kristina29 / social-rl Goto Github PK

Change reward to also maximize PV generation for enhancing this

Integrate NY weather data into citylearn (wind speed + forecasts)

Adapt price time series

$$p_{new}(t)=p(t)+\alpha\frac{E_f(t)}{E_f(t)+E_r(t)}$$

with old price $p$ at time $t$, produced energy with fossil sources $E_f(t)$, produced energy with renewable sources $E_r(t)$ and factor $\alpha$

Include at least share of renewable energy

Normalize the weighted price between 0 and 1 and test how this influences the algorithms and also the new buildings (own data of PV generation)

Increase PV production of buildings in order to match mean of old buildings

Include the PV production that was used in the KPI 1-renewable energy share

Try reward used by Tolovski et al: Advancing renewable Electricity consumption with reinforcement learning

$$r(s,a)=-(Energy_{renewable}-Energy_{demand}(s,a))^2$$

Track how much of the PV generated solar energy of each building is used

Use renewable share as observation instead of absolute amount of renewable/non-renewable energy produced

Just assume everything not renewable is fossil

There is too much energy available: Every building could quite easly shift its loading such that it only uses renewable Energy. Not realistic! -> Limit available energy

Assumption: The total amount of produced energy $E_f(t)+E_r(t)$ in the fuel_mix data is equivalent to the total demand $D(t)$ at this time point $t$: $$E_f(t)+E_r(t)=D(t)$$

Idea:

• Preprocessing:
  • Calculate the share of renewables in the fuel_mix data $\frac{E_r(t)}{E_r(t)+E_f(t)}=\frac{E_r(t)}{D(t)}$
  • Calculate the median $\bar{E}_{b,net}^x$ over $x$ time step simulations of the net energy consumption of building $b$
• Training (+ Evaluation ?): At each time point
  • Calculate $E_r^b(t)=\frac{E_r(t)}{D(t)}\sum_b \bar{E}_{b,net}^x$ over all included buildings $b$ such that the relative share is the same for the reduced absolute demand

Add the factor for weighted price by share of renewable energy production into schema

Also correlate energy price to PV production

Goal: maximize use of renewable energy

Change PV generation of the building files such that they are better correlated to the DHI/DNI

KPI cannot be calculated at building-level (need to assign energy source to buildings)

Integrate fuel mix data in observations etc

Change to Kaiming initialization according to this source