Hugging Face's logo

kozy9
/
GWTCN

Keras
time-series
forecasting
tcn
hydrology
groundwater
Model card Files
xet
 Community
kozy9 commited on
Commit
3f57d10
·
verified ·
1 Parent(s): 03b8a8b

Upload TCN — test RMSE 2.9366 m R² 0.5511

Browse files
Files changed (7) hide show
  1. README.md +11 -11
  2. model_config.json +13 -11
  3. tcn_baseline_training_history.png +0 -0
  4. tcn_forecast_comparison.png +2 -2
  5. tcn_model.keras +2 -2
  6. tcn_residual_analysis.png +0 -0
  7. tcn_tuned_training_history.png +0 -0
README.md CHANGED
@@ -17,7 +17,7 @@ level forecasting using meteorological variables as exogenous inputs.
17
 
18
  | Parameter | Value |
19
  |---|---|
20
- | Architecture | TCN(nb_filters=32, kernel_size=3, dilations=[1, 2, 4, 8]) → Dense(1) |
21
  | Framework | TensorFlow / Keras (keras-tcn) |
22
  | Task | Single-step monthly forecasting |
23
  | Lookback window | 24 months |
@@ -36,22 +36,22 @@ level forecasting using meteorological variables as exogenous inputs.
36
 
37
  | Parameter | Value |
38
  |---|---|
39
- | nb_filters | 32 |
40
- | kernel_size | 3 |
41
- | dilations | [1, 2, 4, 8] |
42
  | dropout_rate | 0.1 |
43
- | learning_rate | 0.001000 |
44
- | Receptive field | 61 months |
45
 
46
  ## Test Set Performance
47
 
48
  | Metric | Value |
49
  |---|---|
50
- | RMSE | 3.5771 m |
51
- | MAE | 2.8901 m |
52
- | MAPE | 4.3138% |
53
- | R² | 0.334 |
54
- | NSE | 0.334 |
55
 
56
  > This model is part of a benchmark study comparing SARIMAX, LSTM, and TCN
57
  > for UK groundwater level forecasting.
 
17
 
18
  | Parameter | Value |
19
  |---|---|
20
+ | Architecture | TCN(nb_filters=128, kernel_size=4, dilations=[1, 2, 4, 8, 16, 32]) → Dense(1) |
21
  | Framework | TensorFlow / Keras (keras-tcn) |
22
  | Task | Single-step monthly forecasting |
23
  | Lookback window | 24 months |
 
36
 
37
  | Parameter | Value |
38
  |---|---|
39
+ | nb_filters | 128 |
40
+ | kernel_size | 4 |
41
+ | dilations | [1, 2, 4, 8, 16, 32] |
42
  | dropout_rate | 0.1 |
43
+ | learning_rate | 0.010000 |
44
+ | Receptive field | 379 months |
45
 
46
  ## Test Set Performance
47
 
48
  | Metric | Value |
49
  |---|---|
50
+ | RMSE | 2.9366 m |
51
+ | MAE | 2.3952 m |
52
+ | MAPE | 3.5738% |
53
+ | R² | 0.5511 |
54
+ | NSE | 0.5511 |
55
 
56
  > This model is part of a benchmark study comparing SARIMAX, LSTM, and TCN
57
  > for UK groundwater level forecasting.
model_config.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "model_type": "TCN",
3
- "architecture": "TCN(nb_filters=32, kernel_size=3, dilations=[1, 2, 4, 8]) \u2192 Dense(1)",
4
  "framework": "TensorFlow/Keras (keras-tcn)",
5
  "task": "Single-step monthly groundwater level forecasting",
6
  "features": [
@@ -16,17 +16,19 @@
16
  "method": "Bayesian Optimisation (Keras Tuner)",
17
  "n_trials": 20,
18
  "best_config": {
19
- "nb_filters": 32,
20
- "kernel_size": 3,
21
  "dilations": [
22
  1,
23
  2,
24
  4,
25
- 8
 
 
26
  ],
27
  "dropout_rate": 0.1,
28
- "learning_rate": 0.001,
29
- "receptive_field": 61
30
  }
31
  },
32
  "data_splits": {
@@ -47,11 +49,11 @@
47
  }
48
  },
49
  "test_metrics": {
50
- "RMSE": 3.5771,
51
- "MAE": 2.8901,
52
- "MAPE_pct": 4.3138,
53
- "R2": 0.334,
54
- "NSE": 0.334
55
  },
56
  "notes": "Scaler fitted on train only. Oracle exog assumption \u2014 contemporaneous met vars used at forecast time."
57
  }
 
1
  {
2
  "model_type": "TCN",
3
+ "architecture": "TCN(nb_filters=128, kernel_size=4, dilations=[1, 2, 4, 8, 16, 32]) \u2192 Dense(1)",
4
  "framework": "TensorFlow/Keras (keras-tcn)",
5
  "task": "Single-step monthly groundwater level forecasting",
6
  "features": [
 
16
  "method": "Bayesian Optimisation (Keras Tuner)",
17
  "n_trials": 20,
18
  "best_config": {
19
+ "nb_filters": 128,
20
+ "kernel_size": 4,
21
  "dilations": [
22
  1,
23
  2,
24
  4,
25
+ 8,
26
+ 16,
27
+ 32
28
  ],
29
  "dropout_rate": 0.1,
30
+ "learning_rate": 0.01,
31
+ "receptive_field": 379
32
  }
33
  },
34
  "data_splits": {
 
49
  }
50
  },
51
  "test_metrics": {
52
+ "RMSE": 2.9366,
53
+ "MAE": 2.3952,
54
+ "MAPE_pct": 3.5738,
55
+ "R2": 0.5511,
56
+ "NSE": 0.5511
57
  },
58
  "notes": "Scaler fitted on train only. Oracle exog assumption \u2014 contemporaneous met vars used at forecast time."
59
  }
tcn_baseline_training_history.png CHANGED
tcn_forecast_comparison.png CHANGED

Git LFS Details

  • SHA256: 095ad9982c82cceceed41399ff4820eae6bc0978c30ebea69a642aa182302513
  • Pointer size: 131 Bytes
  • Size of remote file: 393 kB

Git LFS Details

  • SHA256: c744ba7ea03d6719a9c3c8fa1cc6d2f9a52b7d5c8f6a187adff9ab75593a62d6
  • Pointer size: 131 Bytes
  • Size of remote file: 394 kB
tcn_model.keras CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f7b000cf26bedc98461a4d960aa999747d38e39fe1545ddcf093f0f88cd0f537
3
- size 387592
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9a21237d8f8dfb50926730017db6d464f22784e9990e73ce18ee1bc59a35c5c8
3
+ size 8868973
tcn_residual_analysis.png CHANGED
tcn_tuned_training_history.png CHANGED