Enhancing the reliability of solar PV power modelling through HBA-optimized variational autoencoder and GRU neural networks
TL;DR
This paper introduces a variational autoencoder technique to construct an SPV power imputation model, aiming to fill in gaps in unobserved data, and the effectiveness of the proposed imputation and forecasting models is demonstrated with real-world SPV power data for imputation and forecasting tasks.
Abstract
Solar photovoltaic (SPV) technology is gaining prominence due to its sustainability and minimal carbon footprint. However, the intermittent nature of SPV systems poses a challenge to widespread adoption. Accurate forecasting of SPV power is essential, relying heavily on high-quality data. Meteorological parameters from the National Renewable Energy Laboratory (NREL) often contain missing values, impacting the reliability of power system planning. These missing values can stem from various issues such as sensor malfunctions, communication errors, or database issues. Therefore, this paper introduces a variational autoencoder technique to construct an SPV power imputation model, aiming to fill in gaps in unobserved data. Moreover, the filled SPV power data is used to train a gated recurrent unit neural network for forecasting the SPV power. The hyperparameters of the imputation and forecasting models are optimized using the hummingbird optimization algorithm. Furthermore, the NovoGrad optimizes the hyperparameters of the imputation and forecasting models during training, thereby enhancing their performance and convergence. Furthermore, the proposed imputation and forecasting models are tested for various missing patterns. The effectiveness of the proposed imputation and forecasting models is demonstrated with real-world SPV power data for imputation and forecasting tasks.