Comparative analysis of CNN–LSTM and CNN–BiLSTM hybrid deep learning models for solar radiation prediction

Engineering notes

Engineering notes will be added by the Power for Solar editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为光伏效率、钙钛矿太阳能电池、储能技术、太阳能热利用、BIPV、并网技术等高价值论文补充中文说明。

Original abstract

Solar radiation forecasting is a complex task since the radiation signal is nonlinear, intermittent and is significantly influenced by meteorological variability, which makes it vital for PV planning, renewable energy planning and stability of the smart grid. In this work, a replicable comparison between CNN-LSTM and CNN-BiLSTM models for one-step ahead solar clearness-index forecasting based on multivariate climate variables from NASA POWER dataset for Delhi, India, is presented. Under identical preprocessing, windowing, chronological splitting, and training conditions, CNN-LSTM achieved MAE = 0.0880, RMSE = 0.1100, R2 = 0.3100, EVS = 0.3154, WI = 0.6317, and APB = 1.89%, whereas CNN-BiLSTM obtained MAE = 0.1015, RMSE = 0.1224, R2 = 0.1456, EVS = 0.1998, WI = 0.5261, and APB = 5.98%. The Skill Scores shown and the negative values for direct clearness-index prediction do not imply that the persistence reference was unattainable, but rather reveal that the results are a controlled model-to-model comparison and not evidence of state-of-the-art superiority. Reconstructed all-sky irradiance produced stronger agreement with observations (MAE = 0.4353, RMSE = 0.5417, R2 = 0.7884, EVS = 0.7965, WI = 0.9299, and APB = 3.90%). The main task of CNN-LSTM is to provide a practical balance between accuracy and efficiency in this experimental context, and further testing with other locations, more powerful baselines and probabilistic forecasting techniques is needed.

Links and sources

• PDF from original source

Comments