Abstract: The prediction of system dynamic frequency under the fluctuation of new energy with high permeability is the basis for security situation awareness of receiver network frequency. In this paper, a dynamic frequency prediction method is proposed for new energy power systems based on combined equation-of-state-driven and data-driven bi-directional tree-struct long and short-term memory (CEOSD-BITREE-LSTM) with hybrid measurement and equation of state. First, considering the measurement difference and timing synchronicity of PMU and SCADA, a hybrid measurement fusion strategy is proposed by double-layer and multi-head attention graph neural network. Second, according to the dense sampling characteristics of PMU, a linear time-varying state equation is established, which takes into account the physical connection of the source network and load, and describes the frequency characteristic interaction relationship in physics-data space. Then, considering the uncertain factors such as new energy output and load fluctuation, combined with the topology structure formed by the PMU parallel search of frequency modulation resources, the CEOSD-BITREE-LSTM dynamic frequency prediction model is constructed to achieve the high-precision prediction of system frequency. Finally, an improved New England 10-machine 39-node system and tri-zone interconnection system are taken as an example to verify the feasibility and effectiveness of the method.