Using Machine Learning to Predict Renewable Energy Generation and Smart Grid Designs

Abstract

Traditional electrical grids consist of one way communication between electrical grids and customers. While they are widely used, there are also problems with efficiency. Smart grids serve to address some of these issues with a two-way communication between the utility and customer. One key component of developing smart grids is the ability to accurately forecast energy demand. Additionally, accurately forecasting renewable energy generation allows for the better incorporation of solar and wind energy into the grid, and reduces problems associated with intermittency. This thesis aims to use historical weather, load, and energy generation data to develop machine learning models in order to make these predictions. Two models were developed: one linear regression model, and one deep neural network. With the neural network, three optimizers were used, namely mini-batch gradient descent, Adagrad, and Adam. The models were trained on three New York Independent System Operators regions: New York City, Millwood, and Dunwoodie. The models were evaluated using root mean squared error. It was found that while the deep neural network models were overall better than the linear regression models at predicting load and energy generation at a regional scale, they still showed a lot of room for improvement. Further work will be needed in order to build upon and improve these models.