Table 2 Hyperparameters for CNN-LSTM-DNN Model
From: Unequal spatio-temporal distribution of population-weighted pollution extremes through deep learning
Category | Hyperparameter | Description | Values |
|---|---|---|---|
CNN | Number of Filters | Number of feature detectors | 32 |
Kernel Size | Size of the filter window | 3,3 | |
Stride | Step size of the filter | 1 | |
Padding | Handles border regions | valid | |
Activation Function | Introduces non-linearity | ReLU | |
Pooling Type | Reduces spatial dimensions | Max | |
Dropout Rate | Prevents overfitting | 30% | |
LSTM | Number of LSTM Units | Size of the hidden state | 32, 16 |
Number of LSTM Layers | Stacked LSTMs | 2 | |
Dropout Rate | Prevents overfitting | 30% | |
Recurrent Dropout | Dropout between time steps | 0.3 | |
Activation Function | Used in LSTM cells | tanh | |
Bidirectional | Whether LSTM processes both directions | True | |
Dense | Hidden Units | Size of dense layers | 64, 32, 16 |
Dropout Rate | Prevents overfitting | 30% | |
Regularisation | L1/L2 regularisation | L1 = 1e-5, L2 = 1e-4 | |
General Training | Batch Size | Samples per training batch | 64 (Adaptive) |
Learning Rate | Step size for optimiser | 0.001 | |
Optimizer | Algorithm for weight updates | Adam | |
Loss Function | Measures prediction error | MSE | |
Number of Epochs | Complete training cycles | 100 | |
Early Stopping | Stops training if no improvement | True (patience=15) | |
Learning Rate Reduction | Reduces LR on plateau | factor=0.5, patience=7 | |
Validation Split | Portion used for validation | 20% |
