Cloud computing is an emerging technology and it is growing exponentially. Cloud computing delivers shared computing resources such as servers, storage, databases, networking, applications, platforms and other resources over the internet. These services are provided in the form of Infrastructure as a Service (IaaS), Software as a Service (SaaS), and Platform as a Service (PaaS). The biggest challenge for cloud data centers and servers is how to handle and service user requests that are arriving frequently from end users.rnLoad balancing enables to achieve a maximum resource utilization, providing high quality of services, improving performance and prevents the system from failure by distributing workload across one or more computing resources. Existing load balancing approaches could be categorized as static or dynamic load balancing algorithms. Both static and dynamic load balancing algorithms suffers from virtual machine migration during load balancing process which causes significant performance degradation and overhead.rnRecently, resource management driven by forecasting the future workload is proved to be efficient approach to achieve effective dynamic resource scaling, automatic resource provisioning, energy efficient computing, virtual machine migration and scheduling.rnArtificial neural network (ANN) trained by Backpropagation (BP) has been used as a load balancing mechanism in cloud computing to distribute workloads. However, ANNs trained with gradient descent algorithms such as popular BP are associated with four distinct limitations; i.e., (1) easily trapped in local minima; (2) slow convergence speed; (3) convergence depends on choice of momentum, learning rate and weights; and, (4) vanishing and exploding gradient problems.To avoid the above problems a nature inspired swarm intelligence algorithms has been used to train ANN as an alternative training methods. Continuous Ant Colony Optimization (ACOR) is a non-gradient, meta-heuristic swarm intelligence algorithm that has been used to train ANNs.rnIn this research a load balancing mechanism based on workload forecasting which utilizes historical dataset is proposed. ANN optimized by ACOR algorithm is used to predict future workload based on historical workload information.rnTo evaluate the proposed approach, we designed and implemented a system based on two GWA-T-12 Bitbrains FastStorage and Rnd workload trace datasets. FastStorage, consists of the individual traces of 1,250 VMs and the rnd datasets consist of individual traces of 500 virtual machines. The dataset includes information about CPU utilization, memory usage, network usage and storage usage of virtual machine.rnThe system is tested using four types of information extracted from the datasets using Random forest feature selection algorithm. Based on their scores the top six important features, i.e. (CPU provisioned, CPU usage in megahertz, CPU usage in percentage, memory usage, disk read, and network transmitted throughout) are selected. Using this dataset ten second, fifteen second, and twenty second window size dataset are prepared for evaluation. With respect to prediction window size the performance of the model is tested and the effect of different observation window sizes on the performance of the model is evaluated. The best performance achieved by the proposed system is 96.3% accuracy on Faststorage CPU dataset and 95.2% accuracy on rnd CPU dataset. Finally, we have used other swarm intelligence training mechanism and tested the performance using all versions of the datasets. The best performance is achieved when utilizing Particle swarm (ANN-PSO) on FastStorage CPU datasets. Particle swarm (ANN-PSO) achieved an accuracy of 99.19% on training and 98.9% on testing while using FastStorage CPU datasets.