Search

• A machine-learning approach for studying the e-saving effects of smart information interventions via SEMS- in HK. • An investigation of the statistical correlation between the e-saving effects of three types of smart interventions and household characteristics in HK. • A generic and scalable approach for SVR machine-learning-driven SEMS-induced e-consumption studies across different scales and sample sizes. Smart energy monitors (SEMs), which enable householders to measure electricity usages of different appliances in real-time, have been widely deployed by utilities across many different countries. However, the actual electricity saving effects of smart information interventions via the SEM connected to the smart energy management system (SEMS) remain inconclusive, due to failures of the existing statistical models in capturing non-linear relationships. To address the non-linearity challenge and to observe the effects of smart information interventions on electricity savings among the public housing householders in Hong Kong, we initiate a longitudinal electricity consumption behavioural study in Hong Kong. We propose a machine-learning approach to capture any non-linearity identified from our SVR machine learning model. In particular, we identify the correlation between the different combinations of three smart information interventions and the percentage of electricity savings at the household-level in Hong Kong. Smart Energy Management System (SEMS), consisting of a smartphone app and a SEM installed respectively on the smartphone and the participant household of our participants in a public housing estate in Hong Kong, have been developed and deployed by the HKU AI-WiSe team. An innovative technological intervention cum environmental behavioural study was conducted on representative of 14 households residing in a public housing estate in Hong Kong, across a one-year period, from 2018 to 2019. Three types of smart information interventions were introduced to our household participants, including their (1) current electricity consumption profile (2) historical electricity consumption profile, and (3) ranking in electricity savings as compared to other participating households. Our study concludes that the overall average electricity savings across all 14 households is 7.1%. However, as different households have displayed different electricity consumption characteristics, the electricity savings vary significantly across 14 households, from slightly negative or almost zero savings, to significantly positive savings. Our results show that with respect to the three types of smart information interventions, Type (1) and Type (2) display a stronger electricity saving effect when compared to the ranking-based smart information intervention. We conclude our study by identifying the right electricity policies for the HKSAR Government to promote household electricity savings via SEMs and SEMS in HK. To the best of our understanding, our study represents the very first attempt to capture the non-linear statistical correlation between smart information interventions and household electricity savings via the machine-learning SVR approach. Our approach is generic and scalable; it can be applicable to other related electricity consumption experimental studies, across any geographical scale and sample size. [ABSTRACT FROM AUTHOR]

Highlights • Drivers of domestic users' price responsiveness are identified. • A neural network model is built to accurately estimate baseline consumption. • Variable selection is conducted by combining 3 advanced machine learning techniques. • Feasibility of selecting high potential user based on results is demonstrated. Abstract Time-based pricing for domestic electricity users has been effective in reducing peak demand and facilitating integration of renewable energies. However, high cost, price non-responsiveness and adverse selection present challenges. To tackle these challenges, it would be important to investigate which users exhibit a higher potential to respond to price change, such that time-based pricing can be introduced to such users. Few studies have examined which users are more price-responsive and what drives price responsiveness. This article aims to fill this gap by comprehensively identifying the drivers that determine electricity users' price responsiveness, in order to facilitate the selection of high potential users. We adopt a novel machine-learning approach to select the high potential users, using the Irish smart meter dataset (2009–10), which forms part of the national Time of Use trial. Our methodological novelties cover two aspects: First, using a feed-forward neural network model, we aim to determine precisely the price responsiveness of individual households, and address the nonlinearity of energy consumption and price response attributes. Second, we apply an integrated machine learning methodology to identify the drivers of price responsiveness. Our integrated approach outperforms the traditional variable selection methods by identifying drivers that are reliable. Our empirical results have shown that demographic and residential characteristics, psychological factors, historical electricity consumptionand appliance ownership are significant drivers. In particular, historical electricity consumption, income, household size, perceived behavioural control, and adoption of specific appliances, including immersion water heater and dishwasher, are significant drivers of price responsiveness across the Irish electricity users. Our results have also verified that continual price increase within a moderate range will not drive additional peak demand reduction; there is an intention-behaviour gap, where stated intention does not lead to actual peak reduction behaviours. Based on such results, we have conducted a scenario analysis to demonstrate the feasibility of selecting high potential users of achieving significant peak electricity reduction. [ABSTRACT FROM AUTHOR]