Energy Demand Model Framework For Forecasting Hourly Electricity Consumption

Abstract

Accurate forecasting of electricity consumption is essential for electrical utilities to prevent blackouts, reduce wastage, and ensure efficient resource allocation. This thesis
develops and evaluates predictive models for short-term hourly electricity consumption
using five years of data from a United States energy utility company. The study employs
both traditional time series models, such as 𝑆𝐴𝑅𝐼𝑀𝐴, and advanced regression-based
models, including Gradient Boosted Regression (𝐺𝐵𝑅).
The methodology involves decomposing the original electricity demand time series
into monthly averages and hourly residuals. The monthly averages are modeled using a
𝑆𝐴𝑅𝐼𝑀𝐴 model, while the hourly residuals are addressed with three regression models:
linear regression, multi-layer perceptron (𝑀𝐿𝑃) regression, and 𝐺𝐵𝑅. Among these,
the 𝐺𝐵𝑅 model demonstrated the best performance, with an 81% improvement over
the baseline persistence model, achieving a 𝑀𝐴𝑃 𝐸 of 0.083.
To assess the robustness and generalizability of the models, cross-validation was conducted using data from Malta in the European Union. Despite regional differences, the
𝑆𝐴𝑅𝐼𝑀𝐴 and 𝐺𝐵𝑅 models maintained strong performance, with 𝑀𝐴𝑃 𝐸𝑠 of 0.08 and
0.088, respectively. These findings underscore the models’ versatility and their potential
for implementation across different regions.
A key aspect of this thesis is the implementation of the developed models in a webbased user interface using the Streamlit library. This interface allows users to visualize
and apply the models in real-time, upload their own energy consumption data, and
perform exploratory data analysis (𝐸𝐷𝐴) and forecasting seamlessly.
This thesis provides a reliable framework for short-term electricity consumption
forecasting, offering utilities valuable insights to optimize energy generation and distribution, thereby enhancing grid stability and operational efficiency.

Date of Award	2024
Original language	English (American)
Supervisor	Thomas Ziebermayr (Supervisor) & Armin Veichtlbauer (Supervisor)