> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1 Abstract—5G is expected to bring about disruptive industrial- societal transformation by enabling a broad catalogue of (radically new, highly heterogeneous) applications and services. This scenario has called for Zero-touch network and Service Management (ZSM). With the recent advancements in Artificial Intelligence (AI), key ZSM capabilities such as the runtime prediction of user demands can be facilitated by data-driven and Machine Learning (ML) methods. In this respect, the paper proposes a runtime prediction approach that transforms time series forecasting into a simpler multivariate regression problem with Artificial Neural Networks (ANNs), structurally optimized with a genetic algorithm (GA) metaheuristic. Leveraging on a novel set of input features that capture seasonality and calendar effects, the proposed approach removes the prediction accuracy’s dependence on the temporal succession of input data and the forecast horizon, which is typical in time series forecasting. Evaluation results based on real telecommunications data show that the GA-optimized ANN regressor has better prediction performance compared to 1-day and 1-hour ahead forecasts obtained with state-of-the-art Multi-seasonal Time Series (MSTS) and Long Short-Term Memory (LSTM) forecasting models by an average of ~59% and ~86%, respectively. Furthermore, despite its longer training times compared to the baseline models, the proposed ANN regressor relaxes the monitoring requirements in 5G dynamic management systems by allowing less frequent re- training offline. Index Terms—Artificial Neural Networks, Genetic Algorithm, Network Dynamics, Runtime Prediction, Time Series, ZSM. I. INTRODUCTION VER the years, learning complex system dynamics has maintained significant interest in the research scene, as well as in various industrial domains, for its notable potential in autonomic management and control. At a networking perspective, the runtime prediction of user demands at various management levels (e.g., infrastructure-, service- and application-level) and coverage granularities (e.g., cell-, city- and nationwide-level) would not only facilitate the dimensioning of the network/service/application, but also enable a wide range of management and control mechanisms such as the dynamic (and proactive) provisioning of the underlying resources. Indeed, this is a fundamental component towards enabling Zero-touch network and Service Management (ZSM) [1], as well as the disruptive industrial-societal transformation brought forth by 5G [2]. User demand dynamics are usually recorded as temporal data, which have been widely investigated through time series analysis and forecasting, as well as regression analysis, among others. Moreover, it is worth noting that they naturally have layered seasonality (e.g., daily, weekly, monthly, etc.), and may exhibit the so-called calendar effect (e.g., weekends, holidays, sales, etc.), as well [3]. With the recent advancements in Artificial Intelligence (AI), data-driven and Machine Learning (ML) methods have opened new directions towards predictive analytics. Among others, bio- inspired Artificial Neural Networks (ANNs) have gained particular interests for their ability to model noisy and nonlinear systems, by learning from examples. In fact, numerous works in the literature (e.g., [4]-[7]) exploit past temporal data as input to extend ANNs’ high prediction accuracies to time series forecasting. The training times of ANN-based time series predictors are, however, substantially longer than typical regressors, and such models usually have a limited usable prediction horizon [5]. Furthermore, they depend on the temporal succession of data and, hence, call for re-training as the series are updated with new observations. With this in mind, the main contribution of this paper is a novel input feature set and framework for transforming time series forecasting into a simpler multivariate regression problem, using ANNs (that are structurally optimized with a Genetic Algorithm (GA) metaheuristic, as in [4]) in the context of network activity runtime prediction. In particular, by leveraging a set of seasonal and calendar features, the dependence of the modeling and forecasting steps on the temporal succession in the data is removed. Hence, the proposed approach is capable of predicting any future value based only on the fed inputs, as well as capture possible (ir)regularities in the dynamics brought forth by seasons and special events. ANNs Going Beyond Time Series Forecasting: An Urban Network Perspective Jane Frances Pajo, George Kousiouris, Dimosthenis Kyriazis, Roberto Bruschi, Senior Member, IEEE, and Franco Davoli, Life Senior Member, IEEE O Manuscript received *date*; revised *date*. This work was partially supported by the H2020 Innovation Actions SPIDER and 5G PPP 5G- INDUCE, funded by the European Commission under Grants 833685 and 101016941, respectively. J. F. Pajo, R. Bruschi and F. Davoli are with the Department of Electrical, Electronic and Telecommunications Engineering, and Naval Architecture (DITEN), University of Genoa, and with the National Laboratory of Smart and Secure Networks (S2N) of the Italian National Consortium for Telecommunications (CNIT), Genoa, Italy (e-mail: jane.pajo@tnt-lab.unige.it, roberto.bruschi@unige.it, franco.davoli@unige.it). G. Kousiouris is with the Department of Informatics and Telematics, Harokopio University of Athens, Greece (e-mail: gkousiou@hua.gr). D. Kyriazis is with the Department of Digital Systems, University of Piraeus, Greece (email: dimos@unipi.gr).