Demand Forecasting of Bike-Sharing Based on Isolation Forest and Bayesian Optimization Bidirectional Long Short-Term Memory Model

Abstract

Within bike-sharing system management, accurate travel demand forecasting is important to optimize resource allocation and improve operational efficiency. To this end, this study proposes an optimized deep learning model combining isolation forest (IF) with a Bayesian optimization (Bayes) bidirectional (Bi) long short-term memory (LSTM) network (IF-Bayes-BiLSTM). The model first considered the impact of outliers on the forecasting results, using the IF method to identify outliers in the data. Then, Pearson correlation analysis and linear regression were used to analyze and identify the main factors influencing the demand for bike-sharing. On this basis, forecasting models were constructed using the training set, and the model performance was evaluated using the test set. The results showed that IF-Bayes-LSTM, IF-BiLSTM, and IF-Bayes-BiLSTM had high accuracy in forecasting the demand for bike-sharing compared with traditional recurrent neural network, LSTM, and BiLSTM models. In particular, the IF-Bayes-BiLSTM model had lower values for root mean square error, mean absolute percentage error, and symmetric mean absolute percentage error than the benchmark models, whereas Pearson’s correlation (R²) was higher than the benchmark model. IF-Bayes-BiLSTM not only had a stronger generalization ability but also demonstrated higher stability and could accurately forecast the demand for bike-sharing.