This work concerns analysis of turbulence in the atmospheric boundary layer (ABL) shortly before and after sunset. Based on a large set of Doppler lidar measurements at rural and urban sites, we analyze frequency spectra of vertical wind at different heights and show that they increasingly deviate from Kolmogorov's -5/3 prediction in the measured low-wavenumber part of the inertial range. We find that before sunset, the integral length scales tend to decrease with time. These findings contrast with a classical model of equilibrium decay of isotropic turbulence, which predicts that the scaling exponent should remain constant and equal to -5/3 and the integral length scale should increase in time. We explain the observations using recent theories of non-equilibrium turbulence. The presence of non-equilibrium suggests that classical parametrization schemes fail to predict turbulence statistics shortly before sunset. By comparing the classical and the non-equilibrium models, we conclude that the former may underestimate the dissipation rate of turbulence kinetic energy in the initial stages of decay.