In this study, the problem of short-term prediction of wave energy is approached from an ordinal perspective. For this purpose, we propose the use of a soft labeling approach, which replaces the 0/1 encoding of the classes with soft labels. Specifically, such soft labels or probabilities are obtained from triangular probability distributions, which better distribute the probabilities: the target class receives higher probability than its adjacents classes. Therefore, integrating the soft labeling approach into the loss function modifies the computation of the error during model optimization, now taking into account the ordinal information encoded in the soft labels. For this purpose, an ordinal classification artificial neural network model, termed RNA-T, is implemented and optimized using a categorical cross-entropy loss function that integrates the proposed soft labeling approach. The performance of the RNA-T model is analyzed using two datasets built from reanalysis data and measurements recorded by marine buoys. The RNA-T model is compared, in terms of two ordinal performance metrics, with two standard ordinal classification techniques. The results confirm the superiority of the RNA-T model over the compared techniques.