We continue studying the possibility to predict the so-called long-term evolution of a depth profile in coastal zone. Long-term means at the scale of years, namely 10s years of measurements and several years of prediction. Prediction of coastal profile evolution is done by first calibrating a diffusion-type model, earlier proposed in the literature to describe this phenomenon, and then extrapolating the model's coefficients (by the Support Vector Machine) over time. A realistic prediction is then made by solving numerically a diffusion equation whose coefficients have been obtained by the extrapolated values of such coefficients, numerically obtained using the measured data. In previous works, we studied the model in the 'frequency-type' domain. Observations made over more than 20 years were used to accomplish the aforementioned calibration. We have then analysed the diffusion model directly in the time-space domain. In this paper, we consider the possibility of prediction the coastal profiles evolution for 3 years using only 10 or even 5 years measurements. Numerical experiments made by using the JARKUS dataset have shown that our model is indeed capable to predict the depth profile evolution occurring 3 years from now, within a relative error normally less than 10%.