Page 274 - 《高原气象》2025年第3期

高 原 气 象 44 卷
832
cast system for Beijing city centralized heat-supply energy-saving 究［J］. 计算机仿真， 39（9）： 507-512. Zhang T D， Li Q， Chen
system［J］. Gas & Heat， 28（11）， 23-25. B， 2022. Research on short-term heat load forecasting of thermal
张嘉益， 薛贵军， 2023. 基于混合神经网络的短期供热负荷预测模 power station based on LSTM［J］. Computer Simulation， 39（9）：
型研究［J］. 自动化仪表， 44（5）： 63-68. Zhang J Y， Xue G J， 507-512.
2023. Research on short-term heating load forecasting model 张文倩， 2019. 基于气象因素的集中供暖系统的热负荷预测研究［D］.
based on hybrid neural network［J］. Process Automation Instru‐ 保定： 华北电力大学 . Zhang W Q， 2019. Research on Heat load
mentation， 44（5）： 63-68. forecasting of centralized heating system based on meteorological
张腾达， 李琦， 陈波， 2022. 基于LSTM的热力站短期热负荷预测研 factors［D］. Baoding： North China Electric Power University.

Research on Heating Load Prediction Model based on the Influence
of Multiple Meteorological Elements using Deep Learning

MIAO Rui ， LI Mingcai ， SUN Meiling ， PAN Di ， ZHANG Xifan 3
3
1， 2
1， 2
1， 2
（1. Tianjin Key Laboratory of Marine Meteorology， Tianjin 300074， China；
2. Tianjin Institute of Meteorological Science， Tianjin 300074， China；
3. Tianjin Meteorological Service Center， Tianjin 300074， China）
Abstract： Accurate heating load forecasting is crucial for enhancing the efficiency of district heating systems and
improving indoor comfort in buildings. This study takes Tianjin， a major city in northern China， as a case study.
Based on hourly heating load and meteorological data from the heating season in 2021 -2022， the impacts of
comprehensive meteorological factors， such as temperature， wind speed， relative humidity， and solar radiation，
on heating load is analyzed. An efficient short-term heating load prediction model is constructed using the Nonlin‐
ear Autoregressive with Exogenous Inputs （NARX） neural network algorithm. The results show that the hourly
heating load that displays significant diurnal and monthly variations， has a notably negative correlation with tem‐
perature， weakly negative correlation with solar radiation， while the relationships with humidity and wind speed
vary depending on the season. Compared to the prediction model considering only temperature， the model incor‐
porating temperature， wind speed， relative humidity， and solar radiation together has better prediction perfor‐
mance， reducing the relative error by approximately 1. 4%. By comparing the forecast results with the LSTM
neural network prediction model， the NARX model significantly enhances prediction accuracy with decreasing
the relative error by about 3. 6%.
Key words： meteorological factors； heating load prediction； NARX neural network； centralized heating