基于LSTM-NeuralProphet模型的城市需水预测方法研究

Research Urban Water Demand Forecasting Method Based on LSTM-Neural Prophet Model

城市水资源规划和管理是确保城市可持续发展和居民生活基本需求得到满足的关键环节,城市短期需水预测是城市水资源规划和管理的基础。由于气温、降水量和蒸发量等随季节变化明显,直接影响不同季节的用水峰值、高峰期,导致传统基于时间序列算法的固定时隙预测无法适应时隙的变化,从而不能保证预测精度。针对固定时隙预测精度低的问题,研究了基于四季24 h时间分辨率和夏季15 min时间分辨率的双时间尺度城市短期需水预测模型。该模型使用Anomaly-Transformer模型进行异常值检测,并通过分段曲线拟合对异常值校正,采用主成分分析法对城市短期需水影响因子进行分析提取主成分,在AutoML的标准模型分析中选取三个效果最好的模型作为Stacking模型的基学习器再结合长短期记忆网络(Long Short-Term Memory,LSTM)和Optune框架超参数优化后的NeuralProphet模型对双时间尺度的城市短期需水量进行预测,同时加入安全网机制,以保证LSTM-NeuralProphet模型的精确度。与其他模型(LSTM模型、NeuralProphet模型、BP神经网络模型)相比,LSTM-NeuralProphet模型的平均绝对误差在四季24 h时间分辨率的数据集上降低了0.18%~1.96%,在夏季15 min时间分辨率的数据集上降低了0.45%~11.90%。实验结果表明,LSTM-NeuralProphet模型具有更好的拟合效果和更高的预测精度,能较准确地预测双时间尺度下的城市需水量,可以较好地应用于城市短期需水预测研究中。

Urban water resources planning and management is a key link to ensure sustainable urban development and satisfy the basic needs of residents.Short-term water demand forecasting is the basis of urban water resources planning and management.As temperature,precipita⁃tion and evaporation vary significantly with the seasons,they directly affect the peak of water consumption in different seasons.As a result,the traditional fixed time slot prediction based on time series algorithm cannot adapt to the change of time slot,so the prediction accuracy cannot be guaranteed.Aiming at the problem of low accuracy of fixed time slot prediction,this paper studies a dual-time scale urban shortterm water demand prediction model based on 24 h time resolution in four seasons and 15 min time resolution in summer.This model uses Anomaly-Transformer model to detect outliers and calibrate outliers by piecewise curve fitting.Principal component analysis is used to ana⁃lyze and extract the main components of urban short-term water demand factors.In the standard model analysis of AutoML,three models with the best effects are selected as the base learner of the stacking model combined with the long-term and short-term memory network(LSTM,long-term and short-term memory)and the neural prophet model(NP)after hyperparametric optimization of Optune framework predict the short-term water demand of the city on both time scales,and add the safety net mechanism to ensure the accuracy of the LSTMNeuralProphet model.Compared with other models(LSTM model,Neural Prophet model,BP neural network model),the average absolute error of LSTM-Neural Prophet model is reduced by 0.18%~1.96%on the data set with 24 h time resolution of four seasons.In the summer 15 min time resolution data set,the reduction is 0.45%~11.90%.The experimental results show that the LSTM-Neural Prophet model has a bet⁃ter fitting effect and higher prediction accuracy,and can accurately predict the urban water demand at both time scales,and can be applied to the short-term urban water demand prediction research.