IoT-Driven Energy Consumption Prediction in a Mexican Residence: A Case Study Utilizing Deep Learning with Attention Mechanism
Abstract
The energy consumption patterns exhibited by individuals carry significant implications for both the environment and energy production. Leveraging historical data, inferential models can foresee future consumption trends, assisting energy providers in planning adequate energy generation while encouraging shifts towards more environmentally sustainable practices. This, in turn, fosters a transition towards more eco-friendly behaviors. This paper outlines the development of a comprehensive system based on IoT that gathers real-world energy consumption data from a household in northeast Mexico, employs deep learning models for prediction, and incorporates a visualization tool to present energy demand.
For the prediction, it was conducted a comparative analysis on three advanced deep learning models tailored for sequential data: LSTM, GRU, and Seq2Seq. Additionally, we explored the impact of enhancing each model with an Attention mechanism. Our findings consistently demonstrate that the incorporation of an Attention layer improves model performance, leading to a reduction in error metrics across the board. Specifically, we achieved an average Mean Absolute Percentage Error of 8.83% for daily predictions and 30.44% for hourly forecasts. These results underscore the efficacy of our selected models in accurately predicting energy consumption patterns, marking a notable stride towards informed and sustainable energy management.
For the prediction, it was conducted a comparative analysis on three advanced deep learning models tailored for sequential data: LSTM, GRU, and Seq2Seq. Additionally, we explored the impact of enhancing each model with an Attention mechanism. Our findings consistently demonstrate that the incorporation of an Attention layer improves model performance, leading to a reduction in error metrics across the board. Specifically, we achieved an average Mean Absolute Percentage Error of 8.83% for daily predictions and 30.44% for hourly forecasts. These results underscore the efficacy of our selected models in accurately predicting energy consumption patterns, marking a notable stride towards informed and sustainable energy management.
Keywords
Energy consumption prediction; Deep learning models; Recurrent Neural Networks; Attention mechanism