The available bandwidth of 5G networks is a pivotal factor influencing real-time video services. However, how to achieve accurate predictions in the context of low-latency real-time video services remains a difficult problem. Conventional algorithms for predicting available bandwidth typically rely on data metrics at the application layer and complete the forecast according to the packet transmission strategies. Such algorithms can lead to prediction lag in complex scenarios, thereby significantly impairing the received video quality for users. To address this concern, this study proposed a novel available bandwidth prediction algorithm based on cross-layer multi-dimensional parameters. This algorithm comprehensively integrates pertinent data metrics from the application, physical, and network layers, utilizes multiple dimensions of parameters to enhance the precision of wireless network bandwidth detection. In this paper, deep reinforcement learning was adopted as the model framework to integrate offline prediction and online prediction through cross-layer and multi-dimensional data model learning for different motion scenarios. Furthermore, network packet loss rates, image quality assessments, end-to-end delays, and other link-related factors were introduced as constraints, to realize the real-time adjustment and optimization of the prediction model during the transmission process. Experimental results conducted on a semi-physical platform show that: the prediction performance of the proposed algorithm is better than the traditional prediction method, and the fitting degree of the prediction curve and the actual curve is more than 95.8%; compared with the single-layer prediction algorithm, the packet loss rate of the proposed algorithm in walking and driving scenarios decreases by 47.3% and 30.9%, respectively, and the quality of received video is improved by 12%.