Using Artificial Neural Networks to Forecast Wet Weather Flow in a Sanitary Sewer System

The Milwaukee Metropolitan Sewerage District (MMSD) uses wet weather flow forecasts to aid in the effective operation of their wastewater conveyance and treatment system. Historically, these forecasts have been based on predicted rainfall and antecedent moisture conditions. However, after seven years of operation it became apparent that poor flow forecasts were inadequate for effective system management. A study, started in 2001, attempted to improve forecasts through the use of an Artificial Neural Network (ANN). Two options are available for constructing an ANN for the system, using historical system data or relying on data generated using a computer simulation model of the collection system. This paper describes the development and testing of several ANN models for forecasting flows in the MMSD conveyance system. The effectiveness of the selected model in reducing the frequency and volume of system overflows is also demonstrated. An ANN model was first developed based on measured water levels and rain gauges in the MMSD system. This model has been in use since early 2004. MMSD has recently completed construction of an 88 million-gallon storage and relief sewer. This storage and relief sewer was constructed on the northwest side of the MMSD service area in a separate sewer area. To ensure that the effect of this newly constructed storage facility is included in the forecasts, the ANN model needed to be updated to include this system. Due to the recent construction completion date, measured system water levels are not available with this structure in place. The MMSD system comprehensive Conveyance System Model (CSM) simulation results were utilized in place of measured water levels. The CSM was updated to include the new storage facility. The updated CSM was then used to simulate numerous storm events to create simulated water level data. The resulting set of simulated water level data was used to train the new ANN model. A combination of water levels from three locations within the conveyance system and nine rain gauges was found to effectively predict separate sewer inflows to a central inline storage system six hours in the future. Currently, online testing is being performed to analyze the effectiveness of this ANN model in a real-time online environment.