Solar radiation forecasting using attention-based temporal convolutional network

Solar energy, an inexhaustible and pristine power source, harbors the capability to mitigate the emissions of greenhouse gases and the dependency on fossil fuels, thereby playing a pivotal role in the conservation of our ecosystem. Nevertheless, the process of harnessing solar energy from sunlight is subject to the capricious characteristics of weather conditions, which include variables such as the density of cloud cover, levels of atmospheric moisture, and fluctuations in temperature. Hence, the task of prognosticating solar radiation holds significant importance for the strategic planning and efficient management of solar power systems. The current machine-learning methods for predicting global solar radiation make use of recurrent networks. One major downside of recurrent-based models is that they are exposed to vanishing gradients and stagnant performance over longer available input sequences. The model showcased is an attention-fueled Temporal Convolutional Network (TCN) intertwined with Convolutional Neural Network (CNN). The suggested method merges the advantages of the feature extraction proficiencies of a TCN and the aggregation capabilities of a CNN. The method has been tested for up to 24 hours of future time sequence prediction and it has been noted that its performance is unmatched.