Adarmer: An Adaptive Transformer for Direct Normal Irradiance Forecasting

Concentrated solar power (CSP) technology offers a carbon-free approach to power generation, providing a relatively stable and uninterrupted electricity supply compared to other renewable sources. The effectiveness and success of CSP plants are heavily dependent upon accurate forecasting of the direct normal irradiance (DNI) from the sun. Various DNI forecasting models have been proposed to this end; however, they rely on meteorological factors that are difficult to acquire. A recent study addressed the problem of DNI forecasting without meteorological data for the first time, formulating the problem as a multi-class classification task, and proposed multiple deep neural networks to model it. Extending this research, we propose Adarmer - a novel transformer model that automatically adapts its neural architecture to the complexity of the DNI estimation task. Our findings demonstrate that Adarmer surpasses the performance of previously proposed models across the board.