Table 1 Summary of experimental parameters and model configuration.
From: Comparative analysis of deep learning architectures in solar power prediction
Category | Parameter | Value / Setting |
|---|---|---|
General Setup | Random Seed | 42 |
Framework | Keras 3 with TensorFlow 2.16 | |
Dataset File | “Dataset.csv” | |
Data Processing | Train / Validation / Test Split | 70% / 20% / 10% (via 30% split then 1/3 split of temp set) |
Feature Scaling Method | StandardScaler (z-score normalization) | |
Time Series Conversion | Reshaping input as sequences with shape (features, 1) | |
Feature Engineering | Feature Renaming | Applied for clarity (e.g., temperature_2_m_above_gnd → Temp_2m) |
Feature Selection | Lasso Regression ( alpha = 0.001 ) | |
Optimization | Optimizer | Adam |
Learning Rate | 0.001 | |
Loss Function | Mean Squared Error (MSE) | |
Early Stopping | patience = 10 epochs, restore best weights | |
Learning Rate Scheduler | ReduceLROnPlateau, patience = 5 epochs | |
Training Configuration | Epochs | 200 (with early stopping) |
Batch Size | 32 | |
Evaluation Metrics | Forecast Metrics | RMSE, MAE, MAPE, R² |
Diagnostic Tests | Shapiro-Wilk, Jarque-Bera, Ljung-Box | |
Uncertainty Estimation | Monte Carlo Dropout (100 runs) | |
Visualization | Plots | Loss curves, predicted vs. real scatter, azimuth plots, residual histograms, 95% CIs |
Models Evaluated | Autoencoder | Dense + Bottleneck + Reconstruction |
RNN Models | SimpleRNN, GRU, LSTM | |
CNN | 1D Conv + MaxPooling + Global Avg Pooling | |
TCN | Dilated Causal Convolutions | |
Transformer | Multi-Head Attention blocks with feedforward layers | |
InformerLite | Causal Conv + Attention + Global Avg Pooling |
