Abstract
Epidemic forecasts are unreliable when surveillance data are noisy or incomplete and when underreporting and rapidly changing population behaviour distort observed incidence, degrading the stability of conventional statistical and deep-learning models. We aim to develop an interpretable, uncertainty-aware forecasting pipeline that remains robust under data corruption and is practical for real-time use. We convert COVID-19 incidence into multilayer temporal graphs: global cumulative counts are differenced to daily incidence, normalised, and segmented into overlapping 30-day windows; for each window, we build a visibility graph from the empirical series and a matched-length visibility graph from stochastic simulations (fractional Brownian motion and Lévy-type dynamics) to represent reporting and behavioural randomness. We fuse the graphs (weighted edge averaging), extract compact descriptors (mean degree, clustering coefficient, entropy) and train a lightweight regressor to predict the 7-day-ahead average incidence. On the Johns Hopkins COVID-19 dataset, the method outperforms ARIMA, LSTM and standard GCN baselines (MAE = 0.0558; RMSE = 0.0709). Stress tests with noise and missingness and ablations show that stochastic augmentation and graph fusion materially improve robustness, while a cloud-oriented deployment reduces inference time by >60% and memory usage by 35%, enabling low-latency monitoring for timely public-health decision-making.