Minimalist Deep Learning for Solar Power Forecasting: Transformer-Based Prediction Using Key Meteorological Features

Solar power forecasting is important for energy management and grid stability, yet many deep learning studies use a large set of meteorological and time-based variables because of the belief that more inputs improve model performance. In practice, a large feature set can introduce redundancy, increase computational effort, and reduce clarity in model interpretation. This study examines whether dependable forecasting can be achieved using only the most influential variables, presenting a minimal feature deep learning approach for short term prediction of solar power. The objective is to evaluate a Transformer model that uses only two key variables, solar irradiance and soil temperature at a depth of ten centimetres. These variables were identified through feature importance analysis. A real world solar power dataset was used for model development, and performance was compared with RNN, GRU, LSTM, and Transformer models that use the full set of meteorological inputs. The minimal feature Transformer reached a Mean Absolute Error of 1.1325, which is very close to the result of the multivariate Transformer that uses all available inputs. This outcome shows that essential temporal patterns in solar power generation can be captured using only the strongest predictors, supporting the usefulness of reducing the size of the input space. The findings indicate that selective feature reduction can maintain strong predictive performance while lowering complexity, improving clarity, and reducing data requirements. Future work may explore the adaptability of this minimal feature strategy across different regions and environmental conditions.