A Cross-Domain Short-Term Power Prediction Method Based on Multiple Environmental Sensors in Photovoltaic Systems and Domain Generalization Theory

Accurate forecasting of photovoltaic (PV) power generation is crucial for optimizing the management of renewable energy systems and maintaining the stable operation of the power grid. However, current deep network algorithms perform significantly only when there is sufficient environmental sensor data from the PV system. They face considerable limitations when the target system and its sensor data are completely unknown. To address this, we propose a novel cross-domain short-term PV power prediction method based on domain generalization theory [cross domain forecast method (CDFM)]. This method introduces generalization theory into the task of PV system power prediction, using maximum mean discrepancy (MMD) as a measure of the disparity between environmental sensor data from different domains, and combining adversarial generation to achieve domain generalization of the data. Subsequently, mutual information theory is utilized to quantify the correlations among the PV system’s multidimensional features, facilitating optimal feature selection through feature combination techniques. Finally, the Informer architecture is introduced to learn feature representations from both the source domain and the generalized domain, aiming to predict the PV system’s power output in an unknown target domain. We conducted model generalization studies using PV arrays of different grades in two distinct locations and performed comparative validation with three advanced methods. The results show that CDFM can be successfully applied to the power prediction tasks of unknown target PV systems. The final root mean square error (RMSE)/mean absolute error (MAE) indicators can reach 0.7689/0.4316, and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${R}^{{2}}$ </tex-math></inline-formula> can reach 0.9945.