{
  "model_type": "LSTM",
  "architecture": "LSTM(128) \u2192 Dropout(0.1) \u2192 Dense(32) \u2192 Dense(1)",
  "framework": "TensorFlow/Keras",
  "task": "Single-step monthly groundwater level forecasting",
  "features": [
    "water_level",
    "temperature",
    "precipitation",
    "wind_speed"
  ],
  "target": "water_level",
  "lookback_months": 24,
  "horizon_months": 1,
  "tuning": {
    "method": "Bayesian Optimisation (Keras Tuner)",
    "n_trials": 20,
    "best_config": {
      "n_layers": 2,
      "units_1": 64,
      "dropout": 0.1,
      "lr": 0.0007731576839806804,
      "batch_size": 32
    }
  },
  "data_splits": {
    "train": {
      "start": "1944-01-01",
      "end": "2007-10-01",
      "n_months": 766
    },
    "validation": {
      "start": "2007-11-01",
      "end": "2015-10-01",
      "n_months": 96
    },
    "test": {
      "start": "2015-11-01",
      "end": "2023-10-01",
      "n_months": 96
    }
  },
  "test_metrics": {
    "RMSE": 2.9386,
    "MAE": 2.397,
    "MAPE_pct": 3.671,
    "R2": 0.5505,
    "NSE": 0.5505
  },
  "notes": "Scaler fitted on train only. Oracle exog assumption \u2014 contemporaneous met vars used at forecast time."
}