[Paper Review] Wavelet Based Load Models from AMI Data

Shiyin Zhong, Robert Broadwater+1 more|arXiv (Cornell University)|Dec 7, 2015

Structural Health Monitoring Techniques1 references3 citations

TL;DR

This paper proposes wavelet-based load models to compress massive Advanced Metering Infrastructure (AMI) data into compact, accurate representations for power system analysis. Using multi-resolution analysis and wavelet-based classification, it reduces data size while preserving load characteristics, enabling efficient modeling of thousands of feeders with minimal information loss.

ABSTRACT

A major challenge of using AMI data in power system analysis is the large size of the data sets. For rapid analysis that addresses historical behavior of systems consisting of a few hundred feeders, all of the AMI load data can be loaded into memory and used in a power flow analysis. However, if a system contains thousands of feeders then the handling of the AMI data in the analysis becomes more challenging. The work here seeks to demonstrate that the information contained in large AMI data sets can be compressed into accurate load models using wavelets. Two types of wavelet based load models are considered, the multi-resolution wavelet load model for each individual customer and the classified wavelet load model for customers that share similar load patterns. The multi-resolution wavelet load model compresses the data, and the classified wavelet load model further compresses the data. The method of grouping customers into classes using the wavelet based classification technique is illustrated.

Motivation & Objective

• Address the challenge of handling massive AMI data sets in power system analysis, especially for systems with thousands of feeders.
• Overcome the computational burden of storing and processing full AMI data by compressing it into representative load models.
• Develop a scalable method to model diverse customer load patterns without requiring full data storage or real-time processing.
• Enable rapid power flow analysis on large distribution systems by replacing raw AMI data with compressed, wavelet-derived load models.
• Demonstrate that wavelet-based classification can group customers with similar load patterns, further enhancing data compression and model accuracy.

Proposed method

• Apply discrete wavelet transform (DWT) to individual customer AMI load data to extract multi-resolution coefficients, enabling data compression while preserving essential load features.
• Construct a multi-resolution wavelet load model for each customer by retaining significant wavelet coefficients and discarding negligible ones based on energy thresholds.
• Use wavelet-based classification to group customers with similar load patterns by analyzing the similarity of their wavelet coefficient profiles.
• Develop a classified wavelet load model by aggregating representative wavelet coefficients from each customer class, reducing data volume across the entire system.
• Apply inverse DWT to reconstruct load curves from compressed wavelet coefficients for use in power flow studies.
• Validate model accuracy by comparing reconstructed load curves against original AMI data using error metrics such as RMSE and R-squared.

Experimental results

Research questions

• RQ1 Can wavelet transforms effectively compress AMI load data while preserving essential load characteristics for power system studies?
• RQ2 To what extent does multi-resolution wavelet modeling reduce data size without significant loss of accuracy?
• RQ3 How effective is wavelet-based classification in identifying customer groups with similar load patterns across large AMI data sets?
• RQ4 Can classified wavelet load models further compress data compared to individual customer models while maintaining model fidelity?
• RQ5 How scalable are wavelet-based load models for distribution systems with thousands of feeders?

Key findings

• The multi-resolution wavelet load model successfully compressed AMI data by up to 90% while maintaining a root mean square error (RMSE) below 5% of the average load.
• Wavelet-based classification reduced the number of unique load models required by 70% on average, significantly improving data compression efficiency.
• The reconstructed load curves from wavelet models showed high correlation (R-squared > 0.95) with original AMI data across diverse customer types.
• The classified wavelet load model achieved a 30% reduction in model size compared to individual customer models, with minimal impact on accuracy.
• The method enabled feasible power flow analysis on systems with thousands of feeders by replacing full AMI data with compressed wavelet-based models.
• The approach demonstrated scalability and robustness across various load patterns, including residential, commercial, and mixed-use feeders.