Given a collection of sensors monitoring a flow network, the problem of discovering teleconnected flow anomalies aims to identify strongly connected pairs of events (e.g., introduction of a contaminant and its removal from a river). The ability to mine teleconnected flow anomalies is important for applications related to environmental science, video surveillance, and transportation systems. However, this problem is computationally hard because of the large number of time instants of measurement, sensors, and locations. This paper characterizes the computational structure in terms of three critical tasks, (1) detection of flow anomaly events, (2) identification of candidate pairs of events, and (3) evaluation of candidate pairs for possible teleconnection. The first task was addressed in our recent work. In this paper, we propose a RAD (Relationship Analysis of spatio-temporal Dynamic neighborhoods) approach for steps 2 and 3 to discover teleconnected flow anomalies. Computational overhead is brought down significantly by utilizing our proposed spatio-temporal dynamic neighborhood model as an index and a pruning strategy. We prove correctness and completeness for the proposed approaches. We also experimentally show the efficacy of our proposed methods using both synthetic and real datasets.