Improving forecasting accuracy of renewable energy generation via periodicity-aware deep learning framework with Time2Vec

Accurate forecasting of photovoltaic (PV) and wind power generation is vital for reliable renewable energy operations. However, their differing temporal characteristics, especially in periodicity, pose challenges for unified prediction models. This paper proposes a periodicity-aware hybrid deep learning framework designed to enhance forecasting accuracy for both PV and wind power generation. The framework incorporates Time2Vec (T2V) to extract periodic and non-periodic temporal features and combines Bidirectional Temporal Convolutional Networks (BiTCN) with Bidirectional Gated Recurrent Units (BiGRU) through a flexible serial-parallel architecture. Based on the autocorrelation of the input data, the model selects a parallel structure for strongly periodic PV and a serial structure for weakly periodic wind data. The T2V layer enhances feature representation, while the serial–parallel design provides architecture-level flexibility that improves forecasting accuracy. The proposed T2V–BiTCN–BiGRU model was evaluated on real-world PV and wind datasets from China and Belgium. Experimental results demonstrate that the model consistently outperforms IEDN-RNET, VAM-MTL, and CEEMDAN-EWT-BiTCN-BiLSTM-AT, achieving high R 2 and low MAE, RMSE, and MAPE across datasets. Moreover, the model maintains robust accuracy and stability, confirming its generalizability. This work highlights the effectiveness of combining periodicity-aware encoding with architecture-level hybrid models, providing a novel and flexible approach for high-precision hybrid wind and solar power forecasting. • A periodicity-aware serial–parallel forecasting framework is proposed. • Data periodicity enables architecture-level flexibility in model design. • The proposed model outperforms benchmarks on real-world PV/wind datasets. • The framework maintains flexibility across alternative CNN–GRU architectures.