Multistep Multivariate Scenario Generation and Forecasting for Power Systems using Machine Learning

The penetration of renewable energy in the power system has grown considerably in the past few years. While this use may come with an abundance of advantages, it also introduces new challenges in operating the 100+ years old electrical network. Fundamentally, the power system relies on a real-time balance of generation and demand. However, renewable resources such as solar and wind farms are not available throughout the day. Furthermore, they introduce temporal variability to the generation process due to metrological factors, making the balance of generation and demand precarious. Utilities use standby units with reserve power and high ramp-up, ramp-down capabilities to ensure balance. However, such solutions can be very costly. An accurate scenario generation and forecasting of the stochastic variables (load and renewable resources) can help reduce the cost of these solutions. The goal of this research is to solve the scenario generation and forecasting problems using state-of-the-art machine learning techniques and algorithms. The training database is created using publicly available data obtained from NREL and the Texas-2000 bus system. The IEEE-30 bus system is used as the test system for the analysis conducted here. The conventional generators of this system are replaced with solar farms and wind farms. The ability of four machine learning algorithms in addressing the scenario generation and forecasting problems are investigated using appropriate metrics. The first machine learning algorithm is the convolutional neural network (CNN). It is found to be well-suited for the scenario generation problem. However, its inability to capture certain intricate details about the different variables was identified as a possible drawback. The second algorithm is the long-short term memory-variational auto-encoder (LSTM-VAE). It generated scenarios that are very similar to the actual scenarios indicating that it is suitable for solving the forecasting problem. The third algorithm is the conditional generative adversarial network (C-GAN). It was extremely effective in generating scenarios when the number of variables were small. However, its scalability was found to be a concern. The fourth algorithm is the spatio-temporal graph convolutional network (STGCN). It was found to generate representative correlated scenarios effectively.