In previous research on Computing power networks (CPN) load prediction, many models use channel mixing strategies to process multi-dimensional data. Specifically, these models directly project the vectors composed of each dimension of the time series into the embedding space to achieve the fusion of multi-channel information. The primary benefit of this approach is that it can substantially reduce the complexity of the Transformer model in terms of time and space, thereby enabling it to achieve a relatively ideal prediction effect even when computing resources are scarce. However, this method also has certain limitations because it fails to fully consider the trend dependence characteristics of computing load. A substantial corpus of experimental studies has demonstrated that trend dependence is a pivotal factor in enhancing the accuracy of predictions. In view of the above problems, this chapter proposes a new deep-learning model. This model is distinguished by its innovative approach to trend analysis before executing the Patch operation, which involves the segmentation of data according to its inherent characteristics. To verify the effectiveness of the model, we selected Alibaba’s publicly available cluster-trace-gpu-v2020 and Alibaba-pai-machine-metric datasets, as well as load data collected regularly in the real computing network production environment as training datasets, and carried out systematic model training. It is evident from the experimental results that the proposed method demonstrates excellent performance in computing load prediction. The objectives were successfully met, and a new solution for computing load prediction was provided.