The major objective of resource management systems in the cloud environments is to assist providers in making consistent and cost-effective decisions related to the dynamic resource allocation. However, because of the demand changes of the applications and the exponential evolution of the cloud, the resource management systems are constantly called into question with regard to their ability to guarantee an effective resource provisioning. To tackle these challenges, the future demand prediction is a practical solution that has been adopted in the literature. The prediction has widely relied on the CPU utilization since it is considered as a leading cause of the Quality of Service (QoS) dropping. The successful application of artificial intelligence techniques in forecasting problems motivated us to use the Kohonen Self Organizing Maps (SOM) that tries to capture the gathered empirical CPU load time series in regular behaviors to perform an accurate forecast. The proposed solution is a two-step approach that first classifies the collected data and then predicts the future CPU load. The experimental results show that our proposed system outperforms other models reported in the literature. In addition, we proved that SOM known for its strength in classification is also effective for prediction.