Inside a Physics-Based Synthetic Smart City Dataset: A Reproducible Baseline Analysis

Abstract. This article examines the complete 1,000-row, 64-column IoTSyn v4.4.0 Smart City & Public Spaces baseline dataset rather than relying on its preview. The run covers 250 hours at 15-minute resolution and couples meteorology, occupancy, traffic, air quality, environmental noise, thermal comfort, and an ordinal multi-hazard decision layer. Across the emitted records, mean Universal Thermal Climate Index (UTCI) was 27.76 °C, mean 15-minute A-weighted equivalent sound level was 59.45 dB(A), and mean occupancy was 34.98 persons. The overall action state was caution or higher in 98.5% of intervals, driven by thermal conditions in 56.2% of rows and noise in 43.8%. Air-quality and crowding risk remained low throughout. These results describe the behaviour of a synthetic generator under one frozen configuration; they are not measurements of an actual city and must not be interpreted as epidemiological, regulatory, or causal evidence.

Why analyse this baseline?

Synthetic urban datasets are most useful when researchers can distinguish the simulated physical state, the simulated sensor observation, and the decision labels derived from those variables. This dataset preserves that separation. It therefore supports questions that are difficult to study with a single observational feed: How much sensor noise obscures a slowly varying physical signal? Which hazard controls a conservative multi-hazard alert? How should temporal validation be designed when labels are derived from the same row?

The analysis addresses three questions:

1. Which environmental components drive the operational risk states in this run?
2. Does the dataset expose enough provenance and internal validation to support reproducible methodological work?
3. Which research tasks are defensible, and which claims would exceed what synthetic data can establish?

Dataset, provenance, and analytical method

The dataset was generated by IoTSyn v4.4.0 with master seed 123456. It contains 1,000 intervals of 900 seconds, beginning at 2026-07-14 00:00 in the Africa/Algiers time zone. The normalized scenario identifies coordinates 36.75° N, 3.06° E, a 500-person reporting capacity, a 24-hour warm-up, UTCI as the primary thermal index, 24-hour air-quality decision windows, 15-minute LAeq noise windows, and conservative-maximum risk aggregation. The decision mode is explicitly ex_post.

No rows were removed, imputed, or resampled. Descriptive statistics were calculated over all 1,000 rows. Pearson correlations were used only to describe relationships encoded in this run. Sensor error was evaluated against the corresponding synthetic physical-truth column using bias, mean absolute error, and root mean square error. The downloaded CSV produced the SHA-256 digest 8aba313b54710366d734abfa8378aebb5d99b5ba826c877469c6e293a4b52e58, exactly matching the published manifest. This confirms the integrity of the analysed artifact; it is not, by itself, evidence of independent cross-environment regeneration.

Provenance note. The human-readable catalog label includes qualitative control labels, whereas the normalized scenario records temp_bias_c = 0 and all exposed multipliers as 1. Because the catalog string is marked non-normative, this analysis treats the normalized scenario and emitted rows as authoritative. Researchers should preserve both the scenario and manifest when reporting the run.

What the 1,000 rows show

The period contains 76.8% weekday intervals and 23.2% weekend intervals. Crowd states were normal in 45.7% of rows, quiet in 27.0%, busy in 22.4%, and event in 4.9%. A total of 15,345 arrivals were offered and all were admitted. No rejection, deferral, timeout, or queue was recorded. Maximum occupancy reached 157 persons, or 31.4% of the stated 500-person capacity, so this run does not test capacity saturation.

Thermal conditions drive the highest-risk periods

UTCI combines air temperature, radiant temperature, wind, and humidity into a thermal-stress index. The operational procedure and stress categories are described by Bröde and colleagues in the foundational UTCI publication. In this run, 45.1% of intervals were classified as no thermal stress, 23.7% as moderate heat stress, 29.0% as strong heat stress, and 2.2% as very strong heat stress. There were no cold-stress intervals.

Air temperature and UTCI were strongly correlated (r = 0.958), as were mean radiant temperature and UTCI (r = 0.939). These coefficients are expected within the model and should not be presented as independent empirical findings. They are useful as internal signals that the coupled thermal fields move coherently.

The thermal risk component was low in 45.1% of rows, caution in 23.7%, high risk in 29.0%, and severe in 2.2%. All 22 severe overall-action intervals were thermally dominated. The validation report also flags 22 rows in which at least one UTCI input was handled by the engine's clamp-and-flag domain policy. Importantly, the clamped rows are not identical to the 22 very-strong-heat rows; input-domain warnings and output stress categories are different concepts and should be analysed separately.

Noise creates a persistent caution floor

The dataset represents environmental noise as 15-minute LAeq, with contributions combined in the energy domain. LAeq is an equivalent continuous sound level, and its correct interpretation depends on the declared averaging window. ISO 1996-1:2016 defines core quantities and assessment procedures for environmental noise but does not itself prescribe universal noise limits.

Noise risk was caution in 971 intervals (97.1%), low in 16 (1.6%), and high risk in 13 (1.3%). This near-continuous caution state explains why the overall action layer rarely falls to low even when air quality, crowding, and thermal conditions are low. Traffic flow had a strong positive relationship with LAeq (r = 0.750), while occupancy had a smaller positive relationship (r = 0.270). Because both dependencies are generated by the model, they demonstrate internal coupling rather than real-world effect sizes.

Air quality stays low-risk, but the sensor layer remains useful

The physical PM2.5 signal varied only from 11.676 to 12.072 µg/m³, and the physical NO2 signal from 19.469 to 20.207 µg/m³. The dataset's own operational air-quality profile labelled every interval low risk. These internal bands must not be substituted for jurisdiction-specific regulatory or health guidance.

The narrow physical range contrasts with a visibly noisier raw sensor layer. For PM2.5, the raw sensor error had a mean absolute error of 0.944 µg/m³, root mean square error of 1.191 µg/m³, and bias of +0.035 µg/m³. For NO2, the corresponding values were 1.635 µg/m³, 2.031 µg/m³, and −0.072 µg/m³. Raw-to-physical correlations were only 0.023 for PM2.5 and 0.035 for NO2 because the simulated physical signal changes little relative to observation noise. Calibration is disabled in the normalized scenario, and both calibrated sensor columns are entirely null.

This makes the dataset suitable for testing denoising, uncertainty estimation, or state-reconstruction methods, provided the researcher reports that the target is synthetic physical truth. It is not a suitable standalone demonstration that a calibration method transfers to a real sensor network.

How the multi-hazard decision layer behaves

The conservative-maximum aggregation rule yields caution or higher in 985 intervals (98.5%) and high risk or severe in 315 intervals (31.5%). Thermal conditions were the dominant issue in 562 rows (56.2%), and noise in 438 (43.8%). Air quality and crowding were low in all rows and never became dominant. The compound-risk flag was true in only 10 rows (1.0%).

The episode sidecar groups the 985 caution-or-higher intervals into 10 contiguous episodes. The longest episode contains 287 intervals, or 71.75 hours. Episode detection is therefore useful for testing alert persistence and segmentation logic, but 10 episodes are too few to support strong claims about event-level generalization.

Validation: what passed and what remains untested

The published validation report recomputed data types, ranges, nullability, capacity accounting, audience totals, operational-capacity ratios, UTCI categories, risk vectors, and rolling-window completeness. It reports zero invariant violations across 1,000 rows, complete risk information in every row, no provisional decision rows, and no incomplete air-quality or noise windows after warm-up.

The report's overall state is nevertheless WARN, not an unqualified pass. There are two reasons:

• Twenty-two rows invoke UTCI input clamping. The values are retained and explicitly flagged rather than silently discarded.
• Row-level Kolmogorov–Smirnov tests are not evaluated because the fully specified reference distributions apply to latent draws and innovations that are not exposed as CSV columns.

Accordingly, the validation evidence supports schema conformance and internal consistency. It does not establish external validity against field observations, realism for every city, or statistical equivalence to a measured population.

Recommended research protocols

1. Time-series forecasting

Use blocked or rolling-origin evaluation. A random row split would allow adjacent 15-minute observations from the same generated trajectory to appear in both training and test sets, producing optimistic performance estimates. Report the 24-hour warm-up and retain timestamps and time zone.

2. Sensor reconstruction

Use pm25_physical_ug_m3 and no2_physical_ug_m3 as synthetic targets, and the corresponding sensor_raw fields as observations. Do not silently fill the calibrated fields: their null state records that calibration was disabled.

3. Risk classification

Avoid label leakage. overall_action_level is deterministically derived from same-interval component risk states. Predicting it from those component labels tests rule recovery, not forecasting. A defensible experiment predicts a future action state from earlier physical or sensor observations and evaluates performance chronologically.

4. Multi-scenario generalization

Do not train and test exclusively on this baseline. Combine it with independently generated scenarios or the companion event and multi-hazard datasets, then hold out entire seeds or scenarios. This tests whether the method generalizes beyond one trajectory.

5. Reproducibility package

Archive the CSV, normalized scenario, validation report, episode file, manifest, engine version, and analysis code. Report the artifact hash. The frozen IoTSyn v4.4.0 research-artifact record should be cited alongside the individual dataset when the broader generator artifacts are used.

Limitations

• The data are synthetic. They contain no direct observations of people, streets, instruments, or health outcomes.
• The run spans only 250 hours, approximately 10.4 days, and cannot represent seasonal or long-term urban variability.
• Relationships among traffic, emissions, occupancy, noise, and risk are partly consequences of the generator's equations and parameter choices.
• The capacity mode is report-only, occupancy never exceeds 31.4% of capacity, and no rejection or queue behaviour is observed.
• Air-quality concentrations occupy a narrow, low-risk range in this scenario; models trained here should not be assumed to work during pollution episodes.
• The validation report establishes internal consistency, not empirical fidelity. Twenty-two UTCI rows are flagged for input clamping, and row-level distributional tests of latent quantities are not available in the CSV.
• The qualitative catalog control labels and normalized control values should be reported separately to avoid overstating the scenario configuration.

Conclusion

This dataset is a strong methodological baseline because it makes provenance, physical truth, sensor observations, derived decisions, and validation artifacts available together. Its main analytical lesson is that a baseline trajectory can still be operationally active: persistent noise produces a caution floor, while thermal stress accounts for the highest-risk intervals. The dataset is well suited to reproducibility exercises, temporal validation, synthetic sensor-error studies, and decision-rule auditing. It is not a substitute for field validation, causal inference, regulatory assessment, or claims about actual urban exposure.

Data access, licence, and citation

The individual dataset is available from the IoTDataset catalog record and the IoTSyn source record. It is released under CC0 1.0 Universal. CC0 permits reuse without requesting permission, but scientific citation remains important for provenance, reproducibility, and responsible interpretation.

Suggested dataset citation

IoTSyn Engine v4.4.0. (2026). Smart City & Public Spaces Dataset [iotsyn_6a68ecfebc1429.95638466] [Data set]. IoTSyn Dataset Generator. https://iotsyn.com/view.php?uid=iotsyn_6a68ecfebc1429.95638466

Method references

1. Bröde, P., Fiala, D., Błażejczyk, K., et al. (2012). Deriving the operational procedure for the Universal Thermal Climate Index (UTCI). International Journal of Biometeorology, 56, 481–494. https://doi.org/10.1007/s00484-011-0454-1
2. Oke, T. R. (1982). The energetic basis of the urban heat island. Quarterly Journal of the Royal Meteorological Society, 108, 1–24. https://doi.org/10.1002/qj.49710845502
3. International Organization for Standardization. (2016). ISO 1996-1:2016: Acoustics — Description, measurement and assessment of environmental noise — Part 1: Basic quantities and assessment procedures. https://www.iso.org/standard/59765.html
4. Duhair, A. (2026). IoTSyn v4.4.0 public research artifacts: Scenarios, verified outputs, contracts and external-comparison data [Data set]. Zenodo. https://doi.org/10.5281/zenodo.21480873