Research pick: The digital lineman - "A novel residential electricity load prediction algorithm based on hybrid seasonal decomposition and deep learning models"

A new data-driven model could improve the accuracy of residential electricity demand forecasting on a daily, or even hourly basis. The model, described in the International Journal of Energy Technology and Policy, might help utility companies better manage power grids as they become more reliant on variable renewable energy sources.

Shan Gao, Xinran Zhang, Lihong Gao, and Yancong Zhou of Tianjin University of Commerce, China, suggest that their research has addressed one of the main problems in the transition from conventional to sustainable electricity production supplying diverse residential settings.

Renewable electricity sources, such as solar and wind, are inherently intermittent. They do not produce a steady electricity supply throughout the year and moreover can drop to zero supply at any time during the day depending on weather conditions. And, of course, solar only works during daylight hours. Such variability makes it difficult for electricity grid operators to ensure that supply meets demand at all times. Accurate forecasting of electricity usage, particularly in residential settings where patterns are highly individual and sensitive to social routines, is now seen as essential for the stability and efficiency of so-called smart grids.

The researchers have developed a hybrid model that is less error-prone than other models. The new approach combines a Convolutional Neural network, which is particularly effective at identifying short-term patterns in datasets, with a Long Short-Term Memory network, a type of recurrent neural network well-suited to tracking longer-term dependencies in time series data.

Such hybrid models have been used successfully in other settings previously, but the team has also incorporated an attention mechanism into theirs. This is a tool borrowed from natural language processing that allows the system to prioritize the most relevant parts of its input data when making predictions. This dynamic filtering process improves the model’s ability to respond to variations in household behaviour or external conditions. In addition, in order to account for systematic fluctuations in consumption, the model uses seasonal decomposition, a statistical technique that isolates regular seasonal trends, such as increased winter heating demand or reduced usage during the summer, from the overall dataset.

All of these tools combined allow the new model to recognize and adapt to subtle patterns in electricity demand. For example, it can distinguish between weekday and weekend routines, anticipating delayed morning energy use on Saturdays and Sundays, or adjusting for increased evening usage during colder months. The team has thus demonstrated a mean absolute percentage error as low as 0.76% for daily predictions and just 2.36% for hourly ones. These figures improve on existing models, the team suggests.

Gao, S., Zhang, X., Gao, L. and Zhou, Y. (2025) ‘A novel residential electricity load prediction algorithm based on hybrid seasonal decomposition and deep learning models’, Int. J. Energy Technology and Policy, Vol. 20, No. 5, pp.1–23.