--- type: Dataset title: PreZero glass fill levels description: 11,701 anonymized fill-level readings for 449 glass recycling containers in Heilbronn, covering almost one year across white, green, and brown glass. resource: ../../PreZero/glass_fill_levels_anonymized_f.csv tags: [dataset, prezero, waste-management, fill-level, time-series, glass-recycling] status: stable generated: { by: opencode/deepseek-v4.1-flash, at: 2026-11-13T17:13:56Z } sources: - id: fill-levels resource: ../../PreZero/glass_fill_levels_anonymized_f.csv title: "glass_fill_levels_anonymized_f.csv — glass container fill levels" - id: prezero-readme resource: ../../PreZero/20261113_Readme%20Challenge%20PZ.docx title: "WasteRouting.AI — PreZero challenge README" author: team:prezero --- # Overview This dataset supports the [PreZero WasteRouting.AI challenge](/challenges/prezero-waste-routing.md). It contains fill-level readings for glass recycling containers, intended as the starting point for pattern analysis and fill-level-driven route optimization.[^prezero-readme] # Schema | Column | Type | Description | |---|---|---| | `Location_id` | string (UUID) | Anonymized, stable container identifier. No coordinates are included. | | `material_type` | enum | Glass colour: `white`, `green`, or `brown`. | | `date` | ISO 8601 date (`YYYY-MM-DD`) | Date of the reading. | | `fill_level_perc` | integer or string | Fill level in percent, usually 0–100. The value `overfull` appears as a separate state. | # Coverage | Metric | Value | |---|---| | Readings | 11,701 | | Distinct containers | 449 | | Date range | 2025-10-17 to 2026-09-30 | | White glass readings | 4,432 | | Green glass readings | 4,169 | | Brown glass readings | 3,100 | | `overfull` readings | 1,192 | | Numeric fill-level range | 0 to 100 | # Data quality notes - **No coordinates.** `Location_id` is anonymized and there is no address, latitude, or longitude column. Per-container history and grouping work, but route geometry, inter-point distances, and depot distances require an explicit assumption or external geodata. - **No container specifications.** The file does not include container type, capacity, or maximum volume. These are variables the challenge asks teams to model. - **No fleet or depot data.** Vehicle capacities, weight limits, shift constraints, and depot locations are not part of the dataset. - **Irregular sampling.** Readings are event-like, not a complete daily grid. Each container has a different number and spacing of observations. - **Discrete fill levels.** Values cluster on round percentages (for example 20, 30, 40, 60, 80, 90, 100), suggesting reported or estimated levels rather than continuous sensor measurements. - **`overfull` is not numeric.** Encode it explicitly, for example as `100+` or a separate ordinal category, rather than dropping those rows or coercing them to `NaN` silently. - **Anonymized IDs.** IDs are stable within the file, so joins across rows are safe. Do not assume they map to public container IDs. # Example: load and normalize ```python import pandas as pd df = pd.read_csv("PreZero/glass_fill_levels_anonymized_f.csv", encoding="utf-8-sig") df["date"] = pd.to_datetime(df["date"]) df["is_overfull"] = df["fill_level_perc"].eq("overfull") df["fill_numeric"] = pd.to_numeric(df["fill_level_perc"], errors="coerce") print(df.groupby("material_type")["fill_numeric"].describe()) ``` # Questions the data can answer - Which containers share similar fill cycles, and can they be grouped into service patterns? - How do school vacations, public holidays, and seasons shift fill rates? - Which containers are consistently full or overfull, and which are rarely used? - How irregular is the reading cadence, and how does that limit prediction? # Sources The README describes the challenge context and intended use.[^prezero-readme] All counts and ranges above were computed from the delivered CSV. [^prezero-readme]: PreZero, "The Challenge — WasteRouting.AI", challenge README, 2026.