Prediction, clustering and analysis of supermarket and retail energy demand data using machine learning techniques

Abstract

The need to improve efficiency and reduce the total energy demand in all sectors of the economy is widely recognised. Amongst retail stores, supermarkets have higher intensity energy use due to refrigeration and other in-store services, giving supermarket chains a strong incentive to reduce energy demand across their portfolio of stores, including stores being planned. Predicting energy demand helps planning, on-going energy management, and detecting anomalous use patterns. However, literature about predicting supermarket energy demand is scarce. Using historical hourly electricity data of 213 UK supermarkets (same company), annual electricity daily load profiles of new supermarkets were predicted using regression models, including neural networks and support vector machines. Exploiting various uses by floor area and geographic location, prediction errors varied between 3–20% depending on method, year, supermarket type, season and temperature intervals. Profiles computed for warm periods (cooling required) were better predicted than cold periods (heating required). A reduced-feature method accurately represented the electricity daily load profiles of both the supermarket data-set and a data-set of 641 non-food retail stores. Comparing the clustering and prediction experiments with results obtained using the whole profile, showed that the errors only slightly increased. Thus, the reduced feature set is a concise way to represent load profiles without including small variances that do not add useful information. Finally, the relationship of the urban heat island effect and the electricity demand of 38 supermarkets in Greater London was analysed. In Summer, supermarkets located closer to the city centre had higher area-normalised energy demand than those farther from the centre, suggesting that additional cooling was responsible. The limitations of applying machine learning methods to this real-world problem showed that human expertise for interpretation and understanding were essential. However, performing similar analyses using a solely engineering approach would require significantly more time and resources.