Automatic approaches for Short-Term Load Forecasting

This is an experiment to test accuracy and speed of various forecasting method for Short-Term Load Forecasting (STLF). The experiment use 40 time-series, 20 from GEF2012Com and 20 from Hvaler substations. Code is written in R, tested in jupyter, and run on r-notebooks docker. R packages can be downloaded easily using install.packages() command.

To easily distingush different zones, we name zones in the two datasets differently:

• Zones in GEF2012Com dataset are called zone.1 to zone.20
• Zones in Hvaler dataset are called subs.1 to subs.20

Project Structure:

• GEFCom2012: contains data, predictions, and prediction visualizations for GEFCom2012 dataset
  • load_raw.csv: original dataset provided by GEF2012Com
  • temp.csv: original temperature timeseries provided by GEF2012Com
  • n034_ensemble.csv: prediction of the winner of GEF2012Com, used to impute the missing values in GEF2012Com
  • complete.csv: complete dataset with all 20 zones and 11 temperature timeseries (after running TidyGEFCom2012.ipynb)
  • training_set.csv: complete dataset with testing periods masked by NA values (after running MarkTestingPeriods.ipynb)
  • performance_report.csv: report MAPE for all predictions made for GEFCom2012
  • Predictions/: contains all predictions stored in csv files (and potentially visualization for predictions)
• Hvaler: contains data, predictions, and prediction visualizations for GEFCom2012 dataset
  • top_20.csv: original dataset collected from MySQL with many missing values
  • temperature_2010_2014.csv: raw temperature for Hvaler region (only one timeseries)
  • complete.csv: complete dataset with all 20 zones and 1 temperature timeseries (after running TidyHvaler.ipynb)
  • imputed_complete.csv: using Amelia to impute missing values from complete.csv (after running HvalerImputation.ipynb)
  • training_set.csv: complete dataset with testing periods masked by NA values (after running MarkTestingPeriods.ipynb)
  • Predictions/: contains all predictions stored in csv files (and potentially visualization for predictions)
• Lib: contains all the R functions that used in the experiment
• Visualizations: contains all visualizations of raw data and imputed data
• Reports: contains all report and plots used in the paper, moved here manually
• TidyGEFCom2012.ipynb and TidyHvaler.ipynb: tidy dataset
• HvalerImputation.ipynb: impute missing values in Hvaler dataset
• MarkTestingPeriod.ipynb: mark the testing period by NA values
• VisualizeData.ipynb: visualize data
• MakePrediction.ipynb: call various predict functions in Lib/ to make prediction
• RunExperiment.R: script that receive arguments and run prediction based on selected dataset, zones, and methods. Everything report to RunExperiment.out and RunExperiment.log
• RunExperiment.sh: sh script to demo how you can call RunExperiment.R
• ReportPerformance.ipynb: scan predictions folder in GEFCom2012 and Hvaler to automatically report performance and plots
• ReportRunningTime.ipynb: scan the RunExperiment.log files to produce report and plot on running time of different methods
• subs.1.traingdata.pdf and zone.1.trainingdata.pdf: plot shows data example with testing period marked