Time Series · Model Evaluation
Electricity Forecasting
Compare statistical, neural, and tree-based forecasts of hourly electricity demand across 368 clients grouped into four load-profile clusters.
- Python
- Time Series
- XGBoost
- LightGBM
- LSTM
- Prophet
- n8n
What I built
A chronological 2014 holdout after training on 2011-2013 observations.
- Cleaned UCI electricity readings, aggregated 15-minute data to hourly demand, and clustered client load profiles.
- Compared ARIMA, SARIMAX, Prophet, LSTM, XGBoost, and LightGBM using a chronological holdout.
- Reported cluster MAPE of approximately 3.8%-6.1% for XGBoost and 4.4%-13.3% for LightGBM in the repository analysis.
How it works
- Clean and aggregate the UCI load readings to hourly observations.
- Cluster 368 active clients into four load-profile groups.
- Train statistical, neural, and tree-based models on the historical split.
- Evaluate on 2014 observations and inspect cluster-level forecasting errors.
Evaluation & evidence
Repository-reported MAPE uses 2011-2013 training and a 2014 holdout. Compare models on the same split and inspect cluster-level performance rather than treating the best score as a universal result.
Scope & limitations
These results describe this historical dataset and split. They are not evidence of current production forecasting accuracy or generalization to other markets.