Cross-Validated Time Series Modeling & Forecasting

Overview:

This personal project aims to apply XGBoost in modeling and forecasting Time Series energy consumption data from PJM Interconnect's Eastern seaboard.

File Glossary:

gh_timeseries_project.ipynb :

Jupyter Notebook with python code implementing Pandas, Scikit-Learn, and XGBoost in exploring Time-Series hourly energy consumption (mW) data.

PJME_hourly.csv :

Time-Series hourly energy consumption (mW) data from PJM Interconnection LLC, specifically from the East region of its USA-based Regional Transmission Network.

/PJME_analysis_predictions/ :

Output directory for all assets produced by 'gh_timeseries_project.ipynb', directly uploaded for convenience. The python code lightly handles redundant file generation should you choose to run the repo contents locally.

Data Glossary:

Target Feature:

• ['PJME_MW'] - Energy consumption in 'mW' - TARGET FEATURE FOR MODELING

Time Features:

• ['hour'] - Label features of Time Series data
• ['dayofweek'] - Label features of Time Series data
• ['dayofmonth'] - Label features of Time Series data
• ['dayofyear'] - Label features of Time Series data
• ['month'] - Label features of Time Series data
• ['quarter'] - Label features of Time Series data
• ['year'] - Label features of Time Series data

Engineered Lag Features:

• ['lag1'] - Engineered feature relating a given 'mW' data point to the same Date and Time 1 year ago
• ['lag2'] - Engineered feature relating a given 'mW' data point to the same Date and Time 2 years ago
• ['lag3'] - Engineered feature relating a given 'mW' data point to the same Date and Time 3 years ago

Tools and Approaches:

• Pandas, Numpy - data handling
• Matplotlib, Seaborn - data visualization and assessment of basic data behavior
• SciKitLearn - error metrics, TimeSeriesSplit, cross-validation
• XGBoost - modeling and forecasting w/ XGBRegressor
• 'os', 'shutil' - local device model and visualization storage

Results:

Engineered Lag Features were determined to be highly important factors in our model's ability to reduce error, followed by Hour of Day and Day of Year.

I was able to model hourly energy consumption from 2002-2018 with Root Mean Squared Error (RMSE) of 3457 mW for predictions on a scale of 20000 to 60000 mW. Additionally, I was able to forecast a full year's worth of hourly data from 2018-08 to 2019-08, and save the model for future use or adjustment.