Trees vs Neurons: Generalizability for Electrical Consumption Forecasting in Educational Institutions

Abstract

In this work, the generalizability was evaluated by comparing performance of Random Forest (RF) and Multi-Layer Perceptron (MLP) models on different data training and test. The main goal of this work is to find the best performing approaches from these two models, especially how they perform on unseen data sets. This paper employs Python to implement MLP and RF models on open source meter data from the Great Energy Predictor III (GEPIII) contest on Kaggle. The experimental framework of this study comprises three distinct parts; the first segment, we assess the performance of RF and MLP models through individual building evaluations. The second phase examines the behavior of MLP and RF models across diverse space uses. The final segment is dedicated to evaluate the two models using all buildings and incorporating all space uses. This structured approach offers a detailed analysis of the models' performance particularly concerning their generalizability, and the outcomes contribute to a better understanding of the strengths and limitations of both models in term of generalizability.