Source area partition and quality filtering can improve the dependability of eddy covariance (EC) flux data while reducing its temporal consistency. Here, we constructed an ultra-wide artificial neural network (ANN) structure based on the TensorFlow framework. For the ANN inputting feature information selection, we attempted to establish feature vectors utilizing adequate thermodynamic forcing information of micro-meteorological background. The temporal consistency of EC data was optimized by interpolating with the ANN modeled fluxes, raising the temporal coverage rates from under 0.40 to over 0.98 for the flux data at the lake surface of Yamzhog Yumco. The evaluation of flux simulation performance via 10-fold cross-validation indicates that the bias level exhibits minuscule perturbation over different subsamples, disclosing preferable robustness for the ANNs model. Comparing for the approximately 18.8 W/m²/81.5 W/m² of average value for the observed sensible/latent heat flux, 1.84 mmol/(s·m²) for water vapor flux, the mean absolute errors is 5.4 W/m² for the simulated sensible heat flux, 15.7 W/m² for the simulated latent heat flux, and 0.35 mmol/(s·m²) for the water vapor flux. The results suggest that the combination of ANN structure with variable selecting principle can utilize the micro-meteorological information of field observation more sufficiently to estimate the flux intensity. Consequently, the temporal consistency is efficiently optimized with the analysability of EC flux data enhanced. The optimization method we proposed makes the interpolation of EC flux observation data no longer depend on the calculation of specific micro-meteorological elements such as turbulence transport coefficient. The paper provides a reference idea for improving the data quality of EC flux observation experiments for alpine lakes and other harsh environments.