基于高低频循环神经网络的电气火灾风险预测系统
Electrical Fire Risk Prediction System Based on High and Low Frequency Recurrent Neural Networks

在当今电气系统和设备日益普及的背景下，电器故障和老化等因素引发的火灾事故频繁发生，严重威胁着人们的生命安全和财产。现有的火灾预警方案多数依赖于电气参数与固定阈值的比较，存在响应速度慢、准确性不足等问题，无法有效应对复杂的电气故障情况。为了解决这种问题，提出一种创新的电气火灾预警系统，基于长短期记忆网络(LSTM)技术，结合高频电气参数循环神经网络(HF-LSTM)和低频电气参数循环神经网络(LF-LSTM)进行研究。HF-LSTM深入挖掘线路的温升规律和超温故障特性，而LF-LSTM则用于探索线路温度变化的周期性模式。通过这两种模型的结合，使系统能够精确预测线路温度，实现对电气火灾风险的早期识别和预警。该系统突破了传统模式只依赖某几个参量的数据特征对电气火灾危险性进行计算和研判，忽略了参量间的物理关联，本文采用基于LSTM的动态阈值调整机制，增强了时间序列信息的连续性和相关性，从而提高了预警准确性和响应速度。系统还引入了预警分位的概念，实现了火灾风险的定量评估和分级管理。硬件电路实时采集电流、电压和温度信息，并与物联网平台结合，实现实时监控和自动响应。通过先进算法，系统提高了对微弱信号的识别能力，确保了早期风险感知和预防。实验数据表明，该电气火灾预警系统在预测准确性和响应速度上均显著优于现有方案，能够有效降低火灾发生率，为保障生命和财产安全提供了高效可靠的解决方案。
In the context of the increasing prevalence of electrical systems and devices, fire incidents caused by electrical faults and aging factors are occurring frequently, posing serious threats to people’s lives and property. Most existing fire warning systems rely on comparing electrical parameters with fixed thresholds, which suffer from slow response times and insufficient accuracy, making it difficult to effectively address complex electrical fault situations. To tackle this issue, an innovative electrical fire warning system is proposed, based on Long Short-Term Memory (LSTM) network technology, combining High-Frequency Electrical Parameter Recurrent Neural Network (HF-LSTM) and Low-Frequency Electrical Parameter Recurrent Neural Network (LF-LSTM) for research. HF-LSTM delves into the heating patterns of circuits and the characteristics of overheating faults, while LF-LSTM explores the periodic patterns of temperature changes in circuits. By integrating these two models, the system can accurately predict circuit temperatures, enabling early identification and warning of electrical fire risks. The system breaks through the traditional mode of relying only on the data characteristics of a few parameters to calculate and judge the electrical fire danger, ignoring the physical correlation between the parameters, and this paper adopts the dynamic threshold adjustment mechanism based on LSTM, which enhances the continuity and correlation of the time-series information and thus improves the accuracy and response speed of the early warning. The system also introduces the concept of warning quantiles, allowing for quantitative assessment and graded management of fire risks. The hardware circuit collects current, voltage, and temperature information in real-time, integrating with an Internet of Things (IoT) platform to