Is Feedforward Learning more Efficient than Feedback Learning in Smart Meters of Electricity Consumption?

The most popular way to improve consumers’ control over their electricity cost is by providing frequent and detailed feedback with “in-home displays” (IHD). In this study, we examined alternative ways to train experimental participants to control and optimize their use of electricity by “feedforward” training to map energy consuming behaviors to costs. The participants were trained in one of four experimental conditions, one feedback (“IHD”) and three feedforward conditions before they had to control the electricity consumption in a simulated household. Results showed that one of the feedforward conditions produced somewhat higher utility and as good or better satisfaction of a monthly budget than the feedback training condition, despite never receiving any feedback about the monthly cost, but the generalization to a new budget constraint proved to be slightly poorer. Introduction The use of so-called “smart electricity meters” is rapidly becoming common. It has been estimated that within the European Union alone some 51 billion euro is being invested in smart meters (Faruqui, Harris, & Hledik, 2009). In many countries, household energy consumption is still billed once a month, but smart meters can offer feedback that is detailed and more frequent with so called In Home Displays (IHDs). Intuitively, the latter kind of feedback system seems more beneficial, and, indeed, many early studies suggested energy reductions up to 15%. However, more recent studies point at consumption reductions at 2-4%, few of them being significant (Klopfert & Wallenborn, 2011). In the present study, we focus at in-home displays (IHDs), which only display the electrical consumption at different time intervals, and, unlike smart meters, they do not have a two-way communication with the central system. In a previous laboratory experiment (Guath, Millroth, Elwin, & Juslin, 2012), we investigated how feedback about electricity consumption is best presented to electricity consumers in order to control and optimize their use of electricity. To measure a participant’s energy efficiency in an experimentally controlled environment, the participants took on the role of an inhabitant in a simulated household, performing different types of energy consuming behaviors within a given budget (Figure 1). The goal of decreasing electricity consumption is often emphasized, but the participant’s task is actually an optimization problem that requires an appropriate balance between the cost of the electricity consumed and the benefit or utility obtained. The problem is illustrated in Figure 2, where the utility of electricity consumption is plotted against cost. The maximum utility obtainable at a given cost, assumed to be a decelerating function of the cost, is illustrated by the curve in Figure 2. The hypothetical utility obtained at a cost by a consumer is illustrated with a dot. The task is to move closer to the line for “maximal utility”, however, this is associated with two constraints: achieving sufficient utility to make life bearable and not surpassing a constrained budget. Guath et al. (2012) showed that in a deterministic system, frequent and detailed feedback was advantageous, but in probabilistic system, feedback aggregated over time was better, because it filtered out random noise. The Present Study In the present study, we wanted to evaluate if the same improvement could be obtained by feedforward training, rather than feedback training (as in most IHDs), hence, minimizing the negative effects from feedback interventions as conceptualized in Kluger and DeNisi’s (1996) study. Specifically, we wanted to avoid the decrease of effectiveness when attention is moved away from the task to the self, thus, making the effects of the training short-term. Another motive was to make the mapping task more flexible, not being dependent on the simulated household (Figure 1). Detailed and frequent feedback (an IHD) was compared to three feedforward conditions. Feedforward is defined as a process where knowledge is used to act directly to control the system, thus anticipating the changes that will occur (Basso & Olivetti Belardinelli, 2006). In the present task, partici-