A Spatio-temporal Hybrid Deep Learning Architecture for Short-term Solar Irradiance Forecasting

Abstract

Accurate forecasting of solar irradiance plays an important role in the reliable grid integration of renewable energy. Forecasting of photovoltaic (PV) generation of solar sites is routinely studied in academia and industries, however spatio-temporal weather dependencies amongst regional plants are often ignored. It is intuitively expected that PV generation of neighboring solar plants are correlated and may show similar time-varying patterns. In this paper, we are investigating spatiotemporal correlation between solar sites which are at different geographical locations. We propose a hybrid deep learning model which is a combination of a residual network (ResNet) to learn different representations of data and long short-term memory (LSTM) to capture long temporal dependencies. We measure the performance of the proposed model relative to other deep learning models, such as a convolutional neural network (CNN) and long short-term memory (LSTM), ResNet, and ResNet/multilayer perceptron (MLP) using seventeen years of meteorological data from twelve different sites in Philadelphia. Our experiments show that the proposed architecture integrating special and temporal contexts provides superior performance in solar irradiance forecasting.