Bus Arrival Time Prediction with LSTM Neural Network A. Agafonov, A. - PowerPoint PPT Presentation

Bus Arrival Time Prediction with LSTM Neural Network A. Agafonov, A. Yumaganov Samara National Research University A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 1 / 16

Task definition Public transport arrival time prediction to stops Take into account different factors that characterize the traffic state Develop a distributed prediction model Task Real-time processing High accuracy A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 2 / 16

Initial data. Preprocessing GPS coordinates are obtained every 30 seconds Coordinates are fitted using information about the road network geometry and transport routes Travel times for each road link are calculated A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 3 / 16

Problem formulation S is the set of stops; R is the set of routes; N is the maximum number of route links; t dep the departure time from stop i ∈ S ; i t arr is the arrival time at stop j ∈ S ; j T travel the travel time between stops i and j . ij = t dep t arr + T travel j i ij A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 4 / 16

Feature vector: base factors To estimate the travel time T travel we used the following factors: ij day , time v i − 1 , i - travel speed on the previous route link h r - time headway to the preceding vehicle with the same route T m , r travel time of the preceding vehicle m with the same route r ij ˜ T r ij - weighted travel time of preceding vehicles with the same route: � t − t dep , k � T travel , k � k ∈ N r ω i ij T r ˜ ij = � t − t dep , k � � k ∈ N r ω i A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 5 / 16

Feature vector h any - time headway to the preceding vehicle with any route T m , any - travel time of the preceding vehicle with any route ij T any ˜ - weighted travel time of preceding vehicles with any route ij T hist ij ( t ) - historical average travel time T flow ( t ) - historical average travel time by traffic flow data ij c ij - number of vehicles on the targeted route link � ij , h any , T m , any T any � day , time , v i − 1 , i , h r , T m , r ij , ˜ T r , ˜ ij , T hist , T flow , c ij X i , j = ij ij A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 6 / 16

Long short-term memory (LSTM) cell h t C t C t-1 x + x x tanh f t � i t C o t t forget gate input gate tanh output gate h t-1 x t h t-1 x t h t-1 x t h t-1 x t h t-1 x t h t h t-1 x t A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 7 / 16

LSTM network � travel � travel T T � t t , 1 N 1, N � Output Output LSTM LSTM LSTM LSTM LSTM ... ... cell cell cell cell cell ... ... x 0,1 x 1,2 x t-1,t x t,t+1 x N-1,N . . . . . . . . . . . . . . . A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 8 / 16

Long short-term memory (LSTM) neural network Input data Output data Batch Feature Vector Route Links Mask Batch 0 0 1 ... ... 1 1 Route Links 0 0 1 0 0 0 1 1 0 0 Batch 1 1 1 ... .... 0 0 Route Links A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 9 / 16

Model analysis Comparison: Proposed / Base LSTM models ANN, 1 hidden layer Linear Regression n MAE = 1 � | V t − ˆ V t | , n t = 1 n | V t − ˆ MAPE = 1 V t | � × 100% n V t t = 1 Data set: Five bus routes Average route length is 16 km Travel time observations in 30 days A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 10 / 16

Model analysis. MAE / MAPE 1800 Table: Algorithms Comparison Predicted travel time, sec 1500 MAE MAPE 1200 LSTM 22.12 19.78 900 Base LSTM 23.64 21.24 600 ANN 25.54 23.25 Regression 26.89 25.19 300 0 0 300 600 900 1200 1500 1800 Real travel time, sec A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 11 / 16

Model analysis. MAE / MAPE for routes 35 35 30 30 25 25 MAE, sec MAPE, % 20 20 15 15 10 10 5 5 0 0 route 12 route 50 route 181 route 265 route 330 route 12 route 50 route 181 route 265 route 330 LSTM Linear Regression ANN Base LSTM LSTM Linear Regression ANN Base LSTM A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 12 / 16

Model analysis. MAE / MAPE MAE, seconds MAPE, % 200 60 180 50 160 140 40 120 MAPE, % MAE, s 100 30 80 20 60 40 10 20 0 0 0 5 10 15 20 25 30 0 5 10 15 20 25 30 Accumulated travel time, minutes Accumulated travel time, minutes LSTM Linear Regression ANN Base LSTM LSTM Linear Regression ANN Base LSTM A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 13 / 16

Model analysis. Execution time Intel Core i5-3740 3.20 GHz, 8 GB RAM / Nvidia GeForce GTX 1080 Ti 800 24 700 22 Computation time, ms Computation time, ms 600 20 500 18 400 16 300 14 200 12 100 0 10 0 128 256 384 512 0 128 256 384 512 Batch size Batch size CPU GeForce1080Ti GeForce1080Ti A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 14 / 16

Conclusion The proposed LSTM based arrival time prediction model has the following advantages: Combines different factors describing the traffic situation. It has high prediction accuracy. It has a low computation time. A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 15 / 16

Thank you! Anton Agafonov ant.agafonov@gmail.com The work was supported by the Ministry of Science and Higher Education of the Russian Federation (project no. RFMEFI57518X0177) A. Agafonov, A. Yumaganov Bus Arrival Time Prediction with LSTM Neural Network 16 / 16