This work explores the effectiveness of explainable artificial intelligence in mapping solar photovoltaic power outputs based on weather data, focusing on short-term mappings. We analyzed the impact values provided by the Shapley additive explanation method when applied to two algorithms designed for tabular data—XGBoost and TabNet—and conducted a comprehensive evaluation of the overall model and across seasons. Our findings revealed that the impact of selected features remained relatively consistent throughout the year, underscoring their uniformity across seasons. Additionally, we propose a feature selection methodology utilizing the explanation values to produce more efficient models, by reducing data requirements while maintaining performance within a threshold of the original model. The effectiveness of the proposed methodology was demonstrated through its application to a residential dataset in Madeira, Portugal, augmented with weather data sourced from SolCast.

Keywords: explainable artificial intelligence; feature selection; machine learning; photovoltaic seasonality

Forecasting algorithms for photovoltaic (PV) power generation play an important role in energy management systems. Nevertheless, the precision of machine learning models is significantly compromised when historical data is limited. This situation is challenging for new plants for which a long history of measurements is not yet available. The unpredictable nature of the weather gives the perception that a competitive forecast requires a substantial amount of data and a very complicated algorithm. However, in this manuscript, it was found that using five historical days for the inverse quantification of uncertainty, can implicitly describe complex non-linear relationships between last five-day records and day-ahead PV power generation. The proposed approach learns the emerging patterns across various seasons throughout the year without relying on exogenous data such as air temperature, wind speed, pressure, cloud cover, and relative humidity. Results using real-world data collected at the microgrid of the University of Campinas (UNICAMP) confirm that our proposed model outperforms previous state-of-the-art deep learning models as Long short-term memory (LSTM), Gated Recurrent Unit (GRU) and traditional Autoregressive Integrated Moving Average (ARIMA) statistical model, using limited data. The proposed approach is flexible and can be easily adapted to other PV power generation systems with limited data. The source code is available at https://github.com/byronacunia/Day-Ahead-Photovoltaic-Power-Forecasting-with-Limited-Data.git