In order to guarantee the vehicle safety, it is necessary to clarify the microscopic speed characteristics of the urban expressway merging area and to ensure the coordination and control of the vehicle speed in the area. First, after the full-sample high-precision vehicle trajectory data of typical multi-lane interweaving area were extracted from a wide-area view based on the UAV overhead video, the operational characteristics of vehicle speed, such as cumulative frequency, distribution trend, and characteristic percentile value, were analyzed. Then, the Bi-LSTM vehicle speed prediction model was constructed based on the LSTM model that could effectively capture the change characteristics of forward historical speed data. Considering the significant effect of manual setting of training parameters on the model prediction performance and the long time they take, the Bi-LSTM speed prediction model based on genetic algorithm optimization (GA-Bi-LSTM) was proposed. Finally, a multi-metric fusion evaluation scheme was established with seven types of evaluation metrics, namely, R², Error Mean, Error StD, MSE, RMSE, NRMSE, and Rank Correlation. The results show that the GA-Bi-LSTM speed prediction model performs better, with the fitting indicators R² and Rank Correlation r_s of 0.904 6 and 0.949 5, respectively, and the error indicators Error Mean, Error StD, MSE, RMSE, and NRMSE of 0.004 1,0.447 0,0.199 7,0.446 9 and 0.076 5, respectively. The findings can provide a theoretical basis for speed regulation in merging zones of urban expressways.