Methane is a major greenhouse gas—about 28× the 100-year warming potential of CO₂ and ~80× over 20 years. Oil and gas operations are a large Canadian methane source. The Wood Buffalo region is monitored around the clock by the Wood Buffalo Environmental Association (WBEA); data are public but not always modeled in depth. Luo, Wang, and Gates (2024), Environmental Modeling & Assessment, study multivariate LSTM forecasts at seven fixed stations, comparing methane-only inputs with temperature and wind covariates. The setting is surface air-quality networks in an oil-sands region, not underground coal faces, but the multivariate time-series design is transferable to ventilation and industrial methane monitoring.

1. Motivation

1.1 Beyond univariate ARIMA

Classic ARIMA-style tools often use methane series alone, while ambient temperature and wind are known drivers. Machine learning is widely used for GHG and air-pollution forecasting, yet systematic LSTM studies that compare climate covariates for methane concentration were limited—this paper asks whether temperature, wind speed, and direction improve LSTM forecasts.

1.2 Data context

Hourly records come from the Alberta Air Data Warehouse (WBEA), seven stations with spans roughly 2010–2021 (station-dependent):

| Station | |---------| | Anzac | | Bruderheim | | Buffalo Viewpoint | | Fort McMurray—Athabasca Valley | | Fort McMurray—Patricia McInnes | | Lower Camp | | Mildred Lake |

Methane, temperature, wind speed, and direction share the same hourly sample counts per station, under CEMS and WBEA–AEP QA/QC.

2. Data and preprocessing

2.1 Inputs

| Variable | Role | Note | |----------|------|------| | Methane | Target (ppm) | Hourly | | Temperature | Covariate | °C | | Wind speed/direction | Covariates | (U_x = U\cos\varnothing), (U_y = U\sin\varnothing) (radians) |

Seasonal plots (e.g. Bruderheim) show higher concentrations in colder months on average, with year-to-year exceptions—supporting nonlinear sequence models.

2.2 Cleaning and rolling windows

Outliers: Z-score screening plus Isolation Forest so means/std are not dominated by spikes. Rolling lookback (\Delta t) builds supervised steps from (t-\Delta t,\ldots,t) to predict (t+1); lookback is grid-searched.

2.3 Splits and leakage control

Per station: 80% train, 10% test, 10% cross-validation; no mixing across stations.

Blocked time-series cross-validation with gaps between train and validation and between folds (gap size tied to lookback) to limit lag leakage and memorization.

3. Model: branched LSTM and four input sets

3.1 Topology (conceptual)

Raw series → anomaly removal → rolling window + normalization
                ↓
    One LSTM branch per covariate → concat → dense → methane forecast

Grid search tunes hyperparameters (see original Table 3). Metrics: RMSE and MAE (MAE as (\sum |Y_t-P_t|/Y_t)).

3.2 Four compared setups

| Label | Inputs | |-------|--------| | Methane-only | Methane history | | Methane + Temp | Methane + temperature | | Methane + Wind | Methane + (U_x, U_y) | | Methane + Temp + Wind | All of the above |

4. Key results (from the paper)

4.1 Training and validation

Broad patterns across seven stations:

  • Adding temperature often beats methane-only and methane + wind;
  • Methane + wind frequently worse than methane-only;
  • All covariates win on some stations but lose to methane + temp on others.

Anzac MAE example (train+val, ppm):

| Model | MAE | |-------|-----| | Methane-only | 0.00245 | | Methane + Temp | 0.00207 | | Methane + Wind | 0.00322 | | Methane + Temp + Wind | 0.00168 |

Fort McMurray—Athabasca Valley RMSE example: methane-only 0.000410; +wind 0.000412; +temp 0.000439—site heterogeneity matters.

4.2 Forecast phase (held-out future)

Errors are roughly an order of magnitude larger than in train/validation—deployment needs true forward testing, not in-sample scores alone.

Fort McMurray—Patricia McInnes forecast MAE example: methane-only 0.0478 ppm; +wind 0.0486; +temp 0.0457.

4.3 Variance and wind

Authors report no simple link between methane variance and forecast error across stations. Wind did not reliably help, likely due to stochastic dilution signals versus learnable structure. Temperature more often helped, consistent with gas volume and solubility effects in oil-and-gas contexts.

5. Limitations (authors)

LSTM overfitting, suboptimal hyperparameters, residual anomalies; uneven performance by station; limited benchmark against ARIMA/Prophet in the same table; surface air-quality setting—not interchangeable with underground gas concentration or face emission rate models.

6. Engineering takeaways

  • More covariates ≠ better: wind often hurt here—validate against physics and local data.
  • Per-site models: station-separated training mirrors per-face or per-district mine models.
  • Temporal validation: blocked CV and forecast-phase error jumps warn against over-trusting fit-period metrics.
  • QA pipeline: Z-score + isolation forest parallels outlier rejection + cross-checks in mine monitoring.
  • Open data: via Alberta Airsheds Council for reproduction.

Reference

Luo, R.; Wang, J.; Gates, I. Forecasting Methane Data Using Multivariate Long Short-Term Memory Neural Networks. Environmental Modeling & Assessment 2024, 29, 441–454.