Source-linked AI summary
Short-term traffic flow forecasting with spatial-temporal correlation in a hybrid deep learning framework
Yuankai Wu, Huachun Tan
TL;DR
Short-term traffic-flow forecasting must model both spatial and temporal structure, while existing CNN and LSTM applications to traffic remain limited. CLTFP combines CNN and LSTM components to forecast future flow, and the reported results show encouraging performance alongside interpretable incremental-predictability analysis.
Problem
Short-term traffic-flow forecasting is difficult because traffic data contain spatial-temporal structure, while CNN and LSTM applications in this setting remain limited.
Method
CLTFP combines a deep CNN for spatial features with LSTMs for short-term variation and long-term periodicity, then fuses the features for forecasting.
Results
CLTFP produced encouraging forecasting results, and incremental predictability analysis showed that neural-network forecasting can provide meaningful knowledge about traffic flow.
Takeaways & Limitations
The findings indicate potential for CNN and LSTM applications in transportation and support analyzing black-box forecasting features through incremental predictability.
Takeaways & Limitations
The LSTM-captured features achieved only modest forecasting accuracy, motivating more complex approaches and improvements.
Abstract
from arXiv · showhide
Deep learning approaches have reached a celebrity status in artificial intelligence field, its success have mostly relied on Convolutional Networks (CNN) and Recurrent Networks. By exploiting fundamental spatial properties of images and videos, the CNN always achieves dominant performance on visual tasks. And the Recurrent Networks (RNN) especially long short-term memory methods (LSTM) can successfully characterize the temporal correlation, thus exhibits superior capability for time series tasks. Traffic flow data have plentiful characteristics on both time and space domain. However, applications of CNN and LSTM approaches on traffic flow are limited. In this paper, we propose a novel deep architecture combined CNN and LSTM to forecast future traffic flow (CLTFP). An 1-dimension CNN is exploited to capture spatial features of traffic flow, and two LSTMs are utilized to mine the short-term variability and periodicities of traffic flow. Given those meaningful features, the feature-level fusion is performed to achieve short-term forecasting. The proposed CLTFP is compared with other popular forecasting methods on an open datasets. Experimental results indicate that the CLTFP has considerable advantages in traffic flow forecasting. in additional, the proposed CLTFP is analyzed from the view of Granger Causality, and several interesting properties of CLTFP are discovered and discussed .
1 Introduction
Short-term traffic-flow forecasting matters for intelligent transportation systems because traffic data combine spatial dependencies and temporal patterns. The paper proposes CLTFP, combining CNN-based spatial modeling with LSTMs for temporal features and feature-level fusion.
- Accurate short-term traffic-flow forecasting supports proactive traffic control, route guidance, and location-based services in intelligent transportation systems.
- Classical statistical models can capture uncertainty and probabilistic mechanisms but often struggle with traffic-flow nonlinearity and high dimensionality.
- Neural networks offer flexible non-parametric modeling and nonlinear activation functions, motivating their use for complex traffic-flow forecasting.
- Traffic flow exhibits topological locality across nearby locations and long-term temporal memory linked to recurring travel habits.
- CLTFP combines a deep convolutional network for spatial features with LSTMs for short-term variation and long-term periodicity.
- The model fuses spatial-temporal features in a linear regression layer, adds an l1 weight constraint, and trains the network end-to-end.
- On freeway-corridor traffic data from an open dataset, CLTFP reportedly outperformed state-of-the-art methods and was analyzed using incremental predictability.
2 Related works
Prior traffic-forecasting studies used deep networks or LSTMs, but generally did not jointly exploit topological locality and long-term memory. The paper positions CNN–LSTM combination as a response to this limitation.
- Earlier deep traffic-forecasting approaches included deep belief networks, multi-task learning, stacked autoencoders, and pre-training strategies.
- These approaches achieved some forecasting accuracy but did not exploit traffic-flow topological locality and long-term memory together.
- LSTM-based traffic forecasting captured long-term memory, but its spatial dependency was not fully utilized.
- CNN–LSTM combinations had been applied in visual activity recognition, sentiment analysis, video classification, and image or video description.
- Because traffic-flow data share properties with visual data and language, prior CNN–LSTM successes suggested potential for traffic-flow forecasting.
3 Model
CLTFP combines a 1D CNN with LSTMs to model spatial, short-term temporal, and daily or weekly periodic traffic-flow features. It fuses these representations through a regression layer to forecast future traffic flow.
- Architecture: CLTFP is an end-to-end architecture combining a 1D CNN, two LSTM RNNs, and a fully connected layer.The CNN captures spatial features, while the LSTMs capture short-term and periodic features before fusion.
- Spatial features: The 1D CNN models spatial locality in freeway traffic by learning from traffic flows at locations and their neighbors.The approach treats the time dimension as channels and omits pooling layers.
- Spatial features: For complex transportation networks, conventional 1D CNNs require modification, with graph-structured CNNs suggested as an alternative.The paper notes that city-scale transportation networks can be more complex than a freeway corridor.
- Short-term temporal features: LSTM is used to generate short-term temporal features at each historical time point rather than directly generating traffic-flow predictions.This design differs from prior work that used LSTM directly for prediction and supports a deeper forecasting model.
- Periodic features: Separate LSTM inputs represent daily and weekly periodicity using time points from the previous day and previous weekday.The daily and weekly LSTM connections are added to capture both periodic feature types for forecasting.
- Feature-level fusion: CLTFP concatenates spatial, short-term, daily-periodic, and weekly-periodic features, then uses a regression layer trained with squared-error loss.Sparsity regularization on the fully connected weights is added to reduce the influence of redundant features.
4 Experiments
Experiments evaluate CLTFP on PeMS freeway-corridor data against several forecasting methods and analyze the contributions of spatial, short-term temporal, and periodic features. Results report stronger overall prediction performance for CLTFP and use Granger-Causality analysis to examine feature predictability.
- 4 Experiments: CLTFP is compared with LSTM, SAE, a shallow neural network, and GBRT using PeMS traffic flow data.The experiments also study the performance contribution of CLTFP's spatial, short-term temporal, and periodic features.
- 4.1 Datasets: Experiments use PeMS traffic flow from 33 North-bound I-405 locations, aggregated every 5 minutes, with 110000 past-future pairs for training.The study period runs from 01/04/2014 to 30/06/2015; remaining pairs form the test data.
- 4.2 Implementation: The forecasting setup uses 75 minutes of historical data to predict the next 5 minutes, with daily and weekly periodic inputs sampled around corresponding earlier times.The short-term window contains 15 time steps; periodic inputs use six points before and after 30 minutes in the previous day and weekday.
- 4.2 Implementation: CLTFP combines a three-layer 1D CNN for spatial features with LSTMs for short-term variation and long-term periodicity, followed by feature-level regression.The CNN uses 30 filters per layer, while the regression layer applies an l1 weight regularizer of 0.002.
- 4.3 Comparison results: MAPE, MAE, and ACE evaluate prediction error and spatial-distribution forecasting across multiple locations.MAE complements MAPE because MAPE can be lower for higher traffic volumes; ACE measures correlation error between predicted and actual traffic-flow vectors.
- 4.3 Comparison results: CLTFP achieves better prediction accuracy than LSTM, SAE, the shallow neural network, and GBRT in the quantitative comparison.The paper also visualizes forecasting results for CLTFP and different LASSO models at one time point.
- 4.4 Analysis of Features: Adding more feature types improves prediction results, while spatial near-term information shows the strongest contribution among the analyzed feature groups.The S, T, and P notation denotes spatial, short-term temporal, and periodic features; the combined model can outperform CLTFP on MAE, MAPE, and ACE when a proper regression model is used.
5 Conclusions and future work
The paper proposes CLTFP, a CNN-LSTM forecasting method whose results are encouraging and whose captured features can be examined for meaningful traffic-flow knowledge. It identifies modest LSTM-feature accuracy, missing auxiliary factors, and broader transportation applications as directions for improvement and extension.
- CLTFP combines CNN and LSTM for short-term traffic-flow forecasting, with encouraging forecasting results.
- Incremental predictability analysis is used to examine the black-box forecasting method and shows that neural-network forecasting can provide meaningful traffic-flow knowledge.
- The LSTM-captured features achieve only modest forecasting accuracy, motivating more complex alternatives such as convolutional LSTM.
- Weather, social events, and road state affect traffic flow, so exploiting them as auxiliary information remains future work.
- Applying the model to general transportation networks and similar spatial-temporal data are identified as extensions.