Impact of predictor variables on the performance of future occupancy prediction: Feature selection using genetic algorithms and machine learning

Abstract

This study analyzes the impact of employing different features on the performance of future occupancy prediction models. The aim is to identify the most effective predictor variables for occupancy models in order to enhance the prediction performance. To this end, a multi-objective genetic algorithm (MOGA) is proposed and employed to maximize accuracy while minimizing the number of features utilized in the model. A trade-off between the mentioned objectives is provided by using the technique for order of preference by similarity to ideal solution (TOPSIS). The performance of the MOGA is compared with that of a single-objective genetic algorithm, forward sequential selection, and backward sequential selection methodologies to assess the effectiveness of the proposed method. The results reveal that the MOGA provides superior performance and is able to improve the median accuracy by up to 4.81% from 70.94% for long-term occupancy prediction based on backward sequential feature selection while utilizing fewer features. Based on the TOPSIS method, no more than six features are required for developing the occupancy models with recent occupancy states and day of week selected as the essential features. In some cases, using recent CO2 levels, occupancy duration, lighting states, and previous occupancy states also shows a potential improvement in the prediction performance.