Forecasting

Regional day-ahead wind power forecasting (WPF) requires spatial dependence to adapt to forecast weather. Many existing graph models mainly capture pairwise relations derived from static geography or historical correlations, limiting their representation of multi-farm responses under changing wind fields. This paper proposes AeroSimp-Lite, an NWP-conditioned dynamic simplicial neural network that represents farms, pairwise relations, and three-farm interactions as nodes, edges, and triangles. A fixed sparse complex preserves simplicial closure, while forecast wind vectors dynamically condition interactions at each lead time. Hodge-inspired filtering integrates multi-order information, and a temporal residual decoder corrects a local NWP-based power prior. Regional-consistency and ramp losses constrain aggregate production and rapid changes. Experiments on onshore and offshore clusters demonstrate improved site-level and regional accuracy with moderate computational cost. In the model-order comparison, dynamic edges reduce regional nMAE by 7.1% onshore and 12.0% offshore relative to the node-only model, while adding triangle interactions yields further reductions of 3.8% and 4.5%, respectively, relative to Dynamic-GNN. Ablation results show that dynamic edges provide the main gain, while triangles offer smaller, metric-dependent improvements. However, the lower offshore ramp F1 shows that average accuracy gains do not extend to every objective. AeroSimp-Lite provides a practical approach to modeling weather-dependent higher-order spatial interactions without claiming atmospheric causality.

Wind forecasting in complex coastal terrain, such as the Colombian Caribbean, is hindered by air–sea heat fluxes and nonlinear atmospheric dynamics. This pre-registered systematic review and meta-analysis tests whether advanced architectures (artificial intelligence, hybrid, ensemble) outperform classical ones (physical, statistical). Scopus and Web of Science (last query: 22 September 2026) yielded 100 studies with extractable effect sizes; 76 compared a proposed model with an explicit comparator (primary analysis). Hedges’ g (negative values favor the proposed model) was pooled with REML random effects and Knapp–Hartung inference. The pooled effect was $g=-1.07$ (95% CI [ ]; 95% prediction interval [$-2.77,0.64$]; $I^2=69.5\%$), robust to the estimator and to influential studies, although 59 comparisons were against a baseline of the same family. The estimates favored advanced over classical architectures, but not significantly ($\beta =-0.30$, 95% CI [$-0.99,0.38$]); only 10 studies compared them directly, and 9 favored the advanced model ($g=-0.78$, 95% CI [$-1.93,0.36$]). Only AI ($g=-1.08$, $k=48$) and physical ($g=-0.93$, $k=19$) models had intervals excluding zero. Egger and Begg tests indicated funnel-plot asymmetry, but the Pustejovsky–Rodgers test, robust to the dependence of the standard error on the effect, did not. Proposed models outperform their baselines; the superiority of any architectural class is not established.

Satellite Earth observation (EO) is widely proposed as alternative data for agricultural credit risk, yet reported fusion gains are rarely tested against the protocol producing them. We specify three inexpensive instruments—matched alert volumes instead of a fixed threshold, neutralization of a modality inside a fitted model, and diagnostics of whether a fused branch is used—and apply them to ST-CRAAN, a dual-branch forecaster combining an eleven-channel monthly EO tensor with borrower financial covariates, on 1,496,721 Thai rice-cultivating borrowers with the 2022 crop year withheld. The diagnostics detect silent modality collapse: weight decay drove the recurrent input kernel from 0.0750 at initialization to 7.0 × 10⁻⁸ (about 10⁻⁶ of its initial value), and replacing the EO tensor by its population mean changed test AUC by 0.00000. An ablation traces the collapse to gradient starvation of the branch, which weight decay completes rather than causes. With the branch in use, a factorial of five model families by eight feature sets, refitted under several seeds, finds that EO lowers test AUC in every fit of the three tabular families when it joins financial and spatial covariates, and gives the two neural families no reliable gain. Added to sub-district repayment history alone, it lowers test AUC in every fit of all five families. Ranking borrowers by the realized default rate of their sub-district–year cell bounds any predictor constant within such a cell at AUC 0.67721; every family stays below it, by 0.029 to 0.031. On this panel and across this year boundary, areal EO does not improve borrower-level default forecasting for a lender holding its own records: the binding constraint is resolution, not architecture.