WaveGRU: A framework with frequency-domain spatial attention for accurate solar PV and wind power forecasting

Abstract

Accurate solar and wind energy forecasts are essential for efficient power production, transmission, storage, and distribution to ensure the stability and reliability of the power system. With the rapid development of Artificial Intelligence (AI) and Internet of Things (IoT) technologies, their potential in renewable energy generation forecasting is becoming increasingly evident. This technological trend provides a new direction for electric energy management research, prompting scholars to actively explore wind and photovoltaic (PV) power prediction methods based on AI and IoT technologies. Previous work has focused on time-domain characterization, which cannot capture intertemporal trends and cyclical features. To address this problem, this paper introduces a wavelet learning framework for modeling complex temporal dependencies in time series data. The wavelet domain integrates time and frequency data, allowing local features of the series to be analyzed at different scales. The framework uses a bidirectional gated recursive unit (BiGRU) to mine long-term dependencies between features. However, mapping time series to the wavelet domain introduces redundant features. Therefore, we propose the frequency-domain spatial attention module (FSA), which adaptively adjusts the feature weights to help the framework pay more attention to the most important features, thus improving the model. This paper uses a cross-corroboration training method customized for time series segmentation to forecast solar PV accurately and wind power generation. We conducted experiments on various time series segmentations (1 to 60 min), and the results show that our proposed model outperforms the compared GRU, LSTM, Transformer, and DLinear methods by reducing the MSE metrics by 69.24%, 68.87%, 69.13%, and 68.32%, respectively.