Predicting Electricity Usage Based on Weather and Usage History
Nair, Renji (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025121838404
https://urn.fi/URN:NBN:fi:amk-2025121838404
Tiivistelmä
Electricity consumption forecasting is a key component of modern energy management, particularly for electricity providers aiming to optimize grid operations and design customer-specific contracts. At the household level, electricity demand is highly variable and influenced by historical usage patterns, weather conditions, and household characteristics, making accurate forecasting challenging.
The objective of this research was to develop a data-driven framework for forecasting household-level electricity consumption using historical usage data and weather temperature information. In addition, households were segmented into distinct consumption profiles to improve model interpretability and support targeted energy analysis.
Real-world electricity consumption data from multiple Finnish households was combined with external weather data. A comprehensive preprocessing pipeline was implemented, including data cleaning, resolution unification, and feature engineering. Several forecasting approaches were evaluated, including traditional time-series models, tree-based machine learning models, and a hybrid deep learning architecture combining convolutional, recurrent, and attention mechanisms (CNN-LSTM-Attention). Household segmentation was performed using clustering based on consumption behavior and metadata attributes.
Model evaluation was conducted using MAE and RMSE as the primary accuracy measures. SMAPE was additionally reported to support relative error interpretation, particularly across regions and load regimes, but was not used as the sole performance indicator due to its sensitivity to low-load values. The results demonstrate that segmentation-based forecasting improves interpretability and supports more reliable household-level demand analysis, while deep learning models show potential in capturing temporal consumption patterns.
The objective of this research was to develop a data-driven framework for forecasting household-level electricity consumption using historical usage data and weather temperature information. In addition, households were segmented into distinct consumption profiles to improve model interpretability and support targeted energy analysis.
Real-world electricity consumption data from multiple Finnish households was combined with external weather data. A comprehensive preprocessing pipeline was implemented, including data cleaning, resolution unification, and feature engineering. Several forecasting approaches were evaluated, including traditional time-series models, tree-based machine learning models, and a hybrid deep learning architecture combining convolutional, recurrent, and attention mechanisms (CNN-LSTM-Attention). Household segmentation was performed using clustering based on consumption behavior and metadata attributes.
Model evaluation was conducted using MAE and RMSE as the primary accuracy measures. SMAPE was additionally reported to support relative error interpretation, particularly across regions and load regimes, but was not used as the sole performance indicator due to its sensitivity to low-load values. The results demonstrate that segmentation-based forecasting improves interpretability and supports more reliable household-level demand analysis, while deep learning models show potential in capturing temporal consumption patterns.