outperforms existing methods in terms of increased accuracy and decreased uncertainty in the prediction. The approach performs particularly well for long-term, multiple-time-steps ahead predictions and anomalous hazard conditions in addition to normal operating scenarios.