China's power system is undergoing a dual shift: urbanization is driving up electricity demand, while the growth of solar and wind capacity makes supply more capricious. A seasonal forecast three months ahead could help plan load, maintenance, and fuel purchases. However, dynamical models have long shown weak predictive skill in China, and local practitioners have grown accustomed to relying on statistical methods. This work fits into the broader effort to assess the predictive value of climate models for energy, alongside studies of El Niño, space weather, and local forecasts, but it stands out by directly testing energy-relevant variables over Chinese territory.
The basis was a retrospective version of the GloSea5 global seasonal system with 24 ensemble members, initialized in November and May for the winter and summer seasons from 1992 to 2011. The model was compared against the ERA-Interim reanalysis for three variables: near-surface air temperature, 10-meter wind speed, and incoming shortwave radiation. For each variable, maps of Pearson correlation between the ensemble mean and observations were constructed, then regions important for energy were selected. Within regions, probabilistic forecasts were assessed using reliability diagrams, Brier scores, and ROC curves, with the contribution of the linear trend checked separately.
The greatest skill was found for winter wind speed in the South China Sea: the correlation reached 0.78, and the probabilistic forecast of events above the median gave a Brier score of 0.297 and ROC skill of 0.620. This region is linked to the East Asian winter monsoon: during El Niño, the pressure gradient between the Siberian High and the warm ocean weakens, and winds subside. Winter solar radiation in eastern China was forecast with a correlation of 0.62, likely due to its connection with cloudiness and precipitation driven by El Niño and Indian Ocean temperature. Summer temperature in Beijing had a correlation of 0.57, and after detrending still 0.48, indicating a genuine predictable signal rather than just warming. Meanwhile, summer winds and winter temperature showed little significant skill almost everywhere.
This work shows that even with the overall patchiness of dynamical forecasts in China, there are local areas with high predictive power suitable for practical use. Unlike neighboring studies that often focus on precipitation or global indices, here we test exactly the variables that directly affect energy generation and consumption. This is the first systematic evaluation of GloSea5 for China's energy sector, and it provides concrete geographic targets for future services.
Further skill gains are possible by increasing the ensemble from 24 to 42 members, as in the operational GloSea5, and through statistical calibration based on large-scale atmospheric drivers such as the Arctic Oscillation or the Middle East jet stream. Also promising is the direct comparison of forecasts with actual output data from wind and solar farms, if grid operators share them.
The results could underpin seasonal climate services for Chinese energy companies, helping them anticipate risks of cold winters, hot summers, and weak winds. They are also useful for hydropower, agriculture, and urban planning in regions with confirmed skill.
The next step is to test whether a statistical post-processing layer based on El Niño and Arctic Oscillation indices improves forecasts where direct model skill is weak. In parallel, a pilot service for the South China Sea and Beijing should be launched, involving users in co-development.
For climatology, an open question remains why the predictability of summer winds and winter temperature in China is so weak, despite clear links to slow ocean processes. It is also unclear how much interannual variability in aerosols, especially in urban areas with frequent smog, distorts solar radiation forecasts, since both the model and the reanalysis use fixed climatological aerosols.
The correlation of 0.78 for winter winds in the South China Sea is comparable to the skill of El Niño forecasts, which is considered the gold standard of seasonal prediction. Yet in the same region in summer, the correlation drops to 0.13, meaning the forecast loses all skill.