Water demand forecasting using machine learning on weather and smart metering data
Date: 25 November 2019
University of Exeter
PhD in Water Informatics Engineering
Water scarcity is a global threat due to lifestyle and climate changes, pollution of water resources, as well as a rapidly growing population. The UK water industry’s regulators demand plans from water companies to sustainably manage their water resources, reduce per capita consumption and leakage, and create projections for climate ...
Water scarcity is a global threat due to lifestyle and climate changes, pollution of water resources, as well as a rapidly growing population. The UK water industry’s regulators demand plans from water companies to sustainably manage their water resources, reduce per capita consumption and leakage, and create projections for climate change scenarios. This work addresses critical problems of water demand by expanding the understanding of water use and developing improved forecasting methods. As part of this effort, the influence of the weather is thoroughly investigated, using a disaggregated, big-data statistical analysis. Results show that the weather effect on water consumption is overall limited, non-linear, and variable over time and households. Next, a short-term demand forecasting model is developed, based on Random Forests, that predicts household consumption using several socio-economic, customer and temporal characteristics. This model is of significant value due to its accuracy as well as accompanying methodology that allows the interpretation of results. In order to further improve the forecasting accuracy achieved using Random Forests, a new modelling technique is developed. The new method that uses model stacking and bias correction, outperforms most other forecasting models, especially when past consumption data are not available, as well as for peak consumption days. Finally, a water demand forecasting model based on Gradient Boosting Machines is trained at different levels of spatial aggregation, for different input configurations. Results show that the spatial scale has a strong influence on the best model predictors and the maximum forecasting accuracy that can be achieved. The methodology developed here can be used as a guide for researchers, water utilities and network operators to identify the methods, data and models to produce accurate water demand forecasts, based on the characteristics and limitations of the problem.
Item views 0
Full item downloads 0