This repository use Markov model to predict real-time price bias (compare to day-ahead price) and evaluate battery state-of-charge(SoC) level. With value function of SoC, we optimize battery price abitrage using stochastic dynamic programming.
Markov models for 4 zones in NYISO are pre-trained and saved in csv files. Python scripts are available for other Markov model training. Use Matlab scripts to test different cases.
Use pre_processing_main.py script:
Use Line 10-11 to change training data size by changing year.
Use Line 12-13 to change location by changing real-time price and day-ahead-price file name.
optional_kwargs are in Line 35.
Use main_EV.m script:
For different cases, change parameters in Case Setting section.
location could be ['NYC', 'LONGIL', 'NORTH', 'WEST'].
startDay and endDay could be any dates in test dataset (2019 for current code), the startDay should be at least 1 day ahead of endDay.
Current code have several parameters for different EV usage patterns:
direction could be [0, 1], 0-charge only; 1-bidirection.
pluginPeriod could be [1, 2], which means number of plug-in period per day.
startPeriod and endPeriod are start plug-in time period and end plug-in time period, change numbers between [0,48]. When pluginPeriod = 1, startPeriod2 and endPeriod2 are not used.
eC1 and eC2 are SoC consumption during plug-off period. When eC2 is not used.
e0 is initial SoC, only used in first plug-in period of startDay.
ef is final SoC target, used in all plug-in period.
efp is the penalty for final discharge level.
Pr is normalized power rating wrt energy rating, storage duration is 1/Pr (ex. Pr=0.5, storage duration is 2hrs). Battery size is normalized to 1MWh.
eta is battery one-way efficiency.
c is marginal discharge cost - degradation, only used in bidirection cases.