Time-Series Forecasting and Evaluation on City Sensing Data: Pedestrian and Microclimate Benchmarks

Urban sensing infrastructures are increasingly instrumented with dense networks of pedestrian counters and environmental sensors, yet there is no standardized, reproducible benchmark for evaluating modern time-series forecasting models on real city data. In this work, we introduce a multi-resolution forecasting benchmark built on three open datasets from the City of Melbourne: (i) Argyle Square Microclimate Sensors Data (hourly), (ii) Microclimate Sensors Data (15-minute), and (iii) Pedestrian Counting System Data (hourly). For each dataset, we construct fully specified forecasting tasks with fixed input lengths and horizons, chronological 70%/10%/20% train-validation-test splits, and a leakage-aware preprocessing pipeline that includes timestamp alignment, dataset-specific cleaning, and per-series standardization. We then evaluate three recent state-of-the-art neural architectures, TimeMixer, RAFT, and WPMixer, under a unified training and evaluation protocol, reporting mean MAE and MSE over three random seeds for each dataset-horizon configuration. Our results show that the relative ranking of these models depends strongly on the sensing regime: RAFT performs best on the single-site hourly Argyle Square Microclimate Sensors Data, whereas WPMixer outperforms both TimeMixer and RAFT on the higher-frequency Microclimate Sensors Data (15-minute) and on multi-location pedestrian counts, particularly at longer horizons, while TimeMixer remains consistently competitive despite its compact configuration. Overall, no single architecture dominates across all datasets and horizons, underscoring the importance of multi-dataset, multi-resolution evaluation in urban time-series settings. All code, cleaned datasets, and experiment scripts are released in a public repository to enable end-to-end reproducibility and to support future extensions of the benchmark, including new architectures, additional cities when comparable open data become available, hierarchical forecasting formulations, and the integration of exogenous contextual signals.