This project predicts AC Power output from solar plants using machine learning. It integrates inverter-level generation data and plant-level weather sensor data collected from two solar plants in India over 34 days.
Dataset: Solar Power Generation Data (Kaggle)
The pipeline covers data cleaning, feature engineering, model training, evaluation, and real-time forecasting.
- Generation Data (Plant 1 & 2): Inverter-level power generation.
- Weather Data (Plant 1 & 2): Environmental sensor measurements.
Key Features:
- Power & Yield:
DC_POWER,AC_POWER,DAILY_YIELD,TOTAL_YIELD - Weather Conditions:
AMBIENT_TEMPERATURE,MODULE_TEMPERATURE,IRRADIATION
- Merge generation and weather data by timestamp.
- Handle missing values (forward-fill).
- Convert timestamps and extract time-based features.
- Rolling Features:
- 1-hour mean & standard deviation
- 3-hour maximum irradiation
- Lag Features: Previous interval values (DC power, irradiation).
- Cumulative Features: Daily yield progression.
- Interaction Features: Irradiation × Temperature.
- Efficiency Metric:
DC Power / Irradiation.
Algorithms evaluated:
- Linear Regression
- Decision Tree
- Random Forest
- Gradient Boosting
- K-Nearest Neighbors (KNN)
- Neural Network (MLP)
Training Setup:
- Data split: 70% train, 15% validation, 15% test
- Preprocessing: Imputation + Standard Scaling
- Metrics: R² Score, RMSE
- Select best model by validation R².
- Evaluation includes:
- Scatter plots (Actual vs Predicted)
- Time series comparison
- Residual analysis
- Save artifacts:
best_model.pkl,scaler.pkl,imputer.pkl,feature_cols.pkl. - Supports real-time predictions with history buffer for lag/rolling features.
- Batch forecasting for short-term (few hours ahead) power prediction.
- Best Model: Random Forest (achieved highest R²).
- Performance: Strong alignment between predicted and actual AC power.
- Visualization: Predictions accurately follow solar generation trends.
- Real-time AC power prediction
- Short-term AC power forecasting (few hours ahead)