Project Overview

Project Summary:

Our project aims to predict wind and solar power generation on Jeju Island, addressing the challenge of energy oversupply. Accurate prediction of the model is essential to solve this problem. We were able to implement the model by integrating the “Attention Mechanism” into the LSTM to obtain better results than previous studies. After that, the web was implemented using Open API and Django. On, the web, we provide predicted values through graphs and tables for users to easily understand. We also calculate the required hourly fossil fuel power generation.

Identifying the Challenge

The Social Problem:

When the power supply is temporarily overflowing, the government will limit the output to prevent overloading of the transmission and distribution network. It caused a social loss of 2.2 billion won in 2021. If accurately predicted, it will be possible to reduce the output limit and reduce the thermal power generation. That will contribute to making Jeju a carbon free island, a project Jeju plans to do by 2030.

Innovation and Uniqueness

Why Our Project Stands Out:

Regarding the prediction of solar power generation, LSTM showed the best performance in previous studies. However, previous studies were conducted before the attention mechanism was introduced. The attention mechanism is known to be a model that completely overcomes the limitations of LSTM, and in this project, the attention mechanism is integrated into LSTM to show better performance than previous studies.

Insights and Development

Learning Journey:

Model optimization and hyperparameter tuning were the main components of our development process. The performance of the model with only day and temperature data for the first time was terrible, but it was found that the performance improved when the amount and type of training data were gradually increased. Through this, we learned that if we could naturally receive more detailed and diverse data, we could further improve the model performance.

Development Process:

Data from “Data portal”, “KMA”, and “Air Korea” were used. For the model, LSTM and pre-trained BERT were integrated and implemented. After that, the next day's data was received through the open API, and the web page was implemented in the local environment using React and Django.