摘要
针对高层建筑电气系统因运况复杂多变导致的火灾预警难题,提出了一种融合长短期记忆(Long Short-Term Memory,LSTM)网络与Kolmogorov-Arnold网络(Kolmogorov-Arnold Network,KAN)的电气火灾风险预警方法。通过分析正常工况下电气线路负荷电流与环境温度变化规律,构建LSTM-KAN温度预测模型,计算预测值与实测值残差,并利用高斯核密度函数拟合残差值的概率分布确定预警阈值,最终采用异常情况下的数据进行验证。试验结果表明,与LSTM、双向长短期记忆(Bidirectional Long Short-Term Memory,BiLSTM)网络模型和单维温度输入的LSTM-KAN模型相比,该模型温度预测平均绝对误差降低至0.836℃,均方根误差为1.014℃,预测精度显著提升,且未出现误报情况,实现了电气火灾风险的有效预警。
This study aimed to reduce false alarms and delays in electrical fire risk early warnings for high-rise buildings.We proposed a hybrid model that combines Long Short-Term Memory(LSTM)network with Kolmogorov-Arnold Network(KAN).The model analyzes the relationships between three-phase load current and ambient temperature under normal operating conditions.First,to enhance the model,we replaced the fixed activation functions and linear layers in traditional LSTM with KAN's learnable spline-based functions.This modification allows for dynamic weight allocation,effectively capturing the complex nonlinear interactions between current and temperature.Next,we normalized the time-series data for current and temperature to prepare for model training.To establish adaptive warning thresholds,we utilized Gaussian kernel density estimation to statistically analyze the residuals between predicted and measured temperatures.The 99%confidence interval of the residual distribution was used to determine these thresholds.The experimental dataset comprised 13365 points collected during normal operation from monitoring devices located in a high-rise building's basement distribution cabinet in Anhui,China,covering the period from November 2023 to March 2024.Compared to baseline models—Bidirectional LSTM(BiLSTM),LSTM,and single-dimensional LSTM KAN—our LSTM KAN model achieved a Mean Absolute Error(MAE)of 0.836℃and a Root Mean Square Error(RMSE)of 1.014℃.These results represent improvements of 16.3%,31.8%,and 10.9%in MAE over BiLSTM,LSTM,and single-dimensional LSTM KAN,respectively.Residual fluctuations remained stable within±3℃,and the derived warning thresholds were 2.463℃(upper bound)and-2.517℃(lower bound).In abnormal scenarios,the model triggered warnings at 89.45 h,outperforming LSTM by 0.2 h and both BiLSTM and singledimensional LSTM KAN by 0.5 h.Notably,our method produced zero false alarms,while BiLSTM and single-dimensional LSTM KAN generated 6 false alarms,and LSTM resulted in 18 false alarms.The innovation of our approach lies in integrating LSTM's temporal feature extraction with KAN's dynamic weight allocation mechanism.This combination enhances prediction accuracy by effectively modeling both short-term fluctuations and long-term trends in the temperature-current relationships.Our method not only reduces false alarms but also facilitates the early detection of gradual electrical overheating.
作者
彭曙蓉
李元书
黄浩宇
唐程
王娜
苏盛
PENG Shurong;LI Yuanshu;HUANG Haoyu;TANG Cheng;WANG Na;SU Sheng(School of Electrical and Information Engineering,Changsha University of Science and Technology,Changsha 410114,China)
出处
《安全与环境学报》
北大核心
2026年第2期576-583,共8页
Journal of Safety and Environment
基金
国家自然科学智能电网联合基金项目(U1966207)
国家电网总部科技项目(5400-202320578A-3-2-ZN)。
