Implementation of Machine Learning Models for Transmission Grid Monitoring and Blackout Prevention

Sustainable energy is one of the most important systems in modern society. To achieve the sustainable energy provision goal, Ethiopia has made significant investments in renewable energy resources in the last decades, resulting in a substantial increase in power generation. However, the growing demand for power consumption and uneven spatial distribution of generation and consumption centers have caused electrical network overloads and frequent power outages. These local outages can potentially escalate into widespread blackouts across the country. Solutions for this may be upgrading and construction of new lines, which require substantial investments, and are challenging for a developing economy in addition to ongoing big generation projects. An alternative approach involves optimizing the use of existing transmission lines and infrastructure through the implementation of real-time monitoring systems. This dissertation introduces a real-time current carrying capacity (or ampacity) monitoring known as dynamic line rating (DLR). DLR involves continuously monitoring overhead transmission lines (OHTLs) by taking into account surrounding weather conditions. The method involves adjusting meteorological forecasts to the local weather conditions along the transmission line. Furthermore, weather station placement can be optimized to monitor areas or spots where the transmission line might reach its highest temperatures. The DLR system offers the benefit of increasing ampacity during emergencies or sudden load fluctuations, thus improving power system security (i.e., blackouts). DLR utilizes a machine learning model to assist transmission system operators (TSOs) in congestion and overload management in coupling with grid frequency deviation monitoring for supply and demand balancing. This contributes significant improvement to grid operation, control, and planning. However, analyzing risk avoidance mechanisms in ampacity prediction is essential for grid operators. Most existing DLR ampacity forecasting techniques were based on a deterministic forecasting mechanism, which is prone to errors. In this dissertation, Quantile regression forest (QRF) probabilistic DLR forecasting is used for OHTL ampacity forecasting. QRF is well-suited for short-term predictions like minutes to hours in congestion monitoring. The presented frameworks for determining DLR based on probabilistic forecasts demonstrate that low quantiles can reduce the risk for decision-makers, avoiding overestimation and minimizing losses compared to deterministic DLR forecasting. Maintaining frequency balance is crucial for effective line congestion monitoring and ensuring power system security (i.e., avoiding blackouts). Combining DLR monitoring with frequency deviation prediction can reduce power interruption by maintaining a constant demand-supply power chain. Deviations from the reference frequency of 50Hz (case in Ethiopia), result from imbalances in power supply and demand due to fluctuations in load patterns. This needs a strong real-time monitoring system for demand-supply balancing in addition to transmission line congestion monitoring. In the Ethiopian grid system, the generation scheduling response is usually load following. During high load variation and contingencies, it needs a fast response and requires rapid adjustments to prevent potential power outages and cascading blackouts. To guarantee frequency stability in such a complex and uncertain environment, TSOs intensively monitor the system and allocate expensive control reserves. An improved understanding of the frequency dynamics, line congestion, and its interaction with power demand-supply imbalance could greatly facilitate control efforts and contribute to power system stability. This dissertation presented a cost-effective wireless sensor network design suitable for implementation in Ethiopia and other developing nations. The study focuses on creating energy-efficient wireless mesh networks (LPWMNs) utilizing Semtech's LoRa (long-range) modulation technology. The network is intended for monitoring power grid infrastructures that span tens to hundreds of kilometers. The key objective is to gather data from the grids to support a machine-learning model implemented for DLR purposes. The performance of the proposed protocol is analyzed first by simulation and then with a demonstrator network on the university campus. The findings indicated that the proposed network achieves a sufficiently high packet delivery ratio (PDR) to enable the monitoring of static infrastructures extended over geographical areas. Lastly, a web-based tool was created to visualize real-time generation data, transmission grid and load data, weather data, and transmission lines predicted capacity once the system is operational (i.e., once installed and running). This tool also conducts power flow and transient analyses to assess whether temporary congestion can be managed without exceeding maximum conductor temperatures. It proposes a solution to support a flexible electrical grid for a short-duration plan, which is a crucial component for the optimal utilization of existing infrastructure and a successful energy transition plan.