MUFASA: Multi-Expert Unified Forecasting Architecture with Simplex Aggregation for Multi-Step Hourly Solar Irradiation Forecasting

This paper proposes a multi-expert unified forecasting architecture with simplex aggregation (MUFASA), a development-anchored forecast-combination framework for multi-step hourly solar irradiation forecasting. A nonlinear temporal expert and a regularized ridge expert provide complementary high-capacity and low-variance inductive biases. Development-calibrated shrinkage is followed by Euclidean projection onto the probability simplex, enforcing non-negative unit-sum weights and convex-hull-bounded forecasts. Site-specific hyperparameters are selected by Gaussian process Bayesian optimization using data through 2019 only; 2020 is retained as a held-out evaluation year and does not enter parameter, seed-weight, or aggregation-weight estimation. The framework is evaluated over 11 hourly forecast steps from 08:00 to 18:00 across six major metropolitan areas in South Korea. Under the controlled conditional/oracle-weather protocol, macro RMSE was 0.3430 MJ m−2, macro MAE was 0.2444 MJ m−2, and macro R2 was 0.8740. MUFASA achieved the lowest site-level RMSE at all six sites and ranked first in 62 of 66 site–horizon comparisons against 18 trainable benchmark architectures. Dependence sensitivity remained stable across HAC lags 1–28 and moving-block lengths 3–28; strict simultaneous 11-horizon superiority was supported against 12 of 18 comparators. The findings describe controlled conditional forecasting performance rather than operational NWP-driven day-ahead accuracy.