More information on our Model.

A PyTorch LSTM reads the last 48 hours of sensor and weather data and predicts PM2.5 one full day ahead, for three BC monitoring stations.

2.790 MAE

held-out future test, µg/m³ — best of all four models (RMSE 4.138, also best)

−17.6%

RMSE vs the naive same-hour-yesterday forecast

−3.2%

RMSE vs a tuned XGBoost trained on the same windows

—

live rolling 7-day MAE across stations

Four models, same exam

All scored on the same held-out future test period. Lower is better.

Climatology

typical value for this hour of day

5.294 RMSE 4.094 MAE

Seasonal persistence

same as this hour yesterday

5.019 RMSE 2.879 MAE

XGBoost (tuned)

gradient-boosted trees on the flattened window

4.276 RMSE 2.847 MAE

LSTM — ours

48 h × 13 features → 64 hidden units

4.138 RMSE 2.790 MAE

−17.6% RMSE vs naive−3.2% RMSE vs tuned XGBoostbest MAE and best RMSE

How to read this: persistence has decent MAE but the worst RMSE of the serious models — copying yesterday is catastrophicallywrong exactly when conditions change, which is when a forecast matters. The LSTM's RMSE margin comes from handling those volatile hours, not from shaving decimals on calm days.

Live, in production

Rolling 7-day MAE from the eval loop — no split, no lab conditions.

Prince George

Plaza 400 · sensor 4098

—

LSTM

—

naive

Vancouver

Clark Drive · sensor 9146190

—

LSTM

—

naive

Kelowna

KLO Road · sensor 1325038

—

LSTM

—

naive

The methodology check

The naive baseline's live MAE can be compared directly with its offline test value (2.88) — the closer they track, the stronger the evidence that the offline evaluation reflects production reality rather than fooling itself.

MAE — mean absolute error

Take every forecast, measure how far off it was, average. “Our forecasts are off by 2.8 µg/m³ on average.” The most interpretable metric — same unit as PM2.5 itself. For scale: a clean day is ~2–10, “unhealthy for sensitive groups” starts at 35.4, severe wildfire smoke exceeds 150.

RMSE — root mean squared error

Squares errors before averaging, so being wrong by 10 counts 100× more than being wrong by 1. The gap between RMSE and MAE reveals spikes — and for wildfire smoke, the spike days are exactly the days that matter for health. That's why both are reported.

PyTorchSnowflakedbtWeights & BiasesPostgresOpenAQOpen-Meteo

← Back to the live forecast