Ex-ante analysis and baseline estimation for pilot sites using machine and deep learning

The development of accurate models for load forecast is becoming increasingly important, following the steady electrification of domestic demand and the growing penetration of renewable generation. To this end, this work describes all the steps required for a complete energy baseline estimation, from data retrieval and exploratory analysis to the the development and testing of the actual predicting models. Specifically, we present (i) a multi-variable linear regression model which predicts the baseline energy consumption based on several independent variables, (ii) a uni-variate linear regression model with time series using the lag (time-shift) feature and, (iii) a deep learning Convolutional Neural Network (CNN) model. The root mean squared error has been chosen as the main metric to compare the performance of each model. Results show that the energy demand forecast provided by the CNN model outperforms the linear regression models as the former is able to accurately capture the spatio-temporal dependencies in buildings' load profiles, making it a suitable deep learning model for load forecast problems.