Abstract: Under the goals of carbon emission peak and carbon neutrality, the new distribution network with increasing penetrations of distributed generators has raised the requirements on operational reliability; moreover, how to improve the adaptability of fault location method under scenarios which penetration of renewable energy are rapid changing is an urgent issue to be solved. Therefore, benefiting from the learn-to-learn mechanism of meta learning, this paper proposes a novel stepwise cognitive-discovery based fault location method for new distribution networks. Firstly, a specialized location model is built by neural architecture search algorithm based on data from existing scenario. Then, the knowledge factors are acquired by meta learning during model formation process, and a cognition-discovery library of fault section location is built. Furthermore, with the help of coordinative effect of data stream and knowledge flow, self-evolving of the model under diverse scenarios is achieved. Finally, the proposed method is verified in PSCAD. Results show that the proposed method has the advantages of high positioning accuracy and strong robustness, and has strong generalization ability under diverse penetration scenarios. The research results can provide technical supports for the application of fault location method based on artificial intelligence in practical systems.