Advanced hierarchical control based on extreme gradient boosting (XGBoost) algorithm for hybrid solar-diesel-battery standalone microgrid

Abstract

Forecasting the stochastic behaviour of solar energy is crucial for designing a resilient microgrid. Typically, this capability is employed by the third level of hierarchical control to generate reference signals for the first and second control levels. However, the lack of inertia in the power grid, coupled with climate change dynamics and the increasing integration of renewable energy sources through power converter interfaces, adversely impacts existing high-precision methodologies. Accurate prediction of the stochastic properties of solar energy is essential for ensuring the stable operation of microgrids. This thesis presents a methodology that utilizes multiple linear regression (MLR) and correlation matrices, in conjunction with an advanced AI-driven algorithm (eXtreme Gradient Boosting, XGBoost), to forecast the power output and reliability of renewable energy sources such as solar power. The output of this new algorithm serves as a reference signal for estimating the reference inverter current. The methodology was validated using historical data from operational solar power plants in Germany over three years. Additionally, the performance of this approach was benchmarked against other state-of-the-art neural networks, such as the long short-term memory (LSTM) model, using metrics like mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (CoD). Furthermore, the prediction algorithm was applied to a PV-Diesel hybrid system operating independently of the grid. This system comprises a DC-DC step-down converter that maintains constant DC bus voltage, enabling an inverter to maintain a steady frequency at connection points. This operational setup ensures stability during dynamic fluctuations in energy production caused by rapid changes in solar irradiation or load variations. The efficacy of the virtual inertia control (VIC) developed for the off-grid PV-Diesel hybrid application was validated in real-time using MATLAB Simulink, and a scaled prototype of 2 kW was constructed in the laboratory. Additionally, a virtual inertia system was combined with droop control applied to the power inverter interface to enhance frequency deviation management in the off-grid PV-Diesel hybrid system. Ultimately, the research mitigates the adverse impacts on solar energy systems through the deployment of advanced forecasting techniques and the utilization of virtual inertia technology.

Date	14 Jun 2024
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Ambrish Chandra (Supervisor) & Miloud Rezkallah (Co-supervisor)