serrano-power

Repository for power consumption experiments

Setup

Please cerate a virtual environment and install the requirements:

conda env create -f environment.yml

Usage

For execution of experiments one requires first to create a YAML configuration file. Curently, we support 4 detection methods:

• AdaBoost (see ada.yaml)
• Random Forest (see rdf.yaml)
• XGBoost (see xgb.yaml)
• LightGBM (see lgb.yaml)
• CatBoost (see cat.yaml)

Configuration File

For each method, one can specify the following parameters:

• method - Name of method to be utilized, can be one of: AdaBoostClassifier, RandomForestClassifier, XGBoost, LightGBM, CatBoost

• params - Dictionary of parameters for the method, see example yaml files for examples and consul the documentation for each algorithm

• datafraction - Distribution of training data percentage to be used, each item from the distribution will result in a newly trained predictive model. User has to define start, end and number of values to generate between start and end.

• validationcurve - List of parameters as well as value ranges to test. Each parameter value will undergo cross-validation as configured from the command line.

• rfe - Recurrent Feature elimination, by default starts with 1 feature and incrementally adds rest of features until complete. By default, cross-validation as defined by command line is used. User can define own list of features to be used, for example use ranking based on feature importance.

Command line

• -h, --help show this help message and exit
• -f FILE, --file FILE configuration file (default: cat.yaml)
• -d DATASET, --dataset DATASET choose dataset, can be one of: anomaly, audsome, clean, clean_audsome (default: anomaly)
• -i ITERATIONS, --iterations ITERATIONS number of iterations (default: 1)
• -cv CROSS_VALIDATION, --cross_validation CROSS_VALIDATION number of splits used for cv (default: 5)
• -cvt CROSS_VALIDATION_TEST, --cross_validation_test CROSS_VALIDATION_TEST percent of test set (default: 0.2)
• -e EXPERIMENT, --experiment EXPERIMENT experiment unique identifier (default: exp_6)
• -v, --verbose verbosity (default: False)

##Example

• Run catboost example with anomaly dataset, one iteration and 5 fold Stratified Shuffle Split cross-validation, experiment id exp_1, verbose mode

python sr_energy_experiment.py -f cat.yaml -d anomaly -i 1 -cv 5 -cvt 0.2 -e exp_1 -v

• Run XGBoost example with audsome anomaly dataset, 10 iterations and 5 fold Stratified Shuffle Split cross-validation, experiment id exp_2, verbose mode

python sr_energy_experiment.py -f xgb.yaml -d audsome -i 10 -cv 5 -cvt 0.2 -e exp_2 -v

Note: Each experiment a new folder will be created with the name of the experiment id. In the model subfolder all reports and figures will be stored.