An Iterative Shifting Disaggregation Algorithm for Multi-Source, Irregularly Sampled, and Overlapped Time Series.

Area of Science:

• Data Science
• Time Series Analysis
• Signal Processing

Background:

• Accurate time series forecasting often necessitates higher temporal resolution than available data provides.
• Existing temporal disaggregation methods are limited to single, uniformly sampled time series, hindering real-world applications.
• Multi-source, non-uniformly sampled sensor data streams present challenges for traditional disaggregation techniques.

Purpose of the Study:

• To introduce the Iterative Shifting Disaggregation (ISD) algorithm for processing and disaggregating multiple, non-uniformly sampled sensor data streams.
• To transform low-frequency measurements into a single, coherent high-frequency signal.
• To address the limitations of existing temporal disaggregation techniques in multi-source scenarios.

Main Methods:

• The Iterative Shifting Disaggregation (ISD) algorithm employs an iterative, two-phase process.
• Phase 1: Prediction using multiple linear regression to generate high-frequency series from low-frequency data and correlated variables.
• Phase 2: Update phase redistributes low-frequency observations across high-frequency periods, refining estimates iteratively.

Main Results:

• ISD successfully disaggregates multiple, non-uniformly spaced time series with overlapping intervals into a single daily representation.
• Case studies with natural gas data demonstrated significant improvements over existing methods.
• Achieved 1.4-4.3% WMAPE improvement for billing cycle data and 4.6-10.4% for residential data.

Conclusions:

• The ISD algorithm is effective for disaggregating complex, multi-source sensor data into daily time series.
• ISD offers a robust solution for enhancing time series forecasting accuracy when dealing with non-uniform data.
• The method shows practical applicability and superior performance in real-world energy data scenarios.