Abstract—Vehicular-to-Everything (V2X) communication enables data exchange between vehicles and infrastructure, playing a crucial role in autonomous and connected driving systems. However, growing vehicular density often leads to Radio Access Network (RAN) congestion. To mitigate this, Multi-Radio Access Technology (Multi-RAT) networks offer diverse services and data rates, enhancing throughput and reliability, but they also introduce the challenge of effective Load Balancing (LB) among heterogeneous RATs. Conventional LB approaches, such as those based on Received Signal Strength Indicator (RSSI) or Time-to-Leave (TTL), struggle under dynamic mobility and service variability. This paper proposes two novel LB frameworks: the Multi-Service Caching-Based Load Balancing (MSCLB) and the Intelligent Multi-Service Caching-Based Load Balancing (IMSCLB) schemes. The IMSCLB integrates service-aware caching with a CNN-LSTM predictive model to forecast RAT load and optimize resource allocation adaptively. Simulations conducted across three RATs (IEEE 802.11n, 802.11ac, and 802.11ad) and three service categories (Safety, Traffic, and Information) demonstrate substantial improvements over traditional methods. Specifically, MSCLB improves throughput by 10 Mbps, boosts packet delivery ratio by 10%, and reduces latency by up to 6%, while IMSCLB further enhances performance consistency, achieving satisfaction rates of 99.89%, 99.99%, and 99.99% for 802.11n, 802.11ac, and 802.11ad, respectively. These results validate the effectiveness of integrating service-aware caching and predictive intelligence for efficient load balancing in heterogeneous vehicular networks.