Abstract: With the growing need for energy storage in renewable power, accurate temperature prediction of molten salt thermal storage is key to efficient power dispatch. This study uses data from an 800 MWh molten salt storage system in a solar thermal plant. We propose a Transformer-based model with six encoder layers, multi-head self-attention, and time consistency constraints to predict temperature changes. The model captures non-linear patterns and long-term trends in temperature data. Predictions are then used in a multi-time-scale dispatch framework to improve power scheduling. Results show higher prediction accuracy and better economic and stable performance in real-world operations.