Network function virtualization (NFV) enables network operators to flexibly provide diverse virtualized functions for services such as Internet of things (IoT) and mobile applications. NFV platforms are required to offer stable and guaranteed quality-of-service (QoS)even during dynamically changing resource demands and traffic volumes. To meet the QoS requirements against time-varying network environments, infrastructure providers must dynamically adjust the amount of computational resources, such as CPU, assigned to virtual network functions (VNFs). To provide agile resource control and adaptiveness, predicting the virtual server load via machine learning technologies is an effective approach for proactive control. In this paper, we propose a traffic prediction framework based on ensemble learning, comprising weak regressors trained by ML models, such as recurrent neural networks (RNNs), random forest, and elastic net. It was observed that the prediction error tends to worsen with time because the gap of trends between the past and future traffics becomes wider. Therefore, to reduce the prediction errors, we propose an adjustment mechanism for regressors based on forgetting and dynamic ensemble. The evaluation result with real traffic data verified that the resource adjustment scheme based on the proposed traffic prediction framework keeps the frequency of over and under provisioning low, which is lesser by 45% in comparison to RNNs and autoregressive moving average (ARMA).