The power customer service order data records the demand of power users in text. A reasonable work order classification method is helpful to accurately identify the demand of users and improve the operating efficiency of the power system. To solve the problems of sparse feature data and strong dependency of work order data, this study optimizes the structural model that combines character-level embedded Bidirectional Long-Short-Term Memory network (BiLSTM) and Convolution Neural Network (CNN). Firstly, this model obtains the feature representation of text by noise reduction on the term vectors trained by the Word2Vec model. Secondly, it uses the BiLSTM network to recursively learn the time sequence information of the text to extract the feature information of sentences. Finally, those obtained are input into the double-channel pooled CNN for the extraction of local features. The test experiments on the real work order data set of power customer service demonstrate that the model has good accuracy and robustness in the task of classifying work orders of power customer service.