In the hot rolling process of seamless steel pipes, the energy consumption management of the CPE push bench mill faces challenges, primarily due to the difficulty in extracting feature information from multi-source heterogeneous data such as equipment and process parameters, which in turn affects the accuracy of power consumption prediction. To address the issue of high-dimensional, nonlinear, and time-varying features in power consumption prediction, this paper proposes a prediction model that integrates multi-source heterogeneous data. The model first combines Convolutional Neural Networks (CNN) for spatial feature extraction. Then, a Transformer encoder is used for modeling temporal features. Finally, the data fusion is achieved through a Cross-Attention mechanism (CA). By preprocessing, feature extraction, and fusion of historical data from actual production, a power consumption regression prediction model is established. Experimental results show that the proposed model significantly outperforms traditional methods, with the coefficient of determination (R2) exceeding 98%. This research provides effective technical support for the energy consumption management of the CPE push bench mill and has high engineering application value.