Time Series Forecasting and Stochastic Optimization for the Operation of Decarbonized Power Grids

Abstract

Every aspect of the power grid, from generation to transmission, is influenced by the weather. As climate change amplifies the frequency and severity of extreme weather events, it becomes increasingly difficult to maintain a stable grid. Ironically, climate change mitigation efforts are exacerbating this problem, with the shift toward renewable energy sources strengthening the dependence of power grids on the weather. For Texas – the only state with an independent and deregulated grid, and the nation’s renewable energy leader – there have already been severe consequences in the form of deadly blackouts. To address these issues, grid operators such as the Electric Reliability Council of Texas (ERCOT) must improve grid resilience through energy storage and other infrastructural changes, while implementing better forecasting techniques and risk management practices. Modern machine learning techniques have shown great promise in time-series forecasting tasks. In this thesis, I first train recurrent neural network (RNN) models to make short-term weather and load forecasts. I then build a simulation with stochastic linear programs to explore strategies for optimizing the operation of a decarbonized power grid under forecast uncertainty