Keywords
Time series modelling, water demand, predictive models, Narx networks, UKf
Start Date
27-6-2018 10:40 AM
End Date
27-6-2018 12:00 PM
Abstract
Optimal management operations in water distribution systems (WDSs) improve water utility performance by reducing costs and resource waste. Obviously, the more accurate the water demand forecasting model, the better the operation. Classical artificial neural networks (ANNs) use external variables (e.g. weather and social variables) as inputs for their water demand forecasting processes. Recurrent neural networks (RNNs) use ANN architectures of several layers, and use the output of these layers as inputs for the previous layers, thus generating a closed loop for the regression process, what typically outperforms ANN performance. This work aims to generate even better results than RNNs by additionally including the time-series error in the process. The proposal is to develop a recursive forecasting approach, applying a nonlinear autoregressive model with exogenous networks (NARX), which forecasts water demand by also processing external inputs such as temperature, rain, weekday, and hour of the day. The water demand forecasting model is then updated by processing its associated error through an unscented Kalman filter (UKF), which is specifically useful for near real-time processes. This derives from UKFs’ ability to quickly synthesize highly nonlinear spaces. The proposed method is applied to three water utility district metered areas of a medium-size city. The results show significant performance improvements when compared to classical non-recursive methods (e.g. ANN and support vector regression). The UKF-updated NARX has, thus, huge potential to provide key benefits for WDSs operation and management in near real-time.
Updating recurrent neural networks for near real-time water demand predictive models
Optimal management operations in water distribution systems (WDSs) improve water utility performance by reducing costs and resource waste. Obviously, the more accurate the water demand forecasting model, the better the operation. Classical artificial neural networks (ANNs) use external variables (e.g. weather and social variables) as inputs for their water demand forecasting processes. Recurrent neural networks (RNNs) use ANN architectures of several layers, and use the output of these layers as inputs for the previous layers, thus generating a closed loop for the regression process, what typically outperforms ANN performance. This work aims to generate even better results than RNNs by additionally including the time-series error in the process. The proposal is to develop a recursive forecasting approach, applying a nonlinear autoregressive model with exogenous networks (NARX), which forecasts water demand by also processing external inputs such as temperature, rain, weekday, and hour of the day. The water demand forecasting model is then updated by processing its associated error through an unscented Kalman filter (UKF), which is specifically useful for near real-time processes. This derives from UKFs’ ability to quickly synthesize highly nonlinear spaces. The proposed method is applied to three water utility district metered areas of a medium-size city. The results show significant performance improvements when compared to classical non-recursive methods (e.g. ANN and support vector regression). The UKF-updated NARX has, thus, huge potential to provide key benefits for WDSs operation and management in near real-time.
Stream and Session
B3: Sixth Session on Data Mining as a Tool for Environmental Scientists (S-DMTES-2018)