Table 4 Hyperparameters for TCN + BiLSTM + Attention mechanism Training.

Network Parameters

TCN

Kernel Sizes

[5, 7, 9]

Dilation Rates

[1, 2, 4, 8]

Number of Filters

64 per branch

Dropout Rate

0.2

BiLSTM

Hidden Units

256 (each direction)

Number of Layers

2

Dropout Rate

0.3

Attention

Number of Heads

8

Key Dimension

64

Dropout Rate

0.1

Training Parameters

Optimization

Optimizer

Adam

Initial Learning Rate

3 × 10^−4

Weight Decay

1 × 10^−5

Gradient Clip Norm

1.0

Training Process

Batch Size

64

Gradient Accumulation Steps

4

Number of Epochs

157

Early Stopping Patience

15

Loss Function

MSE Weight (\(\:{\alpha\:}_{1}\))

0.6

Temporal Weight (\(\:{\alpha\:}_{2}\))

0.3

Attention Weight (\(\:{\alpha\:}_{3}\))

0.1

Data Processing

Sequence Length

10,000 timesteps

Sampling Rate

1000 Hz

Window Overlap

50%

Model Settings

Random Seed

42

Cross-validation Folds

5

Test Set Ratio

0.2