To address the pronounced time-varying characteristics of power system inertia under high penetration of renewable energy and the difficulty of achieving both accuracy and robustness using a single estimation method

a hierarchical multi-source fusion architecture for inertia estimation is proposed. At the front layer

weighted least squares and a Conv-Transformer hybrid neural network are integrated to achieve preliminary inertia estimations from mechanism-driven and data-driven perspectives

respectively. At the intermediate layer

a parameter tracking strategy based on adaptive Huber residual-weighted recursive least squares is employed. Through the coordinated use of a dynamic forgetting factor and Huber influence function

the damping coefficient (D) and synchronizing power coefficient (K) are robustly tracked. At the back layer

a dynamic weight extended Kalman filter adaptively adjusts the observation noise covariance according to the discrepancies among multi-source estimates

enabling adaptive fusion and rapid tracking of system inertia. Simulations based on the New England 10-machine system demonstrate that the proposed method exhibits excellent performance under extreme conditions such as vertical steps and parameter drift

with strong capabilities for tracking time-varying inertia and rejecting disturbances. The proposed method offers an effective solution for online inertia estimation in power systems with high renewable energy penetration.