dc.description.abstract | Urban water supply is coming under increased pressure due to urbanisation, water scarcity and climate change. Efficient urban water management can help alleviate this pressure by improving service quality and reducing water loss. Accurate demand and consumption forecasting enables expansion planning, financing, and operation of water distribution systems. Current research often focuses on model-centric approaches where the model is improved to drive forecast accuracy; however, more efficient data usage could be realised as an alternative to model-centric approaches, without incurring additional computation costs. This work investigates the potential of data-centric forecasting approaches, focusing on ways to improve the efficiency of data and computation resource usage for short-term water demand forecasting.
To initiate the investigation, several intrinsically different forecasting models are analysed. A total of four different forecasting models, i.e., Prophet, Autoregressive Integrated Moving Average, Neural Network (NN) and Random Forest (RF) are applied to four demand datasets, i.e., one Chinese hourly demand dataset and three UK 15-minute demand datasets. Various aspects of data and model requirements for optimal performance are investigated. Results obtained from the case studies show that prolonging training data may not be necessary, and that accurate sub-daily water demand forecasting is possible with 10 days of past data for model training. In terms of accuracy, neural network and random forest tend to be better suited towards short-term water demand forecasting over statistical models.
The second part of the work aims to unbox the four black-box machine learning methods – NN, Long Short-Term Memory (LSTM), RF, Extreme Gradient Boosting (XGB) and explain their inner workings using SHapley Additive exPlanations and Local Interpretable Model-Agnostic Explanations, Prophet and ARIMA are excluded due to inferior forecasting accuracy. Results have found that feature requirement depends on data resolution, the forecasting model used and the forecast time of day. Network-based models (NN and LSTM) are more temporally dependent and feature intensive, indicating that they require more feature inputs to produce equal accuracy compared to tree-based models (RF and XGB). High-resolution forecasts can maintain a high level of accuracy with only one feature at the previous point.
The final part of the work analyses the possibility of incorporating Transfer Learning (TL) into the context of water demand forecasting. To evaluate the potential of TL, 18 UK DMAs water demand datasets are used. Experiments are designed to predict water demands in one DMA that has limited or unavailable data, with an aim to anaysing the forecasting ability of models built with alternative DMA data. Results have found that four and eight external DMA datasets are respectively suitable for 15-minute and hourly demand and that limited accuracy gains are achieved from samples size larger than 20,000. Finally, TL-incorporated methods can improve machine learning forecasting accuracy when there is limited data availability.
The results obtained in this study prove the usefulness of data-centric approaches’ ability to improve forecasting accuracy. The data-centric approaches explored in this thesis can be used to guide the development of machine learning-based short-term demand forecasting models for researchers, operators, and utilities. Efficient use of forecasting models and demand data holds further potential in improving forecast accuracy, reducing computation cost, and bettering user confidence in the application of machine learning models. | en_GB |