Investigation of different LSTM-based encoder-decoder neural networks for vehicle speed prediction

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Computer Science

Background:

• Vehicle speed prediction is crucial for traffic management and autonomous systems.
• Long Short-Term Memory (LSTM) networks are commonly used for time-series forecasting tasks like speed prediction.
• Evaluating the trade-off between model complexity and predictive accuracy is essential for efficient AI deployment.

Purpose of the Study:

• To investigate the necessity of complex Long Short-Term Memory (LSTM) encoder-decoder networks for vehicle speed prediction.
• To compare the accuracy and computational requirements of simple versus complex LSTM networks for this task.
• To determine if less computationally intensive models can provide sufficient accuracy for vehicle speed prediction.

Main Methods:

• Utilized simulatively generated data from a Simulation of Urban Mobility (SUMO) traffic simulation of Darmstadt, Germany.
• Investigated various LSTM encoder-decoder network architectures with differing complexity.
• Performed grid searches to analyze hyperparameter sensitivity (mini batch size, learning rate, weight decay, number of LSTM cells).

Main Results:

• Less complex LSTM models demonstrated sufficient accuracy for vehicle speed prediction.
• Complex network architectures did not consistently yield significantly better results.
• Simpler models required less computational power for training, indicating potential for energy savings.

Conclusions:

• Simple and less computationally intensive Long Short-Term Memory (LSTM) networks are viable for vehicle speed prediction.
• Prioritizing simpler models can lead to significant savings in computing power and energy consumption.
• The findings suggest a shift towards more efficient model selection in traffic prediction applications.