Multi-Horizon Solar Power Forecasting for Grid Integration

Machine learning-based solar power forecasting system for grid integration, featuring multi-horizon predictions, uncertainty quantification, and model interpretability analysis.

Problem Statement

Solar power is inherently intermittent - output varies with weather and time of day. Grid operators require accurate forecasts to:

• Balance supply and demand in real-time (AGC)
• Schedule reserves and backup generation
• Optimize economic dispatch of power plants
• Plan unit commitment for day-ahead operations

This project implements multi-horizon forecasting (15-min, 1-hour, 4-hour ahead) with prediction intervals and SHAP-based interpretability to address these operational needs.

Key Results

Model Performance (Test Set)

Horizon	Best Model	R²	RMSE (kW)	nRMSE
15 min	Ridge	0.934	21,078	32.5%
1 hour	LightGBM	0.884	27,872	43.5%
4 hours	CatBoost	0.875	28,570	46.9%

Cross-Validation Stability (1-hour horizon)

Model	CV R² (Mean ± Std)
CatBoost	0.879 ± 0.011
XGBoost	0.871 ± 0.019
LightGBM	0.870 ± 0.015
Ridge	0.857 ± 0.027

Uncertainty Quantification

• 80% Prediction Interval Coverage: 78.5% (target: 80%)
• Adaptive interval width based on forecast uncertainty

Features

Technical Highlights

• Multi-Horizon Forecasting: 15-min, 1-hour, and 4-hour ahead predictions
• 32 Engineered Features: Time-based, solar position, lag, rolling statistics, ramp rates
• 4 Model Comparison: Ridge, XGBoost, LightGBM, CatBoost
• Time Series CV: 5-fold walk-forward validation
• Uncertainty Quantification: Quantile regression for prediction intervals
• SHAP Analysis: Model interpretability and feature importance

Feature Engineering

Category	Features	Description
Time	hour_sin/cos, time_slot	Cyclical encoding for daily patterns
Solar Position	hour_angle, solar_elevation	Sun position approximation
Temperature	temp_diff, temp_efficiency	Thermal derating (~0.4%/°C)
Lag Features	power_lag_1/4/96	Historical values (15min, 1hr, 24hr)
Rolling Stats	mean, std, max	1-hour and 2-hour windows
Ramp Rate	power_ramp, irr_ramp	Rate of change features

Grid Integration Applications

Forecast Horizon	Application	Operational Use
15 minutes	AGC	Real-time frequency regulation
1 hour	Economic Dispatch	Generator scheduling, reserve allocation
4 hours	Unit Commitment	Start-up/shut-down decisions

Project Structure

solar-power-forecasting/
├── data/                                    # Dataset (gitignored)
│   ├── Plant_1_Generation_Data.csv
│   └── Plant_1_Weather_Sensor_Data.csv
├── notebooks/
│   └── 04_multi_horizon_forecasting.ipynb   # Main analysis
├── requirements.txt
└── README.md

Quick Start

# Clone repository
git clone https://github.com/ZeuPark/SolarPowerForecasting.git
cd SolarPowerForecasting

# Install dependencies
pip install -r requirements.txt

# Run the notebook
jupyter notebook notebooks/04_multi_horizon_forecasting.ipynb

Dataset

• Source: Kaggle - Solar Power Generation Data
• Duration: 34 days (May 15 - June 17, 2020)
• Resolution: 15-minute intervals (3,157 samples)
• Plant: 22 inverters aggregated to plant-level

Feature	Description	Unit
DC_POWER	Target variable (total plant output)	kW
IRRADIATION	Solar irradiance	kW/m²
AMBIENT_TEMPERATURE	Air temperature	°C
MODULE_TEMPERATURE	Panel temperature	°C

Technical Stack

Category	Tools
Data Processing	pandas, numpy
Machine Learning	scikit-learn, XGBoost, LightGBM, CatBoost
Interpretability	SHAP
Visualization	matplotlib, seaborn

Methodology

1. Data Preprocessing

• Aggregate 22 inverters to plant-level
• Remove data leakage (AC_POWER, DAILY_YIELD, TOTAL_YIELD)
• Handle missing values from lag/rolling features

2. Feature Engineering

• 32 features capturing temporal patterns, solar physics, and historical trends
• Temperature efficiency factor based on Si panel thermal coefficients

3. Model Training

• Chronological train/test split (80/20)
• Separate models for each forecast horizon
• Hyperparameter tuning via cross-validation

4. Uncertainty Quantification

• Quantile regression (10th, 50th, 90th percentiles)
• 80% prediction intervals for reserve planning

5. Model Interpretability

• SHAP TreeExplainer for feature importance
• Dependence plots for feature interactions
• Waterfall plots for individual predictions

Key Insights

SHAP Analysis Findings

1. Lag features dominate: Recent power values most influential
2. Irradiation matters: Primary physical driver correctly identified
3. Physical consistency: Model behavior aligns with solar physics
4. Time patterns: Hour and time_slot capture daily cycles

Error Patterns

• Higher errors during peak generation hours (10:00-14:00)
• Morning/evening transitions show more variability
• Prediction intervals widen during uncertain periods

Future Work

• Weather forecast integration (NWP data)
• Deep learning models (LSTM, Transformer)
• Real-time deployment with SCADA integration
• Multi-plant ensemble forecasting

Author

Electrical Engineering Student
Focus: Power Systems, Renewable Energy Integration, Machine Learning

License

MIT License