| dc.description.abstract | This thesis investigates the suitability of Artificial Neural Network (ANN) models for forecasting wet-weather flow timeseries within urban drainage networks. Such models are valuable as forecasted flows may be used to optimise the operation of downstream wastewater treatment plants (WWTPs), thereby improving their performance under pressure from growing urban populations and the increasing intensity and frequency of rainfall events associated with climate change.
A critical literature review is used to identify key research gaps in the field that are limiting adoption by practitioners. These include limited approaches for producing confidence levels alongside ANN model predictions, little direct comparison between the most popular ANN model architectures in this field, and a lack of detail in spatial feature selection methodologies for driving ANN forecasting models.
Informed by this literature review, three experimental work packages are developed that tackle these research gaps:
1. Firstly, the performance of four ANN model architectures is evaluated by the how well they replicate the results of a high-fidelity physics based model simulating the response to synthetic rainfall over an artificial catchment. Their performance is compared against three other data driven baselines. The best performing of these architectures and problem specifications, the Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) cells, performs well across a diversity of rainfall and network conditions.
2. Secondly, the use of the Monte Carlo Dropout (MCD) technique for producing confidence intervals on forecasts made by ANN models is evaluated. This is tested by developing an RNN-LSTM forecasting model for a WWTP in Næstved, Denmark, then applying MCD to produce confidence intervals around predictions. The MCD technique is shown to be accurate and performant in comparison with naïve baselines, and the trade-off between increased hyperparameter optimisation costs against improved confidence interval precision and accuracy is detailed.
3. Thirdly, the use of a time-lagged cross correlation analysis for optimising the selection of spatial and temporal input features is tested using data from an adjacent but highly related discipline; the modelling of flows within limestone karst drainage networks, using data from Sleets Gill Cave, UK. This feature selection technique is shown to be a straightforward and effective method for reducing input feature sizes, thereby reducing the problem dimensionality and improving model accuracy and performance.
The results of these three work packages, alongside insights provided by the critical literature review, are used to inform the development of implementation guidelines for practitioners wishing to exploit ANNs for hydraulic timeseries forecasting in urban drainage networks. These provide guidance across the lifespan of such a project, from problem design and data requirements, to network design and implementation, and performance evaluation and enhancement.
Future research opportunities stemming from this work, and research gaps identified in the literature review, are highlighted. Suggested future research areas include; an assessment of stakeholder data-readiness, with a focus on the quality, quantity, and availability of appropriate data; an evaluation of the challenges facing operational implementation and maintenance of ANN models that are limiting their adoption; and an exploration of uncertainty quantification for out-of-sample input data, which is of particular importance due to the disproportional severity of high duration/intensity rainfall events. | en_GB |
