A Novel Forecasting Model for Solar Power Generation by a Deep Learning Framework With Data Preprocessing and Postprocessing

Abstract

Photovoltaic power has become one of the most popular forms of energy owing to the growing consideration of environmental factors; however, solar power generation has brought many challenges for power system operations. With regard to optimizing safety and reducing the costs of power system operations, an accurate and reliable solar power forecasting model would be a significant step forward. This study proposes a deep learning method to improve the performance of short-term one-hour-ahead solar power forecasting, which includes data preprocessing, feature engineering, kernel principal component analysis, a gated recurrent unit network training mode based on time-of-day classification, and postprocessing with error correction. Both historical solar power, solar irradiance, and numerical weather prediction (NWP) data, such as temperature, irradiance, rainfall, wind speed, air pressure, and humidity, were used as the input dataset in this work. As a case study, the measured power from ten PV sites in Taiwan were collected and predicted with a one-hour resolution. The normalized root mean squared error and normalized mean absolute percent error were chosen to evaluate the performance of the forecasting models. Compared with other benchmark models, including ANN, LSTM, XGBoost, and single GRU, the experimental results showed the proposed model's superior performance. Furthermore, the importance of data preprocessing and postprocessing based on error correction was demonstrated.