Date of Award
12-2025
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Electrical Engineering
Committee Chair/Advisor
Johan Enslin
Committee Member
Ramtin Hadidi
Committee Member
Christopher Edrington
Abstract
Increases in distributed generation and changes to predominate load types in modern power systems are driving the need for updated or newly developed dynamic load models. Accurate load modeling is necessary for transmission planning and operation, post-mortem analysis, and inter-regional transfer studies. The most widely used dynamic load model is the WECC Composite Load (CMLD). The CMLD model is an aggregation of substation impedance, under voltage and frequency controls, dynamic motor loads, and static loads. Parameterizing dynamic load models, especially composite models with high parameter counts like the CMLD, is can be difficult. This work proposes a measurement-based, machine learning method to parameterize an aggregated load model for a specific operating condition and contingency, ideal for transmission planning or post-mortem analysis. It uses a stacked LSTM-DNN architecture trained on transient stability data to parameterize a ZIP and induction motor load model aggregated at a single bus. The training data are the real power, voltage, and frequency time series measured at the load bus. The robustness of the method is tested by completing the parameterization process with two modified datasets, the first adds noise to the time series data and the second reduces the sampling rate. Results show the LSTM-DNN is capable of parameterizing the ZIP and induction motor load models with an average $R^2$ score of 0.96, with the lowest individual parameter score being the constant current coefficient at 0.90.
Recommended Citation
Smith, Zach H., "Measurement-based Dynamic Load Model Parameterization Via A Stacked Long Short-Term Memory - Deep Neural Network" (2025). All Theses. 4630.
https://open.clemson.edu/all_theses/4630