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

Abstract

Artificial intelligence-based methods have progressed rapidly to become leading tools for energy analysis. However, information from the petrochemical processes is commonly associated with natural variation, uncertainties, and signal quality degradation of the instruments during operations. In this work, two deep learning models —predictor compensation energy efficiency predictor and internal compensation energy efficiency predictor—are presented to counterbalance the contribution of sensor abnormal behavior by reconstructing the original input and maintaining the dynamic characteristics deploying the long short-term memory as a computational layer. The compensation networks accurately predict the energy efficiency of the vinyl chloride monomer process under 10% and 20% fault variations and provide higher reliability and reproducibility in the model deployment phase. The robustness of the models has been validated through testing with a wide range of fault variations, while achieving an average r-squared value exceeding 0.95 on the 35% fault variation dataset. The action plan reveals a great potential to save energy by 49 GJ per day or 230,000 tonnes of annual utility consumption and a reduction of 4000 tonnes of carbon dioxide emissions per year by performing energy gap analysis and mapping model inputs with manipulated system variables.