Electricity Consumption Forecasting

You are provided with electricity consumption data measured every 15 minutes from 01 January 2010, 01:15 to 20 February 2010, 23:45 (4,896 observations).
The goal is to forecast the next 96 values, corresponding to 21 February 2010 (one full day at 15-minute intervals).

You can take inspiration from an existing R notebook to get a clear idea of the steps, models, and structure to follow, then adapt it to your dates and dataset.

The forecasting must be done under two scenarios:

Without external covariates – models trained only on historical consumption.

With external covariates – models trained on historical consumption plus external explanatory variable temperature (aligned to the same timestamps).

Models to Apply
Scenario 1 – Without Covariates (Univariate time series)
Holt-Winters Exponential Smoothing – to model level, trend, and seasonality (additive or multiplicative).
SARIMA – Seasonal AutoRegressive Integrated Moving Average, both:
Automatic selection via auto.arima()
Manual order selection via ACF/PACF analysis.
NNAR – Neural Network Autoregression to capture nonlinear patterns in past consumption values.

Scenario 2 – With Covariates (Multivariate time series)
ARIMAX – ARIMA model extended with regressors (temperature) to explain part of the variance before modeling residual autocorrelation.
TSLM – Time Series Linear Model, including:
Temperature and temperature² (to capture nonlinear effects)
Trend and seasonality terms (trend, weekday/weekend, etc.).

Optional Model
RNN (Recurrent Neural Network), manage to obtained coherent results using a Deep Neural
Network in R (coherent means comparable to what you obtained with other models).

Data Preparation Steps
Data cleaning: Ensure no missing timestamps,
Time series conversion:
ts or msts object for single or multiple seasonalities (daily seasonality: 96 periods/day).

Train/Test split:
Train: 01/01/2010 → 20/02/2010 (4,896 points)
Test: 21/02/2010 (96 points)

Covariate alignment:
Synchronize temperature data with consumption timestamps.
Scale or normalize if necessary.
Exploratory Data Analysis
Plots: Trend, seasonality, decomposition.
ACF/PACF: Identify autocorrelation and partial autocorrelation patterns for SARIMA.
Seasonal plots: Visualize daily and weekly seasonal effects.
Summary stats: Mean, variance, stationarity checks (ADF test).
Modeling Approach

For each model:
Fit the model to the training set.
Generate forecasts for the test set

Evaluate accuracy with:
RMSE (Root Mean Squared Error)
MAE (Mean Absolute Error)
MAPE (Mean Absolute Percentage Error)

Residual diagnostics:
checkresiduals() to test for white noise.
Ljung-Box test for autocorrelation.
Expected Deliverables
Plots:
Forecasts vs actual test data.
Residuals diagnostics (ACF/PACF of residuals, histogram, QQ plot).
Tables:
Model parameters and fitted equations.
Accuracy metrics (RMSE, MAE, MAPE) for all models.
Discussion:
Which model performed best in each scenario.
How adding temperature impacted forecast accuracy.
Model limitations and potential improvements.
Evaluation Criteria
Correct application of models covered in the course.
Proper handling of covariates in the second scenario.
Sound methodology (train/test separation, residual analysis).
Clarity and quality of the final report (figures, explanations, comparisons).