MATLAB scripts to evaluate the Adaptive Real Time Exploration and Optimization (ARTEO) algorithm using a simulated refrigeration plant.
The refrigeration plant is simulated in Simulink. It has 5 parallel compressors, as shown in the diagram below.
The model was developed by Mehmet Mercangöz and coworkers at Imperial College, London, based on work by K. N. Widell, and T. Eikevik (2010) at Norwegian University of Science and Technology.
This work is part of a research project supervised by Dr. Mercangöz. The results here have been submitted as a conference paper proposal (accepted) with the following title:
- Using Prior Knowledge to Improve Adaptive Real Time Exploration and Optimization
After downloading the code, I recommend running the unittests first (MATLAB command line from the main repository directory):
runtests
Note: There are two tests which may fail. test_run_simulation.m always fails when executed by runtests because it runs a Simulink model. To do this test, open the script and run it the normal way (not as a unit test) in MATLAB. The test_ens_model.m fails because this model has not been implemented yet.
First, run this script to compute the optimal machine loads for all machines (warning: This takes about 15 minutes):
find_optimum_solution
Once this script has finished, the optimum load solutions are saved on file so it shouldn't need to be run again.
Then, generate the random simulation input sequences:
gen_input_seqs
Then, open the script run_simulations.m. In the top part of this script you can choose from a set of test simulations to run:
% Choose simulation sub-directory name where config files, data,
% are located and results will be stored
% sim_name = "test_sim_gpr"; % Gaussian process models
% sim_name = "test_sim_fp1"; % Simple first-principles model
% sim_name = "test_sim_lin"; % Linear model
sim_name = "test_sim_true"; % test optimizer with true system models
The sim_name variable refers to a sub-directory in the 'simulations' directory which contains a config file that defines the
simulation and load optimizer setup.
If you run the above simulation, you should get the following output:
Starting single simulation...
Start time: 11:18:55
Loading simulation configuration from 'simulations/test_sim_true/sim_specs/sim_spec.yaml'
Loading system configuration from 'simulations/test_sim_true/sim_specs/sys_config.yaml'
Loading optimizer configuration from 'simulations/test_sim_true/sim_specs/opt_config.yaml'
Starting simulation...
0 2.128e+06 +6.637e+02 -0.000e+00 = 2.128e+06 [ 93 537 403 795 403]
250 2.128e+06 +6.637e+02 -0.000e+00 = 2.128e+06 [ 93 537 403 795 403]
500 2.128e+06 +6.637e+02 -0.000e+00 = 2.128e+06 [ 93 537 403 795 403]
750 2.128e+06 +6.637e+02 -0.000e+00 = 2.128e+06 [ 93 537 403 403 795]
1000 2.071e+06 +6.232e+02 -0.000e+00 = 2.071e+06 [ 89 537 387 387 795]
1250 2.071e+06 +6.232e+02 -0.000e+00 = 2.071e+06 [ 89 537 795 387 387]
1500 2.071e+06 +6.232e+02 -0.000e+00 = 2.071e+06 [ 89 537 387 795 387]
1750 2.071e+06 +6.232e+02 -0.000e+00 = 2.071e+06 [ 89 537 387 387 795]
2000 2.496e+06 +1.461e+08 -0.000e+00 = 1.486e+08 [ 92 367 399 795 795]
2250 2.496e+06 +1.461e+08 -0.000e+00 = 1.486e+08 [ 92 367 399 795 795]
2500 2.496e+06 +1.461e+08 -0.000e+00 = 1.486e+08 [ 92 367 399 795 795]
2750 2.496e+06 +1.461e+08 -0.000e+00 = 1.486e+08 [ 92 367 795 795 399]
3000 1.935e+06 +6.277e+02 -0.000e+00 = 1.935e+06 [ 98 376 421 421 795]
3250 1.935e+06 +6.277e+02 -0.000e+00 = 1.935e+06 [ 98 376 421 421 795]
3500 1.935e+06 +6.277e+02 -0.000e+00 = 1.935e+06 [ 98 376 421 421 795]
3750 1.935e+06 +6.277e+02 -0.000e+00 = 1.935e+06 [ 98 376 795 421 421]
4000 2.049e+06 +6.081e+02 -0.000e+00 = 2.050e+06 [ 88 537 381 795 381]
Simulation finished.
Simulation results saved to 'simulations/test_sim_true/results/sim_out.mat'.
Max. power limit exceedance: 1 kW
Avg. power limit exceedance: 0 kW
Avg. load tracking errors vs. target: 97 kW
Avg. load tracking errors vs. max.: 1 kW
Avg. excess power used: 0.523976 kW
Avg. excess power used: 0.0% (of total)
Final total model uncertainty: 0.0
Final overall model prediction error (RMSE): 0.0 kW
Number of times optimizer failed: 0
Summary saved to file:
simulations/test_sim_true/results/sims_summary.csv
After running run_simulations.m, you may run plot_model_preds.m to make various plots of the simulation results and the fitted models.
Plot figures are saved to a sub-directory called 'plots' within the simulation sub-directory.
The outputs of the simulation are written to a sub-directory called 'results' in the simulation sub-directory.
Details of every simulation (configuration, optimizer parameters, evaluation results etc.) are appended to a summary file named 'sims_summary.csv'. Note that this file is not over-written, so it is useful for accumulating and comparing results from multiple simulations. To delete past simulation results, just delete all files in the results sub-folder of the simulation directory.
If everything appears to be working, you can try running the main evaluation simulations.
Before running these, you need to generate the simulation specification files, which will be saved in simulations/sim_all_eval/sim_specs/queue by running
gen_sim_specs_all_eval
Then, set sim_name = "sim_all_eval" in run_simulations.m and run it. Warning: There are 110 simulations which take about 5 hours to complete.
Once complete, you can run the following script to produce the box-plot shown in Fig. 10 in the paper:
analyse_results_all_eval.m
- make_comp_curve_plots.m - Fig. 2. Steady-state characteristics of compressors.
- find_optimum_solution.m - Fig. 3. Optimum load allocation and power consumption.
- make_gpr_example_plots.m - Fig. 4. Predictions with different types of model.
- make_popt_w_plot.m - Fig. 5. Effect of w parameter on optimizer performance.
- make_popt_z_plots_all.m - Fig. 6. Effect of z parameter on optimizer performance.
- gen_input_seqs.m - Fig. 7. Target load sequences.
- plot_model_preds_mult.m - Fig. 8. Optimizer simulations – load and power.
- plot_model_preds_mult.m - Fig. 9. Optimizer simulations – evaluation metrics.
- analyse_results_all_eval.m - Fig. 10. Optimizer performance metrics.
- load_opt.m - S function used in Simulink model to simulate load optimizer.
- LoadObjFun.m - Objective function used by load optimizer
- MaxPowerConstraint.m - Constraint function used by load optimizer
- gpr_model_setup.m - Initializes a GP model
- gpr_model_predict.m - Make predictions with a GP model
- gpr_model_update.m - Updates a GP model
- lin_model_setup.m - Initializes a linear model
- lin_model_predict.m - Make predictions with a linear model
- lin_model_update.m - Updates a linear model
- fixed_poly_model_setup.m - Initializes the true system model
- fixed_poly_model_predict.m - Make predictions with the true system model
- fixed_poly_model_update.m - Updates the true system model
Simulink model file:
Other scripts:
- evaluate_models.m - runs Monte Carlo experiments to evaluate the extrapolation behaviour of each model type when trained on small samples of random training points
Utility functions:
- yaml - contains a copy of a package by Martin Koch (2023) for reading and writing Yaml files
- RandPtsInLinearConstraints - contains a function by Cheng (2023) to generate random points in constrained space
- data-utils - various data processing tools - copied from https://github.com/billtubbs/ml-data-utils
- plot-utils - various plotting tools - copied from https://github.com/billtubbs/ml-plot-utils
- gen_seq_brw01.m - generates a bounded random walk (BRW) sequence
To run the unit tests, execute the following command in the main directory from MATLAB.
runtests
- Buse Sibel Korkmaz, Marta Zagórowska, Mehmet Mercangöz (2022), Safe Optimization of an Industrial Refrigeration Process Using an Adaptive and Explorative Framework, arXiv preprint arXiv:2211.13019, (https://arxiv.org/abs/2211.13019)
- Cheng (2023). Generate Random Points in Multi-Dimensional Space subject to Linear Constraints (https://www.mathworks.com/matlabcentral/fileexchange/36070-generate-random-points-in-multi-dimensional-space-subject-to-linear-constraints), MATLAB Central File Exchange. Retrieved February 27, 2023.
- Martin Koch (2023). yaml (https://github.com/MartinKoch123/yaml/releases/tag/v1.5.3), GitHub. Retrieved February 28, 2023.