A Multi-level Attention-Based LSTM Network for Ultra-short-term Solar Power Forecast Using Meteorological Knowledge

Abstract

AbstractDeep learning has been widely applied to modeling, prediction, control, and optimization in smart energy systems, especially in solar energy. Accurate forecasting of photovoltaic (PV) power is essential for power system dispatch. However, PV power output is greatly affected by weather conditions. Therefore integrating domain knowledge to improve power prediction accuracy has become a major challenge. To tackle this problem, this paper proposes an effective deep learning model for solar power forecasting based on temporal correlation and meteorological knowledge. The model adopts an encoder-decoder architecture with multi-level attention machines and long short-term memory units. The encoder is designed to dynamically extract the historical features of the in situ measurements of the PV plant, whereas the decoder is used to capture the temporal features of multi-source variables. The domain knowledge (e.g., clearness index, numerical weather prediction, and short-wave radiation) is integrated into the decoder for forecasting solar PV power. A case study is conducted using the dataset collected from real-world PV plants in different weather conditions. The experiment results demonstrate that the forecasting accuracy of the proposed model outperforms three baselines, and the prediction error is reduced by 9.5% on average compared to others.KeywordsSolar photovoltaic power forecastLong short-term memoryAttention mechanismDomain knowledgeData fusion