# **About SMARTSPACES – Feedback for continuous improvement**

The final page is Feedback for continuous improvement and provides the context of the data, how the website and smiley icons works, and consumption modelling applied to the data sets available.

The data underpinning this website are energy and water consumption values and outside air temperature readings recorded at half hourly intervals. These data sets are combined to generate a performance indicator that compares consumption in the current half hour to expected consumption levels. The expected consumption is determined based on a 365-day baseline period. The baseline period used is continually updated so that new data sets are rolled into the baseline and older data sets discarded. This approach leads to a system that continuously ‘learns' the pattern of consumption

Users are presented with feedback that indicates on average, whether the building is using more or less energy than expected. If energy performance is a building improves then the feedback system will identify and visualise the change through the smileys.

# **But how does it work?**

The result of this complex approach is that the indicator represents energy performance as a comparison of current consumption with that of the latest 365 days. If consumption is exactly in the middle of the expected range then the indicator is equal to 50. If consumption is higher than the expected range then the indicator is equal to 100. If consumption is lower than the expected range then the indicator is equal to 0. An indicator value of 35, for example, can be interpreted as meaning that 35% of equivalent baseline consumption (i.e. that occuring at the same time of week) was below this value and 65% was above this value.

The smiley face scale provides a user-friendly way to understand energy performance. The indicator is calculated on a half hourly basis. These raw indicator values are then averaged over a day or a week to generate aggregated values. The resulting values are then converted into simple smiley faces shown on the home page and building overview page.

The faces shown in the introduction tab indicate the full range of values. A value of fifty produces a yellow, neutral looking face representing no change over time. A value of zero produces a green, happy looking face representing a significant reduction over time. A value of one hundred produces a red, sad looking face representing a significant increase over time.

# **Consumption Modelling**

Determining the expected level of consumption is a complex process. Data are divided into subsets, one for each 'time of the week'. For example, one subset will include only data recorded at 09:00 on Wednesday mornings. A linear regression model is then fitted to all data (consumption vs outside air temperature) in each subset. No model fits perfectly, for each point of data, the residual consumption (i.e. the difference between actual consumption and that predicted by the model) is recorded. This set of residuals captures how far consumption diverged from the baseline model and so represents the expected scatter around the baseline model. The indicator is calculated by observing how far current consumption diverges from the baseline model and expressing that as a percentile score with respect to the baseline residuals.

Thus, consumption that exceeds the model prediction by more than any experienced in the baseline period would achieve a percentile of 100. Similarly, consumption that falls below the model prediction by more than any experienced in the baseline period would achieve a percentile of 0. Between these values, the percentile score represents the proportion of historical residuals that fell below the current residual. A value of 50 would be achieved for consumption that diverged from the model by more than (and less than) 50% of historic consumption values. This can be summarised by saying that the indicator reflects both the value and spread of historic consumption patterns. If a building is highly predictable then consumption need only shift slightly for an effect to be registered. Less predictable buildings will have a wider spread of residuals and so will be less sensitive to changes in consumption patterns.