Adding more predictors to an ensemble does not automatically improve a forecast. Reducing Multi-Horizon Photovoltaic Forecast Error through Tree-Anchored Expert Combination studies whether complementary temporal models can reduce the residual errors of a strong direct tree forecaster.
Let the strongest standalone model remain the anchor
The candidate bank contains LightGBM and seven neural implementations with different temporal representations. The combiner fits mean absolute error using nonnegative weights that sum to one. Unhelpful candidates can receive zero weight, while the tree provides a development-stage fallback.
The main design shares weights across forecast horizons from 15 to 240 minutes. Horizon-specific weights and alternative loss choices test whether additional complexity is useful. This is selective combination, rather than an assumption that every member should contribute equally.
What improved and what did not
On 56,078 windows from two co-located public PVDAQ systems, the historical shared-MAE policy reduces scaled mean absolute error by 3.21% relative to the tree. The latest bounded neural correction does not reduce MAE further than that static combination.
The second finding matters when deciding whether another module is worth maintaining. In this experiment, the additional correction did not improve the stated error measure beyond the simpler weighted combination. The paper supports examining complementary predictions within this candidate bank, rather than counting the number of models as evidence of quality.
Preserve the evaluation limitation
Both analyses reuse a previously inspected evaluation period. The authors explicitly say the results do not establish independent replication, weather-specific causation or universal gains from eight experts.
For a new forecasting project, separate development decisions from a genuinely uninspected test period and compare against the strong tree baseline under the same horizons and error definition. The reported reduction is a result in this historical setting. Whether the policy remains useful at another site, under different weather or in later operating periods needs its own evaluation.
Comments