The existing power quality disturbance classification algorithms based on feature selection have poor robustness and weak anti-noise performance. This paper proposes an improved one-dimensional convolutional neural network model for power quality disturbance classification. Firstly, the eigenvectors of the power quality disturbance signal are extracted by three convolutional neural network sub-models, and then the extracted eigenvector is fused into a new one; finally, the classification is implemented by BP neural network. The result shows that the proposed algorithm has higher robustness, recognition rate and strong anti-noise abilities, because the new eigenvector has greater discrimination, compared with the old one-dimensional CNN models and the existing power quality disturbance classification algorithms. It provides a new idea for classifying power quality disturbance signals. This work is supported by Sichuan University Key Laboratory Open Fund (No. szjj2017-046), Health Management Center of Xihua University (No. DJKG2019-005), National Natural Science Foundation of China (No. 61571371), and Natural Science Foundation of Guangdong Province (No. 2015A030313853).