--- license: cc-by-nc-4.0 task_categories: - tabular-classification - tabular-regression - time-series-forecasting language: - en tags: - synthetic - oil-and-gas - midstream - downstream - inventory - tank-farm - spr - eia-padd - seasonal - xpertsystems pretty_name: "OIL-033 — Synthetic Crude & Product Inventory Dataset (Sample)" size_categories: - 100K inventory_t+1 = clip(0, capacity, inventory_t + inflow_t − outflow_t + disruption_t) > seasonal(d) = 1 + 0.15 · sin(2π · day_of_year / 365) > inflow_t = N(250000, 50000) × seasonal(d) bpd > outflow_t = N(245000, 45000) × seasonal(d) bpd > disruption_t = U(-400000, 400000) with prob 0.005 The sample's **seasonal coupling** (day_of_year ↔ inflow r ≈ -0.36, expected seasonal ↔ inflow r ≈ +0.47) validates the sinusoidal modulation. **`spr_operations.csv`** — **4 real DOE Strategic Petroleum Reserve sites**: | Site | State | Real Capacity | Notes | |---|---|---:|---| | Bryan Mound | Texas | ~245M bbl | Largest SPR site, near Freeport | | Big Hill | Texas | ~160M bbl | Beaumont area | | West Hackberry | Louisiana | ~227M bbl | Near Lake Charles | | Bayou Choctaw | Louisiana | ~76M bbl | Baton Rouge area | Sample reserve level mean ~648M bbl matches **post-2022 SPR drawdown era** (peaked ~727M in 2009; reduced to ~350M after 2022 sales; rebuilding 2024+). **`inventory_labels.csv`** — **feature-coupled ML labels**: > shortage_risk = 1 if utilization_pct < 35 else 0 > optimization_score = clip(0, 100, 100 - |utilization_pct - 72|) The sample's **deviation from 72% optimal ↔ optimization_score r = -1.000000** (deterministic V-shape coupling per generator formula) — **near-perfect feature-coupled label validation**. The shortage_risk binary classifier shows r ≈ -0.52 with utilization, validating EIA tight-inventory threshold. --- ## Suggested use cases 1. **Inventory time-series forecasting** — predict `inventory_bbl` from inflow/outflow features per mass balance accumulation. **Strong physics signal** — within-site dynamics deterministic. 2. **Binary shortage risk classification** — predict `shortage_risk` (util<35%) from inventory + region + storage_type features per EIA tight-inventory tracking methodology. **Strong physics coupling**. 3. **V-shape optimization regression** — predict `optimization_score` from `|utilization - 72|` per API 2350 / EIA mid-fill target. **Near-deterministic** — models can learn exact V-shape. 4. **Seasonal demand pattern ML** — predict seasonal inflow/outflow patterns from day_of_year features per EIA Weekly Petroleum. 5. **5-class disruption event classification** — multi-class classifier on event_type (Hurricane / Pipeline Outage / Refinery Fire / Import Disruption / Tank Failure). 6. **SPR operations forecasting** — predict SPR release events from reserve_level + global market features (extend with OIL-029 prices). 7. **Regional PADD inventory analytics** — aggregate inventory by EIA PADD region per EIA Weekly Petroleum Status methodology. 8. **5-class storage type classification** — predict storage_type from capacity + working_capacity features. 9. **Daily inflow/outflow regression** — predict inflow_bpd / outflow_bpd from seasonal + site features. 10. **Multi-table relational ML** — entity-resolution across the 6 tables via `inventory_id` + `timestamp` for joinable training pipelines. --- ## Loading ```python from datasets import load_dataset ds = load_dataset("xpertsystems/oil033-sample", data_files="crude_inventory_levels.csv") print(ds["train"][0]) ``` Or with pandas: ```python import pandas as pd master = pd.read_csv("hf://datasets/xpertsystems/oil033-sample/inventory_master.csv") crude = pd.read_csv("hf://datasets/xpertsystems/oil033-sample/crude_inventory_levels.csv") refined = pd.read_csv("hf://datasets/xpertsystems/oil033-sample/refined_product_inventory.csv") spr = pd.read_csv("hf://datasets/xpertsystems/oil033-sample/spr_operations.csv") labels = pd.read_csv("hf://datasets/xpertsystems/oil033-sample/inventory_labels.csv") # Multi-table feature engineering for ML: crude_agg = crude.groupby('inventory_id').agg( avg_inventory=('inventory_bbl', 'mean'), avg_utilization=('utilization_pct', 'mean'), net_flow_std=('inflow_bpd', lambda x: x.std() - 0) # placeholder ).reset_index() joined = (master .merge(crude_agg, on='inventory_id') .merge(labels.groupby('inventory_id').agg( avg_shortage=('shortage_risk', 'mean'), avg_opt=('optimization_score', 'mean') ).reset_index(), on='inventory_id')) ``` --- ## Reproducibility All generation is deterministic via the integer `seed` parameter (driving `np.random.seed`). A seed sweep across `[42, 7, 123, 2024, 99, 1]` confirms Grade A+ on every seed in this sample. --- ## Honest disclosure of sample-scale limitations This is a **sample** product calibrated for inventory ML research, not for live tank farm operations or EIA forecasting. Several notes: 1. **No region/storage-type conditioning on capacity.** All sites use `randint(500K, 10M) bbl` regardless of being SPR (real ~150M bbl), Cushing terminal (real ~15-20M), or floating tank (typically 500K-2M). **For type-conditioned ML, normalize by storage type scale**: ```python type_scales = {'SPR': 150e6, 'Cushing': 18e6, 'Refinery': 8e6, 'Terminal': 5e6, 'Floating': 1e6, 'Tank Farm': 6e6} ``` 2. **No region/storage-type conditioning on inflow/outflow.** All sites use ~250K bpd inflow/outflow regardless of being SPR (typically very low daily flux) or refinery (250K-500K bpd realistic). For flux- conditioned ML, **filter to single storage_type** before training. 3. **Inflow ≈ outflow nearly balanced** (250K vs 245K mean). Net flow std 67K bpd dominates the 5K mean drift, so inventory exhibits slow random walk with capacity bounds. **For mass-balance ML, focus on near-term dynamics** rather than expecting trend-following behavior. 4. **Refined products are NOT joined to crude utilization.** Gasoline, diesel, jet fuel are independently sampled from N(120K, 25K), N(95K, 20K), N(60K, 15K) per site/day without coupling to crude inflow or refinery throughput. Real refined product inventory tracks refinery utilization with ~1-2 week lag. **For product-yield ML, derive your own coupling** or use the full product. 5. **SPR site distribution is per-day random** rather than per-event. The generator samples `spr_site` independently each day, so the 4 SPR sites appear roughly uniform (24-27% each) over the 365-day period even though release events are rare. For SPR-site-specific ML, **filter to release events only** (`release_rate_bpd > 0`). 6. **SPR reserve level changes very little** (~647-650M range across 365 days) because only ~1% of days trigger releases. Real SPR inventory changes more dramatically (~120M bbl reduction in 2022). **For SPR drawdown ML, use the full product** or augment with historical 2022 release events. 7. **Disruption magnitude includes positive values** `U(-400K, +400K)`, which is physically odd (disruptions should typically reduce supply). The sample treats positive values as "anti-disruptions" (e.g., emergency receipts). **For supply-shock ML, filter to negative disruption values** or use `abs(disruption)` as severity. 8. **Capacity ↔ utilization is uncoupled** (r ≈ 0.04). Real markets show smaller tanks have more variable utilization (higher turnover cycles relative to capacity). **For capacity-conditioned ML, use normalized utilization** (e.g., daily change / capacity). 9. **Working capacity % is uniform U(0.65, 0.95)** without conditioning on storage type. Real SPR working capacity is ~95%+ (low heel), while floating roof tanks are ~80% (shell heel + sludge). For type-specific ML, **derive type-conditioned working capacity**. 10. **Inventory mean 78.67% is elevated** vs EIA optimal 72% target. The generator's random walk drifts upward over 365 days due to `inflow - outflow = 5K bpd net positive`. **For optimal-target ML, filter to days near 72%** or augment with historical EIA reference levels. --- ## Where physics IS strong (use these for ML) Six coupling signals in this sample are **physically valid and ML-useful**: | Signal | Result | Source | |---|---:|---| | **Deviation from 72% ↔ optimization score** | r = -1.000 | Generator V-shape formula (deterministic) | | **Utilization ↔ shortage risk** | r = -0.522 | Generator binary threshold | | **Expected seasonal ↔ inflow** | r = +0.466 | sin(2π·day/365) modulation | | **Mass-balance inventory accumulation** | Deterministic per site | Tank conservation law | | **Day of year ↔ inflow** | r = -0.363 | Seasonal phasing | | **SPR reserve mean** | ~648M bbl | DOE post-2022 drawdown | --- ## Cross-references to other XpertSystems OIL SKUs This SKU is the **second storage/inventory SKU** in the catalog — complementing OIL-028 (tank operations) with **multi-site portfolio + SPR + seasonal dynamics**: | Storage layer | SKU | Focus | |---|---|---| | Tank operations | OIL-028 | API 650 mass-balance inventory + 6 product types × 3 tank types (single-site granularity) | | **Portfolio inventory** | **OIL-033** | **EIA PADD regions + 4 DOE SPR sites + seasonal dynamics + feature-coupled labels** *(this SKU)* | **OIL-033 vs OIL-028**: OIL-028 simulates **individual tank operations** (per-tank hourly mass balance, product types, integrity). OIL-033 simulates **portfolio-level inventory** across multiple PADD regions + SPR sites with daily granularity + seasonal patterns. Use OIL-028 for **single-tank ML**, OIL-033 for **regional/national inventory analytics**. **Natural integrations**: - **OIL-033 + OIL-029** → EIA inventory levels ↔ WTI prices for fundamentals- driven trading - **OIL-033 + OIL-030** → portfolio inventory ↔ global supply/demand - **OIL-033 + OIL-028** → portfolio rollup ↔ individual tank operations - **OIL-033 + OIL-031** → inventory levels ↔ tanker arrivals at terminals --- ## Full product The **full OIL-033 dataset** ships at **5,000 storage sites × 730 days** (prod mode) producing tens of millions of rows with **EIA PADD-tier-weighted capacity** (SPR sites ~150M bbl, Cushing ~18M, refineries ~8M), **type- conditioned inflow/outflow rates** (SPR ~50K bpd vs refinery ~400K bpd), **realistic SPR drawdown events** (2008/2011/2022 historical scenarios), **crude-refined product coupling** via refinery throughput ML linkages, **signed disruption events** (negative for outages only), **multi-year seasonal cycles with weather-driven anomalies**, and **PADD-aggregated EIA weekly inventory reports** matching real EIA Friday release schedule — licensed commercially. Contact XpertSystems.ai for licensing terms. 📧 **pradeep@xpertsystems.ai** 🌐 **https://xpertsystems.ai** --- ## Citation ```bibtex @dataset{xpertsystems_oil033_sample_2026, title = {OIL-033: Synthetic Crude & Product Inventory Dataset (Sample)}, author = {XpertSystems.ai}, year = {2026}, url = {https://huggingface.co/datasets/xpertsystems/oil033-sample} } ``` ## Generation details - Sample version : 1.0.0 - Random seed : 42 - Generated : 2026-05-23 13:46:01 UTC - Storage sites : 235 - Simulation days : 365 (1 year) - Regions : 5 (USGC, Midwest, East Coast, West Coast, Cushing per EIA PADD) - Storage types : 5 (Tank Farm, SPR, Refinery, Terminal, Floating) - SPR sites : 4 (Bryan Mound, Big Hill, West Hackberry, Bayou Choctaw — real DOE Gulf Coast salt domes) - Disruption types : 5 (Hurricane, Pipeline Outage, Refinery Fire, Import Disruption, Tank Failure) - Capacity range : 500K - 10M bbl (API 650 mixed portfolio) - Calibration basis : EIA Weekly Petroleum Status, EIA Petroleum Supply Annual, EIA Storage Capacity Report, DOE SPR, API 650, API 653, API 575, API 2350, PADD classifications, OECD Oil Stocks, JODI, EPA AP-42, NFPA 30 - Overall validation: 100.0/100 — Grade A+