Multivariate electricity consumption prediction with extreme learning machine

In this paper, Extreme Learning Machine (ELM) is demonstrated to be a powerful tool for electricity consumption prediction based on its competitive prediction accuracy and superior computational speed compared to Support Vector Machine (SVM). Moreover, ELM is utilized to investigate the potentials of using auxiliary information such as electricity-related factors and environmental factors to augment the prediction accuracy obtained by purely using the electricity consumption factors. Furthermore, we formulate a combinatorial optimization problem of seeking an optimal subset of auxiliary factors and their corresponding optimal window sizes using the most suitable ELM structure, and propose a Discrete Dynamic Multi-Swarm Particle Swarm Optimization (DDMS-PSO) to address this problem. Experimental studies on a real-world building dataset demonstrate that electricity-related factors improve accuracy while environmental factors further boost accuracy. By using DDMSPSO, we find a subset of electricity-related and environmental factors, their respective window sizes, and the number of hidden neurons in ELM which leads to the best prediction accuracy.