Context-Aware Synthetic IoT Dataset for Smart Waste Management in Dhaka, Bangladesh: 43.8 Million Hourly Observations
Description
This dataset provides a one-year, context-aware synthetic IoT benchmark for smart waste-management research in Dhaka, Bangladesh. It represents 5,000 persistent synthetic smart-bin digital twins monitored hourly from 1 January to 31 December 2025, yielding exactly 43,800,000 hourly observations. The simulation incorporates land-use-specific activity, synthetic catchment population, population density, institutional operating status, market and recreation patterns, source-specific waste composition, informal material recovery, seasonal environmental behaviour, cultural-event scenarios, collection accessibility, overflow dynamics, and IoT sensor behaviour. The simulated urban contexts include residential and high-density settlements, fish markets, vegetable markets, general markets, commercial areas, restaurant and food zones, educational institutions, offices, garment factories, parks, playgrounds, recreation areas, and informal settlements. Waste composition is represented using source-dependent components rather than a single generic waste category. Waste mass and volume are modeled separately, enabling composition-dependent bin-filling and overflow behaviour. Contextual temporal patterns include institutional closure effects, market activity periods, recreation-area weekend behaviour, Friday-related activity, Ramadan and Eid scenarios, and annual weather seasonality. IoT-related variables include true and sensor-observed fill states, waste-generation and composition variables, collection and accessibility variables, overflow-related variables, battery telemetry, controlled sensor noise, packet-loss indicators, and anomaly indicators. The dataset is entirely synthetic and scenario-based. It is intended as a reusable urban digital-twin benchmark and should not be interpreted as field-observed municipal smart-bin measurements from Dhaka. The public release is distributed in Parquet format. The deposited archive contains the partitioned representation of the complete 43.8-million-observation dataset, and all partitions together constitute one logical dataset.
Files
Steps to reproduce
1. Simulation Environment & Framework: The dataset was generated using a custom-built Digital Twin Simulation Engine developed in Python 3.x. The primary libraries used for data synthesis and structural organization were Pandas for dataframe manipulation and NumPy for stochastic modeling and noise injection. The dataset was produced using the deterministic Python simulation pipeline included in the accompanying documentation package. The framework creates 5,000 persistent synthetic smart-bin entities and simulates hourly observations for the complete 2025 calendar year. The principal generation sequence is: land-use subtype → catchment population and density → activity status → temporal and event effects → waste composition → informal material recovery → residual waste mass and volume → bin fill → collection accessibility → overflow dynamics → IoT sensor observation. To reproduce or inspect the generation pipeline, extract the documentation archive, install the dependencies listed in requirements.txt, and review/run generate_v4_0.py. Validation utilities and the recorded validation report are also provided in the documentation package. The released observation data are stored as partitioned Parquet files. All partitions should be treated as components of a single logical dataset and may be read using PyArrow Dataset, DuckDB, Polars, pandas, Apache Spark, or another Parquet-compatible framework.3. Temporal and Seasonal Modeling: The simulation ran for 84 consecutive days at an hourly resolution. To ensure contextual realism, three specific temporal layers were applied: Weekly Cycle: Waste generation rates were multiplied by a factor of 8-10x for bins near mosques during Friday prayer hours (12:00–14:00). Annual Seasonality: A simulated festival window (Eid-ul-Adha) was modeled with a 12x multiplier for residential areas to mimic ritualistic waste spikes. Weather Effects: Hourly temperature and humidity were generated based on Dhaka’s tropical climate, directly influencing the battery_level discharge rate. 4. Stochastic Generation & Noise Injection: Waste accumulation followed a base growth rate influenced by the area_type. To simulate real-world IoT sensor behavior: Gaussian Noise: A normal distribution ($\sigma = 1.5$) was added to the fill_level. Outlier Generation: A 0.1% probability was set for sensor malfunctions (e.g., stuck values or impossible readings like -10% or 150%). Packet Loss: A 0.5% random data dropout was simulated to mimic network latency. 5. Export & Formatting: The final 10-million-row dataset was exported in both CSV (for universal access) and Apache Parquet (columnar format) to optimize for Big Data processing and Machine Learning workflows.