Deep learning model sharpens short-term electricity forecasts for individual homes

Forecasting the electricity use of a single household is one of the hardest problems in smart-grid research, and a new study in Neural Computing and Applications reports a deep learning architecture designed specifically for that task. The model, called TriLSTMWavenet, combines a three-level wavelet decomposition with a convolutional neural network and three layers of long short-term memory (LSTM) units, and the authors report an RMSE of 0.09 when predicting individual household electricity loads.

The work, authored by R. Mathumitha, P. Rathika and K. Manimala, appears in Volume 38 of the journal as an original research article. Its motivation lies in the transformation of power grids: as renewable energy sources grow, grids are becoming more flexible and more reliant on real-time data to optimise performance, a shift enabled by advanced communication technologies that allow the grid to operate as a cooperative system.

Smart meters are a central part of that infrastructure. They provide real-time consumption data that can support accurate load predictions, and as grids depend increasingly on such data, the authors argue that accurately predicting short-term electricity demand for individual homes will become crucial for grid operation and future design.

Individual household forecasting is markedly more difficult than community-wide forecasting.

Machine learning and deep learning offer a powerful approach to predicting electricity demand, but the authors emphasise that further development is needed to ensure their accuracy. TriLSTMWavenet is their contribution to that development: an innovative deep neural network architecture built to handle the noisy, volatile signals typical of household-level data.

Technically, the model integrates a 3-level wavelet decomposition with a CNN and three LSTM layers for robust feature extraction and accurate prediction. The wavelet stage breaks the input signal into multiple frequency components, and the architecture is designed to efficiently capture and integrate information from these multi-level decomposed inputs, combining them before the final prediction.

According to the authors, this design enhances the model’s ability to analyse and predict complex patterns and fluctuations in energy consumption over time.

The model’s efficacy was thoroughly assessed against the most advanced models currently in use for predicting a household’s short-term electricity consumption, with evaluation carried out using standard error metrics. Comparing the proposed model to other competing methods, the results were noticeably superior, and the reported RMSE of 0.09 indicates a significant reduction in errors when predicting individual household electricity loads.

The study’s scope is household-level short-term load forecasting, and its reported performance figures come from the authors’ own evaluation described in the abstract.

As power grids become more reliant on real-time data, the authors argue that accurately predicting short-term electricity demand for individual homes will be crucial for grid operation and future design. TriLSTMWavenet represents one attempt to meet that demand, and the full article, published on 23 October 2026, is available via subscription access from Springer.

Subject of Research: A deep learning architecture for short-term electricity load forecasting in individual households.

Article Title: TriLSTMWavenet: An innovative individual household load forecasting with an integrated wavelet and CNN-based deep learning fusion

Article References: Mathumitha, R., Rathika, P., & Manimala, K. (2026). TriLSTMWavenet: An innovative individual household load forecasting with an integrated wavelet and CNN-based deep learning fusion. Neural Computing and Applications, 38(20), Article 831. https://doi.org/10.1007/s00521-026-12633-8

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12633-8

Keywords: load forecasting, deep learning, smart grids, smart meters, LSTM, wavelet decomposition, CNN, renewable energy, household electricity, Neural Computing and Applications, TriLSTMWavenet, innovative