Optimizing Solar Energy Predictions with a Scaled Dot-Product Attention-LSTM-TCN Framework

Since the increasing use of solar power and the innovative application of machine learning in various domains, there have been significant advancements in solar power generation forecasting. Like many other instances, this development has started from some simple, single models, such as recurrent neural networks (RNN) to hybrid models by employing different types of recurrent neural networks (RNN) and convolutional neural networks (CNN). Recently, incorporating attention mechanisms into hybrid models has gained considerable interest due to their ability to enhance model performance by leveraging the strengths of these mechanisms. Thus, the aim of this study is to examine a new solar photovoltaic power forecast technique, which consists of Scaled Dot-product Attention (SDPA), Long Short-term Memory (LSTM) and Temporal Convolutional Networks (TCN). A comprehensive evaluation on four different datasets has been conducted to test the flexibility and effectiveness of the proposed model. Additionally, a systematic analysis of error metrics—Root Mean Square Error (RMSE) and Mean Absolute Error (MAE)—is performed to compare our model with other single and hybrid architectures. The outcomes present a maximum reduction of $45.31 \%$ in RMSE when comparing our model with other ones. The results generally indicate that our model outperforms most of these models across the datasets.