Table 1 Parameter settings for the LSTM-CNN model.
Serial number | Parameter | Unit | Parameter setting |
|---|---|---|---|
1 | Optimizer | – | RMSProp |
2 | Learningrate | – | 0.01, decayed by 0.5 every 100 epochs, minimum 1e-6 |
3 | Variable initialization | – | Xavier initialization |
4 | L1 regularization | – | 0.001 |
5 | L2 regularization | – | 0.0001 |
6 | Activation function (conv/fc) | – | ReLU |
7 | Activation function (LSTM cell) | – | tanh |
8 | Activation function (LSTM gate) | – | sigmoid |
9 | Activation function (output) | – | Linear |
10 | Training epochs | epochs | 2000 (early stopping with patience = 50 epochs) |
11 | Batch size | samples/batch | 64 |
12 | Sliding window length | days | 30 |
13 | Sliding step | days | 1 |
14 | Prediction horizon | days | 1 |
15 | Feature dimensions | features/timestep | 6 |
16 | Optimizer momentum | – | 0.9 |
17 | Optimizer decay rate | – | 0.0001 |
18 | Data split ratio | – | Training: 70%; Validation: 15%; Test: 15% |
