Predicting Power Usage For Renewable MicroGrid Communities
Summary
Developed and implemented a machine learning model (LSTM) to accurately predict power usage for renewable microgrid communities
Reduced prediction error by 77% compared to the original model.
Mission
The project's mission was to enhance the accuracy of predicting power usage for renewable microgrid communities, specifically for Block Energy. The goal was to develop a machine learning model capable of more accurately forecasting power demand, especially during peak times, compared to the current moving box average algorithm. This is crucial for optimizing energy dispatch and ensuring economic viability for microgrid communities.
Background
Block Energy is focused on adapting to the evolving energy demands of residential communities, especially with the increasing use of electric vehicles and large appliances. The existing model used by Block Energy—a moving box average—has limited ability to predict peak loads. This project explored the use of advanced machine learning techniques, such as Long Short-Term Memory (LSTM) neural networks, to improve load prediction accuracy.
My Role
My role involved applying machine learning methods learned from the MIT course 2.C51 to develop a model that could better predict energy consumption patterns. While my teammates worked on data exploration, preprocessing, and clustering, I implemented an LSTM model to predict load profiles based on historical data and tested its performance against the baseline moving average model.
The Long Short-Term Memory (LSTM) model is a type of recurrent neural network (RNN) designed to effectively learn and remember long-term dependencies in sequential data. It is especially powerful in applications such as speech recognition and time-series prediction.
Results
The LSTM model significantly outperformed the baseline, reducing the Mean Squared Error (MSE) by 77% for single-house predictions. However, the model did not perform as well on a larger collective of houses, indicating the need for further refinement. Clustering data into more homogenous groups showed potential for improved prediction accuracy. Future work includes exploring different clustering methods and model architectures to enhance prediction capabilities for diverse household groups.
Huge thanks to my teammates Nat Wong, Rachael Rosko, our advisors Professor George Barbastathis, Difei Zhang, and Christina Ji, and our sponsor Cesar Maeda.