.png){kind=icon}
.png){kind=icon}
Return
Comprehensive Energy Forecasting of Smart Homes: A Lightweight Multi-Task Neural Network via Sparse Neuron Reuse
.png){kind=icon}
PRE
AI
Z
Z
徐
DOI：10.1109/TSG.2026.3682527.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Smart homes integrate diverse energy sources, such as flexible appliances, photovoltaic generation, and electric vehicles, etc. Traditionally, the energy forecasting of these sources is through separate machine learning models, which require excessive computational resources at the edge. To address this challenge, this letter designs a lightweight neural network model for comprehensive energy forecasting. In the proposed method, we train a single neural network to sequentially learn multiple energy forecasting tasks by selectively reusing neurons with low activation across tasks. Neurons are organized into modular packages, restructuring the network into a directed acyclic graph (DAG) that supports task-wise specialization and neuron reuse. During sequential training, highly activated packages in current task are identified and frozen to preserve task-specific knowledge, while low-activation packages remain adaptable for subsequent tasks. For hardware deployment, only task-relevant packages are activated, offering a compact and efficient inference model. Simulation results demonstrate that, compared with standalone neural network models, the proposed model is significantly lightweight and easier to deploy in resource-constrained smart home applications. In addition, the model supports task-wise modular training and execution while maintaining satisfactory accuracy.
Keywords:
Sparse neuron reuse
multi-task deep neural network
smart home
directed acyclic graph
Journal
IF:
9.8
Papers:
5.6K
Citations:
4.3W
Organization
Cited Papers