Feature Selection and Evolutionary Rule Learning for Big Data in Smart Building Energy Management

Abstract

Since buildings are one of the largest sources of energy consumption in most cities of the world, energy management is one of the major concerns in their design. To ameliorate this problem, buildings are becoming smarter by the incorporation of intelligent supervision and control systems. Data captured by the sensors can be interpreted and processed by rule-based computation methods of biological inspiration (such as genetic fuzzy systems, GFS) for predicting the future behavior of the building in a knowledge-based interpretable human-like manner. GFS are computational models inspired in human cognition which use evolutionary computation (inspired in the natural evolution) to automatically learn fuzzy rules which contain explicit imprecise knowledge about a system or process. This knowledge, represented using fuzzy rules that involve fuzzy linguistic variables and values, is used to perform approximate reasoning on the input values for obtaining inferred values for the output variables. In energy management of buildings, these rules allow a smart control of the system actuators to reduce the building average energy consumption. However, the large amount of data produced on a per second basis complicates the generation of accurate and interpretable models by means of traditional methods. In this paper, we present an evolutionary computation-based approach, namely a genetic fuzzy system, to build scalable and interpretable knowledge bases for predicting energy consumption in smart buildings. For accomplishing this task, we propose a cognitive computation system for multi-step prediction based on S-FRULER, a state-of-the-art scalable distributed GFS, coupled with a feature subset selection method to automatically select the most relevant features for different time steps. S-FRULER is able to learn a fuzzy rule-based system made up of Takagi-Sugeno-Kang (TSK) rules that are able to predict the output values using both linguistic imprecise knowledge (represented by fuzzy sets) and fuzzy inference. Experiments with real data on two different problems related with the energy management revealed an average improvement of 6% on accuracy with respect to S-FRULER without feature selection, and with knowledge bases with a lower number of variables.