Continual Learning Strategies for Non-Stationary Data Streams in Smart Cities

Abstract

Smart cities run on sensor streams -- traffic loops, air-quality stations, smart meters, transit and camera networks -- whose
underlying statistics are rarely stable for long, and models trained on yesterday's regime routinely degrade as conditions drift.
Periodic full retraining is the conventional answer, but it is computationally expensive, batch-oriented, and, for neural models
updated naively on recent data alone, prone to catastrophic forgetting of earlier-learned regimes. This article synthesizes the
continual-learning literature relevant to this problem, reading two largely separate traditions together: the classical
data-stream-mining work on concept drift and single-pass streaming classifiers, and the deep-learning continual-learning
literature's regularization-based, rehearsal-based, and architecture-based strategies for retaining old knowledge while
acquiring new. We trace the core catastrophic-forgetting phenomenon and the taxonomy of methods developed to address it,
then turn to a small but growing body of work applying this toolkit directly to smart-city forecasting, including verified
double-digit forecast-error reductions reported for continual-learning-regularized traffic and urban spatio-temporal
forecasters. We close by identifying where the evidence base remains thin: benchmark realism relative to genuinely streaming
urban data, the limited integration of explicit drift detection with forgetting mitigation, memory and privacy tradeoffs among
method families at city scale, and the capacity ceilings architectural approaches face under an effectively unbounded
adaptation horizon.