This repository provides the computational experiments to accompany the paper: Lee, Jonathan T., Anderson, Sean, Vergara, Claudio, and Callaway, Duncan. "Non-Intrusive Load Management Under Forecast Uncertainty in Energy Constrained Microgrids." Electric Power Systems Research, (In Press).
This code is made available under the MIT license. We request that any publications that use this code or follow the methodology therein cite the paper above. Please contact Jonathan Lee by email at jtlee@berkeley.edu or through Github if you have difficulty accessing a copy of the paper.
Please use the Github issue tracker to post any problems encountered using the project, and we will address them.
The code requires having MATLAB 2018A or higher. The full functionality also requires having Gurobi 9.0 installed with a license and configured for use in MATLAB. Earlier versions used CVX version 2.1 in development, but the latest version does not depend on CVX.
Please be advised that this repository includes more than 1 GB of data from simulation results associated with the paper so cloning may take some time.
This repository depends heavily upon microgrid-dispatch-simulator for simulation and upon computational-experiment-matlab for the design of computational experiments. The specific branches of these projects are included here as sub-repositories. Note that to clone these, you need to clone with git clone --recurse-submodules.
Be sure to use git clone --recurse-submodules when cloning to include the microgrid dispatch simulator and computational experiment projects, i.e. "git clone --recurse-submodules https://github.com/Energy-MAC/pscc2020-load-limiting".
At the start of a session, load the relevant directories to the MATLAB path by running init.m.
The main entry point to run an experiment is bin/runExperiment.m. As an example, run:
runExperiment('controllerPerformance','sample')
This will execute a reduced experiment used in the paper, but with only 10 trials, which should complete in a couple of minutes. The full experiment is run with runExperiment('controllerPerformance'), which will likely take multiple days.
To execute both the performance and timing experiments, and reproduce the paper results you can run scripts/runExperimentsPaper.m. To run a reduced sample, you can run scripts/runExperimentsSample.m.
To improve performance and execute experiments standalone, you can build an executable runExperiment.exe using the script build.m. This requires the MATLAB mcc compiler.
This repository includes the output data used in the paper. Reproduce the figures in the paper with scripts/generatePaperFigures. This will write new figures into the figures directory.
This repository provides the computational experiments to accompany the above paper, and also serves as an example usage of microgrid-dispatch-simulator and computational-experiment-matlab, with the caveat that this repository uses specific versions of these projects.
For context on the principles used in creating these experiments and the terminology used in describing them, we recommend reviewing the methodological paper: Lara, José Daniel, Lee, Jonathan T., Callaway, Duncan, and Hodge, Bri-Matthias. "Computational Experiment Design for Operations Model Simulation." Electric Power Systems Research, (In Press). This repository provides specific instances of computational experiments in the directory experiments which inherit from the abstract experiment defined in computational-experiment-matlab. In the terminolgy of the experiment design paper, microgrid-dispatch-simulator for simulation and upon computational-experiment-matlab provides the "emulator model". For details on this model, please refer to the Non-Intrusive Load Management paper and the microgrid-dispatch-simulator README.
The experiment ControllerPerformanceExperiment.m declares the experiment workflow, which includes the data process, the simulation process, and the results and reporting process.
The Experiment superclass from computational-experiment-matlab executes the workflow by calling these methods in loops, performing the iterations over trials, saving output data, managing the random number generator, etc.
The raw data used to define the experiments are defined in .csv files in the data directory.
Specific data for each type of experiment are defined in data/experiments/inputs/[experimentName]. In this case, there are two experiments ControllerPerformance and ComputationTime, and thus a directory for each. Each of these experiments has a default instance, or a case. Cases are used to run the same experiment workflow, but with different input data. The cases given here are sample, whose purpose is to provide an example with a reduced number of trials and shorter simulations that will run quickly, and test, which is even more reduced for the purposes of testing. So, for the ControllerPerformance experiment, the main experiment parameters for the default case are defined in data/experiments/inputs/controllerPerformance/key_values.csv. For the sample case, they are defined in data/experiments/inputs/controllerPerformance/sample/key_values.csv. Additional experiment parameters, which are in list format, are defined in separate .csv files. In this case, the only ones are in the controllers.csv file, which simply defines the names of the controllers to compare as independent variables (data/experiments/inputs/controllerPerformance/controllers.csv. Finally, there are experiment parameters in the data/common directory that apply to all parameters. These include definitions of parameters for the power systems equipment, customer models, and weather data.
The three key steps in parsing the data are generating the experiment parameters, the confounding variables, and the independent variables (also called the treatment variables). This is done in the methods setupAdditionalParameters, generatingConfoundingVariables, and generateTreatmentVariables, respectively. These methods parse the data found in the data folder and apply additional logic to generate synthetic data or expand the raw data set programmatically. Some of these expansion processes are random, but are seeded so they can be reproduced.
The simulation is executed by calling the method simulateTreatment in the ContorllerPerformanceExperiment class. This takes a test-set as an input, comprised of the specific confounding variables for that trial and independent/treatment variables (the general experiment parameters, which are unchanging, are stored in the property ExperimentParams).
This calls simRHC from microgrid-dispatch-simulator to simulate the receding horizon control process and returns the output data. Please see this repo for details on using the simulation package and examples to run it in other contexts.
A M x 2 cell array Metrics defines each of M metrics as a pair of a string defining the name, and a function handle for computing the metric. Each of these functions require the output data and its associated test-set, and return a scalar metric. Thus, each of the metrics can be computed for each trial and for each independent variable. These are the values used to generate the plots shown in the paper. A metric can return NaN if it cannot be computed for that particular test-set.