Improved flush detection and classification in combined sewer monitoring

During rain events, rainwater reaches the combined sewer system and causes additional hydraulic and pollutant load. Due to the limited capacity of the sewer system and the wastewater treatment plant, overflow structures are constructed to reduce the discharge and thus create a potential hazard for the environment. For optimal management of these structures, it is necessary to know the runoff and pollutant load of the events and their distribution over time. When these distributions have a significant peak, they are often referred to as a flush, the best-known phenomenon being the first flush at the beginning of a rainfall event. This knowledge can be used for the design of retention facilities and the calibration of sewer models. The flush phenomena are mainly caused by the erosion of contaminants on the surface as well as the remobilisation of sediments in the sewer network.

Although many papers have investigated the first flush, no common pattern for the occurrence of these flushes has been identified. While the concentration of the flushes in rainwater sewers can be measured directly, the rain flushes in combined sewers are mixed with more polluted wastewater, which leads to a reduction in signal strength.

The sensor site for the used measurement data is located in a combined sewer overflow in the western part of Graz, Austria with a catchment area of 460 ha, consisting mainly of residential areas and with about 19500 inhabitants.

This work aims to separate and classify pollution flush signals from rainfall events in combined sewer systems to better understand the relationship between these signals and rainfall event characteristics.

For this reason, the continuous hydraulic and pollution data are first analysed to determine the representative dry weather contribution. By subtracting the dry weather contribution from the combined wastewater volume and the mass flux, the stormwater contribution and thus the flushes can be estimated. In addition, automatic event detection of combined sewer events was done.

Next, the wet weather events are classified by clustering the stormwater runoff-induced pollutant distribution (flush signals) and the event parameters. For the clustering of the temporal pollutant load distribution of events of different duration, the events are normalised by the mass-volume curves. To obtain the best possible clustering result, the dimension of the mass-volume curves is reduced by a principal component analysis. Different clustering methods, such as partitioning or hierarchical methods, are applied and compared.