The laser wireless power transmission (LWPT) technology offers advantages such as long-range capability, high directivity and high power density, showing great application promise in the fields such as unmanned aerial vehicle (UAV) charging. However, the environmental disturbances including atmospheric scattering, absorption and turbulence induce strong nonlinearity and time-varying characteristics in a medium transmission state. The conventional mechanism models rely on environmental parameters, so they are unsuitable for online awareness. To solve this problem, a feature-attention enhancement neural network (FAENN) is proposed, which enables high-accuracy online prediction of transmission efficiency by only using easily measurable electrical signals such as laser diode current and photovoltaic current. The architecture integrates feature-wise gating with residual connections to balance the modeling capacity and training stability. Experimental results show that this model achieves an R² of 0.995 54 on the test set, significantly outperforming the linear regression, support vector regression and feedforward neural network. On a 6-m LWPT experimental platform, a mean absolute error of only 0.036 kW/m² in irradiance power density prediction was obtained. The proposed method requires no atmospheric sensors, and it has high accuracy, low latency and strong generalization capability, providing reliable state awareness support for intelligent LWPT systems.