Ensuring the health and longevity of satellites is paramount for modern communication, navigation, and earth observation. This study explores the application of advanced time-series models—SARIMAX, Long Short- Term Memory (LSTM) networks, and Transformers—for anomaly detection in satellite telemetry data, specifically focusing on temperature readings. Anomalies in temperature data can be early indicators of potential malfunctions, making robust detection methods critical. In this regard, the effectiveness of both the SARIMAX and Transformer models was evaluated using the Satellite Monitoring and Analysis Platform dataset with over 105,400 temperature readings taken within 20 days. The LSTM autoencoder model was further assessed on univariate datasets containing 14,836 readings within 10 days. The SARIMAX model was then fitted with a percentile thresholding method for catching deviations, taking into consideration seasonal and trend components of data. Later, this model turned out to be comparable in performance with a Transformer model that has already shown really good results on sequential data by catching complex temporal dependencies and can perform anomaly detection leveraging the attention mechanism. In contrast, the LSTM autoencoder model detected anomalies using two methods: percentile- based thresholding and moving average methods. The results showed that the LSTM autoencoder model works well using the running average method; it is more accurate for some types of data. In this respect, the present research shows that integrating a machine learning approach with one statistical model can enhance the detection of satellite time-series data anomalies. This will be followed in future work, dedicated to modifying such models and investigating transformer-based model applicability in multivariate datasets.