A Multimodal Fusion Framework for Solar Forecasting using Dynamic GNNs and Mixture of Experts

• Sarvarbek Erniyazov,
• Mohammed Abdul Jaleel,
• Chang Gyoon Lim,
• Seung Bum Ha,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.012, Download Paper (Free):

• PV forecasting
• deep multimodal fusion
• sky imagery
• Satellite Images
• graph neural networks (GNN)
• mixture of experts (MoE)

Abstract

Accurate solar PV forecasting across minutes-to-day horizons is difficult due to multi-scale cloud dynamics, non-stationary irradiance, and heterogeneous data sources. We introduce a multimodal deep-learning framework that fuses sky imagery, satellite patches, ground irradiance/meteorological time series, and mesoscale model fields from nearby grid nodes. The model (i) predicts clear-sky-index residuals to reduce diurnal/seasonal non-stationarity, (ii) performs an adaptive fusion that dynamically readjusts the importance of each data source (e.g., sky images, satellite) on a per-sample basis at inference time, (iii) uses a spatial GNN with meteorology-driven dynamic adjacency to share context across stations, and (iv) employs a mixture-of-experts head with load-balanced, differentiable routing to jointly produce all horizons. Evaluated on a comprehensive multi-year dataset, our framework consistently outperforms a suite of state-of-the-art deep learning baselines, including various Transformer architectures (PatchTST, Crossformer, Informer) and Temporal Convolutional Networks (TCN). For challenging day-ahead forecasting, our model reduces the average Root Mean Square Error by nearly 15% compared to the next-best Transformer model, achieving a forecast skill of over 49% relative to the NAM physical model, especially under ‘rampy’ (high variability) weather conditions. This advantage is maintained across shorter time scales, demonstrating a 9.5% error reduction on intra-day horizons and a clear performance lead for intra-hour forecasts. These results highlight the framework's ability to leverage complementary data sources through its novel fusion and routing mechanisms, establishing a new state-of-the-art for robust and accurate forecasts across diverse weather conditions and time horizons.

Statistics

Cite this article