Table 2 Components and descriptions for DNN and CNN models with Unencoded and Encoded features.

From: Deep learning framework for hourly air pollutants forecasting using encoding cyclical features across multiple monitoring sites in Beijing

Model type

Component

Description

DNN (Unencoded)

Input layer

Shape: (10, 10)

Flatten layer

Flatten the input to a 1D array

Hidden layers

Dense layer 1: Units = 32, Activation = ReLU

Dense layer 2: Units = 16, Activation = ReLU

Dense layer 3: Units = 8, Activation = ReLU

Output layer

Units: 1, Activation: Linear

Training configuration

Loss Function: (MSE)

Optimizer: Adam

Learning Rate: 0.005

Metrics: (RMSE)

Epochs: 100

CNN (Unencoded)

Input layer

Input Shape (10, 10)

Convolution layer

Filters: 32

Kernel Size: 2

Activation: ReLU

Flattening layer

Flattens the input data

Hidden layers

Dense Layer 1: Units = 16, Activation = ReLU

Dense Layer 2: Units = 8, Activation = ReLU

Output layer

Units = 1, Activation = Linear

Training configuration

Loss Function: (MSE)

Optimizer: Adam

Learning Rate: 0.005

Metrics: (RMSE)

DNN (Encoded)

Input layer

Shape: (10, 7)

Flatten layer

Flatten the input to a 1D array

Hidden layers

Dense layer 1: Units = 32, Activation = ReLU

Dense layer 2: Units = 16, Activation = ReLU

Dense layer 3: Units = 8, Activation = ReLU

Output layer

Units: 1, Activation: Linear

Training configuration

Loss Function: (MSE)

Optimizer: Adam

Learning Rate: 0.005

Metrics: (RMSE)

Epochs: 100

CNN (Encoded)

Input layer

Input Shape (10, 7)

Convolution layer

Filters: 32

Kernel Size: 2

Activation: ReLU

Flattening layer

Flattens the input data

Hidden layers

Dense Layer 1: Units = 16, Activation = ReLU

Dense Layer 2: Units = 8, Activation = ReLU

Output layer

Units = 1, Activation = Linear

Training configuration

Loss Function: (MSE)

Optimizer: Adam

Learning Rate: 0.005

Metrics: (RMSE)

Back to article page