Time Series Forecasting – Second Submission

Author: Lorena Romeo
Date: 2025-01-01

Purpose of the Project

The goal of this project is to forecast electricity consumption (kW) for one day using two different approaches:

1. Without using outdoor temperature
2. Including outdoor temperature as a feature

The forecasts consist of 96 rows (one for each 15-minute interval) and two columns (one per forecast).

In addition to traditional time series methods, a Neural Network (NN) model was also developed and is attached separately.

Data Investigation

• Dataset: 15-minute interval readings of power consumption and temperature from 01/01/2010 to 20/02/2010
• Variables:
  • Timestamp – date and time of the reading
  • Power_kW – electricity consumption in kilowatts
  • Temp_C – temperature in Celsius

Key Steps

• Libraries loaded: readxl, lubridate, dplyr, forecast, ggplot2, xts, zoo, randomForest, xgboost, vars, Metrics, etc.
• Timestamp cleaned using a custom function to handle Excel and character formats
• Zero values in power consumption treated as missing data (NA) and interpolated

Data Exploration & Cleaning

• Time series object created for power consumption (elec_ts) with 96 observations per day
• Decomposition of the series revealed:
  • Strong daily seasonality
  • Slight decreasing trend
  • Residuals with regular spikes, indicating autocorrelation
• Outliers (power = 0) replaced by interpolated values
• Differencing applied (1–2 times) to remove non-stationarity

Modelization

1. SARIMA Models

• Auto ARIMA: ARIMA(5,0,0)(0,1,0)[96]
• Manually Tuned SARIMA: ARIMA(1,1,7)(1,1,1)[96] – best performing model without temperature
• Residuals checked; RMSE, AIC, BIC calculated for comparison

2. Holt-Winters Exponential Smoothing

• Additive Model: RMSE = 11.531, AIC = 944.89, BIC = 954.67
• Multiplicative Model: Slightly better RMSE (11.295) but additive chosen due to stable seasonal variations

3. Including Temperature

• SARIMA model extended with temperature as a regressor: ARIMA(1,1,7)(1,1,1)[96] with xreg = Temp_C
• RMSE slightly improved; residuals analyzed for autocorrelation

4. Linear Regression

• Simple regression Power ~ Temperature tested but residuals showed patterns; model discarded

5. Machine Learning

• Random Forest: Used Timestamp and Temp_C as features
  • RMSE = 45.83 → worse than SARIMA/HW models
• Neural Network (NN) model: Developed separately and attached for further comparison

Forecasting

• Forecasts generated for 21st February 2010 using:

  1. Holt-Winters Additive (best traditional model) without temperature
  2. SARIMA with temperature as regressor
  3. NN model – see separate file for implementation and predictions

• Forecast plots visualized alongside historical data

Conclusion

• Best traditional approach without temperature: Holt-Winters Additive model
• Best approach with temperature: SARIMA with temperature regressor
• Machine Learning/NN models: Useful as alternative; performance evaluated in separate attachment
• Forecasts are ready for deployment and comparison