Enhanced photovoltaic power generation forecasting for newly-built plants via Physics-Infused transfer learning with domain adversarial neural networks

Accurate anticipation of photovoltaic (PV) power generation in advance is crucial for renewable energy development, infrastructure planning, efficient power grid operations, and energy management. Emerging data-driven methods represented by deep learning have provided effective solutions for PV power generation forecasting. However, conventional data-driven forecasting algorithms rely heavily on extensive historical data, making it challenging to predict the output of newly-built PV plants (NPP) due to limited data availability. To address this concern, a novel physics-infused transfer learning model is proposed for short-term cross-plant PV power prediction, which leverages the public prediction knowledge learnt from relevant PV plants to develop a predictor compatible with NPP. Key innovations include: (1) Firstly, we propose a forecasting-oriented Domain Adversarial Neural Network (DANN) that incorporates Wasserstein distance, enabling the reduction of the discrepancy in the feature vector distribution between source and target plants through iterative adversarial training. The feature vectors from NPP would be compatible with a predictor trained on the feature vectors from data-rich PV plants. (2) Secondly, this framework implicitly transfers the intricate regression patterns from data-rich PV plants to NPP with limited historical measurements, allowing for the capture of short-term fluctuations using real-time data as input. (3) Subsequently, a well-calibrated physical model chain enables the refined and stabilized numerical calculations in the conversion from solar irradiance to PV power output, thus extending the forecasting horizon. (4) Finally, the Bayesian combination model (BCM) is deployed to coordinate the two sub-predictors, enabling simultaneous prediction of trends and short-term fluctuations. The hybrid framework is formulated and validated adopting real-world PV data from four PV plants in North China. Simulation results indicate that the prediction accuracy outperforms all selected benchmark models. Compared to standalone models trained on the scarce data of NPP, the transfer learning-based model can reduce prediction errors by approximately 20% to 68%. Moreover, the proposed model demonstrates excellent generalizability and robustness, effectively mitigating the effects of geographical disparities among source and target domains, the performance of cross-geographic region transfer forecasting task is improved by 30% − 40% compared to the traditional alternative.