Short-term load forecasting

Scope:

Forecast 24hr horizon electrical load at 15-min intervals.

Overview:

1) Data Preprocessing

• Filled in missing values for temperature:

  Time-aware linear interpolation
  

• Fixed erroneous load data:

  Replaced zero values with day-of-week / month mean for specific interval.
  

• Dataset partitioning:

  Divided data in a) training-validation and, b) test set
  

2) Model Formulation

• EDA:

  Explored relationship of time - temperature - load
  

• Model Selection:

  Random Forest -- Perform good with non-linear data, easy tuning (given time constraint) and implementation, scalable for this application.
  
  Working hypothesis about the model: Multiple independent models perform better than an aggregate since long-term exposure would introduce trend bias.
  

• Feature creation:

  Temporal features (ie. previous day interval, morning / afternoon peak, min / max temperature)
  

3) Model tuning

• Built methodology to cross-validate

  See notebooks/02 for details.
  

4) Test model

• Evaluation class

  Takes forecaster class and test set as parameter for out-of-sample validation.
  
  Used Mean Absolute Percentage Error
  

• Sample out-of-sample prediction: MAPE = 1.8 %

5) Ongoing thoughts

• Still high variance, cross-validate features, tree parameters, increase robustness of special days.

• Temperature should be estimated through 2 variable interpolation of load-temperature.

• Explain static load on weekdays.