Research On The Combined Forecast Model Of Urban Water Supply Based On ARIMA-LSTM

Water supply forecast can provide data for water supply system dispatching,and can also provide reference for water companies and customers to maintain the balance between supply and demand,so as to save energy consumption and improve the utilization rate of water resources.In order to further improve the prediction accuracy of urban water supply,this paper takes the measured data collected from two water plants of different scales in a city as the research object,and carries out the combined prediction model research of urban water supply based on ARIMA-LSTM from the two time steps of day and month.The relevant research results and conclusions are as follows:(1)Preprocessing the original water supply time series based on data pretreatment technology: Stabilizing the water supply time series and processing the missing values.Firstly,the original time series is decomposed according to the characteristics of the original time series.Secondly,the traditional difference algorithm and seasonal difference algorithm are used to stabilize the non-stationary part of the original time series.Finally,the interpolation method is used to further process the missing data.In this paper,a total of 4 sets of data(two groups of daily water supply and two groups of monthly water supply)are all from two water plants of different scales,among which the daily water supply number is 0# for No.1 Water Plant,1# for No.2 Water Plant,the monthly water supply number is 2# for No.1 Water Plant,and 3# for No.2Water Plant.Through the individual analysis of the water supply time series of the example,it is found that it has the characteristics of linear and nonlinear periodicity and large local fluctuation.Therefore,according to the characteristics of the water supply time series of each group in this case,ARIMA model and LSTM model are selected as the basic model of the combined model.(2)Research based on integrated autoregressive moving average(ARIMA)prediction model: as a traditional time series prediction model,ARIMA has good linear trend prediction effect,fast convergence speed and strong robustness,and it is also the most commonly used time series model at present.Therefore,the ARIMA model was used in this study to predict the time series of water supply in four groups.The results show that:(1)The ARIMA model can track the overall variation trend of water supply time series,but the local prediction effect has a large space for improvement.(2)The MSE values of the ARIMA model are 0#0.0299,1#0.0331,MAPE values are 0#0.0335,1#0.0414,respectively.The predicted MSE values of the monthly water supply were 2#0.0317 and 3#0.0286 respectively.MAPE values were2# 0.0320 and 3# 0.0401,respectively.(3)From the general trend,the prediction effect of daily water supply time series was worse than that of monthly water supply time series,and the MSE and MAPE values of daily water supply 0# and 1# were both higher than that of monthly water supply 2# and 3#.ARIMA model has poor prediction effect in the nonlinear part,which will further affect the overall prediction error.(3)Research on the prediction model based on LSTM;As a new prediction method,LSTM has the ability of long term memory pattern.Its advanced structure relies on feedback connection,and it can remember the characteristics of water supply time series.In this study,the theory and modeling steps of LSTM model were briefly analyzed,and the models were established to predict the four groups of water supply time series.The results show that:(1)Compared with ARIMA,LSTM model can predict the nonlinear trend of time series better and its overall fitting effect is also better.(2)Compared with ARIMA model,the prediction accuracy of LSTM model is improved to a certain extent.For the prediction of daily water supply,MSE value is0#0.0124,1#0.01731,MAPE value is 0#0.0118,1#0.0136,respectively.The predicted MSE values of the monthly water supply were 2#0.0194 and 3#0.0247,respectively.MAPE values were 2#0.0199 and 3#0.0202 respectively,indicating that both MSE and MAPE of LSTM model were smaller than those of ARIMA model in predicting the time series of daily or monthly water supply.(3)Because LSTM model can remember the characteristics of water supply time series in the long and short term,it can predict the period of water supply time series more accurately.However,the LSTM model cannot show all the characteristics of the data,and there may be some missing parts in the overall prediction.(4)Research on ARIMA-LSTM combination prediction model: In view of the problems of low accuracy and limited prediction range in a single prediction model,ARIMA model and LSTM model were selected as the basic models for series and parallel combination in this study.Firstly,the ARIMA model and the LSTM model are combined in series mode.The LSTM model is used for the error correction of the ARIMA model,that is,to predict the nonlinear part of the ARIMA model,and then the final prediction results of the ARIMA model are obtained.Secondly,the final prediction results of ARIMA model and LSTM model are combined in parallel by using fixed weights respectively to get the prediction results of ARIMA-LSTM series and parallel weighted combination model.Finally,the combined model was used to predict and verify the water supply of four groups of examples,and the prediction results of the combined model were compared with the single model(ARIMA model and LSTM model).The prediction results show that:(1)The series and parallel weighted combination prediction model can track the overall trend of four groups of water supply time series;(2)Compared with each single model,the ARIMA-LSTM series and parallel weighted combination prediction model greatly improves the prediction accuracy and reduces the prediction error.The MSE and MAPE of the ARIMA-LSTM combined model are about 0.06 and 0.09 higher than that of the ARIMA single model,and about 0.013 and 0.011 higher than that of the LSTM single model.(3)The ARIMA-LSTM parallel weighted combination prediction model combines the advantages of linear and nonlinear models on the basis of tracking the overall variation trend of four groups of water supply time series,and has better prediction effect when there are more data sets.(4)The combined model has better stability and shorter operation time,which proves that the combined model is feasible for the prediction of water supply time series.The ARIMA-LSTM parallel weighted combination prediction model proposed in this paper provides a new idea for urban water supply prediction and other related research.