Multivariate Variable-Based LSTM-AE Model for Solar Power Prediction

This study proposes a multivariate-based LSTM-Autoencoder (LSTM-AE) model for short-term photovoltaic power generation prediction. The LSTM-based encoder-decoder structure effectively learns multivariate relationships and time series dependencies between significant environmental and power-related variables. Input variables include DC voltage, DC current, DC power, ambient temperature, solar radiation, and environmental factors, and they are preprocessed through scaling to increase learning efficiency. The encoder compresses multivariate time series data to the latent space, and the decoder restores the corresponding sequence to learn the complex time series patterns of the data. The normalization technique was applied to the algorithm to prevent overfitting and improve the model's generalization performance. The prediction accuracy evaluation was made through mean absolute percentage error (MAPE), mean square root error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). As a result of the experiment, the proposed LSTM-AE model outperformed the existing model in capturing nonlinear and long-term dependence. The results of this study suggest that the LSTM-AE architecture can contribute to the development of renewable energy prediction fields, contributing to the development of more accurate and reliable photovoltaic prediction systems.