Prolonged overheating of high-voltage cables may lead to insulation thermal breakdown， consequently affecting the stability of the power grid. However， the existing research primarily focuses on traditional prediction models， and ignores the complexity and dynamic characteristics of temperature data. To address this limitation， a cable temperature prediction model based on Multi-Scale Patch and Convolution Interaction （MSP-CI） was proposed. Firstly， the input dimension was reduced using a channel resampling method， and a multi-scale patch branch structure was constructed， so as to decouple the complex time series. Then， macroscopic information from coarse-grained patches and microscopic information from fine-grained patches were extracted， respectively， through the combination of sequence decomposition and convolution interaction strategies. Finally， an attention fusion module was constructed to balance the weights of macroscopic and microscopic information dynamically and obtain the final prediction results. Experimental results on real high-voltage cable temperature datasets demonstrate that compared to the baseline models such as TimeMixer， PatchTST （Patch Time Series Transformer）， and MSGNet （Multi-Scale inter-series Graph Network）， MSP-CI achieves a reduction of 7.02% to 34.87% in Mean Squared Error （MSE）， and a reduction of 5.15% to 32.04% in Mean Absolute Error （MAE）. It can be seen that MSP-CI enhances cable temperature prediction accuracy effectively， providing a reliable basis for power dispatching operations.
