Multi-step short-term solar radiation prediction based on empirical mode decomposition and gated recurrent unit optimized via an attention mechanism

The utilization of sustainable and renewable energy for building heating is pivotal to achieving sustainable development goals. However, the volatility of solar energy poses serious stochastic perturbation problems for heating systems. High-accuracy solar radiation prediction contributes to building energy systems' economic and reliable operation and provides a basis for the decision-making of investors. To enhance the solar radiation prediction performance, a novel EMD-GRU-Attention (EGA) model through the hybridization of the empirical mode decomposition (EMD), gated recurrent unit (GRU), and attention mechanism is proposed. First, the Kalman filter (KF) is used to denoise the original solar radiation time series, and the EMD method is employed to decompose the denoised sequence into multiple subsequences to minimize the complexity and unpredictability of the solar radiation data. Afterward, the subsequences are modeled by the GRU model optimized via an attention mechanism (GRU-A) to predict solar radiation. To rigorously verify the multi-step prediction performance of the developed hybrid model, extensive comparative experiments are conducted by building several deep learning models, including RNN, GRU, EMD-GRU, and GRU-A. Afterward, a seasonal analysis is also conducted. The results show that the proposed EGA model has satisfactory performance in four seasonal datasets, especially in the autumn dataset. In terms of the one to three-step solar radiation forecasting task, the R2 of the proposed model is 0.983, 0.972, and 0.960, respectively. Thus, the feasibility and superiority of the proposed model in the solar radiation prediction task are confirmed. This study is anticipated to contribute to a reliable reference for the practical regulation strategies of solar-based building energy systems.