Accurately predicting solar irradiation(SI) is crucial for power scheduling and photovoltaic site selection. With the development of high-performance computing and large-capacity storage devices

data-driven deep learning models have gained widespread attentions in the SI prediction domain. However

the lack of physical interpretability due to the “black-box” nature of deep learning models restricts their credibility in specific scenarios. To enhance the interpretability of the model on the premise of maintaining prediction accuracy and keeping the model structure unchanged

and without increasing computational complexity

a model based on long short-term memory(LSTM) neural network is constructed

demonstrating an 8.07% performance improvement over the conventional neural networks and showing superior outlier handling capabilities. By employing layer-wise relevance propagation(LRP) algorithm

factors influencing the model output are scored from both temporal and spatial dimensions

enhancing the model's interpretability. The research results indicate that the model possesses good interpretability under the premise of ensuring performance

with historical solar irradiation

time-related features(such as hour

day

week

month)

solar altitude information(such as sunrise and sunset times)

cloud cover

radiation time

temperature

and dew point temperature being the main factors influencing SI prediction.