Combining meteorological and power information of station-measurement and model-prediction with the hybrid CNN-Transformer and CNN-BiLSTM for ultra-short-term photovoltaic power forecasting
The photovoltaic (PV) industry is rapidly developing and plays a key role in the renewable energy sector. PV power is affected by meteorological factors, which are cyclical and stochastic. Accurate prediction of PV power depends on meteorological forecasting techniques, which is crucial for safe ope...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-10-01
|
| Series: | International Journal of Electrical Power & Energy Systems |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S0142061525005575 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The photovoltaic (PV) industry is rapidly developing and plays a key role in the renewable energy sector. PV power is affected by meteorological factors, which are cyclical and stochastic. Accurate prediction of PV power depends on meteorological forecasting techniques, which is crucial for safe operation of power system. This paper proposes an ultra-short-term PV power prediction method based on a hybrid deep learning model that integrates meteorological and power information of station measurement and model prediction. First, we process the meteorological model data for future time steps. The CNN-BiLSTM single-step prediction model is then employed to generate short-term forecasts. Then, historical meteorological observation and historical power output at station are introduced to align with short-term prediction. Finally, matched data are input into CNN-Transformer multi-step prediction model. In this architecture, short-term prediction corresponds to Value, and historical data at station corresponds to Query and Key. By leveraging cross-attention mechanism of Transformer, ultra-short-term output forecast with higher accuracy is realized. Four PV stations in Jiangsu, China are selected for testing. The results show that compared with short-term forecasting, MAE and RMSE of proposed model are improved by about 16.68% and 26.14% respectively, across all stations, indicating higher accuracy and superior generalizability of the proposed hybrid CNN-Transformer and CNN-BiLSTM framework for the ultra-short-term PV prediction. |
|---|---|
| ISSN: | 0142-0615 |