Household Power Consumption Forecasting
Multi-model time series on 2.9M UCI records (2006–2010). ARIMA, SARIMA, Prophet, LSTM on Global_active_power. STL reveals daily+weekly patterns. Ensemble with inverse-RMSE weighting across all models.
UCI Household Power: 2.9M measurements, 1-min resolution, 4 years
STL decomposition → stationarity → ARIMA/SARIMA/Prophet/LSTM → ensemble
Time series forecasting for household electricity on the UCI dataset.
Dataset
- ▸2,075,259 measurements (Dec 2006 → Nov 2010), 1-minute resolution → resampled hourly
- ▸1.25% missing values → linear interpolation
- ▸Target: Global_active_power (kW)
Time Series Analysis
- ▸ADF test: stationary after differencing
- ▸STL decomposition: daily pattern (morning/evening peaks) + weekly cycle (weekend dips)
- ▸ACF/PACF: clear lag-24h and lag-168h autocorrelations
Models Evaluated
| Model | Strengths |
|---|---|
| ARIMA | Auto p,d,q selection, AIC |
| SARIMA | Seasonal period=24h |
| Prophet | Trend + yearly + weekly + daily seasonality |
| LSTM (encoder-decoder) | 24-step ahead sequential prediction |
| Ensemble | Inverse-RMSE weighted combination |
Key Insight The seasonal component explains ~60% of variance. People follow routines — a "same time yesterday" baseline is hard to beat. Advanced models target the irregular residual component (vacations, guests, appliance failures).
STL Decomposition Findings Trend: slowly declining over 4 years (efficiency improvements). Seasonal: strong daily (double peak: 7–9am + 6–9pm) + weekly (weekends 15% lower). Residual: 1.25% anomalies correlate with holidays.