<p>Wireless network congestion lowers latency and throughput as traffic grows. A Double Deep Q Network for adaptive routing, a hybrid LSTM MLP model for congestion prediction, and an automated feature engineering pipeline for performance optimization comprise our integrated congestion management system. The routing component lowers reinforcement learning overestimation in dynamic environments, improving policy update stability and reliability. The prediction model captures traffic flow temporal and spatial correlations to detect congestion early. Automatic pipeline module cuts training overhead and enhances feature quality, enhancing prediction and decision making. Latency reduction, throughput, and prediction fidelity gains across several datasets demonstrate the practicality and scalability of the congestion control strategy for wireless networks. It takes network state representations as input and give out optimum routing paths with a 15–20% reduction of latency and an increase of 10–15% in throughput compared to traditional Q-learning. In the proposed method, LSTM and MLP are combined to make an accurate congestion predictor. Temporal dependencies, as captured by the LSTM, complement the ability of the MLP to deal with nonlinear transformations in a better way so that the prediction of congestion is made effectively. Inputs to the model enforce both historical traffic data and current network metrics, obtaining prediction accuracies of 85–90% with 20–25% fewer false positives compared to MLP standalone models. TPOT is employed to automatically engineer the features and select the best models. TPOT is set up with the task for feature optimization and model pipeline in a typical iterative fashion to help up the prediction accuracy by 10–15% and reduce by 20–25% in training delays.</p>