Abstract: To meet the real-time simulation requirements of hundred-megawatt-scale grid-following and grid-forming hybrid photovoltaic power stations, this paper addresses the problems of high computational cost caused by nonlinear solving and state updating in detailed mechanism models under multi-unit parallel operation and multi-control-loop coupling, limited real-time computational margin under fixed-step simulation, and the difficulty of purely data-driven models in maintaining grid-connected dynamic consistency. A lightweight data–mechanism hybrid modeling method is developed. In this method, the nonlinear energy conversion process of the photovoltaic array is selected as the lightweight modeling object, and a physics-guided nonlinear autoregressive with exogenous input (PG-NARX) neural network is used to establish a dynamic surrogate model, reducing the online computational burden of source-side nonlinear equations. Meanwhile, the key mechanism structures of the converter and filter are retained to maintain the physical consistency of grid-interaction characteristics such as system inertia, damping, and impedance. Based on a single grid-forming photovoltaic unit and a 300 MW grid-following/grid-forming hybrid photovoltaic power station, real-time computational performance, transient and steady-state responses, and wide-frequency impedance characteristics are verified, and the operating characteristics under different grid-forming ratios are analyzed. The results show that the average steady-state NMAE of key variables in the single-unit hybrid model is lower than 0.72%, and the average single-step computation time is reduced by approximately 9.0% compared with the detailed mechanism model under the same operating conditions. For the station-level hybrid model, the average NMAE of key variables does not exceed 2.63%, and the CPU utilization of each real-time task remains below 60% under a fixed time step of 50 μs, meeting the requirements of multi-area parallel real-time simulation at the station level. Time–frequency-domain verification results show that the proposed model maintains the main dynamic responses and wide-frequency impedance characteristics of the mechanism model. Under the weak-grid condition with SCR = 3 and the test scenarios considered in this paper, the 20% grid-forming ratio shows coordinated characteristics among steady-state tracking accuracy, transient support capability, and wide-frequency stability. The results can provide a reference for lightweight real-time simulation modeling, dynamic support capability evaluation, and grid-forming ratio configuration of grid-following/grid-forming hybrid photovoltaic power stations.