Different-resolution solar radiation prediction with accompanying microclimate data and long-term monitoring experiment by LSTM, RF, and SVR: a case of campus area

Abstract

Solar radiation plays a pivotal role in urban microclimate, significantly, impacting building energy consumption and renewable energy utilization. Utilizing a comprehensive three- and-a- half-year dataset of high-precision urban microclimate, this study proposes a three-step approach to solar radiation prediction: 1) Data Preprocessing and Characterization: The microclimate data are meticulously preprocessed and categorized, with solar radiation values categorized into three resolutions: total daily solar radiation, hourly solar radiation, and 30-min solar radiation. 2) Feature Selection: Various climatic parameters are carefully selected for solar radiation prediction at different resolutions. 3) Machine Learning Models: Three machine learning models, namely Long Short-Term Memory (LSTM), Random Forest (RF) and Support Vector Regression (SVR), are employed for solar radiation prediction. The results show that all three models exhibit prediction accuracy with closely matched errors. Particularly noteworthy findings include SVR’s superior performance in predicting total daily solar radiation with a minimal error rate of about 6.0% and an impressive coefficient of determination of 0.9789. LSTM exhibits adaptability, achieving R2 values of 0.9906 and 0.9956 for hourly and 30-min radiation predictions, respectively. This study offers valuable insights into solar radiation and its implications for building energy research. It underscores that large datasets enable accurate solar radiation predictions, even when excluding certain parameters like daily maximum temperature and relative humidity.

Keywords

• urban microclimate
• solar radiation prediction
• long short time memory
• random forest
• supporting vector regression