Abstract:Building load forecasting is essential for optimizing the architectural design and managing energy efficiently, enhancing the performance of Heating, Ventilation, and Air Conditioning systems, and enhancing occupant comfort. With advancements in data science and machine learning, the focus on predicting building loads through data analysis has significantly intensified as a research domain. However, previous studies have typically faced challenges such as data scarcity, improper feature extraction methods, and weak model generalization capabilities. To gain a deeper understanding of these issues, a comprehensive review of data processing, feature selection, and model selection methods in previous research is conducted from the perspective of the entire load forecasting process. The aim is to identify the most suitable methods for each step of load forecasting to enhance prediction accuracy. This review surveys the research progress of statistical learning methods, traditional machine learning methods, deep learning methods, and hybrid methods in different application scenarios of building load prediction. Then, it emphasized the critical role of data preprocessing and focused on techniques like data fusion and transfer learning to overcome data shortages and bolster the models’ ability to generalize. Moreover, the obtainment of significant features from building characteristics, weather data, and operational statistics to boost prediction accuracy is explored. A notable contribution of this review is the proposed technical framework for EnergyPlus model generation using LLM-based Retrieval Augmented Generation (RAG) technology and room- level load prediction with Spatio-Temporal Graph Neural Networks. This framework utilize architectural design drawings to achieve an “end-to-end” prediction process, aiming to reduce the professional threshold of load prediction and provide technical support for fine-grained regulation of building operation. Exploratory experiment is conducted using a single-zone building model to verify the feasibility of LLM-generated EnergyPlus models, with IDF simulation file generation taking only 196 s. Room-level load forecasting with LLMs remains to be explored further. It is reasonable to believe that the methods proposed in this review hold promise for advancing data-driven building load forecasting technologies.

A General Framework for Load Forecasting based on Pre-trained Large Language Model

LFLLM: A Large Language Model for Load Forecasting

Short-term Load Forecasting Based on Multi-model Fusion of CNN-LSTM-LGBM

A Method of Load Forecasting Based on Temporal Convolutional Network

Data-Driven Building Load Prediction and Large Language Models: Comprehensive Overview

A Unifying Framework of Attention-based Neural Load Forecasting

Deep Learning Based Densely Connected Network for Load Forecasting

TimeGPT in Load Forecasting: A Large Time Series Model Perspective

A Deep Learning Framework for Short-term Power Load Forecasting

Large Language Models for Forecasting and Anomaly Detection: A Systematic Literature Review

A New Multi-Scale Model for Medium-Term Load Forecasting

Application of LSTM-LightGBM Nonlinear Combined Model to Power Load Forecasting

Deep Ensemble Learning Model for Short-Term Load Forecasting Within Active Learning Framework

Deep LSTM and GAN Based Short-term Load Forecasting Method at the Zone Level.

Load Forecasting with Deep Learning: Critical Day Matters

MindLLM: Pre-training Lightweight Large Language Model from Scratch, Evaluations and Domain Applications

Short-Term Load Forecasting Based on Deep Learning Bidirectional LSTM Neural Network

Multivariate Load Forecasting of Integrated Energy System Based on GBLA

Power Load Forecasting Using BiLSTM-Attention