5G wireless networks are a major technological advancement milestone that redefines how we experience connectivity with high bandwidth, ultra-low latency, and massive device density. However, these advances have various challenges that make it difficult to maximize their potential; these include unpredictable user demand and inefficient resource allocation. To address these issues, this research proposes a predictive analytics framework that combines Long Short-Term Memory model and traditional heuristic methods. The 5G traffic dataset used reflects diurnal trends and burst events; data were normalised and features engineered from latency, throughput, and packet loss. Models were trained with MSE loss; evaluation used MAE, RMSE, latency, throughput, packet loss, resource utilisation, QoS, energy efficiency, and a stability index. In simulation with five base stations, three slices, and 100 users over 500 time units, the hybrid approach reduced median latency and packet loss versus AI-only and heuristic-only baselines, while improving throughput and QoS. Heuristic-only briefly exceeded AI-only in QoS/throughput during sharp spikes, but the hybrid method achieved the best overall metrics by combining predictive foresight with reactive correction. Results support hybrid control as a practical path to self-optimising 5G traffic engineering.