To improve the accuracy of highway traffic flow prediction and provide effective support for the dynamic control induction of highway management departments, this paper takes the minimum prediction error of real-time traffic flow as the goal and divides the highway data into four different time intervals by cleaning and normalizing. The data set was divided into training data sets and test data sets. Genetic algorithm (GA) was used to optimize the parameters of data time window step as well as long-term and short-term memory (LSTM) neural network of hidden layer, training times, and dropout, and analyze the influence of four parameters on model optimization. In keras, GA-LSTM model utilized Tensorflow as a background for training and fitting. Experiments show that the GA-LSTM model had a fast search speed. Compared with the SVM, KNN, BP, and LSTM neural networks in the traditional prediction algorithm, the GA-LSTM had the minimum mean square error and root mean square error for data prediction, and the model exhibited better predictive performance.