18 Capstone: Forecasting the electric power consumption of a household

 

This chapter covers

  • Developing deep learning models to predict a household’s electric power consumption
  • Comparing various multi-step deep learning models
  • Evaluating the mean absolute error and selecting the champion model

Congratulations on making it this far! In chapters 12 to 17, we dove headfirst into deep learning for time series forecasting. You learned that statistical models become inefficient or unusable when you have large datasets, which usually means more than 10,000 data points, with many features. We must then revert to using deep learning models, which can leverage all the available information while remaining computationally efficient, to produce forecasting models.

Just as we had to design a new forecasting procedure in chapter 6 when we started modeling time series with the ARMA(p,q) model, modeling with deep learning techniques required us to use yet another modeling procedure: creating windows of data with the DataWindow class. This class plays a vital role in modeling with deep learning, as it allows us to format our data appropriately to create a set of inputs and labels for our models, as shown in figure 18.1.

18.1 Understanding the capstone project

18.1.1 Objective of this capstone project

18.2 Data wrangling and preprocessing

18.2.1 Dealing with missing data

18.2.2 Data conversion

18.2.3 Data resampling

18.3 Feature engineering

18.3.1 Removing unnecessary columns

18.3.2 Identifying the seasonal period

18.3.3 Splitting and scaling the data

18.4 Preparing for modeling with deep learning

18.4.1 Initial setup

18.4.2 Defining the DataWindow class

18.4.3 Utility function to train our models