Wind and photovoltaic generation are strongly affected by geographic heterogeneity and meteorological disturbances, resulting in nonlinear and nonstationary dynamics that hinder cross-regional power forecasting. In renewable-integrated power systems, forecasting errors directly propagate to operational decisions, such as day-ahead scheduling, reserve allocation, real-time balancing, and curtailment mitigation, thereby influencing system reliability and operating efficiency for system operators and renewable asset owners. To address cross-domain forecasting under domain shifts, we propose a Cross-Domain Invariant Feature Evolution Network (CDIFENet) with two modules: (i) a Geo-Aware Decomposition Engine (GADE), which combines Grey Wolf Optimizer (GWO) based tuning of Variational Mode Decomposition (VMD) with Seasonal-Trend Decomposition (STL) to extract more stationary subsequences, and (ii) a Spatiotemporal Evolution Predictor (STEP), which serially couples LSTM and iTransformer to capture local temporal dynamics and global dependencies. Experiments on datasets from 11 wind and photovoltaic power stations spanning plains, deserts, and mountainous regions show that CDIFENet consistently outperforms six baselines in cross-domain forecasting without retraining. At site S5 in the northwestern mountains, for example, the R2 rises from 89.94% to 98.05%, while RMSE, MAE, and MBE decrease by 56.1%, 70.8%, and 30.9%, respectively.