Dual Approximate Dynamic Programming (DADP) for Stochastic Unit Commitment

Authors: Jagdish Ramakrishnan, James Luedtke

This repo contains the code and data for the paper: http://www.optimization-online.org/DB_HTML/2018/06/6672.html . The code is written in Matlab and optimization problems are solved using CPLEX 12.6 through the GAMS modeling language.

Data instances

The generator data instances are stored in the .gdx file format for ease loading into GAMS. The files can be found in the /data generator directories. The instances have the form

uc{1}_{2}md_{3}sig.gdx,

where {1} is the number of generators, {2} is a scaling factor for the mean of the demand scenarios (\mu * 10 in the paper), i.e., {2} * TotCap/10 is the mean, and {3} is the percentage variation in the mean demand (\sigma in the paper).

Running the code

Example run

This would run all algorithms (Adelman \lambda_t approach, DADP approach, and perfect information approach) for the data instance uc15_4md_15sig.gdx. Enter into Matlab console, make sure all data files are in directory, and run this command:

>>> lb_main('uc15_4md_15sig.gdx',250,50/(4 * 0.1 * 4660),(0.99).^(0:249))

The above runs stochastic subgradient method with maximum iterations of 250, a constant gradient multiplier of 50/(4 * 0.1 * 4660), and decaying learning rate of 0.99. Note here TotCap = 4660.

Short description of the main files

lb_main.m : main file which runs all three algorithms to generate bounds (i.e., Adelman \lambda_t approach, DADP \lambda_t(v_t) approach, and perfect information bound). Need to provide instance filename, max iterations / \rho constant for stochastic subgradient method, and step size vector.

max_Lg_r.m : algo for state independent Adelman \lambda_t approach

max_Lg_r_d.m : algo for DADP approach

run_per_info_mip.m : runs perfect information approach MIP

Contact

If you have questions, feel free to reach out to me. My contact details are available at https://jramak.github.io/