AI-Based Forecasting: Parameters Summary
Forecasting Parameters
| Parameter Type | Details |
|---|---|
| Forecast Horizons | 15 min to 7 days |
| Measurement Frequency | 10 seconds |
| Data Duration | Jan 2020 – Nov 2024 |
| Total Data Points | 17,297,280 |
Input Features
- Group 1: First derivatives – 𝑑GHI, 𝑑GHIcs, 𝑑kcs
- Group 2: Second derivatives – 𝑑²GHI, 𝑑²GHIcs, 𝑑²kcs
- Group 3: Third derivatives – 𝑑³GHI, 𝑑³GHIcs, 𝑑³kcs
- Group 4: Basic – GHI, GHIcs, kcs
- Group 5: Seasonal – hour, day, month (sin/cos encoding)
Performance Evaluation Metrics
| Metric | Purpose |
|---|---|
| MAE | Mean Absolute Error |
| RMSE | Root Mean Square Error |
| nMAE | Normalized MAE |
| nRMSE | Normalized RMSE |
| R² | Coefficient of Determination |
| 95% CI | Confidence Interval |
Model Architecture
- Model: LSTM (Long Short-Term Memory)
- Layers: 3 LSTM layers with 300 units each
- Final Layer: Dense (1 unit)
- Activations: Sigmoid, Tanh
- Training: Early stopping, time-series CV
- Framework: Keras (Python)
Data Preprocessing Steps
- Missing value handling
- Outlier detection
- Clear-sky irradiance with PVLib
- Remove night-time data
- Compute kcs (clear sky index)
- Compute 1st–3rd derivatives
- Generate seasonal features
- Downsampling for multiple horizons
Feature Engineering Settings
- Polynomial Order: Varies (1–5) based on horizon
- History Window (h): 5 or 10 time steps
- Feature Selection: Grid search based
✅ Advantages
| Aspect | Details |
|---|---|
| High Accuracy | LSTM-based model achieved low MAE and RMSE. |
| Long-term Dataset | 6 years of high-frequency data (17M+ samples). |
| Comprehensive Features | Derivatives up to 3rd order, seasonality, clear sky indices. |
| Multi-Horizon Forecasting | From 15-minute to 7-day forecasts. |
| Open Dataset & Code | Shared via GitHub for reproducibility. |
| Preprocessing Pipeline | Includes normalization and feature engineering. |
⚠️ Limitations
| Limitation | Explanation |
|---|---|
| High Computational Demand | Training requires substantial hardware. |
| Data-Dependent | Performance may drop outside original region. |
| Limited Explainability | Deep LSTMs are black-box models. |
| No Real-time Integration | No deployment for MPPT/load shifting. |
| Lack of Ensemble Comparison | No hybrid baselines included. |
| Scalability Constraints | Architecture not designed for microgrids. |
🔧 Possible Modifications & Improvements
| Suggestion | Benefit |
|---|---|
| Use Hybrid/Ensemble Models | Improve generalization and interpretability. |
| Deploy in Real-Time Systems | Smart grid decisions based on forecasts. |
| Use Transfer Learning | Better generalization in new regions. |
| Explainable AI | Use SHAP/LIME to interpret model. |
| Edge Deployment | Enable Raspberry Pi/Jetson deployment. |
| Integrate Weather APIs | Boost forecast accuracy in real-time. |