The pronounced nonlinearity of synchronous generator short-circuit currents makes rapid current zero prediction difficult

posing a challenge to phase-controlled interruption. This paper proposes a fast zero prediction algorithm

ELM-KF

which integrates an extreme learning machine (ELM) with a Kalman filter (KF). A lightweight ELM with only a few hidden nodes is used to optimize the initial priors of the KF

thereby accelerating the iterative prediction process. This design alleviates the slow convergence and insufficient accuracy of conventional KF in the early fault stage. It also mitigates the limitations of neural-network methods

including high computational cost

limited robustness

and the lack of dynamic error correction. A generator short-circuit model is established for evaluating ELM-KF under different conditions and comparing it with conventional KF and recursive least squares (RLS). Results show that with only 0.5 ms of data

ELM-KF reduces the first zero error to 0.2 ms and achieves R2 = 91.75%

exceeding the accuracy of the comparison algorithms with a 10 ms window; when the fault angle varies

ELM-KF yields smaller average errors and the highest R2; under different SNRs

its accuracy is comparable to KF and better than RLS; and it is insensitive to the number of hidden nodes while maintaining a lightweight advantage. RTDS experiments with a DSP-based system further validate real-time performance

with an open-loop tripping-command error within 0.4 ms and a closed-loop breaker response delay of 0.1 ms.