Enhancing Decision-making in the Electric Power Sector with Machine Learning and Optimization

The electric power system infrastructure is essential for modern economies and societies, as it provides the electricity needed to power homes, businesses, and industries. It is of critical importance that the operation of the electric power system is optimized to serve the electricity demand reliably and sustainably. Advances in machine learning and optimization have enabled the potential to enhance decision-making in the electric power sector by gaining insight into the vast amount of data stored digitally. The operation of electric power systems poses many challenges, such as the rising integration of renewable energy sources, energy storage, and the aging transmission infrastructure. This thesis explores machine learning and optimization techniques to enhance decision-making concerning decarbonization targets, integration of renewable energy sources, cost savings, and reliable power supply. The first work presents a framework for predicting electricity demand. Comparing statistical and machine learning models for short- and medium-term forecasting revealed that machine learning methods provide higher accuracy and demonstrate good transferability. This highlights the importance of choosing the appropriate model to accurately predict the electricity demand, especially where historical data may be scarce. Next, we examined the electricity transmission grid using machine learning classification techniques to identify causes of power distribution network disturbances. Besides indicating variables that explain fault occurrences on average, identifying specific variables for each fault is essential. To address this challenge, we used a technique called Integrated Gradients for interpreting the decision process of a deep learning model, emphasizing the value of detailed insights into specific fault occurrences. In the third work, we adopted probabilistic forecasting to account for the the uncertainty when predicting electricity generation from wind power. As point forecasts don't account for uncertainties in the predictions, relying on probabilistic forecasts is necessary. We showed that deep learning models can provide accurate day-ahead probabilistic forecasts and discovered that including historical weather data and numerical weather predictions as exogenous variables improves forecast accuracy. In the fourth and fifth works, we modeled the electric power system using optimization techniques. The fourth work analyzed the benefit of using a low-cost thermal energy storage unit called Thermal Energy Grid Storage (TEGS) for balancing solar energy system's intermittent generation, highlighting storage's crucial role for grid reliability. In the fifth and final work, we optimized the engineering design of TEGS to minimize the cost of decarbonization in electric power systems. The findings show that TEGS enables cost-effective grid decarbonization and improves reliability compared to a baseline scenario where TEGS is not an available technology