Project 2 ML4Science - Time series / Event generation
Hi there!
We are doing a ML4Science project with a transportation lab and we are looking for advice about time series forecasting. We spoke with Prof. Jaggi today and he told us to post on the forum for both help here and for reference for the next lab session.
Our objective is to generate artificial data for a traffic simulator based on real traffic data. The data is a log from a sensor placed on a motorway bridge. The sensor adds an entry to the log every time a truck passes over the bridge. The entry contains a timestamp, the speed of the vehicle and some other data. We need to generate a similar log for a time in the future, which will be fed as is into the simulator.
Most time series resources we have found focus on predicting a value, such as a stock index, over a certain time step. For us, this currently implies resampling our discrete events to fixed time steps (say 5 minutes) and predicting the vehicle counts&avg speed over that time step. Once that is done, we would decompose the aggregated values into discrete events again. We evaluate performance by computing the mean absolute error between the predicted and true counts/speeds.
We would greatly appreciate any pointers for the following questions:
we just saw generative models in class today, would this be a good idea for create the fake traffic log? How would evaluation be done?
is our current method a good one? how can we make the decomposition work?
another method we thought about was to transform the timestamp column ts into [cos(ts.hour), sin(ts.hour), cos(ts.week), sin(ts.week), cos(ts.year), sin(ts.week*)] more columns. This would allow the timestamp to serve as both an input and an output. How would we extrapolate data to new time periods and how would we evaluate model performance?
Project 2 ML4Science - Time series / Event generation
Hi there!
We are doing a ML4Science project with a transportation lab and we are looking for advice about time series forecasting. We spoke with Prof. Jaggi today and he told us to post on the forum for both help here and for reference for the next lab session.
Our objective is to generate artificial data for a traffic simulator based on real traffic data. The data is a log from a sensor placed on a motorway bridge. The sensor adds an entry to the log every time a truck passes over the bridge. The entry contains a timestamp, the speed of the vehicle and some other data. We need to generate a similar log for a time in the future, which will be fed as is into the simulator.
Most time series resources we have found focus on predicting a value, such as a stock index, over a certain time step. For us, this currently implies resampling our discrete events to fixed time steps (say 5 minutes) and predicting the vehicle counts&avg speed over that time step. Once that is done, we would decompose the aggregated values into discrete events again. We evaluate performance by computing the mean absolute error between the predicted and true counts/speeds.
We would greatly appreciate any pointers for the following questions:
Thank you in advance for your help!
Add comment