| With the passing of the passenger flow cultivation period and the advent of the era of networked operation,urban rail transit has gradually become the traffic artery of large cities.And passenger flow forecasting is the key technology to realizing safe operation,intelligent dispatch,efficient tranship,and humanized service of urban rail transit.Based on situation awareness methodology and time series forecasting theory,this paper proposes a complete set of short-term inbound passenger flow forecasting models through methods such as data mining,feature engineering,time series decomposition,and deep learning,et al.Firstly,the related concepts of short-term passenger flow are clarified,and the advantages,disadvantages and limitations of the existing passenger flow influencing factors are analyzed.From the perspective of probability theory,the effectiveness of time series decomposition in improving prediction accuracy is proved,and the paradigm of predicting short-term passenger flow with one step size(5 minutes)based on historical values and situational characteristics is determined.In addition,a "pyramid" conceptual model of time series is proposed,which includes "components explainable by historical time series","components explainable by external features",and "random interference components",specifically.Secondly,clean the "dirty data" and redundant information in the four original datasets of Shenzhen,and recover the missing station names through data mining.Based on station coordinate crawler,station coverage calibration,distance and azimuth calculation on earth surface,68 traffic situation features around 166 rail transit stations are extracted by using Python’s multi-process.Then,the outlier processing,detrending and standardization are carried out for each feature sequence,and the manifold learning algorithm,t-SNE,is used to reduct feature dimension.Subsequently,the modal aliasing problem of the EMD algorithm is verified and the reconstruction error contradiction of the EEMD algorithm is proposed.Thus,the complete CEEMDAN algorithm is selected to decompose the time series to separate the "components explainable by historical values" and "components explainable by external features".After decomposition,the former is concentrated in subsequence IMF3 and subsequent smooth sequences,the latter is concentrated in high-frequency sequences like IMF1 and IMF2 which has strong volatility.At the prediction level,the ARIMA model based on statistical learning,different recurrent neural networks,and different convolutional neural networks are deeply modeled,and their learning ability and generalization performance are analyzed.Under the condition of no situation features,the CEEMDAN-ARIMA/RNN model is proposed.Through order calibration or training of different models,the prediction errors are:IWMAPE=11.70%,RMSE=35.07,MAE=19.62,MAPE=18.65%.Among them,the RNN model is applied to IMF1 and IMF2,and the ARIMA model is applied to IMF3 and later sequences.However,gated recurrent networks such as LSTM or GRU,which can avoid gradient disappearance,have some degrees of overfitting.Take the situation features into consideration,through the research on the residual structure and the "Squeeze-Excitation" module in SENet and DRSN,a "Cube-Learn" algorithm based on the attention mechanism is designed.The deep framework has both channel weighting and noise resistance capabilities,and is good at dealing with time series prediction problems instructed by multi-dimensional features,which can further improve the prediction accuracy of IMF1 and IMF2.Through cross-validation,regularization,early stopping and other strategies,the prediction errors of the CEEMDAN-ARIMA/Cube-Learn model are: IWMAPE=10.56%,RMSE=32.31,MAE=17.44,MAPE=17.76%.Finally,the effectiveness,necessity and robustness of the method are further confirmed by ablation experiments. |
