<p>Accurate prediction of cloud workload is essential for efficient resource allocation and capacity planning in large-scale distributed and supercomputing environments. While short-term forecasting has made notable progress, long-term prediction remains challenging due to the need to simultaneously model both short-term and long-term dependencies and capture multi-scale features. Moreover, many existing models rely on complex architectures that result in high inference overhead, limiting their practical applicability. To overcome these challenges, we propose the Wavelet Dual-Branch Network (WDBNet), a long-term cloud workload prediction model that integrates wavelet decomposition with a dual-branch channel strategy to balance accuracy and efficiency. Leveraging the observation that low-frequency components exhibit strong cross-channel correlations while high-frequency components behave more independently, WDBNet employs a dual-branch architecture: the long-term branch captures global trends and inter-channel dependencies, while the short-term branch models local variations within each channel. Extensive experiments on the Alibaba, Azure, Bitbrain, and Google datasets demonstrate that WDBNet improves prediction performance in terms of mean squared error (MSE) by 1.3%, 3.1%, 3.3%, and 2.0%, respectively, compared to state-of-the-art methods, while reducing the number of parameters by 41.3%. These results validate the effectiveness of WDBNet in achieving high predictive accuracy with reduced computational complexity. We also integrated it into auto-scaling scenarios for experimentation. The results demonstrate that WDBNet significantly reduces the number of active machines, thereby improving resource utilization efficiency.</p>