The multi-scale fluctuation characteristics of crude oil prices make time series forecasting a complex but important task. This research suggests a hybrid model that combines Transformer and convolutional neural network (CNN) to overcome the shortcomings of conventional approaches in capturing nonlinear characteristics and long-term and short-term interdependence. While the CNN module employs multi-scale convolutional kernels to extract local information, the Transformer module successfully models global long-range relationships using self-attention techniques. To fully use their own advantages, the two collaborate via a fusion approach. Several baseline models are used for comparison verification, and the experiment is based on historical WTI and Brent crude oil price data. According to the findings, the suggested model performs noticeably better than other approaches in terms of mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE). Ablation experiments further demonstrate the complementarity and necessity of the Transformer and CNN modules in improving model performance. The research in this paper not only provides new ideas for modelling complex time series in theory, but also provides reliable technical support for crude oil market forecasting and decision-making.