Source code for the paper
Joint Control of Manufacturing and Onsite Microgrid System via Novel Neural-Network Integrated Reinforcement Learning Algorithms
by Yang, J., Sun, Z., Hu, W. and Steimeister, L.
Accepted at Applied Energy.
The paper with Supplementary Materials is available here as the file MDP_paper_20220220_AppliedEnergyERevise.docx
The run files are
- experiments_comparison.py
compares the efficiency of optimal solution selected by reinforcement learning, by mixed-integer programming routine strategy and by benchmark random policy.
- mip_plot.ipynb, plot_average_experiments.ipynb
plot the comparison of total energy cost and total production throughput in units for the optimal policy and mixed-integer programming policy; also plot the average over 3 times of these experiments.
The main files are
- microgrid_manufacturing_system.py
simulates the joint operation of microgrid and manufacturing system.
- reinforcement_learning.py
reinforcement learning via two layer fully connected neural network.
- Simple_Manufacturing_System-Pure_Q-Learning.py, 1st_on.npy, 2nd_on.npy, both_off.npy, both_on.npy
learn the microgrid-manufacturing system using pure Q-learning. This is to compare with our new method.
- Simple_Manufacturing_System_routine_strategy.py
learn the microgrid-manufacturing system using routine strategy via linear mixed-integer programming.
- mip-solver.xlsx
solving the mixed-integer programming total cumulative energy cost and total production units given the mixed-integer programming solution.
The auxiliary files are
- projectionSimplex.py
proximal operator to the simplex D^c={(x_1, x_2), 0\leq x_i\leq 1, x_1+x_2\leq 1}.
- SolarIrradiance.csv, WindSpeed.csv, rate_consumption_charge.csv
1 year data in 8640 hours (360 days * 24 hours) for solar irradiance, wind speed and rate of consumption charge.
- real-case parameters-experimental-use.xlsx
the scaled real-case parameters for the manufacturing system and the microgrid used in the experiment.