IoT applications generate multidimensional sensor data that are processed through analytical queries in cloud data centers, where variations in query complexity can cause uneven server workloads. This paper presents a machine-learning-based simulation framework for IoT data-cube workload management. The framework generates synthetic IoT data, constructs a multidimensional cube, extracts pre-execution query features (R, D, Qr , Sc), and applies a Random Forest classifier to predict Low, Medium, and High workload classes. The classifier achieves 94.00% accuracy on the held-out test set. Predicted classes are mapped to training-derived service-demand estimates and used by a workload-aware scheduler to assign tasks based on each server’s predicted remaining load. Across 20 paired simulation runs, the proposed method shows its strongest gains under High load, reducing delay by 14.64%, increasing throughput by 2.52%, improving utilization by 2.53%, and reducing load-balance dispersion by 47.59% relative to Round Robin. Under Overload, it reduces task rejection by 5.37% and delay by 1.14%. Paired tests with Holm correction confirm the significance of these principal improvements. The framework provides a reproducible basis for workload-aware IoT query scheduling in cloud data centers.