Accurately predicting future task demands based on historical data to enhance resource usage and decrease operating costs has become a critical challenge in computing power networks. Most existing models for task demand prediction in these networks rely on localized perspectives and static assumptions, neglecting the complexities of highly dynamic changes and multidimensional interdependencies under diverse task types. To address these limitations, we propose a transformer model named Beformer, its core innovation lies in the behavior-enhanced Attention (BEA) mechanism and the spectral-temporal decomposition (STD) module. The BEA mechanism, by innovatively combining behavioral perception with the channel attention mechanism, can conduct more in-depth modeling of multi-dimensional behavioral features in the input data. Meanwhile, the STD module utilizes the joint analysis of frequency and time domain to decompose the input sequence into periodic and trend components. For evaluation, we conducted experiments using the public datasets to forecast future task demands. Results demonstrate that our method achieves the lowest prediction error overall, outperforming existing approaches in both mean square error (MSE) and mean absolute error (MAE) metrics.