Simulate Behavior: Current or next timestep ?

[nicolas-lair]

Environment

• Grid2op version: 1.11.0
• System: windows

Bug description

obs.simulate() seems to simulate on current step and not the next one as the documentation suggests here

Thanks for the insight!

How to reproduce

Code snippet

I noticed that on rte_case_14_realistic that does not seem to support simulation further timesteps.

import grid2op
import pandas as pd
from pathlib import Path

env = grid2op.make("rte_case14_realistic")
# Simulate
sim_obs, *_ = env.get_obs().simulate(env.action_space({}))

# Load the forecast from files
load_p_forecasted = pd.read_csv(Path(env.chronics_handler.path) / f"{env.chronics_handler.get_name()}/load_p_forecasted.csv.bz2", sep=";", compression="bz2", dtype="float32")

# We would expect that when looking for the matching row, it would be row 1 that would match sim_obs.load_p but it is row 0 actually
print(np.where((load_p_forecasted == sim_obs.load_p).all(1))[0])

Current output

[0]

Expected output

[1]

[BDonnot]

Hello,

As said in the doc (eg here https://grid2op.readthedocs.io/en/latest/available_envs.html#other-environments-test-only) this environment is "test only" and it's better not to use it. It is bugged.
Please use "l2ron_case14_sandbox" instead.

Also, the behavior you describe (though not documented) sounds correct. Forecast at row 0 corresponds to the forecast available for observation 0.
Forecast at row 1 corresponds to forecasts for observation 1 etc.
This is why the forecasts dataframes and the environment dataframes have the same shape.

I might be wrong but the behavior you show sounds totally correct to me.

To confirm 100% I would advise against using the raw data of the environment and compute for example the error between the output obs.simulate and env.step for different steps and see if they "align better" (*eg reduced error) with the current step or with the next one.

[nicolas-lair]

Hello,

Thanks for the quick answer! I checked as you suggested and it tends to confirm the behavior you described.
I was indeed making the wrong assumptions that forecast at row 1 was the forecast of (and not for) observation 1.

Code

import grid2op
import numpy as np

env = grid2op.make("l2rpn_case14_sandbox")
obs = env.get_obs()
done = False
curr_error, next_error = [], []

while not done:
    sim_obs, *_ = obs.simulate(env.action_space({}))
    next_obs, _, done, _ = env.step(env.action_space({}))
    curr_error.append(np.nanmean(np.abs(obs.rho - sim_obs.rho)))
    next_error.append(np.nanmean(np.abs(next_obs.rho - sim_obs.rho)))
    obs = next_obs
    
print(f"Error to current obs: {np.mean(curr_error)}")
print(f"Error to next obs: {np.mean(next_error)}")

Output

The error is slightly

Error to current obs: 0.0049073757
Error to next obs: 0.0038013617

[BDonnot]

Hello,

Thanks for this confirmation.

And to be exhaustive on the matter, if the environment supports "multiple steps" forecast, the "forecast data" also have the same "layout" as what I described, but instead of having "only 1 row" there are multiple rows.

What I mean is that, if the environnement supports say "5 steps ahead forecasts" then the top 5 rows of the "load_p_forecast.csv" will be the forecast for observation 0, then the 5 next rows will store forecast for observation 1 etc.

So if you want, in generatl, forecast data for observation with obs.current_step = k with an environment with a forecast window of h (in general h=1, but it is not mandatory) then you need to find the [k*h:(k+1)*h]` rows.

Let me know if you have other question regarding this "issue".

If you manage to find some time to add a piece of documentation to describe this behaviour in the "grid2op.Chronics.GridStateFromFile" and / or "grid2op.Chronics.handlers.FromCSV" classes that would be awesome. Otherwise i"ll try to do it later :-)

Thanks