Motivated by the results in [ 1] on the self-similarity of the downlink interference in heterogenous service DS-CDMA' networks, in this paper, we propose a model-based linear adaptive-predictive method to estimate the level of interference for optimizing the system throughput and minimizing the delay for non-real-time data transmission. We use a fractional Gaussian noise (fGn) model in an appropriate time-scale to represent the self-similarity in the downlink interference. The estimated interference is utilized to allocate the available power to non-real-time services. In doing so, we use a utility-based optimization scheme and dynamic programming for time-domain optimal scheduling of non-real-time traffic. Simulation results validate the fGn model and show a substantial improvement in the delay fairness and a significant increase in the average cell throughput using our proposed scheme; and confirm that the interference model is valid for a broad range of arrival rates of non-real time traffic.