In high solar irradiance regions of Northwest China, fluctuations in Direct Normal Irradiance (DNI) induce transient thermal shock effects on the receiver of Solar Power Tower (SPT) systems, thereby impairing the operational safety and service lifetime of the receiver. A robust decision-making basis for the regulation of aiming strategies in SPT systems is provided by high-precision DNI data, which functions as a critical feedforward signal. In this study, a hybrid prediction model (termed AVMD-LSTM-Attention) is proposed, which integrates Adaptive Variational Mode Decomposition (AVMD), Long Short-Term Memory (LSTM) networks, and the Multi-Head Attention mechanism. Specifically, the original DNI time series is first adaptively decomposed via AVMD to mitigate the non-stationarity of the series. Subsequently, the LSTM network is employed to deeply capture the long-range temporal evolution characteristics of each decomposed component, while the Attention mechanism is incorporated to calculate the weights of hidden layer states—this enhances the model’s capacity to focus on critical information and suppresses noise interference effectively. Finally, the prediction results of all components are aggregated to yield the final DNI prediction output. This proposed method achieves high-precision DNI prediction, furnishing reliable data support for the control strategy of the heliostat field in SPT systems and enabling the collaborative optimization of solar-thermal coupling control for SPT systems.