Time-Shifted Gramian Angular Field and Recursive Plot Convolutional Neural Network to Adapt Solar Cell Dataset Overfitting in Hybrid Power Generation
Abstract
Solar cell power generation has been employed widely. Battery storage becomes an important part to ensure output availability. However, battery price and technologies are costly. Hybrid generation for daylight supply becomes one of the solutions. This paper proposed hybrid solar cell-diesel power generations. Diesel power generation should maintain overall power requirement to fulfill demand. Since irradiation changes relative to sun positions to the earth, solar cell power output varies overtime. As the solar cell output changes overtime, the diesel power generation should be determined. The differences between the predicted solar power and the energy demand determine the energy amount should be provided by the diesel generator. In order to provide prediction, convolutional neural network utilizing multi-layer perceptron with hyper-parameter optimizations and statistics transformation, as well as data transformation are employed. However, solar cell data is highly overfitted. The folded and time-shifted gramian angular field and recursive plot are then proposed. As results, the proposed methods are able to reduce overfitting on solar cell datasets to increase the prediction performances.