1. Executive Summary & Data Audit

To provide a reliable business-ready forecast for the 2014 residential power load, the dataset was examined for quality and structural integrity. The raw data spans 192 months (January 1998 – December 2013).

The audit identified:

  • One missing value (NA) in September 2008
  • A significant data recording anomaly in July 2010 that required correction to prevent a skewed budget projection

2. Pre-processing & Imputation Techniques

Date Standardization

The character-based “YYYY-MMM” strings were converted into a formal month-year index to create a structured time series (tsibble).

Imputation by Interpolation

For the missing value in September 2008, a Time Series Linear Model (TSLM) was used.
This method estimates the missing point based on:

  • The long-term trend
  • The 12-month seasonal cycle

Rather than using a simple average.

Outlier Treatment (July 2010)

The data contained a massive crash in July 2010 (~770k KWH), which is 90% below typical summer demand. This was identified as a recording error.

To normalize the series:

  • The value was set to NA
  • It was re-imputed to a realistic summer peak (~8.3 Million KWH)

3. Exploratory Data Analysis (EDA)

Time Series Plot (Overall Trend)

The plot shows consistent growth in residential consumption until roughly 2008. After 2008, consumption stabilizes, likely due to improved energy efficiency standards.

Seasonal Plot (Monthly Rhythms)

Two distinct annual surges appear:

  • Primary peak: July/August (Air Conditioning usage)
  • Secondary peak: January/December (Heating demand)

ACF Plot (Statistical Evidence)

Repeating peaks at lags 12, 24, and 36 confirm a strong 12-month annual seasonal cycle.

4. Modeling Methodology

ARIMA (Auto-Regressive Integrated Moving Average)

Used to handle:

  • Complex correlations within monthly data
  • Non-stationary trends through differencing

ETS (Exponential Smoothing)

Used to:

  • Manage additive seasonal patterns
  • Weight recent history more heavily

Model Selection

The ARIMA model was selected due to a significantly lower AICc:

  • ARIMA: 5327
  • ETS: 6141

This indicates a more efficient fit for the energy grid.

5. Techniques Not Used

To maintain forecast quality, simpler methods were excluded:

  • Simple Mean: Rejected due to strong seasonality
  • Zero-Filling: Would falsely imply complete power failure
  • SNAIVE (Seasonal Naive): Cannot account for long-term trend changes

6. Final Forecast Interpretation (2014)

Cyclical Demand Management

Projected Total Annual Load (2014):

94,621,199 KWH

  • Winter Peak: January 2014 (~10.38M KWH)
  • Summer Peak: August 2014 (~9.93M KWH)

Shoulder Season Planning

Lower demand expected during:

  • April (~6M KWH)
  • October (~6M KWH)

These represent mild weather periods with minimal climate control usage.

7. Technical Appendix: Statistical Validation

Residual Diagnostics (Ljung-Box Test)

p-value = 0.67

Since p-value > 0.05, the residuals are considered White Noise, confirming the model captured all relevant patterns.

Stationarity Testing (KPSS Test)

p-value = 0.01

The low p-value confirms the raw data was non-stationary, justifying the use of ARIMA differencing.

8. Conclusion

By identifying and correcting the 2010 recording anomaly and applying a Seasonal ARIMA model, a robust 2014 forecast was generated.

Seasonal swings remain the primary driver of residential demand, requiring higher capacity reserves in January and August to maintain grid stability.