Photovoltaic forecast: anchoring a tree model and keeping only useful networks reduces the error by 3.21%

On 56,078 windows of two solar installations, combining LightGBM with seven weighted neuron networks is better than the simple average or the indistinct addition of modules.

To predict solar production from 15 minutes to 4 hours, a good decision tree model is often the best predictor in isolation. A Sino-German team, including Alois Knoll (TU Munich), investigated on arXiv whether complementary temporal models could still reduce its average absolute error without the worst of them degrading the result. Their candidate bank brings together LightGBM and seven neural networks with different time representations, combined with positive weights whose sum is one: a useless candidate can receive zero weight, while the tree model serves as an anchor and fallback.

Tested on 56,078 windows from two public PV systems from the PVDAQ database installed in the same location, the version that shares the same weights for all horizons reduces the normalized error by 3.21% compared to the tree model alone. Specific weights per horizon or additional neural correction do not add anything. The authors conclude that it is better to selectively combine complementary forecasts than to average everything, while stressing the limitations of the study: an evaluation period already examined before, no independent replication or analysis by type of weather.