To address the difficulty of coordinating frequency support, device constraints, and operating burdens in flywheel–battery hybrid energy storage connected to weak grids, a hierarchical grid-forming control method is investigated. A virtual synchronous generator generates the basic support demand. An autoregressive model updated by recursive least squares predicts the measured net power imbalance 0.36 s ahead. Low-pass filtering and finite-candidate-set economic model predictive control perform constrained power allocation. Simulations are conducted under multiple operating conditions, supplemented by a 10-kW switching-level scaled model.Under medium disturbance, the maximum frequency deviation and the integral of absolute frequency deviation are reduced by 14.56% and 75.64%, respectively, compared with the scheme without prediction. In the 10-kW model, the maximum frequency deviation is reduced by 8.22%, with no hard-constraint violations. During long-term operation, improved frequency performance is accompanied by increased storage energy throughput.The proposed method coordinates prediction-assisted frequency support and device constraints under the tested conditions. However, prediction benefits depend on operating conditions, and a trade-off exists between frequency performance and storage operating burdens.