Empirical Evaluation of Deep Learning Models for Long-Term Solar Power Generation Forecasting

Solar power generation (SPG) forecasting is imperative to optimize solar power integration into the electrical grid, improve energy management, and support renewable energy planning and operations. However, the variability and uncertainties associated with SPG, along with the influence of various factors on its generation, present significant challenges for accurate forecasting. Developing models that can generalize across different environmental situations and provide long-term predictions is especially difficult. This paper conducts an empirical study to evaluate the effectiveness of state-of-the-art long-term prediction models for SPG forecasting under various environmental conditions. To facilitate this evaluation, we introduce a new dataset, SPG-SL, collected from an SPG site in Piliyandala, Sri Lanka. Our empirical study utilizes the SPG-SL dataset in conjunction with another dataset from the National Solar Radiation Database in the USA. We examine existing deep learning models, including PatchTST, DLinear, Linear, NLinear, Informer, and Autoformer. Our analysis reveals that models like PatchTST exhibit promising performance in SPG forecasting in localized settings, but their effectiveness across varied geographic and temporal scopes is inconsistent. This study highlights the limitations of utilizing state-of-the-art forecasting models for SPG, particularly regarding long-term accuracy and robustness in diverse environmental conditions. Our findings emphasize the need for novel methods in SPG forecasting that enhance predictive accuracy and adapt to the complex challenges posed by the dynamic nature of solar power. This will enable developing more reliable and efficient solar power management systems.