Adversarial multi-source domain generalization approach for power prediction in unknown photovoltaic systems

Abstract

Accurate forecasting of power output for previously unseen photovoltaic installations is of critical importance to the reliability and efficiency of renewable-energy management systems. Existing data-driven PV prediction techniques rely primarily on historical measurements from familiar systems, which constrains their applicability to new sites without prior observations. To address this limitation, we introduce a Generative Adversarial Domain enhanced Prediction Network (GADPN). GADPN employs an adversarial generator to synthesize diverse pseudo domain samples that mitigate distributional discrepancies between source and target domains. Through an alternating optimization regime, the framework enforces both semantic consistency and manifold regularization constraints to align synthesized and empirical feature representations, while a Transformer-based predictor captures local and global temporal dynamics. We evaluate the proposed approach on nine geographically and capacity-diverse PV systems (ranging from 2.16 kW to 45.78 kW) under a zero-sample setting. Experimental results demonstrate that GADPN achieves coefficients of determination exceeding 0.97 in eight of the nine cases and attains a peak coefficient of determination of 0.9993, outperforming state-of-the-art baselines. These findings confirm GADPN’s effectiveness for robust, zero-sample generalization in PV power forecasting.