A Feature Selection-Based Predictive-Learning Framework for Optimal Actuator Control in Smart Homes

Abstract

In today’s world, smart buildings are considered an overarching system that automates a building’s complex operations and increases security while reducing environmental impact. One of the primary goals of building management systems is to promote sustainable and efficient use of energy, requiring coherent task management and execution of control commands for actuators. This paper proposes a predictive-learning framework based on contextual feature selection and optimal actuator control mechanism for minimizing energy consumption in smart buildings. We aim to assess multiple parameters and select the most relevant contextual features that would optimize energy consumption. We have implemented an artificial neural network-based particle swarm optimization (ANN-PSO) algorithm for predictive learning to train the framework on feature importance. Based on the relevance of attributes, our model was also capable of re-adding features. The extracted features are then applied as input parameters for the training of long short-term memory (LSTM) and optimal control module. We have proposed an objective function using a velocity boost-particle swarm optimization (VB-PSO) algorithm that reduces energy cost for optimal control. We then generated and defined the control tasks based on the fuzzy rule set and optimal values obtained from VB-PSO. We compared our model’s performance with and without feature selection using the root mean square error (RMSE) metric in the evaluation section. This paper also presents how optimal control can reduce energy cost and improve performance resulting from lesser learning cycles and decreased error rates.