Self directed learning based workload forecasting model for cloud resource management

Abstract Workload prediction plays a vital role in intelligent resource scaling and load balancing that maximize the economic growth of cloud service providers as well as users’ quality of experience (QoE). Numerous approaches have been discovered to estimate the future workload and machine learning is being widely used to improve the forecast accuracy. This paper presents a self directed workload forecasting method (SDWF) that captures the forecasting error trend by computing the deviation in recent forecasts and applies it to enhance the accuracy of further predictions. The model utilizes an improved heuristic approach based on the blackhole phenomena for the training of neurons. The efficacy of the proposed method is evaluated over six different real world data traces. The accuracy of the model is compared with existing models that use different state-of-art approaches including deep learning, differential evolution and back propagation. The maximum relative reduction in mean squared forecast error is observed up to 99.99% compared with existing methods. In addition, the statistical analysis is also carried out using Friedman and Wilcoxon signed rank tests to validate the efficacy of the proposed forecasting model.