Using a Predictive Model to Reduce Emissions/Energy Costs with Virtual Power Plants in North India

Abstract

This project aims to reduce the emissions and energy costs in a pollution riddled country like India by optimizing the energy output from power plants. The emissions/costs are minimized using Mixed Integer Linear Programming (MILP), an optimization tool that accounts for linear constraints and objective function. Utilizing the MILP results allows for the creation of Unit Commitment and Economic Dispatch (UC/ED). UC determines a schedule for which power plants should be on or off at which times while accounting for constraints such as startup/shutdown costs and ramping flexibility. Economic Dispatch optimizes power generation levels for each specific power plant while considering constraints such as meeting the energy demand and minimum/maximum generation limits to be found.

Virtual Power Plants (VPPs) are the real-world application of our optimized UC/ED findings as multiple decentralized energy sources, power plants in this case, can aggregate resources and function as a singular plant. It can recalibrate energy usage based on data such as hourly power demand, weather, fuel source, etc., and apply UC/ED by using constantly updated real-time data. A Random Forest Regressor machine learning model which predicts the gains in terms of CO2 emissions of having VPP assets in the power grid was used. The accuracy at 70F, 73F, and 76F cooling points were 98.89%, 97.76% and 87.67% respectively. A machine learning model is used instead of the MILP model as it works much faster and can be feasibly used for day-to-day operations.