Data drift is a change in the distribution of incoming data relative to the reference data used to train a model. In a deployed digital twin, the change can make predictions less reliable even when the model itself has not changed. section 10.3.2
Solar production changes
A solar forecasting model trained on one year’s data may encounter different production values later as panels collect dirt, trees cast new shadows, or photovoltaic cells degrade. Comparing current production values with the training reference can reveal a distribution shift. section 10.3.2
Data drift differs from concept drift, where the relationship between variables changes, and sensor drift, where the data collection mechanism degrades. These problems can surface silently, so production systems need monitoring beyond the model itself. section 10.3.2 section 8.5.3
Table 10.4 Types of drift that affect digital twin models
|
Type
|
Description
|
Example
|
|---|---|---|
|
Data drift
|
The statistical properties of incoming data change beyond what the model was trained to handle.
|
Solar panel degradation or dirt accumulation shifts the distribution of production values over time.
|
|
Concept drift
|
The relationship between variables changes fundamentally.
|
Replacing a gas cooktop with an electric induction one alters the relationship between time of day and energy consumption.
|
|
Sensor drift
|
The data collection mechanism itself degrades.
|
A low-cost temperature sensor slowly loses calibration, drifting by 0.1 °C per month.
|
Compare data, concept, and sensor drift through concrete digital twin examples. section 10.3.2
Monitor distributions rather than treating every changed value as a fault. For example, surface a sudden 20% increase in average daily energy consumption to an operator. If it represents a new normal rather than a problem, the period can be labeled and used to trigger retraining. section 10.3.2
Listing 10.3 Using Evidently to track drift in live telemetry data against a known state
import pandas as pd
from evidently import Report
from evidently.presets import DataDriftPreset
reference = pd.DataFrame({ #1
"power_kw": [1.8, 2.0, 1.9, 2.1, 2.0, 1.95, 2.05]
})
current = pd.DataFrame({ #2
"power_kw": [2.4, 2.5, 2.6, 2.3, 2.7, 2.5, 2.6]
})
report = Report([DataDriftPreset()])
result = report.run(current, reference)
result.save_html("power_drift_report.html")
print(result.json())
Compare reference and current power readings with Evidently’s DataDriftPreset, then save an HTML report and print the result as JSON. section 10.3.2
A single-feature comparison is a starting point; the approach can extend to multidimensional sensor streams and continuous monitoring pipelines. Production governance should also document training data provenance, version feature pipelines, establish performance baselines, and define retraining triggers and approval workflows. section 10.3.2 section 8.5.3

















