This project develops and evaluates a series of machine learning models to predict and forecast the power generation of solar power plants based on time-series weather data. This project showcases an end-to-end data science workflow, demonstrating a progression from foundational regression models to advanced deep learning techniques for time-series forecasting.
The workflow begins with data curation and exploratory data analysis (EDA). A baseline regression model using XGBoost is established and compared against a standard neural network. The project then advances to a more complex problem: forecasting power output for the next hour using a Long Short-Term Memory (LSTM) recurrent neural network, a specialized architecture for sequential data.
- Irradiation as the Primary Driver: EDA confirmed a strong, positive, and mostly linear relationship between Solar Irradiation and AC Power Output, identifying it as the most critical predictive feature.
- Temperature's Impact: The analysis also revealed a clear "temperature derating" effect, where panel efficiency decreases at very high module temperatures.
- Superior Regression Model: For the task of predicting instantaneous power output, a well-tuned XGBoost Regressor (R² ≈ 0.96) significantly outperformed a standard neural network, proving to be the most effective model for that specific problem.
- Successful Time-Series Forecasting: The more advanced LSTM model successfully learned the temporal patterns in the data. It was able to forecast the next hour's power generation by tracking the daily "bell curve" shape of solar output, demonstrating a powerful capability beyond simple regression.
- Solar Power Generation Data: The dataset, sourced from Kaggle, contains 34 days of 15-minute time-series data from two separate solar power plants, including both power generation and weather sensor data.
- Data Curation: Raw data from four separate CSV files was loaded, cleaned, and aggregated. Inverter-level power data was summed to the plant level for each timestamp to match the granularity of the weather data.
- Feature Engineering: Time-based features—
MONTH,DAY_OF_YEAR,HOUR, andMINUTE—were extracted from theDATE_TIMEcolumn to help models learn daily and seasonal cycles. - Part A: Regression Modeling:
- The data was prepared for standard regression (train-test split, scaling).
- An XGBoost Regressor was trained as a baseline and compared against a Feed-Forward Neural Network. Both were evaluated on their ability to predict instantaneous power output.
- Part B: Time-Series Forecasting (Advanced):
- The data was re-prepared using a "windowing" technique, where sequences of the last 6 hours of data were used to predict the next hour's output.
- An LSTM Recurrent Neural Network was built using TensorFlow/Keras to learn from these sequences and forecast future values.
For the task of predicting instantaneous power output, the XGBoost model was the clear winner.
| Metric | XGBoost (Baseline) | Neural Network |
|---|---|---|
| R-squared ( |
0.9578 | 0.9403 |
| MAE (kW) | 587.20 | 846.80 |
The LSTM model was able to successfully forecast the rise and fall of power generation over the course of a day. The plot below shows a sample of the model's forecast (orange) closely tracking the actual power output (blue).
01_Solar_Data_Curation.ipynb: Details the full process of loading, cleaning, and merging the source data.02_Solar_EDA_and_Modeling.ipynb: Contains the exploratory data analysis and the comparative analysis of the regression models (XGBoost vs. standard NN).03_Solar_Advanced_Forecasting_LSTM.ipynb: Contains the time-series data preparation and the development and evaluation of the LSTM forecasting model.solar_curated_data.pkl: The clean, analysis-ready dataset.app.py: A Streamlit script that deploys the best regression model (XGBoost) into an interactive web app.
- Python 3
- Pandas & NumPy
- Scikit-learn
- XGBoost
- TensorFlow (Keras)
- Streamlit
- Matplotlib & Seaborn