关键词: 谐波分类识别, 特征提取, 长短期记忆网络（LSTM）, 多标签分类, 深度学习

Abstract: Long Short-Term Memory network （LSTM） is introduced for the wide access of power electronic equipment， and harmonic pollution is more complicated and the impact is greater. This paper proposes a composite harmonic disturbance classification model that combines LSTM and multi-label classification algorithms. The model firstly extracts the feature values through LSTM， then performs feature dimensionality reduction processing through the fully connected layer， and finally performs multi-label classification and recognition through the classification layer. This model overcomes the shortcomings of artificial feature selection， the slow convergence speed and easy to be limited to local optimum in traditional neural network training. Experiments show that the algorithm model can effectively classify and recognize complex harmonic disturbance under different noise conditions.