A Nature study by researchers from Peking University and Alibaba Group's DAMO Academy used a deep‑learning model

A Nature study by researchers from Peking University and Alibaba Group’s DAMO Academy used a deep‑learning model on sub‑metre satellite imagery and 7.

Researchers from Peking University and Alibaba Group’s DAMO Academy published a study in Nature this week that uses AI to produce the first complete, high‑resolution inventory of a nation’s wind and solar infrastructure. The model identified 319,972 solar photovoltaic facilities and 91,609 wind turbines across China, creating a unified dataset intended to improve system‑level coordination and planning. This nationwide view matters because it lets operators and planners evaluate how existing assets interact, rather than relying on scattered or modelled assumptions.

To build the map the team trained a deep‑learning model on sub‑metre satellite imagery and processed 7.56 terabytes of data. The resulting dataset spans 1,915 Chinese counties and includes a full range of installations, from rooftop panels in coastal cities to utility‑scale wind farms on the Mongolian plateau. Handling such geographic and deployment diversity required the model to distinguish many installation types under varying terrain and image quality.

Using the inventory, the authors quantify solar — wind complementarity across real, geolocated assets and show that pairing complementary sites substantially reduces generation variability. The analysis finds that the effectiveness of complementarity improves as the geographic scope of pairing expands, meaning wider spatial integration can yield stronger smoothing of supply. Because the dataset is grounded in observed facilities rather than hypothetical deployments, it enables empirical assessment of how complementarity performs under present infrastructure patterns.

The paper contrasts this national dataset with current coordination practices in China, which largely operate at the provincial level, and argues that national‑level integration would better stabilise supply and reduce curtailment. Liu Yu of Peking University described the map as giving operators a “God’s‑eye view,” a perspective the authors say could help avoid wasted renewable generation by revealing where complementary resources and load centres are mismatched.

Technically, the study’s advance rests on training a model capable of coping with wide heterogeneity in installations, terrain and image quality — challenges that prior complementarity studies typically avoided by using hypothetical or modelled deployments. By identifying and geolocating actual assets at scale, the dataset converts complementarity from a theoretical benefit into a testable, actionable planning input. For grid builders and operators the inventory supplies concrete levers: it enables pairing distant but complementary sites to reduce variability, flags regions where growing data‑centre demand overlaps with renewable capacity, and provides a basis for moving coordination from provincial to national planning to lower curtailment.

coordinate rapidly expanding renewables with new, electricity‑intensive loads.