Weather Maps as Tokens: Transformers for Renewable Energy Forecasting

TL;DR

This paper addresses the challenge of renewable energy forecasting by integrating rich spatial weather context with temporal evolution. It introduces Weather Maps as Tokens, encoding hourly weather maps into 128-dimensional tokens via a lightweight CNN and processing sequences with a transformer to forecast a $45$-hour horizon. Compared with ENTSO-E operational forecasts, the approach achieves substantial RMSE reductions for wind (~63%) and solar (~21%), with improved bias stability and generalization. The method is computationally efficient, modular across regions, and supports near-term grid operation improvements for renewable integration.

Abstract

Accurate renewable energy forecasting is essential to reduce dependence on fossil fuels and enabling grid decarbonization. However, current approaches fail to effectively integrate the rich spatial context of weather patterns with their temporal evolution. This work introduces a novel approach that treats weather maps as tokens in transformer sequences to predict renewable energy. Hourly weather maps are encoded as spatial tokens using a lightweight convolutional neural network, and then processed by a transformer to capture temporal dynamics across a 45-hour forecast horizon. Despite disadvantages in input initialization, evaluation against ENTSO-E operational forecasts shows a reduction in RMSE of about 60% and 20% for wind and solar respectively. A live dashboard showing daily forecasts is available at: https://www.sardiniaforecast.ifabfoundation.it.

Figures (5)

• Figure 1: Weather Maps as Tokens architecture.
• Figure 2: MAE and RMSE progressions across the 45-hour forecast horizon for wind (left) and solar (right) power prediction.
• Figure 3: Hourly MAE for wind (left) and solar (right) over the 45-hour horizon with the proposed model (blue) and ENTSO-E (orange)
• Figure 4: Wind power forecasting across systematic seasonal examples in 2024. The CNN-Transformer model (blue circles) consistently demonstrates superior tracking of actual wind generation (purple diamonds) compared to ENTSO-E operational forecasts (orange crosses) across diverse seasonal and meteorological conditions.
• Figure 5: Solar power forecasting across systematic seasonal examples in 2024. The model accurately captures diurnal solar patterns across all seasons, from reduced winter generation (January, November) to peak summer output (May, July). The CNN-Transformer model maintains closer alignment with actual generation compared to ENTSO-E forecasts. September 15th illustrates successful modeling of variable cloud conditions affecting solar irradiance.