基于多维动态隶属度的模糊时间序列的天然气月负荷预测

Natural Gas Monthly Load Forecasting Based on Fuzzy Time Series of Multi-dimensional Dynamic Membership

准确的天然气负荷预测对于完善城市燃气供需系统与提高能源利用效率都大有裨益。由于燃气负荷序列受到多种不确定因素的影响,为了捕捉月度负荷的模糊性和非线性等复杂特征,本文结合模糊理论和长短时记忆神经网络(LSTM)的特性,提出了一种基于多维动态隶属度的模糊时间序列的预测新方法。首先,利用模糊C均值聚类(FCM)从原始数据中构建多维隶属度序列;其次,利用LSTM对多维隶属度序列同时进行预测,得到其动态隶属度;最后,去模糊化得到燃气负荷的预测值。应用该模型对四川成都某地区的天然气月度负荷进行了未来三个月的预测,并与经典模糊时间序列(FTS)、ARIMA模型、BP神经网络(BPNN)、LSTM等模型进行对比。实验结果表明,新模型的MAE、RMSE以及MAPE均优于其他模型。因此,本文提出的模型可对城市燃气供给和调度提供有价值的参考。

Accurate gas load forecasting is beneficial for both improving the city gas supply and demand system and improving energy use efficiency.Since gas load series are affected by various uncertainties,in order to capture the complex characteristics of monthly load such as fuzziness and nonlinearity,this paper combines the characteristics of fuzzy theory and long and short term memory neural network(LSTM)to propose a new method for prediction of fuzzy time series based on multidimensional dynamic affiliation.First,a multidimensional affiliation sequence is constructed from the original data using fuzzy C-mean clustering(FCM);second,the LSTM is used to predict the multidimensional affiliation sequence simultaneously to obtain its dynamic affiliation;finally,the predicted value of gas load is obtained by defuzzification.The model is applied to forecast the monthly gas load of a region in Chengdu,Sichuan for the next three months and compared with the classical fuzzy time series(FTS),ARIMA model,BP neural network(BPNN),and LSTM.The experimental results show that the MAE,RMSE,and MAPE of the new model are better than the other models.Therefore,the model proposed in this paper can provide a valuable reference for city gas supply and dispatch.