The prediction of satellite temperature telemetry data has important research and application value for monitoring satellite status and fault warning in ground operation and maintenance systems. However, traditional prediction methods have limitations such as low accuracy, insufficient robustness, and inability to provide probabilistic interval expressions. Therefore, this study proposes a multivariate temporal probabilistic prediction model TFM Diff for satellite temperature telemetry sequences. Firstly, a hybrid architecture based on gated recurrent units and discrete cosine transform was constructed to more accurately identify time-frequency domain dynamic patterns in telemetry data. Next, by integrating multi granularity features, complex modeling of short-term fluctuations and long-term trends in temperature telemetry data can be achieved, effectively analyzing the multi-scale characteristics of satellite temperature data. Finally, by combining the denoising diffusion model to comprehensively analyze the potential distribution patterns of the data, the probability interval expression of the prediction results can be achieved. Experimental verification based on four sets of real satellite temperature datasets shows that the continuous ranking probability score sum index for probabilistic prediction has improved the predictive performance of the proposed model by 6.26% to 27.77% compared to other mainstream methods, verifying its superior predictive performance, good applicability, and universality in space application scenarios.