Due to the complex space environment and the influence of the satellite itself, satellites in-orbit anomalies and failures occur from time to time. It is necessary for the ground system to predict the changing trend of future spacecraft performance through satellite telemetry parameters to ensure the safe operation of satellites. Aiming at the characteristics of satellite telemetry data with trends, local fluctuations and being easily affected by other environmental factors, this paper proposes a prediction model based on improved LSTM-Attention. The global model and local model are used to obtain the trend component and local fluctuation component of the telemetry sequence respectively. Moreover, the model will learn more sequence information by using the covariates that have an impact on predicting the target sequence, which improves the prediction accuracy. This model can provide both interval forecasting and point forecasting results for telemetry series. The real telemetry parameter data set of scientific satellite and the public data sets of time series are used for experimental verification. The results show that the method has achieved good results compared with benchmark algorithms.