@ar1 sorry for the late response; here are my answers:
- No you don need to use RL models for forecast track.
- For the forecast track you don’t have a use for the LSTM models whose parameters are stored in in the .pth files. For the control track, it is used to predict the building temperature based on some provided cooling energy every time step. The controller provides an action for the building’s cooling_device (heat pump), the action is used to calculate the timestep’s cooling demand and then that cooling demand is used as part of the input variables in the LSTM to predict the building’s temperature. You can reference this work for more information about the methodology. You don’t need to use it for anything as it is just the environment that needs it for its inner processes.
@max_langtry please, can you answer the 3rd question?
@ar1, my apologies, I’m not sure I quiet understand your 3rd point.
At each time instance your prediction model needs to take a dictionary of observations as its only argument, and return time series of predicted values over the following 48 hours for each prediction variable. The example prediction model provided in the forecasting track starter kit details the required formatting and shows you what a prediction looks like.
Does this answer your question?
Thanks for the answers
…your prediction model needs to take a dictionary of observations as its only argument…
Q1: In the baseline solution observations are (1 x Num_observations)
are we allowed to use past history of observations, let say past 24 steps, i.e. input shape (24 x Num_observations) ?
(I think the question above is about this issue)
Q2: the online evaluation (scoring) follows the same “steps” as in local_evaluation.py right? env.timesteps=720, (3600 sec1h/step)
A1: If you want to use a past history of observations then you can use a buffer which stores the observation provided at each time step, but that’s something you’ll have to implement in your prediction model
A2: The steps in the online evaluation are in exactly the same format, but I can’t remember how many their are. Though I’d highly recommend implementing your prediction model so that it can handle an arbitrary numbers of time steps. Also, in the final time step you’re still predicting for a 48hr horizon, so there are no complications around reaching the end of the simulation