In the Internet of Things network, sending a packet from the origin to the destination and optimal routing is a fundamental challenge because the packets must be sent in the optimal path with minimum delay and error rate. One of the challenges of optimal routing is the need for more consideration of the problem of congestion in the paths of sending packets in most researches. This manuscript presents a new approach based on low congestion route prediction and an optimal route selection approach with swarm intelligence for optimal routing in the Internet of Things. The proposed method uses learning based on neural network majority voting, decision tree, and random forests to predict the paths with low congestion. The proposed method uses the Jellyfish Search Algorithm(JSA) to optimize the route of sending packets. The role of the mermaid optimization algorithm is to find optimal routes in the network with minimum delay, queue length, minimum transmission error, and maximum transmission rate in the desired route. The implementation of the proposed method has been done in MATLAB software, and simulated data in the Internet of Things has been used to predict congestion. The research findings show that the routing objective function of the proposed method is more minimized than the gray wolf optimization algorithm(GWO), whale optimization algorithm(WOA), and firefly algorithm(FA), which means that the proposed method improves routing more compared to these algorithms. The remaining energy in the network nodes in the proposed method is consumed later than the GWO, WOA, and FA algorithms, and the network life in the proposed method is more extended than these methods. The proposed method for congestion prediction has accuracy, sensitivity, and precision of 97.63%, 96.85%, and 97.19%, respectively. The use of majority voting causes the accuracy of the proposed method to increase by 6.39%, 1.42%, and 3.40%, respectively, compared to the decision tree, random forest, and multilayer artificial neural network. Experiments showed that the proposed method is more sensitive and accurate in predicting traffic congestion in selected routes from decision trees, repeated trees, random trees, and clustering methods.