In this paper, we present our research that goes through two steps: (1) using meta-heuristic optimization for global space search; (2) applying the proposed optimization to multivariate workload modeling and prediction. In the first step, we pay attention to the improvement of the Queuing Search optimization by the space-walk combination of Levy-flight trajectory to improve population diversity and Opposition-based learning to speed up the convergence process. To evaluate our solution’s effectiveness, we compare it with six well-known optimization algorithms using CEC 2014 benchmark functions. The achieved results show the significant effect of our nQSV designs in avoiding local optima and speed up the convergence process. In the second step, to prove the feasibility of solving real problems, we apply nQSV to train a neural network to model multiple variables of distributed workload simultaneously. The model is called nQSV-Net as the whole. The gained outcomes from extensive experiments with three real datasets show the accuracy and stability of nQSV-Net as a solution in the domain.