A hybrid modelling approach in the simulation of integrated urban wastewater systems
Systematic methods and tools for managing the complexity
Process Simulation & Optimization - III (T4-9c)
Keywords: Integrated modelling, ANN, WWTP, hybrid modelling
This work combines Artificial Neural Network (ANN) modelling and traditional (mechanistic) modelling of the urban wastewater system, in order to obtain a tool that allows for fast evaluations of the performance of such systems. Simulation of the integrated urban wastewater system is typically a computationally-demanding, and consequently a time-consuming task. Thus model-based evaluation of integrated urban water systems consists much of waiting for the simulation than actually processing and evaluating simulation data. In case of long-term simulations (= several years of data) and time-series analysis, this problem becomes even more accentuated.
The high computational demand of the traditional integrated urban wastewater models is a consequence of the high overall complexity of these models, and thus of the individual complexity of the submodels making up the integrated model. Indeed, a mechanistic model of the integrated urban wastewater system typically consists of submodels for the sewer system, the wastewater treatment plant (WWTP) and the receiving waters. Reduction of simulation times can be achieved by speeding up one or more of the submodels of the integrated system. In this paper the use of a fast neural network model instead of the mechanistic model of the WWTP is proposed for this purpose. The neural network is trained on a sequence of treatment plant input/output data generated by the mechanistic model of the WWTP. In other words, the trained ANN becomes a “fast copy”, a reduced model of the original WWTP model. This ANN is then coupled to a mechanistic river model – representing the receiving waters – and to an influent model. The influent model generates time series of typical WWTP influent flow rate and pollutant concentration disturbances. As a result of substituting the mechanistic WWTP model by the ANN, a reduction of the simulation time by a factor of 23 was achieved. The results presented in this paper show that the errors of the hybrid model are of an acceptable level when compared to the results of the purely mechanistic model, confirming the practical usefulness of the proposed method.
Presented Wednesday 19, 11:20 to 11:40, in session Process Simulation & Optimization - III (T4-9c).
