Source-linked AI summary
Multi-output Bus Travel Time Prediction with Convolutional LSTM Neural Network
Niklas Christoffer Petersen, Filipe Rodrigues, Francisco Camara Pereira
TL;DR
Urban bus travel-time prediction must remain accurate and reliable despite congestion and other sources of variability that complicate traditional methods. The paper uses a multi-output, multi-time-step ConvLSTM system built from convolutional and LSTM layers to exploit urban spatio-temporal correlations. Its empirical evaluation reports superior performance to the compared methods and practical scalability for route-level retraining, while identifying route independence as a limitation.
Problem
Urban congestion and variability make accurate, reliable bus travel-time prediction difficult, although such predictions support passenger information, trip planning, and advanced ITS applications.
Method
The paper converts AVL-derived GPS data into link travel times and applies a multi-output, multi-time-step deep network combining convolutional and LSTM layers with an encoder-decoder architecture.
Results
The proposed model significantly outperforms all compared methods and can detect small irregular peaks in bus travel times quickly.
Takeaways & Limitations
The approach supports accurate predictions further into the future and can be retrained daily for an urban route using commodity hardware.
Takeaways & Limitations
The current system is route-independent and does not yet handle correlations between different routes.
Abstract
from arXiv · showhide
Accurate and reliable travel time predictions in public transport networks are essential for delivering an attractive service that is able to compete with other modes of transport in urban areas. The traditional application of this information, where arrival and departure predictions are displayed on digital boards, is highly visible in the city landscape of most modern metropolises. More recently, the same information has become critical as input for smart-phone trip planners in order to alert passengers about unreachable connections, alternative route choices and prolonged travel times. More sophisticated Intelligent Transport Systems (ITS) include the predictions of connection assurance, i.e. to hold back services in case a connecting service is delayed. In order to operate such systems, and to ensure the confidence of passengers in the systems, the information provided must be accurate and reliable. Traditional methods have trouble with this as congestion, and thus travel time variability, increases in cities, consequently making travel time predictions in urban areas a non-trivial task. This paper presents a system for bus travel time prediction that leverages the non-static spatio-temporal correlations present in urban bus networks, allowing the discovery of complex patterns not captured by traditional methods. The underlying model is a multi-output, multi-time-step, deep neural network that uses a combination of convolutional and long short-term memory (LSTM) layers. The method is empirically evaluated and compared to other popular approaches for link travel time prediction and currently available services, including the currently deployed model in Copenhagen, Denmark. We find that the proposed model significantly outperforms all the other methods we compare with, and is able to detect small irregular peaks in bus travel times very quickly.
1. Introduction
Urban bus travel-time prediction supports passenger information, trip planning, and advanced ITS applications, but congestion and other influences make reliable forecasts difficult. The paper proposes a multi-output model combining convolutional and LSTM networks to exploit spatio-temporal correlations in urban bus traffic.
- Real-time arrival and departure information is widely used through digital boards, smartphone apps, and in-vehicle infotainment systems.
- Accurate predictions support route choices, warnings about unreachable connections, and advanced applications such as bus holding and connection assurance.
- Urban bus travel times are difficult to predict because congestion, special events, roadworks, weather, and passenger demand influence traffic flow and variability.
- The model combines convolutional and LSTM neural networks to learn non-static patterns across space and time, producing travel times for all links over multiple future steps.
- The method is evaluated against popular link-prediction approaches, Copenhagen’s deployed model, and Google Traffic.
2. Related work
Earlier bus travel-time approaches include historical averages, regression, Kalman filters, and neural networks, but they often model links or correlations independently. The paper’s ConvLSTM approach combines spatial and temporal learning with multi-step, multi-output prediction while retaining practical training scalability.
- Historical-average and linear-regression models are established baselines, but historical averages respond slowly to short, impactful external influences.
- Kalman filters remain computationally cheap because they are linear, while artificial neural networks have been found to outperform independent Kalman-filter models.
- Prior hybrid and LSTM approaches address adaptation or multi-step prediction but can lose long-distance, cross-link, or fine-grained temporal correlations.
- The proposed approach combines shared convolutional filters and LSTM cells so learned patterns generalize across network links and time steps.
- Its encoder-decoder recurrent structure predicts multiple future steps with more complex temporal patterns than a fully connected layer splitting outputs.
- Training remains computationally demanding, but route models can be retrained daily on commodity hardware and distributed across computational instances.
3. Convolutional LSTM neural networks
Convolutional LSTM networks combine LSTM memory and convolutional filters to model temporal dependencies and spatial relationships across bus links. Convolutional filters reduce learned parameters while enabling shared spatial pattern detection.
- LSTM networks: LSTM networks maintain a cell state across input sequences using input, forget, and output gates to control information.The cell output is denoted h_t, and the gates produce state variables i_t, f_t, and o_t.
- LSTM networks: Stacking multiple LSTM layers enables learning more complex patterns of sequential information.The paper describes temporal patterns as an example of sequential information captured by stacked layers.
- Convolutional networks: CNNs capture local spatial relationships through shared weights applied over fixed-size input windows.This reduces learned weights relative to fully connected networks and transfers detected patterns across space.
- Convolutional LSTM: ConvLSTM applies convolutional filters in both input-to-state and state-to-state LSTM transitions.The convolution operator replaces fully connected transformations in the recurrent structure.
- Convolutional LSTM: Convolutional filters substantially reduce parameters compared with pure LSTM layers, allowing deeper networks.Each desired output uses four input-to-state and four state-to-state filters.
4. Multi-output model
The proposed model uses ConvLSTM layers in an encoder–decoder architecture to predict travel times for all links across multiple future time-steps. Inputs are arranged as fixed-resolution spatio-temporal tensors and normalized using link-, weekday-, and time-of-day-specific statistics.
- The model uses a ConvLSTM layer for multi-output, multi-time-step bus travel time prediction.
- Network topology: Its encoder contains two ConvLSTM layers, while the decoder contains two ConvLSTM layers followed by a fully connected layer.The decoder receives the last k values of the encoded sequence.
- Network topology: The architecture applies one-dimensional convolutional filters across links at each time-step and in LSTM state transitions.The two ConvLSTM blocks use filter sizes of 10×1 and 5×1 for their respective layers.
- Data preparation: Inputs are 4D tensors of size N × w × u × 1, and outputs are 4D tensors of size N × k × u × 1 for predictions over all links and future time-steps.
- Data preparation: Travel times are normalized around expected link-specific patterns by weekday and time of day, with predictions using historical rather than real-time normalization statistics.The mean and standard deviation should be calculated solely from the training set.
5. Experiments
The experiments evaluate the proposed model on Copenhagen bus-line data while examining time-resolution choices, evaluation procedures, and sliding-window testing. The setup uses 15-minute aggregation, a 32-step input window, and predictions up to 45 minutes ahead.
- Dataset: Only the first 32 links are used because the route is congestion-sensitive and splits into different destination patterns southeast of central Copenhagen.
- Time resolution: A 2-minute resolution leaves 89% of time-steps without measurements, while finer resolutions also increase training time and missing values.
- Time resolution: 15-minute aggregation yields an average of 1.7 measurements per training time-step for the 4A line, whose daytime mean headway is 7.5 minutes.
- Model setup: The fixed window uses w = 32, equivalent to 8 hours, and k = 3 enables predictions up to 45 minutes into the future.
- Evaluation: Performance is assessed with MAE, RMSE, and MAPE after summing predicted and true travel times across links, producing one vector of size k.The summed quantity represents total travel time across all 32 links.
- Evaluation: Training uses 23 weeks of data and testing uses one week at a time across four test weeks in a sliding-window procedure.
6. Results and discussion
The ConvLSTM model outperforms the evaluated baselines overall and is particularly robust during peak periods and irregular traffic patterns. Its higher accuracy comes with greater computational expense, but the authors report practical scalability.
- Overall comparison: The ConvLSTM model outperforms the historical average, Movia’s current model, pure LSTM, and Google Traffic overall.The evaluation uses daytime predictions accumulated downstream at the journey level.
- Interpretation and scalability: Improved accuracy can be larger on individual irregular journeys than averaged evaluation errors suggest, especially during vulnerable peak periods.The authors argue that complex models are scalable despite being more computationally expensive than historical averages.
- Peak-hour evaluation: Morning-peak performance does not degrade relative to daytime results, while baseline errors worsen by up to several minutes and MAPE rises roughly one third.This evaluation covers weekdays from 7h–9h.
- Peak-hour evaluation: Google Traffic narrows its afternoon-peak error gap to less than a minute, although ConvLSTM still improves over the baseline methods.ConvLSTM also decreases relative to its overall results in the afternoon peak.
- Irregular traffic patterns: On a Thursday with worse-than-normal peak traffic, historical-average models underestimate travel time and miss a small afternoon peak.The current model behaves similarly because it is based on the historical average.
- Irregular traffic patterns: Both Google Traffic and ConvLSTM approach ground truth during peaks and detect the irregular afternoon peak to some degree.Google Traffic is more accurate in the afternoon peak, while ConvLSTM is more accurate in the morning peak.
- Irregular traffic patterns: The proposed model remains best on a more typical Friday, though its advantage over the historical and current models is smaller.The Friday more closely follows the normal average pattern, especially around the afternoon peak.
7. Conclusion
The paper concludes that a multi-output, multi-time-step ConvLSTM system captures spatio-temporal variability in urban bus travel times and outperforms the compared methods. The approach is reported as scalable for independent routes, while cross-route correlations remain a limitation and future extension.
- Conclusion: The proposed system combines convolutional and LSTM layers to capture non-static spatio-temporal correlations in urban bus travel times.Its multi-output, multi-time-step design predicts across space and time.
- Conclusion: The encoder/decoder approach enables multi-time-step predictions further into the future than independently predicting subsequent time-steps.The authors describe this approach as new in bus travel time prediction.
- Conclusion: Empirical results show that the proposed model outperforms the compared state-of-the-art methods, including Google Traffic and Movia’s deployed model.The accuracy advantage is reported as more significant during peak hours.
- Conclusion: Retraining the experimental route takes less than 20 minutes on commodity hardware, supporting daily retraining and scalable deployment on independent urban bus routes.The method is more computationally expensive than simple historical-average models.
- Limitations and future work: The current route-independent design does not model correlations between different bus routes.The authors identify cross-route modeling as a research opportunity using multidimensional or graph convolutions.
- Limitations and future work: Future work includes integrating predictions with control strategies, adding contextual road-network features, and combining the model in an ensemble.Control-strategy integration is described as organizationally and technically challenging.