Using a hybrid approach to improve rainfall prediction for water management

UNESCO-IHE INSTITUTE FOR WATER EDUCATION

May 2007 Flooding England

Using a hybrid approach to improve rainfall prediction for water management
CARLOS ARTURO MARTINEZ CANO MSc Thesis WSE-HI 09-14 September 2009

Using a hybrid approach to improve rainfall prediction for water management

Master of Science Thesis By CARLOS A. MARTINEZ C.

Supervisors: Dr. Ir. Yunqing Xuan, MSc.(UNESCO-IHE) Dr. Ir. Gerald Corzo, MSc. (UNESCO-IHE)

Examination committee Prof. Dr. D. P. Solomatine (UNESCO-IHE), Chairman Dr. Ir. Shreedhar Maskey, MSc. (UNESCO-IHE) Dr. Ir. Yunqing Xuan, MSc. (UNESCO-IHE)

This research is done for the partial fulfilment of requirements for the Master of Science degree at the UNESCO-IHE Institute for Water Education, Delft, the Netherlands

Delft, TheNetherlands September 2009

The findings, interpretations and conclusions expressed in this study do neither necessarily reflect the views of the UNESCO-IHE Institute for Water Education, nor of the individual members of the MSc committee, nor of their respective employers.

“Oh Mary, conceived without sin, Pray for us who turn to you”. Amen

Abstract
In the last decades the increasedavailability of hydrological spatial models and operational online systems have boosted the necessity for more accurate short term spatial rainfall forecast in water resources management. In short term forecast (hours or minutes), radar observations and numerical weather models are the most common sources of spatial rainfall information. Rainfall prediction based on weather radars presents betterperformance than numerical weather prediction (NWP) models over the first hours. This can be explained by the fact that radar does not use important physical information, like wind, temperature, pressure and others, to calculate the rainfall information. Moreover, the forecast information obtained from radar is commonly based on tracking spatial patterns of rain. Therefore, on extended forecast horizonthe high variability of physical variables (e.g. wind, humidity and others) are contemplated by the NWP and not by the tracking of radar rainfall information. On the other hand, NWP models do not generally capture well the initial rainfall distribution. Nevertheless, over longer time scales, NWP models would perform better than weather radars as they resolve dynamically the large scale flow. Inthis study the integration of radar tracking forecast and NWP models results have been explored. The NWP model used is the meso-scale model five (MM5), and was set-up for a region in Southeast England. Radar tracking forecast based on spatial correlation of the radar patterns over consecutive time steps was applied. The Nimrod radar data from the British Atmospheric Database Centre (BADC) is usedas measured rainfall. Data driven modelling techniques such as model trees (M5), linear regression model (LR), artificial neural networks (ANN) and modular neural networks (MNN) have been used for the integration of both forecasting models. These techniques have been applied to three cases using different precipitation events. • • • Case 1 using data sampling in the process of building the DDM.Case 2 using consecutive rainfall events selected on expert knowledge Case 3 using a rule based integration of models

The results showed an overall improvement in almost all data-driven modelling techniques applied. The techniques were compared with those obtained from weather radar and MM5 model. It is important to highlight that although an improvement was found in the overall region, radartracking and NWP models still have low performance at highly retailed spatial resolution. The use of improved models or inclusion of weather atmospheric variables might bring an important alternative that needs to be tested. The use of modular models is probably the best alternative to integrate the NWP and radar tracking; however, we found limitations in its application due to the low number of...