Fault tolerance and transferability of short-term traffic forecasting hybrid AI models

This chapter presents an evaluation of the fault tolerance and transferability performance and robustness of hybrid deep neural network short-term traffic forecasting models. The chapter focuses on the model’s performance when the input data required for model operation is missing or becomes corrupt due to field communications or transmission errors. Specifically, the tolerance of the model was evaluated in generating predictions when random, series, or mixed input data were missing using variable noise percentages up to 80%. To achieve this, the chapter first evaluates multiple AI models for both flow and speed data with full data integrity and without any corrupted or missing inputs. Field observations were collected from multiple detectors on the Eastern Freeway in Australia. The results showed that BiLSTM models performed better for variable prediction horizons for both speed and flow. Then, the chapter presents evaluations of deep and hybrid BiLSTM/LSTM models and validates their performance and potential for transferability to other detector locations along the freeway. The results showed an improved accuracy when using 4-layer BiLSTM networks. This best-performing model was then selected and tested without re-training in situations when model input variables were faulty, missing, or unknown. The results showed that when speed input data were corrupted by random noise, the model’s performance remained robust for noise levels below 40%. Performance then deteriorated when the noise level increased to 40–80%. For flow prediction, the accuracy was reliable with a noise level below 30% but started to deteriorate when the noise level increased between 30% and 80% for most detectors in both travel directions. However, when the input data were corrupted by noise, the model’s performance remained robust for noise levels below 60% but started to deteriorate with higher noise levels. For flow prediction, the prediction accuracy was reliable with noise levels below 30% but started to deteriorate when the noise levels increased to between 30% and 80% for most detectors in both travel directions. Moreover, when the input data were corrupted by mixed (random and series) noise, the model’s performance remained robust for noise levels below 60% but started to deteriorate at higher noise rates. For flow prediction, the accuracy was reliable at noise levels below 30% for most detectors in both travel directions. These results provide good insights for road operators on the impacts of incomplete data on model performance.