<p>Lane-level traffic prediction requires high spatial resolution and is easily affected by micro-driving behavior and lane changes. It is difficult to fully model such complex dependencies by relying solely on time series and spatial graph networks. This paper applies an adaptive multi-scale spatio-temporal convolutional network and reinforcement learning collaborative optimization model AST–RLM (Adaptive Spatio-Temporal Reinforcement Learning Model), which integrates heterogeneous traffic data and constructs a dynamic graph structure to achieve precise prediction of complex lane-level traffic flow. In view of the spatio-temporal dependency characteristics of heterogeneous data, AST–RLM extracts features through different convolution kernels and dynamically adjusts them to adapt to diverse traffic modes. At the same time, based on the DQN (deep Q-network) framework, the model parameters are optimized through environmental interaction, and the prediction error is used as a reward signal to gradually improve the accuracy. Finally, the spatio-temporal convolutional network and reinforcement learning algorithm are integrated, and continuous optimization is combined with the feedback mechanism to realize the optimal prediction control of lane-level traffic flow. The outcomes show that AST–RLM has significant improvements in prediction error and stability compared to existing methods. Its prediction MAE (mean absolute error) is only 0.033 during the evening peak.</p>