Sewage Volume Forecasting on a Day-Ahead Basis – Analysis of Input Variables Uncertainty

Abstract

Water consumption and resulting sewage volume (both strongly impacted by meteorological parameters) are of key importance for an efficient and sustainable operation of waterworks and sewage treatment plants. Therefore, the objective of this research is to analyze the potential impact of input variables uncertainty on the performance of sewage volume forecasting model. The research is based on a real, three-years long, daily time series collected from Torun (Poland). The used time series encompassed: sewage volume, water consumption, rainfall, temperature, precipitation, evaporation, sunshine duration and precipitation at a six hours interval. As a forecasting tool a multi-layer perceptron artificial neural network has been selected. First a simulation model for sewage volume was created which considered above mentioned earlier time series as exogenous variables. Further its performance was tested assuming that all non-historical input variables are prone to their individual forecasting errors. Analysis was dedicated firstly to each variable individually and later the potential of all of them being uncertain was tested. A lack of correlation between input variables error was assumed. The research provides an interesting solution for visualizing the quality and actual performance of forecasting models where some or all of input variables has to be forecasted.

Keywords

• artificial neural network
• error forecasting
• exogenous variable uncertainty