1Dept. of Computer Science and Engineering, P.E.S. College of Engineering Mandya, Karnataka, India. Email: ashwinimc@pesce.ac.in
2Dept. of Information Science and Engineering, P.E.S. College of Engineering Mandya, Karnataka, India. Email: minavathi@pesce.ac.in
Background: Real-time meteorological monitoring via IoT sensors faces the challenge of concept drift, where evolving statistical properties render static models unreliable. While standard detectors signal the presence of drift, they fail to characterize its structural nature, leading to suboptimal adaptation. This paper proposes a dual-layer, computationally lightweight framework for prescriptive drift management.
Methodology: The first layer, Dynamic Model Selection (DMS), maintains a pool of three online learners - Hoeffding Trees (HT), K-Nearest Neighbors (KNN), and Adaptive Random Forests (ARF), routing inference to the model minimizing the Windowed Mean Absolute Error (W-MAE). The second layer identifies drift typology (Sudden, Incremental, Gradual, or Stable) by fitting a first-degree Ordinary Least Squares (OLS) regression to the recent error sequence. This O(W) heuristic enables diagnostic intelligence on microcontroller-class edge hardware without requiring neural networks. The framework was evaluated on a 58-year NOAA archive (106,382 records) from ten Indian stations using strict 5-fold chronological cross-validation.
Results: ARF achieved superior generalization (MAE: 1.50±0.25, R²: 0.73±0.08), outperforming the HT baseline by 20%. Hardware profiling confirms HT operates at 24 µs inference latency with a 0.07 MB memory footprint, while the OLS-slope analyzer correctly identified long-term climate dynamics without manual tuning.
Conclusion: This work provides a deployable pathway for edge IoT systems to not only detect model degradation but autonomously diagnose its cause and prescribe appropriate adaptation strategies.
Keywords: Concept Drift Detection, Drift Typology Classification, Online Learning, Data Stream Mining, Adaptive Random Forest, Hoeffding Tree, Dynamic Model Selection, Windowed Mean Absolute Error, Heuristic Anomaly Detection, OLS Slope Heuristic, Edge Computing, IoT Stream Analytics, Meteorological Forecasting, Chronological Cross-Validation, Non-Stationary Distributions.
How to cite this article: Ashwini MC, Minavathi. Prescriptive Concept Drift Management in Non-Stationary Meteorological Streams. Int J Drug Deliv Technol. 2026;16(13s): 520-529. DOI: 10.25258/ijddt.16.13s.56
Source of support: Nil.
Conflict of interest: None