kozy9/GWSarimax

SARIMAX Groundwater Level Forecasting — UK

A SARIMAX model trained to forecast monthly groundwater levels (GWLs) using historical water level data and meteorological variables.

Model Details

Parameter	Value
Architecture	SARIMAX(2, 1, 1)x(2, 0, 2, 12)
Seasonal period	12 months
Target	Water level (m)
Exogenous variables	Temperature (°C), Precipitation (mm), Wind Speed (km/h)
Feature engineering	None — raw exog only
Training period	1944-01-01 → 2015-10-01 (862 months)
Test period	2015-11-01 → 2023-10-01 (96 months)

Hyperparameter Tuning

Bayesian search : 50 trials | criterion: validation RMSE ranked by out-of-sample validation RMSE.

Test Set Performance

Metric	Value
RMSE	5.154
MAE	4.1969
MAPE (%)	6.5844
R²	-0.3826
NSE	-0.3826

This model is a statistical baseline for benchmarking against deep learning approaches (LSTM, TCN).

Important Note

Contemporaneous meteorological variables are used as exogenous inputs at forecast time (oracle assumption). Future met values are treated as known. This matches the experimental setup used for LSTM/TCN comparisons in this study.

Repository Contents

├── sarimax_model.pkl     # Fitted model (joblib)
├── model_config.json     # Parameters, metadata & metrics
├── inference.py          # Load model & generate forecasts
└── README.md             # This file

Quick Start

from huggingface_hub import hf_hub_download
import joblib, pandas as pd, numpy as np

model_path = hf_hub_download(repo_id='kozy9/GWSarimax', filename='sarimax_model.pkl')
model = joblib.load(model_path)

idx   = pd.date_range(start='2024-01-01', periods=12, freq='MS')
X_fut = pd.DataFrame({
    'temperature'  : [...],
    'precipitation': [...],
    'wind_speed'   : [...],
}, index=idx)

fc   = model.get_forecast(steps=12, exog=X_fut)
pred = fc.predicted_mean
ci   = fc.conf_int()