Adaptive ProbSparse-Attention with Generative Decoding for Cloud Workload Prediction

With the exponential growth of cloud computing, traditional resource scheduling strategies suffer from severe resource fragmentation due to ineffective workload forecasting. To address the limitations of existing methods in modeling nonlinear/non-stationary time series and high computational costs, this paper proposes Adaptive ProbSparse-Attention with Generative Decoding (APA-GD) - a hybrid deep learning framework integrating:An LSTM encoder capturing multi-scale temporal dependencies; An adaptive ProbSparse-attention mechanism dynamically prioritizing critical intervals via time-decay factors and top- <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$k$</tex> sampling, reducing complexity from <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{O}\left(L^{2}\right)$</tex> to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{O}\left(n_{\text {top }} \cdot U_{\text {part }}\right)$</tex>; A generative decoder enabling probabilistic multistep forecasting through autoregressive top- <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$k$</tex> sampling. Extensive experiments demonstrate significant improvements over state-of-the-art methods: On Google Cluster Trace: APA-GD reduces MAE by 46.9 % against Informer, MSE by 67.4 %, and achieves MAPE of 24.2824 %; On Alibaba Cluster Data: Our framework decreases MAE by 38.2 % compared to FEDformer, MSE by 65.9 %, while attaining 46.0 % lower MAPE. The framework establishes a new paradigm for accurate and efficient cloud resource management, reducing prediction error by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$&gt;38 \%$</tex> while maintaining 37 % faster training than Transformer-based baselines, making it scalable for industrial-scale deployments.