Predicting the energy consumption of a house (BDTR - E1d)

July 23, 2021
24-hour advance prediction of the electricity consumption of an average home in Spain. The Boosted Decision Tree Regression (BDTR) algorithm has been used. Given the characteristics of the experiment, other algorithms gave worse performances.
ML that predicts the daily energy consumption of a dwelling for four residents (two with presence tracking and two without tracking; this property and the number of residents can be easily modified once downloaded to your workspace by modifying the model training variables). Several regression algorithms were fitted to train the model. The algorithm that best fit the prediction data are [BDTR][1] and Decision Forest Regression ([DFR][2]), as opposed to [Linear Regression][3] or [Neural Network Regression][4]. For privacy reasons, some data from the published experiment have been removed (those related to the presence of residents; you can add as many residents as you wish and include the presence data as appropriate; you must fill in the experiment data table with the presence and number of residents you wish and modify the experiment variables according to the settings you have required; you can also delete these columns to omit the presence, but the prediction will be worse). Training with the complete data has provided excellent accuracy (**Coefficient of Determination 0.9842**; **Relative Absolute Error: 0.1085**; **Mean Absolute Error: 452.399**). The BDTR algorithm is susceptible to over-fitting, so care must be taken when configuring it. The experiment datasheet contains a large number of variables and records over an extended period to improve prediction. Pending further description of the data: 1. Data sources and explanation 2. Data processing 3. Feature engineering 4. Model description 5. Results and evaluation of model performance [1]: [2]: [3]: [4]: