REFIT · aggregate and monitored appliance power in W
Similar curves.
Different activity?
Find a similar whole-house interval, then look beneath the aggregate. Do the monitored appliance traces resemble each other too?
Figure 1. Whole-house power
Two days of one-minute summaries. Use the controls to select a query. Gray bands mark minutes ineligible for matching.
Checking REFIT extract…
Figure 2. Compare aggregate patterns
Uncheck “Reveal appliance evidence” for an aggregate-only baseline using the exact same REFIT windows.
Figure 3. What is hidden by the total?
Every card compares the same query and match. Each channel has its own vertical scale, shared by both lines in that card. Compare timing and values within a card; line heights across cards are not comparable.
Labels follow the matching House 1 metadata. “Television Site” may include several devices. Only monitored loads are represented; aggregate minus their sum also includes measurement differences and cannot identify a particular unmonitored appliance.
Table 1. Channel comparison
Mean power is averaged across minute summaries; differences describe measurements, not inferred on/off events or savings.
| Monitored channel | Query mean W | Match mean W | Change W | Trace RMSE W |
|---|
Evidence you can inspect
Table 2. Coverage and cleaning audit
No raw-release cross-check was performed. The cleaned release does not identify which individual zeros or repeated readings were replacement values. No true appliance-state classifier is implemented.
Source cleaning documentation, unchanged
Processing, selection and reproducibility
The date range was fixed to 1–2 January 2014 before examining matches. It is a small holiday-period illustration, not a representative household sample. Query options start every 15 minutes; candidates start every minute and never overlap the query. The five displayed candidates can overlap one another. Both query and candidate require every minute to pass screening.
Shape mode separately z-normalizes each aggregate window and ranks their RMSE; constant windows are excluded. Raw mode compares mean power in W. Lower distance means closer under that metric, not a probability or a validated similarity threshold. Charts always retain original units. Matching uses aggregate values only; appliance channels provide subsequent inspection.
Finite nonnegative readings and unique increasing Unix timestamps were checked. UTC comes from Unix seconds; source Time values were verified against UTC in this January extract. Raw samples are irregular, not assumed to lie on an exact eight-second grid.
Notice a peak.
Explore the pattern.
Compare recorded electricity intervals, find similar shapes, and inspect the evidence behind each comparison.
Electricity over time
Select a window below, or click the chart to set its start.
Checking the embedded source extract…
Two windows, one common scale
Search for a similar period to compare it with your query.
A similar shape does not identify an appliance or explain household behavior.
Why this match? Method and coverage
Energy by date and time of day
A secondary Sankey view of the same recorded energy: total → recorded date → time of day. Ribbon width is proportional to kWh.
Loading recorded energy groups…
These ribbons regroup measurements; they do not trace physical electricity flows or appliance use. Each column represents the same total—do not add columns together. Dates have partial coverage, and time-of-day groups use original timestamp labels, not a verified local time zone.
Read the grouped values and definitions
Night: 22:00–05:59; morning: 06:00–11:59; afternoon: 12:00–16:59; evening: 17:00–21:59. Each record belongs to its labeled calendar date, including readings after midnight. Groups with no positive energy have no ribbon.
| Date | Time of day | Readings | kWh |
|---|
Inspect and download the 100 source readings
Original timestamp labels; time-zone convention remains unconfirmed. No browser-local conversion.
| Timestamp | Energy (kWh) | Context |
|---|
A focused temporal redesign
Tencent’s dense technology display and the museum’s connected illuminated modules motivated selective highlighting and stepwise explanation. They did not reveal employees’ actual decisions. Temporal comparison fits these whole-house measurements; appliance-flow attribution does not.
Research connection
Inspired by TiVy and its grouped temporal patterns. This is a small evidence-linked redesign, not a reproduction of TiVy’s scalable algorithm. The event-sequence task framework informs explicit selection and comparison criteria.
Proposed collaboration
Academic researchers evaluate interpretation. Industry monitoring specialists review measurement quality. Energy educators and residents shape useful, accessible explanations. These roles are prospective.
Individual authorship
Enwei Fan · INFOSCI 301. Jiaxin Wang identified the nonlocal-data limitation. Prof. Luyao Zhang provided design and governance guidance. Codex assisted implementation and documentation. Human review and community evaluation remain necessary.