Abstract
Rapid integration of photovoltaic (PV) systems introduces significant power quality challenges, including voltage sags and harmonic distortion, yet existing approaches either perform post-hoc classification or rely on computationally expensive meteorological forecasting unsuitable for real-time edge deployment. In this paper, a resource-efficient framework is developed for short-term power factor sag prediction using only on-line electrical measurements via a dual-track methodology. A regression model is first established to define the predictive limits of direct value estimation. The analysis of drop-event-specific metrics reveals that sensor features are lagging indicators incapable of regressing immediate drop magnitude, motivating a classification re- framing. A binary Long Short-Term Memory (LSTM) classifier is then developed to predict the probability of a power factor drop within a configurable look-ahead window, with a post-hoc decision threshold optimized per phase via F1-Score maximization. A physics-informed feature ablation study further reduces the input space from 80-dimensional telemetry to a 3-variable subset (Voltage, Current, Power Factor), An architectural sweep across eight configurations spanning 121 to 2.2M parameters confirms that model complexity has negligible impact on performance. Cross-dataset validation on multiple public PV datasets confirms that temporal deep learning provides measurable improvement when atmospheric features are available, but provides no advantage on electrical-quality features alone. Controlled ablation on over 850K samples confirms this pattern: converting power output to an efficiency ratio (power/irradiance) on the same data with the same labels eliminates temporal learning. The primary finding is that temporal learnability is determined by feature type rather than model architecture or complexity, establishing the feasibility of deploying high-fidelity fault prediction on resource- constrained edge controllers without reliance on cloud infrastructure or weather data.
Publication Date
5-2026
Document Type
Thesis
Student Type
Graduate
Degree Name
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering
College
Kate Gleason College of Engineering
Advisor
Bing Yan
Advisor/Committee Member
Andres Kwasinski
Advisor/Committee Member
Dongfang Liu
Recommended Citation
Aiken, Ian, "An LSTM-Based Framework for Predictive Power Quality Monitoring" (2026). Thesis. Rochester Institute of Technology. Accessed from
https://repository.rit.edu/theses/12643
Campus
RIT – Main Campus
