Open DogeEngineer opened 7 months ago
How is the occupancy being calculated? I don't think it should be 0 unless there's no traffic. There is a caveat to consider if you're calculating occupancy from hi-res data, that is the detector delay/extend time alters when the controller records the start/end of the actuation (detector delay artificially deflates occupancy while extension inflates it). However, if you retrieve the occupancy directly from the controller (NTICP object 1.3.6.1.4.1.1206.4.2.1.2.5.4.1.2) then I think it should be correct.
How is the occupancy being calculated? I don't think it should be 0 unless there's no traffic. There is a caveat to consider if you're calculating occupancy from hi-res data, that is the detector delay/extend time alters when the controller records the start/end of the actuation (detector delay artificially deflates occupancy while extension inflates it). However, if you retrieve the occupancy directly from the controller (NTICP object 1.3.6.1.4.1.1206.4.2.1.2.5.4.1.2) then I think it should be correct.
Depending on which sensor you are using for occupancy. Usually system loops and advance loops show very little occupancy because vehicles just pass them without dwelling. However, stop bar detectors usually have higher and more consistent occupancy values.
Agreed, whether or not the detector only sees free flow traffic or has a queue stopped on it would make a big difference. The length of the detector also matters. In your LSTM model, were volume and occupancy modeled together, or separately? I haven't tried to do any forecasting with occupancy yet, so this is interesting to hear about.
For highway detectors, usually radars, we can predict volume, speed, and occupancy in one LSTM model. For arterial detectors, usually induction loops that handle actuation calls, we can only predict volume.
My team was able to use LSTM and GTS algorithm to predict volume, but we are not very successful in predicting occupancy. For system loops, the occupancy is usually 0 unless it is very congested, thus the lack of data to build a robust model.