Background <p>Compliant mechanisms hold significant application value in high-precision sensing and actuation. However, their inherent mechanical resonance characteristics under broadband excitation can induce severe nonlinear distortion in the output response. Existing compensation methods struggle to simultaneously ensure accuracy and robustness under conditions of small sample sizes and high-frequency operation. </p> Methods <p>This paper first analyzes the dynamic characteristics and distortion mechanisms of a hybrid compliant actuator from a mechanistic perspective. Subsequently, an open-loop feedforward dynamic compensation strategy based on an Extreme Learning Machine (ELM) is proposed. This method establishes a nonlinear dynamic mapping from excitation frequency to output amplitude. By randomly initializing the input weights and hidden layer biases of the ELM and analytically determining the output weights via the generalized inverse matrix, the approach enables rapid training with minimal samples, eliminating the need for iterative optimization. A feedforward compensation voltage is then generated inversely based on the trained model to predistort the original excitation signal, thereby effectively suppressing resonance-induced amplitude amplification. Comprehensive simulations and experiments validate the effectiveness of the proposed method, demonstrating satisfactory compensation performance across high, medium, and low frequency bands within a broad frequency spectrum. </p> Results and Conclusion <p>Utilizing only 25 frequency-displacement training samples, the method achieves excellent fitting performance (R² = 0.99998). After compensation, the maximum steady-state amplitude error of the system in the frequency range of 100–1100 Hz was significantly reduced from 782.01% to 3.91%, demonstrating the significant advantages of ELM in combining high accuracy and strong robustness under small sample conditions.</p>