MiENet: A Mixture Expert Network for Dynamic Computing Resource Prediction

Computational resource load prediction is crucial for improving system efficiency and stability in cloud computing, edge computing, and grid computing interface integration. Traditional methods have limitations in multimodal feature fusion, adaptation to dynamic non-stationary environments, and long-range dependency capture, making it difficult to meet the high accuracy and real-time requirements of intelligent scheduling. This paper proposes MiENet, a computational resource load prediction method based on the Mixture of Experts (MoE) and Selective State Space (Mamba) mechanism. This method first utilizes the Mamba attention mechanism (Mamba-Att) to efficiently extract long-range dependencies and dynamic weight distributions among multi-source temporal features, achieving high-dimensional feature refinement and low-complexity modeling. Subsequently, a Mixture of Experts network composed of a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and a transformer is constructed. A gating network is used to dynamically route and weightedly fuse different load modes, enhancing the differentiated prediction capability for complex tasks. Finally, a closed-loop mechanism integrating rolling prediction, exponential smoothing correction, and model adaptation is designed to ensure the robustness of the prediction and continuous optimization. Experiments demonstrate that MiENet achieves consistently superior predictive performance, reducing error metrics by over 10% on average compared with strong baselines, thereby providing reliable support for intelligent scheduling and resource optimization.