Advances of machine learning in multi-energy district communities‒ mechanisms, applications and perspectives

Energy paradigm transition towards the carbon neutrality requires combined and continuous efforts in cleaner power production, advanced energy storages, flexible district energy demands and energy management strategies. Applications of cutting-edge machine learning techniques can improve the system reliability with advanced fault detection and diagnosis (FDD), automation with agent-based reinforcement learning, flexibility with model predictive controls, and so on. In this study, a comprehensive review on artificial intelligence applications in carbon-neutral district community, has been conducted, from perspectives of energy supply, energy storage, district demands and energy management. Classifications and underlying mechanisms on ML techniques have been demonstrated, including supervised, unsupervised, reinforcement and deep learning. Afterwards, practical applications of ML have been reviewed, in respect to renewable energy supply, hybrid energy storages, district energy demand and advanced energy management. Results indicate that, supervised learning was mainly applied in classification and regression, and unsupervised learning was mainly applied in clustering. The reinforcement learning is mainly applied in on-line optimal scheduling for building energy management. With respect to clean energy supply, ML in solar and wind energy systems mainly include solar irradiance forecasting, wind resource forecasting, PV power prediction, maximum power point tracking (MPPT) for smart control, fault detection and diagnosis. ML in fuel cells mainly includes performance prediction, material selection, combination and so on. Furthermore, in respect to hybrid energy storages, ML in electrochemical battery includes dynamic thermal/ electrical behavior, battery sizing and optimization, state-of-charge prediction, battery lifetime estimation, fault detection and diagnosis analysis. ML in sensible energy storages mainly include load prediction and storage capacity sizing, dynamic scheduling for cost saving, thermal stratification analysis and dynamic performance prediction. Advances in energy management with ML mainly include dispatch on stochastic and intermittent renewable power, microgrid adaptive control, smart energy trading with controls and decision-marking. Research tendency over the recent past several years indicates that, critical areas for low-carbon energy systems transit from the only renewable systems (59.4% in 2016) towards both renewable energy supply and energy storages (35.1% and 34.1%, respectively), such as battery, capacitors/supercapacitors, sensible/latent heat storages, compressed air storage and hydrogen storage. This study can provide a holistic overview and in-depth thinking on artificial intelligence in the carbon-neutral district transition.