An interpretable deep learning model for solar power generation forecasting in a grid-connected hybrid solar system

Abstract

Solar energy adoption is rapidly growing as a sustainable option, with solar panels used on residential buildings, commercial properties, and large-scale farms. However, the unpredictable nature of solar power can lead to suboptimal energy generation from photovoltaic (PV) panels. Despite the high effectiveness of deep learning (DL) models in forecasting PV power, they often struggle with the perception of being “closed boxes” that lack clear explanations for their prediction results, which fail to highlight the key features for PV prediction. To address the critical issue of full transparency, this study explores a well-known DL model named lightweight deep neural network (LWDNN) in PV power forecasting, along with the application of explainable artificial intelligence (XAI) tools like Shapley Additive exPlanations (SHAP) and local interpretable model-agnostic explanations (LIME). Real-time data collected from a grid-connected solar PV system located in Dhaka were utilized to perform the prediction. By enabling XAI model interpretation, we identified feature contributions and explained individual predictions, reducing training computational demands without compromising accuracy. The reliability of the LWDNN model is assessed using both complete and reduced feature sets through performance metrics such as root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). The test results show that the proposed LWDNN model based on SHAP analysis outperforms conventional schemes by achieving RMSE = 6.180 kW, MAE = 1.939 kW, and R2 = 0.988. Finally, the model was implemented on a Raspberry Pi for low-power solar forecasting, demonstrating the feasibility of edge deployment.