some questions from power consumption data

[JunPyoL22]

Hi,
I have some questions and mysterious things about the results on power consumption data(whole power consumption, ITE power consumption, HVAC power consumption).

Here is my graphs.

These datas are from eplusout.csv after running the "2ZoneDataCenterHVAC_wEconomizer_Temp_Fan.idf" file with CA weather and Chicago weather for the purpose of checking how the power consumption is changed as the episode goes on.
(three episodes in the graph are 1st, 15th, 30th)

My first thought was the ITE power consumption is stable every episodes and HVAC doesn't because rl-agent just can control the HVAC system and ITE is just operated by the schedule defined in the idf.

But, the results are totally different.

What do you think about these results?

Does data on The HVAC power consumption from eplusout.csv after applying rl-agent's actions or just internal (enegyplus) result (not applying rl-agent's actions) ?

Why the ITE power consumption is not stable or does not show the same pattern every episode?

[takaomoriyama]

@JunPyoL22, thank you for reporting. It's quite interesting.
First of all, could you confirm how you plotted the result because your result looks very different from my case show below, which is the result of training of 15 episodes with CA data.

How many days of simulation result is shown on each of your graph ?

[JunPyoL22]

I ran the program with --num-timesteps 2000000
The data is from "home/eplog/openai-2018-10-xx-xx-xx-xx-xxxxx/output/episode-000000/eplusout.csv.gz" and averaged by date.
I plotted the mean value of those power consumptions.
I am not sure the HVAC power computation from the path is the result to which is applied rl-agent's action.
And I confirmed that the data center server power defined in the 2ZoneDataCenterHVAC_wEconomizer_Temp_Fan.idf is a function of temperature.
If then, would the ITE power power consumption vary due to the fact that the temperature is differ from each episode?

Data Center Servers Power fLoadTemp, !- Name
-1.0, !- Coefficient1 Constant
1.0, !- Coefficient2 x
0.0, !- Coefficient3 x2
0.06667, !- Coefficient4 y
0.0, !- Coefficient5 y2
0.0, !- Coefficient6 x*y
0.0, !- Minimum Value of x
1.5, !- Maximum Value of x
-10, !- Minimum Value of y
99.0, !- Maximum Value of y
0.0, !- Minimum Curve Output
99.0, !- Maximum Curve Output
Dimensionless, !- Input Unit Type for X
Temperature, !- Input Unit Type for Y
Dimensionless; !- Output Unit Type

[takaomoriyama]

Thank you for your comment.
I computed the average temperature of the inlet node of West Zone, which we can find in output/episode-NNNNNNNN/eplusout.csv.gz file. Then computed "Data Center Servers Power fLoadTemp" following your suggestion.

Method	West Zone Inlet Temp.	f(Load, Temp)	Reduction
baseline	16.19	1.079
RL(#39)	13.58	0.823	-23.7%

The result may imply that the source of power reduction comes from use of cooler air. I think it is worth investigating further.