Identified BTM customer owned PV anomalies by considering load profiles, voltage data and load forecasts using AMI, customer and weather data

Abstract

Increased energy demand and the introduction of advanced smart grid technologies such as solar and wind energy have significantly expanded and challenged the electricity grid in recent decades. in addition, the integration of electric vehicles into the grid has provided infrastructure challenges and added complexity to utilities requiring innovative solutions for grid stability and reliability. A key challenge in this developing landscape is the rapid growth of behind-the-meter (BTM) photovoltaic (PV) systems. These decentralized energy resources, each with distinctive structural characteristics, pose challenges for grid operators and engineers in managing the transitions and gaps in renewable energy resources. This complexity highlights the need for advanced forecasting models to accurately forecast load generation and detect abnormal events at customer solar owned installations.This dissertation strategically focuses on developing forecasting models to be used by electric utilities industry and present a methodology to identify unusual events at customer solar locations. The accurate prediction of load generation plays an important role in the optimization of power systems and operations. This optimization is essential for utilities to effectively meet the challenges posed by PV systems, thereby reducing operation and maintenance costs. By addressing these key aspects, the goal is to enhance the overall resilience and efficiency of the power grid amidst the evolving landscape of energy and technological advancements.To forecast the solar generation load, the approach incorporates combinations of K-means clustering, Automated Meter Infrastructure (AMI) real-world PV load generation, weather data, and calculated solar positions. To identify the unusual events, the approach is to utilize the Z-score methodology, which is a statistical technique to evaluate the data points and calculate the standard deviations. This multidimensional approach ensures a thorough consideration of the several factors affecting load generation at customer locations.The proposed solution utilizes different solar generation forecasting methods including Linear Regression, Deep Neural Networks and Long-Short Term Memory. The prediction horizon is considered short-term. These models are used to forecast loads over a period of up to 168 hours, providing a detailed understanding of power demand over an extended period. Although all the methods have acceptable performance, LSTM showed more promising results. The research findings highlight the effectiveness of the proposed model, achieving an impressive 5.7% Mean Absolute Error between actual and predicted generation load for 2022 data test. This level of accuracy demonstrates the model's robustness in capturing the nuances of the evolving energy landscape. Consequently, the proposed forecasting model emerges as a valuable and strategic tool for utilities, providing insights that can significantly enhance the resilience and efficiency of the power grid amidst the dynamic context of energy and technological advancements.