Inferring Elective Activity-Trip Chaining Behaviour with Space-Time Constraint and LSTMs to Handle Structural-Zeros Problem

Population synthesis methods generally suffer from two types of missing data: sampling-zeros and structural zeros. Sampling-zeros occur when individuals exist in the population but are not recorded in the input data. Structural zeros are non-existent persons generated by the model. In the travel demand application, generating an accurate synthetic population with activity-trip chains behavior is crucial to represent population mobility. In this particular study, the utilization of the Long Short-Term Memory (LSTM) model is suggested as a means of producing activity-trip sequences. The model operates by forecasting a new sequence through the input of preceding forecasts, which then generates sequences that share similar attributes to the training set. A space-time prism concept was used to assign secondary activity when impossible sequence combinations are detected. Additionally, open-source datasets on individual characteristics and mobility were used for experimentation. This study sets out to omit the generation of structural-zeros from the synthetic population.