For a long time,tariff recovery risk of large power customers is a hot spot for electric power company.But because of the lack of external data support and efficient early warning model,tariff recovery risk early warning has become a difficult problem.Based on electricity consumption and electricity charge data,combining with the relevant business,tax and court information of the enterprise,the paper established a serial of tariff recovery risk indexes for large power customers.Secondly,the entropy method (EM) is adopted to evaluate tariff recovery risk assessment of customers'electricity bills,and the customer risk level is divided according to the distribution of risk score.Weak influence indexes were filtered by weight coefficients and overlapping indexes were dropped by correlation analysis.Tariff recovery risk early warning model was carried out by Long Short-Term Memory (LSTM) network.Numerical example results show that the proposed risk early warning model was accurate and effective,and the result gained by LSTM is better than Logistic regression in accuracy,precision and recall.The tariff recovery risk early warning results can accurately locate high risk customers and improve tariff recovery efficiency.