Study on load prediction methods of embedded container-based applications

面向嵌入式容器应用的负载预测方法研究

¹ School of Software, Northwestern Polytechnical University, Xi'an 710072, China
² AVIC the First Aircraft Institute, Xi'an 710089, China

Received: 16 May 2025

Abstract

Currently, the use of virtual container architectures in embedded computing environments is becoming increasingly popular, offering new possibilities for resource scheduling to balance load fluctuations. However, effective scheduling relies heavily on accurate load prediction, and existing research in this area suffers from a lack of dedicated datasets and inadequate adaptation of existing prediction methods to the characteristics of embedded applications. Focusing on an avionics embedded application scenario, a dataset tailored to containerized embedded environments is constructed. To address the issues of low prediction accuracy and computational inefficiency, a lightweight load prediction model is proposed by integrating the CEEMDAN algorithm with the Informer model. The CEEMDAN algorithm enhances the modeling accuracy by decomposing the time series data, while the Informer model reduces the computational complexity and memory consumption through a sparse self-attention mechanism. Experimental results demonstrate that, comparing with the mainstream time series prediction methods, the present model achieves an average reduction of about 10% in prediction errors and is well-suited for embedded application scenarios.

摘要

当前随着虚拟容器架构在嵌入式计算环境中的广泛应用, 调度资源优化得以实现对负载动态变化的自适应, 但其效果高度依赖容器负载预测的准确性, 目前相关研究面临数据集匮乏和已有预测方法不适应嵌入式应用特征等问题。面向航空机载嵌入式应用场景, 构建了符合嵌入式容器化环境的应用数据集; 针对负载预测精度和计算效率较低的问题, 提出结合了CEEMDAN算法和Informer模型的轻量级负载预测模型。CEEMDAN算法通过分解时间序列数据, 提升了建模的精度, 而Informer模型利用稀疏自注意力机制, 有效降低了计算复杂度和内存消耗。实验结果表明, 所提出的模型与主流时序预测方法对比, 各项误差指标平均下降约10%, 适合嵌入式应用场景。