Construction and optimization of a time-attention-enhanced intelligent prediction model for correlation analysis of solar thermal storage efficiency and cost

Abstract The coupled evolution of thermal efficiency, operating cost, and operating state in concentrating solar thermal storage is difficult to represent with conventional models because thermal accumulation, delayed cost transmission, and state switching occur on different time scales. This study develops a time-attention-enhanced intelligent prediction model (TAE-ECM). Multi-scale temporal encoding extracts meteorological, thermal, charging/discharging, storage-state, and cost features; a temporal-attention mechanism identifies influential historical intervals; and a shared dual-task head jointly predicts thermal-storage efficiency and unit effective heat-release cost. A rolling residual-feedback scheme provides incremental correction under distribution shift, while bounded outputs and an energy-balance feasibility check preserve physical plausibility. On the 2023–2024 operating dataset, TAE-ECM achieved efficiency RMSE and MAPE values of 0.0189 and 1.71%, respectively, and cost RMSE and MAPE values of 2.37 CNY/GJ and 3.31%, with a composite score of 0.927. Incremental correction reduced efficiency and cost RMSE by 28.22% and 24.11%, respectively. The results indicate improved prediction accuracy and stability across multiple operating conditions, although switching periods remain the most challenging regime.