DDF-Net: A Cloud Computing Load Forecasting Method Integrating Spatiotemporal and Time-Frequency Domain Information

Cloud computing load trend forecasting is a crucial component in the construction of load balancing systems. However, existing methods fail to sufficiently extract sequence information and capture correlations among multidimensional metrics due to the non-periodicity and non-stationarity of load data. We propose the Dual-domain Dynamic Forecasting Network (DDF-Net), which designs a univariate evolution model to dynamically adjust and guide multivariate predictions through a multi-task joint learning approach. First, a spatiotemporal graph convolutional network with an additional graph learning layer is employed to capture and extract correlations among multivariate metrics, addressing spatiotemporal domain information. Second, an improved empirical wavelet transform and a dissimilar moving averages component are designed to effectively decompose load features, and a BiGRU with an integrated attention mechanism is used to capture long-term dependencies, addressing time-frequency domain information. Ultimately, an attention mechanism component that integrates latent representations from both the spatiotemporal and time-frequency domains is designed. Experiments conducted on load datasets from three real-world platforms demonstrate that the proposed model achieves impressive performance in both single-step and multi-step forecasting tasks.