基于LSTM神经网络的地铁变压器绕组温度预测

传统的地铁状态监测系统仅能反映变压器绕组当前的温度状态及其历史温度趋势，当绕组温度超过阈值时系统报警，但不能对绕组未来的温度变化进行预测。绕组温度受设备运行功率和环境温度等多重因素影响，其变化呈现非线性和周期性，传统预测方法精度难以提升。本文基于长短期记忆网络(long short-term memory,LSTM)算法预测变压器绕组温度，选取绕组温度、环境温度、运行功率、运行电流作为输入变量，收集变压器历史状态数据构成训练数据进行离线训练，通过训练完成的绕组温度预测模型反映多重影响因素与绕组温度的变化关系。最后将算法应用于某地铁站动力变压器，收集样本数据进行训练得到温度预测模型，将测试数据输入模型中，计算绕组温度真实值和预测值之间的相对温差，分析验证算法可行性与模型准确度。结果表明：LSTM算法面对大数据量样本可充分挖掘多重影响因素与绕组温度之间的深层关系，温度预测模型可准确预测绕组温度的变化。

Temperature Prediction of Subway Transformer Windings Based on LSTM

Traditional subway condition monitoring systems can only reflect the current temperature state of a transformer winding and its historical temperature trend. When the winding temperature exceeds the threshold, the system provides an alarm. However, the system cannot predict future temperature changes in the transformer or changes in the equipment operating state. The winding temperature is affected by multiple factors, such as operating power and ambient temperature, and its change is nonlinear and periodic. However, improving the accuracy of traditional prediction methods is challenging. With the development of deep-learning technology, the transformer winding temperature has been predicted using a long short-term memory (LSTM) algorithm. The winding temperature, ambient temperature, operating power, and operating current were selected as the input variables, and the historical state data of the transformer were collected to form a large amount of training data for offline training. The transformer temperature prediction model reflects the relationship between multiple influencing factors and winding temperature. Finally, the algorithm was applied to a subway power transformer. Sample data were collected for training to obtain the temperature prediction model. The test data were input into the trained model. The relative difference between the real and predicted values of the winding temperature was calculated to analyze and verify the feasibility and accuracy of the algorithm. The results showed that the LSTM algorithm can fully mine the relationship between multiple influencing factors and winding temperature in the face of large data samples. The temperature prediction model can accurately predict changes in the winding temperature.