VAPOR: A Novel Approach to Power Forecasting in a Photovoltaic Microgrid

Abstract

Localized microgrid systems require accurate forecasting of solar photovoltaic generation to ensure reliable and cost-effective operation. To reduce the burning of fossil fuel energy reserves, this research introduces VAPOR, a deep learning approach to energy generation forecasting. VAPOR implements a novel Softmax Liquid Attention Matrix (SLAM) combined with convolutional, long short-term memory, and fully connected neural network layers. SLAM, which linearly projects a multivariate one-dimensional input into a two-dimensional weighted attention matrix, allows VAPOR to develop stronger correlations between model inputs that enhance model performance and interpretability. Unlike other state-of-the-art methods, SLAM continuously adapts to data inputs even after model training due to VAPOR's segmented model structure and multivariate input. As a result, VAPOR can more accurately forecast fluctuations in energy generation. Forecast results from UC San Diego's 42 MW microgrid demonstrate how VAPOR's forecasting accuracy outperforms several other state-of-the-art methods, thereby enhancing the reliability and profitability of microgrid systems.