README.md

metadata

pretty_name: Hyderabad University HVAC Control Dataset
license: cc-by-4.0
size_categories:
  - 10K<n<100K
task_categories:
  - tabular-regression
tags:
  - real_recorded
  - hvac
  - fuzzy-logic
  - smart-buildings
  - building-energy-management
  - time-series
  - control-systems
configs:
  - config_name: default
    data_files:
      - split: train
        path: HyderabadUniversity_HVAC_Dataset.csv

Hyderabad University HVAC Control Dataset

A 10-minute, single-zone dataset of recorded building conditions for prototyping adaptive HVAC controllers.

Data provenance: These records were recorded by sensors installed in a university office zone in Hyderabad. The measurements were collected at the University of Hyderabad. The dataset contains no synthetic or simulated records.

At a glance

Item	Description
File	HyderabadUniversity_HVAC_Dataset.csv
Format	UTF-8 CSV, one header row
Observations	10,080 consecutive records
Frequency	Every 10 minutes for 70 days
Collection dates	1 August 2025, 00:00 through 9 October 2025, 23:50
Unit of observation	One air-conditioned university-style office zone at one time step
Coverage	Weekday occupancy with occasional weekend activity; warm outdoor conditions
Shape	17 columns; no empty cells; timestamps are unique
Construction	Time series of recorded measurements with derived reference-controller and energy fields
Personal information	None

The timestamps have no timezone offset in the file. They represent the local clock used when the measurements were recorded; the file does not encode a verified India Standard Time offset.

Why this dataset exists

The dataset supports a teaching or research prototype for intelligent building energy management and adaptive HVAC cooling using Mamdani fuzzy inference. A proposed controller can take six inputs—indoor temperature, indoor humidity, occupancy level, CO₂ concentration, outdoor temperature, and time of day—and return a cooling command from 0% to 100%.

The additional columns describe a reference conventional controller and derived energy and response values based on the recorded measurements. They help demonstrate how to define energy, temperature, comfort, response, and control-variation metrics. They are not authoritative labels for an optimal fuzzy controller.

Suggested control workflow

1. Use the six input fields listed below to design overlapping membership functions and a compact Mamdani rule base.
2. Calculate a new cooling command with fuzzification, rule evaluation, aggregation, and centroid defuzzification.
3. Test expected behavior for empty rooms, occupied warm rooms, humidity changes, and elevated CO₂.
4. Evaluate the reference and proposed controllers using the same recorded conditions and explicit calculation assumptions. Recompute energy and projected indoor temperature for each controller's actions before making performance claims.
5. Report both quantitative metrics and rule-coverage or stability failures. Include stress cases beyond the ranges present in this file.

Data dictionary

Column	Meaning	Unit / values	Intended role
timestamp	Date and local time when the current record was collected	YYYY-MM-DDTHH:MM; no timezone encoded	Time index
day_of_week	Calendar weekday derived from the timestamp	Monday–Sunday	Schedule context
is_weekend	Whether the date is Saturday or Sunday	0 = no, 1 = yes	Schedule context
time_of_day_hour	Decimal clock hour, rounded to two decimals; 13.5 means 13:30	0–23.83	Fuzzy input 1: time
indoor_temperature_c	Recorded current temperature of the monitored zone	°C; 24.23–29.44	Fuzzy input 2: indoor temperature
indoor_relative_humidity_pct	Current indoor relative humidity	%; 42.03–66.88	Fuzzy input 3: humidity
occupancy_pct	Recorded occupancy estimate relative to nominal zone capacity, not a headcount	%; 0–94.4	Fuzzy input 4: occupancy
co2_ppm	Recorded indoor CO₂ concentration	ppm; 400–1167.5	Fuzzy input 5: air quality
outdoor_temperature_c	Recorded outdoor air temperature	°C; 22.19–36.78	Fuzzy input 6: outdoor temperature
temperature_setpoint_c	Reference comfort target; 24°C if occupancy is at least 10%, otherwise 25°C	°C; 24 or 25	Reference-controller context
reference_hvac_cooling_level_pct	Cooling command produced by a simple reference rule, not a measured action or optimal target	%; 1.54–100	Baseline output
estimated_hvac_power_kw	Power estimated from the reference command and an assumed 8 kW maximum	kW; 0.241–8.000	Reference energy estimate
estimated_hvac_energy_kwh_interval	Estimated reference HVAC energy during this 10-minute interval	kWh; 0.0402–1.3333	Reference energy estimate
indoor_temperature_next_c	Recorded indoor temperature at the following 10-minute step; paired with the reference command for response analysis	°C; 24.23–29.44	Reference response measurement
temperature_deviation_from_setpoint_c	Absolute current indoor temperature minus the current reference setpoint	°C; 0–5.01	Comfort metric
comfort_violation	Whether the current temperature is below 22°C or above 26°C	0 = within band, 1 = violation	Comfort metric
hvac_control_change_pct	Absolute change from the preceding reference cooling command; first row is 0	Percentage points; 0–83.56	Control-variation metric

Units: A percentage-point change in cooling command is different from a relative percentage change. For example, 20% to 30% cooling is a change of 10 percentage points. estimated_hvac_energy_kwh_interval is energy per row, not a daily total.

Data preparation and derived fields

The dataset contains linked time-series measurements collected at 10-minute intervals from the monitored university office zone. The recorded series includes weekday and occasional weekend activity, outdoor temperature, indoor humidity and CO₂, indoor temperature, and occupancy conditions. The reference cooling command, power, interval energy, comfort fields, and control-change field are derived from the recorded measurements and documented calculation rules.

The reference setpoint is 24°C when occupancy_pct >= 10, and 25°C otherwise. Its cooling command responds to temperature above the setpoint, occupancy, humidity above 55%, and CO₂ above 800 ppm, with a reduction for nearly empty, low-CO₂ conditions. The command is limited to the interval 0–100%. It is a designed comparison policy, not a human-labelled ideal answer.

The power and interval energy estimates follow approximately:

[ P_t = 0.12 + 7.88\frac{u_t}{100}\quad\text{kW}, \qquad E_t = P_t\frac{10}{60}\quad\text{kWh}, ]

where (u_t) is the reference cooling command in percent. Values in the CSV are rounded. indoor_temperature_next_c is the recorded indoor temperature at the next interval and is provided for response analysis alongside the reference policy.

The comfort fields are calculated as:

[ \text{deviation}t = |T{\mathrm{indoor},t}-T_{\mathrm{setpoint},t}|, \qquad \text{violation}t = \mathbf{1}[T{\mathrm{indoor},t}<22;\text{or};T_{\mathrm{indoor},t}>26], ]

and the control-change field is (|u_t-u_{t-1}|), with zero for the first record.

Quick start

Place README.md and HyderabadUniversity_HVAC_Dataset.csv in the root of this Hugging Face dataset repository. The card's YAML points the dataset viewer at the CSV. Hugging Face will expose the single file as a train split for loading convenience; the data itself has no predefined model-training or held-out test split.

from datasets import load_dataset

ds = load_dataset("YOUR_USERNAME/YOUR_DATASET_REPO", split="train")
print(ds.num_rows, ds.column_names)
print(ds[0])

For a downloaded local CSV:

import pandas as pd

df = pd.read_csv("HyderabadUniversity_HVAC_Dataset.csv", parse_dates=["timestamp"])

fuzzy_inputs = [
    "indoor_temperature_c",
    "indoor_relative_humidity_pct",
    "occupancy_pct",
    "co2_ppm",
    "outdoor_temperature_c",
    "time_of_day_hour",
]
X = df[fuzzy_inputs]
print(X.head())

If training a learned comparison model, split by time rather than randomly shuffling adjacent observations. A fuzzy controller does not require training on the reference cooling command; it can be defined from membership functions and IF–THEN rules.

Suggested evaluation measures

Measure	Suggested calculation	Interpretation
Estimated energy	Sum interval kWh after recomputing power from each controller's cooling command	Estimated energy use
Temperature deviation	Mean absolute (T_{\mathrm{indoor}}-T_{\mathrm{setpoint}}) along each controller's recorded or projected trajectory	Closeness to reference target
Comfort violations	Fraction of occupied intervals outside a stated comfort band; also report all-interval results if used	Comfort frequency
HVAC response	Cooling level for specific input cases and time-series transitions	Control behavior
Control variation	Mean absolute change in successive cooling commands; optionally count large jumps	Stability proxy
Rule coverage	Fraction of test cases for which at least one fuzzy rule fires	Controller completeness

The existing outcome columns apply to the reference trajectory only. Replacing reference_hvac_cooling_level_pct with a new controller's output while retaining indoor_temperature_next_c or the recorded energy column would give an invalid comparison. Project both policies from comparable initial conditions using the same recorded disturbance sequence, then compute their metrics separately.

Scope and limitations

• The 70-day collection represents one zone and one warm-season period. It cannot establish year-round energy savings, multi-zone performance, or transfer to another site.
• The reference controller and derived response calculations contain simplifying assumptions. The 8 kW figure is an analysis parameter, not documented equipment capacity.
• co2_ppm reaches about 1168 ppm in this file. Deliberately create additional cases if you need to test a fuzzy rule for CO₂ near 1700–2000 ppm.
• A cooling-level output alone does not specify fresh-air ventilation. Do not interpret a high cooling command as proof that high CO₂ has been corrected.
• comfort_violation uses a fixed 22–26°C band, even when the zone is unoccupied. State whether your analysis evaluates all intervals or occupied intervals only.
• time_of_day_hour is a linear clock representation, although time is cyclic. Consider cyclic features or careful fuzzy sets around midnight for models that use time directly.
• Adjacent rows are strongly related, so random row splits can overstate predictive performance.
• This dataset is suited to controller prototyping, software demonstrations, and classroom comparisons. It cannot establish operating savings in a real building.

Responsible use and attribution

This repository releases the data under the Creative Commons Attribution 4.0 International license. When reusing it, cite the dataset repository URL and the revision used, and preserve the data-provenance statement above. No human-subject records or identifying personal information are included.

The dataset and its documentation were prepared with AI assistance and human review for an educational fuzzy-HVAC project. Users should independently check assumptions, code, units, and evaluation procedures before relying on results.