Accurate solar irradiance forecasting is essential for the secure integration of photovoltaic energy into power grids, especially in tropical regions with high intermittency [1]. While deep learning models like Long Short-Term Memory (LSTM) networks have been applied to this task, the potential of enhanced architectures, specifically Bidirectional LSTM (BiLSTM) and models incorporating attention mechanisms, for univariate hourly forecasting in tropical climates remains a significant research gap. This study addresses this gap by investigating whether these advanced AI architectures improve prediction accuracy in the context of Northeastern Brazil. We developed and rigorously compared four deep recurrent neural network models: LSTM, LSTM with attention, BiLSTM, and BiLSTM with attention. The models were trained on hourly data (2022–2024, 7:00–22:00 UTC) from João Pessoa, PB, preprocessed with Min-Max normalization [0.1, 0.9], and evaluated using standard metrics (RMSE, MAE, and R²) [2]. Performance was assessed robustly through 10 executions per model with 12- and 24-hour lookback windows. Contrary to expectations from recent AI trends, the results demonstrate that while BiLSTM architectures consistently outperformed standard LSTM, the integration of attention mechanisms degraded forecast performance across all configurations. This key finding suggests that for univariate solar irradiance series in tropical Brazil, the added complexity of attention gates may not capture beneficial additional temporal dependencies. This study provides empirical evidence for optimal model selection (BiLSTM) in regional grid integration applications and challenges the assumed universal utility of attention mechanisms in time-series forecasting for renewable energy.

References:

[1] Ministério de Minas e Energia. Plano Nacional de Energia 2050. Empresa de Pesquisa Energética: Brasília, Brazil, 2020.
[2] de O. Santos, D.S., Jr.; et al. Solar Irradiance Forecasting Using Dynamic Ensemble Selection. Appl. Sci. 2022, 12 (7), 3510.