The Practical Byzantine Fault-Tolerant (PBFT) consensus algorithm faces challenges such as random master node selection, high communication overhead, and a lack of incentive and penalty mechanisms, which undermine its efficiency and security in large-scale network environments. To address these issues, this paper proposes an improved PBFT consensus algorithm, RF-PBFT, which integrates a random forest-based node grouping mechanism with a dynamic reputation system. The algorithm leverages key behavioral data of nodes during the consensus process to train a random forest model, partitioning nodes into a consensus set and a candidate set. Nodes with superior predictive performance are selected to participate in the consensus process, thereby reducing redundant communication overhead. Furthermore, a differentiated reputation mechanism is established for the consensus group and the candidate group to encourage positive behavior and deter malicious actions. The master node is elected through a combination of prediction results and voting, enhancing system reliability and security. Simulation results demonstrate that, compared with traditional PBFT and other state-of-the-art variants, RF-PBFT significantly reduces communication overhead, decreases consensus latency, and improves throughput, validating its effectiveness in multi-node blockchain systems.