Abstract: With the rapid growth of electric power communication networks, the importance of predicting the working state of online equipment is increasing as well. Since the running data of typical communication devices always come from heterogeneous resources, the prediction models have to be learned from features with high dimension, high sparsity as well as repetitive patterns. This problem severely restricts the performance of conventional machine learning approaches. This paper proposes a novel state prediction model based on the integration of attention mechanism and LSTM (Long Short-Term Memory). By a two-stage learning strategy, the attention mechanism can achieve both dimensionality reduction and feature extraction of original input. Meanwhile, most related features are extracted for final prediction from the end-to-end learning. Extensive experimental results on practical running data of electric power communication networks demonstrate that, the proposed method has high performance in the working state prediction problem.

Key words: attention, LSTM, neural networks, state prediction of equipment