📘 Resource Overview & Documentation Details

You are viewing the technical directory listing for Personalized Energy Feedback with Smart Meter Data. This engineering resource is curated under our Cyber Physical System division to support automation specialists, control systems engineers, and manufacturing technicians looking for verified reference materials.

🔍 Contextual Abstract & Technical Scope

Share Summary:LinkedInEmail

Personalized energy feedback depends on more than showing a household its own electricity use. The comparison group matters too: a message is more relevant when it compares a home with others that have similar consumption patterns and housing characteristics. Smart meter data and machine learning offer a way to create and update these behavioral reference groups without relying on household surveys or repeated manual clustering.

A study using data from 2,248 households in Holland, Michigan, evaluated an energy-cyber-physical system (e-CPS) designed to assign households to reference groups for recurring feedback. Its results offer a practical example of how energy-use profiling can support scalable, data-driven messaging—and show why careful feature selection and periodic model updates matter.

How personalized energy feedback uses reference groups

Normative feedback presents a household’s consumption alongside a group norm. Descriptive feedback communicates what members of the comparison group use; injunctive feedback signals approval or disapproval of behavior. The premise behind personalization is that comparisons may be more meaningful when the group resembles the recipient in relevant ways, including energy-use behavior and housing size.

Smart meters can provide consumption histories without requiring residents to actively report their habits. Combined with housing information from public records, those histories can be used to identify patterns and form reference groups. The system evaluated in the study was designed to repeat that process as new data arrived, allowing group assignments to reflect changing patterns rather than remaining fixed indefinitely.

Inside the energy-cyber-physical system

The e-CPS links data collection, analysis, classification and message delivery. Its data-entry function receives smart-meter consumption data and housing characteristics. An identification module uses historical data to establish representative groups; a classification module assigns new profiles to those groups; and a messaging module can use the resulting assignments to prepare household feedback.

Before comparing consumption profiles, the system organized homes into five categories by floor area. This step aimed to make comparisons more relevant because housing size is an important factor in household energy use. The analysis also divided each day into four six-hour periods—morning, afternoon, evening and night—to represent daily load profiles in a more compact form.

Rather than relying only on total consumption, the researchers tested three ways to describe load shape:

  • Gradient: captures changes in consumption between consecutive time points.
  • Normalization: scales each household’s values between its own minimum and maximum.
  • Cumulative share: represents the proportion of daily energy used by each point in the day.

These transformations help distinguish when and how consumption occurs, even when households differ in total use. Hierarchical clustering was then used to identify groups with similar profiles. The number of groups was selected using the Davies-Bouldin Index, which evaluates within-group similarity relative to separation between groups; lower values indicate better clustering under this measure.

Why load shape mattered for classification

Once the reference groups had been identified, the classification module used historical profiles to assign new daily profiles to a group. The study compared decision trees, discriminant analysis, support vector machines and k-nearest neighbors. It also tested two kinds of model inputs: load shapes and load indices, which summarize consumption characteristics such as the relationship between average and peak demand.

Performance depended strongly on the features supplied to the classifiers. Load shapes supported much higher average accuracy than load indices across the tested models. The best-performing combination was a one-nearest-neighbor classifier using load shapes. This method assigns a new profile according to the group of the most similar historical profile.

For daily profile classification, this combination achieved average accuracy of 97.7%, with average precision, recall and F1-score of 97.6%. These measures describe different aspects of performance: precision reflects how often a group assignment is correct, recall how many profiles belonging to a group are identified, and F1-score balances the two. Considering them alongside overall accuracy is useful when group sizes are uneven.

Performance across weekly and monthly feedback cycles

For practical messaging, a correct assignment over one day may not be enough. The study therefore assessed whether predictions remained correct across entire feedback periods. Average classification accuracy ranged from 94.7% to 95.9% for weekly feedback and from 89.9% to 93.1% for monthly feedback. In the tested data, using a larger share of available records for training also improved accuracy, and training on the complete dataset across seasons generally performed better than using seasonal data alone.

These findings suggest that smart meter data analytics can support frequent group assignment without rerunning a full clustering analysis for every billing cycle. The distinction is important: clustering establishes the reference groups, while classification provides a repeatable way to assign incoming profiles to them.

Deployment considerations and limits

The results demonstrate performance on one year of data from one Michigan community; they do not establish that the same classifier will perform equally well in other climates or regions. Energy-use patterns can vary with seasonal conditions, and the researchers note that additional regional data would be needed to assess broader applicability. A deployment should therefore validate its groups and classifiers against local data rather than assume a model transfers unchanged.

Reference groups and housing characteristics may also change over time. The proposed approach anticipates periodic updates to historical groups and retraining, rather than treating the original classification model as permanent. Uneven group sizes are another concern: when some groups contain far fewer profiles than others, classification can become biased toward larger groups. Precision, recall and F1-score by group can help reveal that issue.

Finally, accurate assignment does not by itself establish that feedback will change behavior. The study evaluates classification, not the effect of messages on household consumption. Message design also matters: descriptive comparisons can have unintended effects if recipients interpret lower-than-average use as permission to consume more. How people identify with a reference group, and how descriptive and injunctive elements are presented, remain important considerations.

Conclusion

For energy program designers and technical teams, the practical lesson is to treat personalized energy feedback as a data pipeline, not simply a message template. Smart-meter load profiles can identify meaningful behavioral groups, and a classifier can refresh household assignments as new consumption data arrives. The study’s strong weekly and monthly results are promising for scalable delivery, while its regional scope and behavioral uncertainties make local validation, periodic retraining and thoughtful message design essential.

Access Official Technical PDF