Work

Interpretable Physics-Informed Neural Networks for energy consumption prediction using IoT sensors

Year: 2025

Type: article

Abstract: Inefficient energy use in households contributes to higher domestic cost, underscores the need for effective management and predictive maintenance. Conventional forecasting techniques often lack real-time performance and accuracy that limit their practical utility. Our study introduces a novel IoT based framework that integrates machine learning (ML) and Physics Informed Neural Networks (PINNs) and explainable artificial intelligence (XAI) to enhance energy consumption prediction and interpretability. We evaluated our approach against 18 different models. The proposed PINNs model achieved an R-Square of 0.9972, with a Mean Absolute Error (MAE) of 0.07 and Root Mean Square (RMSE) of 0.03. These results outperform traditional machine learning and deep learning models, which typically range between R-Square of 0.89 to 0.94. The system leverages non-invasive IoT sensors for data collection which further processed via a Flask-based web sever and mobile app. Moreover, we used XAI tools (SHAP, LIME) to explain our model’s predictions by showing the contribution of each feature. The final deployment on web and mobile platforms ensures user accessibility, engagement and secure remote access through ngrok. Our research demonstrates PINNs’ effectiveness in delivering accurate, interpretable predictions and supporting sustainable smart energy management for residential settings.

Source: Array

DOI: 10.1016/j.array.2025.100552

Authors: Sadia Hossain, Atiqur Rahman, Samsuddin Ahmed, Sadia Enam, Md. Toukir Ahmed

Institution: University of Frontier Technology, Bangladesh

Language: English

Topic: Smart Grid Energy Management

Subfield: Electrical and Electronic Engineering

Field: Engineering

Domain: Physical Sciences

SDGs: Affordable and clean energy, Industry, innovation and infrastructure

Open Access status: gold

Estimated APC paid: $1,680