A Multi-Branch Deep Feature Fusion Framework Utilizing TCN and ANN for Short-Term Building Load Forecasting Under Pre-Scheduled Constraints

Accurate short-term load forecasting (STLF) is essential for the efficient operation of Building Energy Management Systems (BEMS) and localized microgrid scheduling. However, building load behavior is highly volatile due to the nonlinear interactions among scheduled load baselines, temporal patterns, and dynamic weather conditions. Traditional forecasting models often rely on a single-stream feature representation, which may distort training signals and increase the risk of overfitting, particularly when only limited historical data are available. This paper proposes a multi-branch deep feature fusion TCN-ANN hybrid framework for hourly building energy forecasting using a single year of historical operational data (8,760 samples). The proposed architecture processes diverse input features through separate learning pathways. A deep one-dimensional causal dilated Temporal Convolutional Network (TCN) branch extracts sequential representations from temperature, relative humidity, and 24-hour load lag trajectories, while a parallel multi-layer Artificial Neural Network (ANN) branch learns tabular relationships from scheduled load profiles and temporal variables. The resulting feature representations are fused through channel-wise feature combination and subsequently processed by a shared dense fusion layer. The effectiveness of the proposed framework was evaluated using a real-world building dataset with a strict chronological 70% training, 15% validation, 15% testing split and uniform Min–Max normalization. The model achieved an RMSE of 20.03 kW and a coefficient of determination (R²) of 0.9721, demonstrating excellent forecasting performance. Furthermore, the proposed framework outperformed the standalone TCN benchmark RMSE by 25.75 kW, indicating that deep feature fusion effectively captures complex interactions among temporal, environmental, and operational variables while reducing the risk of overfitting in small-scale time-series forecasting applications.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11032768/v1
Primary Topic
Energy Load and Power Forecasting
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Multi-Branch Deep Feature Fusion Framework Utilizing TCN and ANN for Short-Term Building Load Forecasting Under Pre-Scheduled Constraints

Husam S. Samkari, Alvin John Meng Siang Lim, Siti Nur Afifah Mohd Suhaimi, Nor Aira Zambri et al.
Research Square
Energy Load and Power Forecasting
preprint

A Multi-Branch Deep Feature Fusion Framework Utilizing TCN and ANN for Short-Term Building Load Forecasting Under Pre-Scheduled Constraints

Husam S. Samkari, Alvin John Meng Siang Lim, Siti Nur Afifah Mohd Suhaimi, Nor Aira Zambri, SY YI SIM, Muhamad Azfar Azri Azhari
preprint en

Abstract

Accurate short-term load forecasting (STLF) is essential for the efficient operation of Building Energy Management Systems (BEMS) and localized microgrid scheduling. However, building load behavior is highly volatile due to the nonlinear interactions among scheduled load baselines, temporal patterns, and dynamic weather conditions. Traditional forecasting models often rely on a single-stream feature representation, which may distort training signals and increase the risk of overfitting, particularly when only limited historical data are available. This paper proposes a multi-branch deep feature fusion TCN-ANN hybrid framework for hourly building energy forecasting using a single year of historical operational data (8,760 samples). The proposed architecture processes diverse input features through separate learning pathways. A deep one-dimensional causal dilated Temporal Convolutional Network (TCN) branch extracts sequential representations from temperature, relative humidity, and 24-hour load lag trajectories, while a parallel multi-layer Artificial Neural Network (ANN) branch learns tabular relationships from scheduled load profiles and temporal variables. The resulting feature representations are fused through channel-wise feature combination and subsequently processed by a shared dense fusion layer. The effectiveness of the proposed framework was evaluated using a real-world building dataset with a strict chronological 70% training, 15% validation, 15% testing split and uniform Min–Max normalization. The model achieved an RMSE of 20.03 kW and a coefficient of determination (R²) of 0.9721, demonstrating excellent forecasting performance. Furthermore, the proposed framework outperformed the standalone TCN benchmark RMSE by 25.75 kW, indicating that deep feature fusion effectively captures complex interactions among temporal, environmental, and operational variables while reducing the risk of overfitting in small-scale time-series forecasting applications.

Research Square
University of Tabuk (SA), Tun Hussein Onn University of Malaysia (MY)
Energy Load and Power Forecasting
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon