LSTM Neural Network Architecture Diagram

I need a professional architecture diagram for an optimized LSTM-based neural network used in a water demand forecasting system. The diagram will be included in a research paper, so it must meet academic publication standards (IEEE/Elsevier style).

Requirements for the Diagram
Input
A 14-day sequence of 34 engineered features, including:

Rolling window statistics (mean, std, min, max over 7, 14, and 30 days)

Lag features

Cyclical time encodings using sine and cosine for hour, day, and month

LSTM-Based Architecture

Layer 1: LSTM (128 units), unidirectional

Layer 2: BiLSTM (64 units), bidirectional

Layer 3: LSTM (32 units), unidirectional

Regularization and Normalization

Dropout after each LSTM layer (0.3, 0.25, 0.2)

Batch Normalization after each LSTM layer

RobustScaler used for input normalization

Dense Layers

Use Swish activation: Swish(x) = x · σ(x)

Loss Function

Huber Loss (δ = 1.0)

Optimizer

AdamW with weight decay = 0.001

Training Enhancements

Early stopping (patience = 30)

ReduceLROnPlateau

Model checkpointing