Neural Network Model-Based Real-Time Traffic Prediction System in Malaysia: An Intelligent Solution to Alleviate Urban Congestion

The neural network model-based real-time traffic prediction system in Malaysia aims to address the increasingly severe traffic congestion issues in the country's urbanization process. The system integrates multi-source heterogeneous data (historical traffic flow, real-time sensors, external factors, etc.) and adopts advanced architectures such as LSTM, CNN, and Spatio-Temporal Graph Neural Networks (ST-GNN) to provide traffic flow predictions for the next 15 minutes to 1 hour. It supports applications like intelligent signal control, navigation planning, and public transport scheduling, offering data-driven intelligent support for traffic management decisions while providing a reference solution for similar traffic challenges in other developing countries.

As a rapidly developing economy in Southeast Asia, Malaysia's urbanization is accompanied by severe traffic congestion: the capital Kuala Lumpur and its surrounding areas are consistently among the top in global congestion rankings, with peak-hour speeds below 20 km/h, causing billions of Malaysian Ringgit in economic losses annually and exacerbating environmental pollution and declining quality of life. Traditional traffic management relies on fixed traffic lights and manual guidance, which struggles to adapt to dynamic traffic flow. With the rise of Intelligent Traffic Systems (ITS), data-driven prediction and optimization have become key paths to alleviate congestion. This project develops a specialized real-time traffic prediction system targeting local characteristics of Malaysia such as tropical climate, diverse road networks, and unique travel patterns.

Data Collection and Preprocessing

The system integrates multi-source data:

• Historical traffic flow data: Obtained from operators like PLUS and LITRAK, including indicators such as vehicle speed, traffic volume, and occupancy rate;
• Real-time sensor data: Collected via induction loops, video surveillance, and radar to capture real-time status;
• External factors: Weather (tropical rainstorms have a significant impact), holidays, major events, traffic accidents, etc. Preprocessing includes missing value imputation, anomaly detection (with weather-sensitive mechanisms), time alignment, and feature standardization.

Neural Network Model Design

Explored architectures include:

• LSTM: Captures temporal dependencies and periodic patterns;
• CNN: Extracts spatial features and models congestion diffusion;
• ST-GNN: Combines graph convolution and temporal modeling to capture spatio-temporal dependencies simultaneously;
• Attention mechanism: Automatically learns spatio-temporal weights to improve interpretability.

The model outputs traffic flow, vehicle speed, and congestion index for the next 15 minutes to 1 hour, and provides a RESTful API interface to support integration.

Training Strategy

• Sliding window sample construction: Input historical data from the past N steps and output predictions for the next M steps;
• Loss function: Combines MSE and MAPE to balance accuracy across road segments with different traffic flows;
• Hierarchical training: First pre-train with large-scale mixed data, then fine-tune for specific regions to improve generalization ability in data-sparse areas.

Evaluation Results

Uses metrics such as RMSE, MAE, sMAPE, and hit rate:

• ST-GNN achieves an accuracy of over 85% in 15-minute predictions, significantly outperforming ARIMA and pure LSTM;
• Maintains acceptable accuracy even under extreme weather conditions.

Intelligent Signal Control

Input prediction results into an adaptive system to dynamically adjust traffic light timing, achieving preventive optimization (e.g., increasing green light duration in advance).

Navigation and Route Planning

Integrate with Waze and Google Maps to recommend optimal routes based on future traffic conditions, avoiding sections that will soon be congested.

Public Transport Scheduling

Provide passenger flow predictions for operators like RapidKL to optimize vehicle scheduling and shifts, improving service levels.

Emergency Response

When accidents or road closures occur, quickly recalculate traffic flow distribution and provide diversion suggestions to shorten congestion dissipation time.

Data Quality Issues

Sensor coverage is insufficient in some road segments; spatio-temporal interpolation is used to fill missing values, cross-validation is performed using correlations between adjacent segments, and floating car GPS data is explored for supplementation.

Emergency Event Handling

An online learning mechanism is introduced; when abnormal prediction errors are detected, the model is triggered to update quickly to adapt to sudden changes such as accidents and severe weather.

Computational Efficiency Optimization

Through model quantization, knowledge distillation, and edge computing deployment, inference latency is controlled at the second level to meet real-time requirements.

Future development directions include:

1. Multimodal data fusion: Integrate social media, mobile signaling, and Vehicle-to-Everything (V2X) data;
2. Reinforcement learning optimization: Model signal control as a Markov decision process to achieve end-to-end optimization;
3. Regional collaborative prediction: Establish a cross-city prediction network to support regional-level traffic coordination;
4. Enhanced interpretability: Develop visualization tools to explain the model's decision logic and improve trust.

The neural network-based real-time traffic prediction system provides an intelligent tool for traffic management in Malaysia. Through multi-source data fusion, advanced deep learning architectures, and optimization for local characteristics, it effectively predicts traffic flow changes and supports preventive decision-making. With the improvement of data infrastructure and technological progress, the system will play a greater role in alleviating congestion, improving travel efficiency, and supporting sustainable urban development. Its experience also provides a reference solution for similar traffic challenges in other developing countries.