基于孤立森林算法贝叶斯优化的Bi-LSTM交通流量预测

Traffic Flow Prediction Based on Isolated Forest Algorithm and Bayesian Optimization of Bi-LSTM

摘要: 精确的交通流预测是建设智慧城市的主要问题之一，能为城市发展、城市建设以及交通管控等提供更科学的依据和参考。本文针对短时交通流预测模型中存在的一些不足点进行改进，设计了一种基于孤立森林算法优化下的Bi-LSTM多步交通流预测模型，该模型充分考虑了在日常交通环境下存在某些不确定因素对于相邻天数中同样时间的交通流量影响，对于原始交通流数据集使用孤立森林算法进行异常值判别和清洗，提升了原始交通流数据集的可靠性，使用贝叶斯优化算法寻找网络模型的超参数，最后利用双向长短记忆网络(Bi-LSTM)对时间序列数据的连续性和周期性的特点进行多步交通流预测。采用江苏省某地级市的道路交通流数据进行试验，试验结果表明基于孤立森林算法优化下的Bi-LSTM具备精确的交通流预测能力，在预测T + 1时刻的交通流数据上，与BP、LSTM、Bi-LSTM、EMD-LSTM和PSO-BiLSTM相比，平均绝对误差(MEA)下降了16.3265、15.5185、12.6785、6.2758和0.44，均方根误差(RMES)下降了22.7283、22.4023、14.9303、7.6905和0.5509，判定系数R²提升了7.604%、7.391%、3.701%、1.182%和0.025%。在预测T + 3时刻的交通流数据上，平均绝对误差(MEA)下降了14.9669、17.3019、8.6179、1.4647和0.1997，均方根误差(RMES)下降了21.4897、23.9447、11.9737、1.7354和0.8882，判定系数R²提升了6.589%、7.893%、2.602%、0.211%和0.11%。

Abstract: Accurate traffic flow prediction remains a crucial concern in the advancement of smart urban settings, offering a more empirical foundation and guidance for urban development, city planning, and traffic management strategies. This research tackles some of the inadequacies within short- term traffic flow prediction models and introduces an enhancement through the formulation of a Bi-LSTM multi-step traffic flow prediction model, optimized using the Isolation Forest algorithm. This model takes into comprehensive consideration the uncertainties inherent in daily traffic conditions, which influence the traffic volumes during corresponding time intervals across consecutive days. By employing the Isolation Forest algorithm, the raw traffic flow dataset undergoes anomaly detection and cleansing, thereby augmenting the robustness of the original data. The Bayesian optimization algorithm is applied to identify optimal hyperparameters for the network model. Harnessing the bidirectional long short-term memory network (Bi-LSTM), this model capitalizes on the inherent continuity and periodicity attributes of time series data to forecast multi-step traffic flow patterns. Empirical trials are conducted utilizing road traffic flow data from a prefecture-level city in Jiangsu Province. Empirical results affirm that the Isolation Forest algorithm-enhanced Bi-LSTM model demonstrates noteworthy prowess in accurate traffic flow prediction. When forecasting traffic flow at time step T + 1, in contrast to conventional BP, LSTM, Bi-LSTM, EMD-LSTM, and PSO-BiLSTM models, there has been a reduction in Mean Absolute Error (MAE) by 16.3265, 15.5185, 12.6785, 6.2758, and 0.44, respectively, as well as a decrease in Root Mean Square Error (RMSE) by 16.3265, 15.5185, 12.6785, 6.2758, and 0.5509, respectively. The coefficient of determination, R-squared (R²), has shown an improvement of 7.604%, 7.391%, 3.701%, 1.182%, and 0.025%, respectively. For the prediction of traffic flow data at time step T + 3, the MAE has been lowered by 14.9669, 17.3019, 8.6179, 1.4647, and 0.1997, while the RMSE has decreased by 21.4897, 23.9447, 11.9737, 1.7354, and 0.8882, respectively. Additionally, the R-squared (R²) value has increased by 6.589%, 7.893%, 2.602%, 0.211%, and 0.11%, respectively.