Predictive Adaptive Optics (AO) control is a promising technology for AO applications in high-disturbance and low-signal environments such as directed energy, optical communication, and astronomical seeing. Predictive AO utilizes future state predictions of an optical wavefront propagated through a turbulent medium to drive correction, thereby mitigating the limits imposed by inherent latency in the AO system. In this work, we present a novel Artificial Neural Network (ANN) approach for embedding the flow dynamics for a range of Airborne Aero-Optics Laboratory (AAOL) datasets into a single turbulent flow prediction model. As the angle of the laser beam through the hemispherical AAOL turret changes, flow characteristics vary greatly according to statistics such as mean advection speed, direction, and scale, as well as the presence of different turbulent structures and shock waves. As a result, a predictive model trained on a single look angle and flow condition will likely have poor performance when conditions change, for instance, by slewing the turret look angle during AO operation. In our approach, this limitation is mitigated by introducing the model to flow data from a range of look angles during training. We analyze this combined model’s ability to forecast turbulent wavefronts from look angles included in the training set to establish baseline model performance. We then consider performance on measured AAOL wavefront sensor data from holdout look angles entirely excluded from the training wavefront data to demonstrate the generalization capability of the resulting model, and consider the implications for ANN-based AO correction for dynamic, high-speed, turbulent flows.