While MEC alleviates infrastructure pressure and improves service quality on temporarily congested roads, a surge in requests for in-vehicle video services poses significant challenges. The limited onboard resources often lead to low cache hit rates and high latency. We therefore investigate a collaborative caching strategy that integrates user preference prediction to enhance caching efficiency in MEC. Concurrently, we propose a k-means-based Particle Swarm Optimization (KM-PSO) algorithm to solve the critical prior problem of optimal MEC deployment. A joint optimization problem is then formulated for the cooperative caching strategy in MEC, incorporating constraint related to user preferences predicted via federated meta-learning (FML), with the goal of minimizing latency and energy consumption. The problem of obtaining the globally optimal strategy is solved using our proposed Double Actor TD3 (DA-TD3) algorithm. The effectiveness of this approach is validated through experiments based on a video recommendation use case. Results indicate significant improvements in reducing average request delay, improving cache hit rate, and increasing the number of content feedbacks.