Demand Forecasting of Shared Bicycles at Subway Entrances using SSA-LSTM-RF

Abstract

Shared bicycles are vital for solving the “last-mile” problem in green and sustainable urban development. However, inaccurate demand forecasting causes severe supply-demand imbalances, especially around subway stations. To address this, we propose an improved sparrow search algorithm-long short-term memory neural network-random forest (SSA-LSTM-RF) hybrid model for accurate shared-bicycle demand prediction at subway entrances. The proposed predictive model is founded on the LSTM neural network and RF. SSA is effectively employed to optimize the hyperparameters of these models, and, finally, the least squares weighting method is utilized to integrate the two models effectively. Taking the shared-bicycle order data from Zhu Guang Station in Nanshan District, Shenzhen, China, as a case study, the SSA-LSTM-RF model was benchmarked against traditional seasonal autoregressive integrated moving average and light gradient boosting machine models. The experimental results revealed that the SSA-LSTM-RF model exhibited the lowest mean absolute error and root mean square error, which were 9.8783 and 14.0117 respectively, and the highest coefficient of determination of 0.9943. These results demonstrate the superiority of the SSA-LSTM-RF model in predicting shared-bicycle demand. The model was further applied to predict the shared-bicycle demand at all subway entrances within the district. Results indicated that the model achieved a prediction accuracy of over 99% for high-demand stations, and that it performed well on datasets with high demand and significant temporal variations. This research offers theoretical support for shared-bicycle system planning, improving subway services and the urban transportation ecosystem.