Reducing Multi-Horizon Photovoltaic Forecast Error through Tree-Anchored Expert Combination
This paper addresses multi-horizon photovoltaic power forecasting from 15 to 240 minutes by asking whether residual error from a strong gradient-boosted tree can be reduced through selective, MAE-aligned combination with neural temporal predictors. The core idea is a constrained simplex mixture where the tree anchors the ensemble and weaker candidates can be assigned zero weight, with shared coefficients keeping the combiner parsimonious. The contribution matters because it isolates genuine aggregation value against a competitive baseline rather than inflating model count or comparing across mismatched tasks.
The paper makes a carefully bounded empirical case: the accepted shared-MAE policy achieves a 3.21% historical scaled MAE reduction over the standalone tree, and a separate retraining check without future weather yields a 1.78% reduction. Yet the authors candidly note that "both analyses reuse a previously inspected evaluation period" and "do not establish independent replication, weather-specific causation or universal gains from eight experts" (Abstract). The principal merit is methodological honesty and a protocol designed to test whether fitted combination adds value beyond a strong direct baseline, not whether eight experts are universally necessary.
Experimental discipline is the paper's strongest feature. The authors insist that "LightGBM is the reference because a method that adds several neural networks should first demonstrate value over a competitive tree under the same information and target definitions" (Section 1), and they enforce a simplex-weighted combiner that permits zero weights to avoid dilution by poor candidates. The $A(w)$/$C(w)$ error accounting decomposes aggregate gains into member-error reduction versus cancellation, offering a transparent descriptive diagnostic without overstating causality.
A critical weakness is research-history contamination: the 2016 evaluation period "had already been inspected during earlier research" and "freezing this execution does not remove the earlier research history" (Section 2.2), undermining confirmatory claims. Co-location further erodes external validity because "overlapping targets and seeds are not independent site replications" (Section 5.2). Evidence for added complexity is also fragile: horizon-specific weights introduce 105 extra coefficients, but the gain is sensitive to block length and "the 7-day interval is slightly positive and the 28-day interval spans zero" (Section 4.3), suggesting the extra parameters may not be warranted.
The evidence supports the narrow claim that selective combination outperforms equal averaging within this specific bank, and the authors properly avoid unfair literature comparisons by stating they "do not compare their published error values directly with ours" (Section 1). The descriptive error accounting with $A(w)$ and $C(w)$ is consistent with measured gains, but it is explicitly post-hoc: "These quantities describe realized loss, not changes in the individual frozen predictors or evidence of a causal mechanism" (Section 4.4). Because the latest bounded neural correction fails to beat the static combination, the paper successfully argues against indiscriminate module addition; however, it cannot establish that any reduction would generalize to new sites or issue-time weather.
Hyperparameters and solver details are documented precisely, and the authors report that "The matched retraining records include a frozen feature/split protocol, code and data hashes, three seed predictions, an independent metric audit and hashes for 126 remote model/configuration artifacts" (Data and Code Availability). Despite this thorough protocol description, they explicitly warn that "Public archival release and redistribution checks remain pending; no public reproducibility deposit is claimed" (Data and Code Availability). Without open code or frozen artifacts, independent reproduction is presently blocked.
A strong direct tree model may remain the best standalone photovoltaic forecaster, yet a weighted combination can reduce its residual errors. We examine whether complementary temporal predictors can lower the tree's mean absolute error (MAE) over 15--240 minutes without weaker ensemble members degrading performance. The candidate bank contains LightGBM and seven neural implementations with different temporal representations. A combiner fitted to MAE has nonnegative weights that sum to one. Unhelpful candidates can receive zero weight, while the tree anchors the combination and provides a development-stage fallback. The main design shares weights across horizons; horizon-specific weights and alternative losses test whether further complexity is needed. On 56,078 windows from two co-located public PVDAQ systems, the historical shared-MAE policy reduces scaled MAE by 3.21\% relative to the tree. The latest bounded neural correction does not reduce MAE further than this static combination. The results support selective combination of complementary predictions in this bank rather than equal averaging or indiscriminate addition of modules. Both analyses reuse a previously inspected evaluation period. They do not establish independent replication, weather-specific causation or universal gains from eight experts.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.