<p>Oil temperature prediction in power transformers is a crucial task, encompassing the forecasting of variations in insulation oil temperature, essential for monitoring and analyzing the condition of these transformers. Existing long-term sequence forecasting methods exhibit two key limitations. Firstly, most models concentrate solely on correlating limited time periods, neglecting potential interactions between subsequences, and resulting in inaccurate trend predictions. Secondly, existing modeling approaches overlook the significance of periodic variations with diverse cyclic thermal characteristics that are fundamentally important for forecasting future oil temperature trends. To overcome these limitations, we propose MSTransformer (transformer-based multi-scale session-enhanced model), a novel approach for accurately predicting the temperature of the oil of the power transformer. Our model addresses the limitations by segmenting long time series into session-level sessions and applying self-attention transformer framework to these sessions. This enables effective capture of semantic information across long sequences while reducing model complexity. Furthermore, we introduce a multiscale module to account for both short- and long-term sequence dependencies, including periodic changes. This module enables simultaneous encoding of fine- and coarse-grained sequence patterns. By incorporating these periodic variations, our model enhances the accuracy and stability of the temperature predictions. Extensive experiments validate our method’s excellent predictive capabilities in accurately forecasting power transformer oil temperature. Meanwhile, to validate the model’s generalization capability, we conducted extensive experiments on multiple publicly available multivariate time series datasets. Our code is available at <a href="https://github.com/Sereinzmy/MSTransformer">https://github.com/Sereinzmy/MSTransformer</a>.</p>