What question did this study set out to answer?

The aim is to develop an optimized Multilayer Perceptron Neural Network for short-term forecasting of residential electricity consumption.

April 29, 2026Open Access

Optimized Multilayer Perceptron Neural Network Using Adam Optimizer and Huber Loss Function for Electricity Consumption Forecasting in Residential Buildings: A Case Study of Isfahan City

Key Points

The aim is to develop an optimized Multilayer Perceptron Neural Network for short-term forecasting of residential electricity consumption.
Data preprocessing using Min–Max Normalization.
MLP-NN trained with early stopping and optimized using Adam and SGD algorithms.
Feature importance analysis conducted to refine model inputs.
Achieved a test Huber loss of 0.002594, indicating superior accuracy over competing models.
Refined model used 6 features, outperforming the full 9 features MLP-NN.
Demonstrated better performance than Random Forest, SVR, XGBoost, and other machine learning models.

Abstract

Forecasting electricity consumption in residential buildings is a critical area in building energy management systems (BEMS). This study develops an optimized Multilayer Perceptron Neural Network (MLP-NN) for accurate short-term forecasting of residential electricity consumption in Isfahan City. Data were preprocessed using Min–Max Normalization (MMN). The model utilizes 9 input parameters: HVAC power consumption, lighting, household appliances, energy stored in a 5-kWh supercapacitor, ambient temperature, solar irradiance, hour of the day, holiday indicator, and seasons. The model predicts electricity consumption for the next 15-minute interval. Datasets comprised 15-minute interval records over one year from the Isfahan Electricity Distribution Company and ambient temperature data from the Iranian Meteorological Organization. The MLP-NN was trained with early stopping to prevent overfitting and optimized using Adaptive Moment Estimation (Adam) and Stochastic Gradient Descent (SGD) algorithms. Performance was compared with several machine learning models (Random Forest, SVR, XGBoost, Gradient Boosting, Extra Trees, and Linear Regression), with hyperparameters tuned by Random Search (10 iterations). Additionally, feature importance analysis using Chatterjee's Xi correlation coefficient, combined with physics-informed feature engineering, led to the selection of a refined set of 6 features. This reduced-dimensionality model, trained with the Adam optimizer and Huber loss function, achieved the best performance across all configurations on the test sets (test Huber loss = 0.002594), outperforming both the full 9 features MLP-NN and all benchmark machine learning models. This study demonstrates the superiority of the proposed MLP-NN, enhanced by temporal features, early stopping, and feature selection, for reliable short-term electricity consumption forecasting in residential buildings.

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper