Adaptive time-variant control for piezoelectric smart isolation systems using deep reinforcement learning

Frequent and intense earthquakes demand adaptive seismic protection systems capable of responding to varying ground motion characteristics in real time. Conventional control strategies generally rely on fixed parameters and simplified near-fault and far-field classifications, which may lead to inefficient control and excessive structural responses under strong earthquakes. To address these limitations, this study proposes a novel time-variant Deep Reinforcement Learning (DRL)-based adaptive control framework for a Piezoelectric Smart Isolation System (PSIS). The main novelty lies in developing a Deep Deterministic Policy Gradient (DDPG)-based intelligent controller that learns dynamic control policies and continuously adjusts the PSIS control parameters according to the evolving characteristics of seismic excitations. This enables effective response mitigation for both near-fault and far-field earthquakes without relying on fixed control rules. Comparative evaluations demonstrate that the proposed DRL-based controller outperforms conventional fuzzy logic and Non-Sticking Friction (NSF) controllers in reducing seismic responses. The results confirm its superior adaptability, stability, and energy dissipation capability, indicating strong potential for intelligent seismic protection of critical structures.