Unsupervised domain adaptation methods for photovoltaic power forecasting

Abstract

The accurate forecasting of photovoltaic (PV) power generation is of great significance in renewable energy systems, as it enables optimal energy management and grid stability. Despite the importance of this issue, substantial limitations still exist in the majority of existing research initiatives, which employ shallow machine learning algorithms. Recently, some studies have proposed employing convolutional and long short-term memory neural networks (LSTMs) in conjunction with transfer learning techniques; however, these approaches require that the production of PV systems is known during training. To overcome these limitations, we present the first study in the task of PV power forecasting utilizing unsupervised domain adaptation methods. Specifically, we employ two unsupervised methods, namely Domain Adversarial Neural Network and Margin Disparity Discrepancy. Both approaches use a source and a target domain during training, where the target labels of the target domain are unknown during training. We use production and weather data from seven PV systems with nominal capacities ranging from 23.52 kW to 271.53 kW, located in different areas. The findings demonstrate that our proposed architectures improve root mean squared error (RMSE), normalized RMSE, and R2 scores over the smart persistence model across all the PV systems used for testing. Furthermore, our approaches improve the performance of the smart persistence model, with a forecast skill index reaching up to 45.35%. Our extensive experiments demonstrate that our introduced approaches offer valuable advantages over state-of-the-art ones, as the target variable of the target domain is unknown during training. We also demonstrate the robustness of our approaches by conducting a series of ablation experiments.