Renewable energy sources for electrical power generation are increasingly in demand as the world shifts toward sustainable solutions. Among these, solar energy has become a prominent option, with solar irradiance data being widely utilized to estimate potential energy generation. However, the ongoing impacts of climate change have introduced significant challenges in accurately measuring the availability of solar resources. To address this issue, forecasting models can play a critical role in predicting the feasibility and efficiency of solar energy generation. This paper presents a comparative approach involving two distinct neural network architectures designed for forecasting solar irradiance, which serves as a key input for photovoltaic energy systems. Specifically, the study examines the performance of long short-term memory (LSTM) networks and Transformer models to determine which approach delivers superior results under the given conditions. A first scenario proposed to forecast the irradiance with information from three previous weeks, by employing windows of 24, 48, 72, 84, 96, 120, and 144 h. The second scenario used a cross-validation technique. The analysis revealed that using data from a three-day period combined with a transformer model equipped with eight attention heads produced the most accurate and reliable forecasting outcomes. This finding underscores the potential of advanced neural network architectures in optimizing solar energy forecasting and enhancing renewable energy applications.