Improving energy efficiency prediction under aberrant measurement using deep compensation networks: A case study of petrochemical process