A Bi-LSTM-GAN Based Method for Photovoltaic Data Imputation

This study addresses data loss in photovoltaic (PV) power generation systems resulting from factors such as adverse weather and sensor failures. To obtain more accurate and reliable PV data, we propose a data imputation method based on a Bidirectional Long Short-Term Memory Generative Adversarial Network (Bi-LSTM-GAN). In this model, the Generative Adversarial Network (GAN) serves as the overarching framework, while the Long Short-Term Memory (LSTM) and its bidirectional variant, the Bidirectional Long Short-Term Memory (Bi-LSTM), form the core components for learning and reconstructing missing data sequences. The key innovation of this method lies in replacing the traditional fully connected layer in the GAN with a Bi-LSTM-based architecture, which enables the model to effectively capture the latent temporal information in PV power generation data. The temporal correlation module is designed to capture the temporal dependencies and the characteristics of event series. Furthermore, by integrating the Bidirectional Gated Recurrent Unit (BiGRU) and the temporal attention mechanism, the model enhances its ability to identify the temporal relationships in the data, thereby improving the accuracy of data imputation. Experiments on real PV datasets show that the proposed method effectively recovers missing data under both random and continuous missingness, significantly enhancing the integrity and reliability of PV data. Comparative evaluations against prevalent data imputation algorithms confirm the superiority of the proposed method.