This project utilizes Time Series Forecasting techniques to predict electricity consumption using the XGBoost algorithm. The dataset contains hourly electricity usage from PJM West, which is used to train the model. Through feature engineering and hyperparameter tuning, I have aimed to create an accurate model that forecasts future electricity demand.
The dataset consists of hourly electricity consumption values over time. Here's a small sample:
| Datetime | PJME_MW |
|---|---|
| 12/31/2002 1:00 | 26498 |
| 12/31/2002 2:00 | 25147 |
| 12/31/2002 3:00 | 24574 |
| 12/31/2002 4:00 | 24393 |
| 12/31/2002 5:00 | 24860 |
After feature engineering, the dataset looks like this:
| Feature | Data Type |
|---|---|
| Datetime | datetime64[ns] |
| PJME_MW | float64 |
| Year | int32 |
| DayOfMonth | int32 |
| Month | int32 |
| Hour | int32 |
| DayOfWeek | int32 |
| DayOfYear | int32 |
| Quarter | int32 |
- XGBoost Regressor For Time Series Forecasting.
- Hyperparameter Tuning using RandomizedSearchCV to optimize model performance.
Some of the charts used in this analysis:
- Actual vs Predicted Consumption 📉
- Electricity Consumption Trend Over Time 🕒
- Residuals of Predictions 🧮
- Additional Features: Incorporate weather forecasts, holiday data, and other external factors.
- Model Comparisons: Test other machine learning algorithms such as LSTM or ARIMA.
- Refinement of Hyperparameters: Use GridSearchCV for more exhaustive hyperparameter optimization.
By leveraging XGBoost and robust feature engineering, we successfully predicted future electricity consumption. This model can serve as a valuable tool for planning electricity demand and supply management.
-
Clone the repository:
git clone https://github.com/hemilshah99316/ELECTRICITY_USAGE_PREDICTION_USING_ML.git
OR Download ZIP File
-
Run the Jupyter Notebook:
jupyter notebook forecasting.ipynb