Published at Solar Energy – A clustering-based contrastive ensemble LSTM model for receiver thermal power prediction in CSP plants

Abstract:
Accurate prediction of receiver thermal power in concentrated solar power (CSP) plants is crucial for maintaining efficient coordination between power plants and the grid, and for facilitating optimal energy storage and generation planning. However, existing solar prediction models often treat weather conditions as discrete categorical variables, assigning fixed weather types to each time period, which inherently restricts further improvements in prediction accuracy. To overcome this limitation, this study proposes an ensemble long short-term memory (EnLSTM) model incorporating a clustering-based contrastive autoencoder (CCAE). The CCAE module utilizes clustering loss to ensure latent space encodings follow a Gaussian mixture model distribution and contrastive loss to enable latent variables to capture continuous weather patterns, thereby learning interpretable weather representations that determine the weighting scheme for the EnLSTM model. Experimental validation using data from a Chinese CSP plant demonstrates the superior performance of our approach through comprehensive ablation studies and comparative analyses with benchmark models. The proposed algorithm achieves a mean absolute error of 0.0977, representing significant improvements over conventional BiLSTM (0.1113) and TCN (0.1616) models. This work establishes a novel methodological foundation for accurate thermal energy prediction in CSP plants by transitioning from discrete to continuous weather pattern representations.

Ziheng Xu, Xiaoling Mi, Wenjun Huang, Jianxiang Jin, A clustering-based contrastive ensemble LSTM model for receiver thermal power prediction in CSP plants, Solar Energy, Volume 308, 2026, 114392, ISSN 0038-092X, https://doi.org/10.1016/j.solener.2026.114392