Hydrologic performance modelling of green roofs using conceptual and data-driven modelling tools
Abstract
With climate change and rapid urbanization, the amount and intensity of stormwater flow are expected to increase, leading to frequent failures and overflows of conventional stormwater infrastructures. To address this issue, green infrastructures provide a sustainable alternative and complementary approach to conventional stormwater infrastructures. By promoting natural water balance restoration, green infrastructures, such as green roofs and bioretention cells, decrease runoff, improve stormwater quality, and preserve ecosystems through increased green surface areas. Green roofs, in particular, are an effective way to combat stormwater issues, as they provide a range of benefits, including reducing energy costs, enhancing biodiversity, and improving air quality.
Hydrological models play a crucial role in evaluating the effectiveness of green roofs with respect to stormwater management. Particularly, they can be used to estimate the retention (i.e., reduction of stormwater) and detention (i.e., attenuation of outflows) of green infrastructures. These models can be divided into physically-based, conceptual and data-driven and . Physically-based models are associated with high complexity, data requirements and computational cost. In addition, recent studies highlighted the limitation of current physically-based models in simulating runoff from multilayered green roofs with complex drainage mats.
Conceptual models can provide accurate simulation of green roofs with a lower level of complexity. However they require calibration with measured data which limits their use when data are not available for calibration. In addition, transferring calibrated model parameters from similar green roofs located in different locations were found to yield poor simulation accuracy in recent studies. On the other hand, the use of data-driven models in estimating the hydrological performance of green roofs, was found to be only limited to simple regression models which are site specific and not transferable. More powerful data-driven models, such as Machine learning (ML), which can be more accurate and transferable, were not evalauted in simulating runoff from green roofs to evaluate their performance in previous literature. In addition the effects of roof geometries (i.e., slope and length) on the hydrological performance of green roofs were not well studied in the literature and it is not clear if the current conceptual models of green roofs can account for these effects.
This PhD thesis focuses on enhancing the accuracy of hydrological performance estimation of extensive green roofs using data-driven and conceptual modelling approaches. A dataset of 31 extensive green roofs in three countries was utilized to evaluate and compare the effectiveness and transferability of conceptual and ML models for estimating the hydrological perfromance of green roofs (i.e., retention and detention). Four ML algorithms were tested in this study, namely artificial neural network (ANN), M5 model tree, long short-term memory (LSTM) and k nearest neighbour (kNN). In addition, this PhD thesis evaluated a reservoir routing model, developed in this study, and the low impact development module of the stormwater management model (SWMM). Different calibration strategies were tested tested for enhancing the transferability of conceptual models which include calibrating models using multiple data (i.e., runoff and soil moisture) and multisite model calibration. Furthermore, a laboratory test bed was constructed to investigate the impact of slope and length on green roof detention and retention, and to assess the ability of conceptual models to account for these factors.
ML models were found to accurately simulate runoff from green roofs, yielding estimates of green roof retention that were close in accuracy to a proven conceptual model of green roofs. The trained ML models demonstrated transferability across cities with similar rainfall characteristics, indicating their potential for estimating green roof retention performance in different locations. However, it is recommended to use uncalibrated conceptual models for estimating green roof retention in cases where measured data are not available, as they can provide accurate results with lower levels of complexity compared to the ML models. The LSTM was found to be the most suitable ML model for simulating runoff from green roofs with a high temporal resolution of 5 minutes, which is important for estimating green roof detention. The study found that LSTM models with simple structures to outperform those with more complex structures, providing more accurate results at lower computational costs. Regional LSTM models, which were trained on data from multiple green roofs, showed promising results in predicting the hydrological performance of new green roof configuration. These findings highlights the transferability of LSTM models and their potential for wider application in the hydrologic deisgn of green roofs.
The reservoir routing model was found to accurately simulate runoff from green roofs after proper calibration. The calibration strategy was found to affect the values of model parameters, which in turns influenced the transferability of model parameters to similar green roofs. The calibration of the conceptual model using multiple data reduced model equifinality and yielded parameter values that could be related to physical properties of green roofs. These results led to the development of a set of guidelines for estimating model parameters in the absence of measured data for calibration. However, some model parameters varied across cities, reflecting the important influence of local climatic conditions at the green roof cites.
The multisite model calibration approach is effective in generating transferable parameters for different locations. The results indicate that the nonlinear parameterization of the reservoir routing is the most appropriate model for transferability between locations using parameters obtained from multisite calibration. The transferred nonlinear model provides the most accurate estimates of peak runoff values, which are crucial in determining the green roof detention system’s performance.
Through a set of laboratory experiments, the slope and length of green roofs were found to have a significant influence on the retention and detention performances of the drainage mats. The slope of the roof impacts the detention and retention performances of the drainage mats for low and high intensity rainfall events, while the length of the roof affects the detention of drainage mats for low intensity rainfall events. The influence of slope and length cannot be accounted for by the current conceptual models of green roofs. Hence, it is recommend adjusting the values of model parameters manually to account for these effects when using a model that was calibrated on a green roof with different slope and length values. The study attempted to develop guidelines that practitioners can use to estimate the retention and detention performance of green roof using hydrological models.