Machine Learning and Optimization Model Development for Northern Community Energy Planning

This thesis investigates two key areas of northern energy planning through mathematical modelling. Firstly, a predictive machine learning model is developed to estimate stream velocity, a key variable for hydrokinetic power assessment. A generalizable Random Forest model, trained on 4,313 observations with novel geometric parameters, predicts stream velocity with a mean absolute percent error of 24%. This model improves on existing models, which require field data collection or were incompatible with smaller streams suitable for community-level energy planning. Secondly, a new, multi-phase, mixed integer linear programming generation expansion planning model for a comprehensive community energy system is developed which meets both thermal and electricity demand through a single electric load profile. An optimal investment plan consists of wind, solar, and battery storage, at an annualized cost of $13,525. A 20-kW wind turbine was found to lower the cost by 24% compared to using commercially available 100-kW wind turbines.