Embedding Cyclical Information in Solar Irradiance Forecasting

TL;DR

This paper shows that embedding temporal information, like timestamps, significantly improves the accuracy of solar irradiance forecasting models that use neural networks.

Contribution

The study demonstrates the effectiveness of incorporating cyclical temporal features into neural network models for solar irradiance prediction.

Findings

Embedding timestamp information improves forecast accuracy

Models with temporal features outperform those without

Data from NTU Singapore confirms the benefit of temporal embedding

Abstract

In this paper, we demonstrate the importance of embedding temporal information for an accurate prediction of solar irradiance. We have used two sets of models for forecasting solar irradiance. The first one uses only time series data of solar irradiance for predicting future values. The second one uses the historical solar irradiance values, together with the corresponding timestamps. We employ data from the weather station located at Nanyang Technological University (NTU) Singapore. The solar irradiance values are recorded with a temporal resolution of $1$ minute, for a period of $1$ year. We use Multilayer Perceptron Regression (MLP) technique for forecasting solar irradiance. We obtained significant better prediction accuracy when the time stamp information is embedded in the forecasting framework, as compared to solely using historical solar irradiance values.