基于LSTM-ANFIS串联网络的建筑环境温湿度预测与自适应控制

Building Environment Temperature and Humidity Prediction and Adaptive Control Based on LSTM-ANFIS Cascaded Network

摘要: 建筑环境温湿度控制过程受非线性特性、时滞效应及多源扰动交互影响，传统单一建模手段在预测精度与控制响应速度方面往往难以达到理想效果。本文构建了一种融合长短期记忆网络(LSTM)与自适应神经模糊推理系统(ANFIS)的串联预测控制架构。在预测阶段，采用LSTM对温湿度时序演变进行建模，以捕捉历史数据中的长期依赖特征；在控制阶段，将LSTM输出与当前环境仿真数据一并作为ANFIS的输入，凭借其模糊推理与自学习机制完成温湿度协同调控。以2024年夏季典型高温高湿气象条件为应用场景开展仿真性能测试，结果表明：相较于Elman神经网络、单独使用ANFIS以及PSO-Elman方法，本文所提LSTM-ANFIS串联方案的温度预测MAE为0.31℃、RMSE为0.42℃，温湿度控制最大偏差分别下降62.3%和58.7%，稳态调节时长缩减41.2%，在预测准确度与控制动态性能方面均表现出显著优势。

Abstract: The process of controlling temperature and humidity in building environments is influenced by nonlinear characteristics, time-delay effects, and the interaction of multiple disturbances. Traditional single modeling methods often struggle to achieve ideal results in terms of prediction accuracy and control response speed. This paper constructs a tandem predictive control framework that integrates a Long Short-Term Memory network (LSTM) with an Adaptive Neuro-Fuzzy Inference System (ANFIS). In the prediction phase, LSTM is used to model the temporal evolution of temperature and humidity, capturing long-term dependency features in historical data; in the control phase, the LSTM output, together with current environmental measurements, is used as input to ANFIS, which achieves coordinated regulation of temperature and humidity through its fuzzy inference and self-learning mechanisms. Performance tests are conducted under typical high-temperature and high-humidity weather conditions in the summer of 2024. The results show that compared with the Elman neural network, standalone ANFIS, and PSO-Elman methods, the proposed LSTM-ANFIS tandem scheme achieves a temperature prediction MAE of 0.31˚C and RMSE of 0.42˚C, with maximum deviations in temperature and humidity control decreasing by 62.3% and 58.7%, and steady-state adjustment time reduced by 41.2%, demonstrating significant advantages in both prediction accuracy and dynamic control performance.