The freight volume data between regions of highway has mutual influence, and the temporal and spatial problems cannot be handled simultaneously by traditional long short-term memory (LSTM) models. An improved LSTM model (TS-LSTM) based on spatiotemporal information fusion was designed, and a method for reconstructing the dataset was proposed according to the importance of spatiotemporal information. To verify the effectiveness of the model, the highway toll system data (25 563 256 in total) for a consecutive 12-month period in a certain region was used as the original dataset, and TS-LSTM was compared and analyzed with a time-based LSTM model (T-LSTM), a space-based LSTM model (S-LSTM), a fully connected neural network, an unidirectional LSTM, a bidirectional LSTM, and the Transformer. Results showed that the performance of TS-LSTM varied across different regions, and compared to other machine learning models, the reduction range of the mean absolute error was between 40% and 85%. The mean absolute error of TS-LSTM was 10% lower than that of Transformer, and the mean absolute percentage error was 21 percentage points lower. The prediction performance of TS-LSTM were superior to those of the comparison model.