Skip to main content

Spatio-temporal prediction of wind behaviour about the Bristol microclimate for use in an early warning system

Date

 

Robert Clay, Nikolaos Nikitas, Antonio Abellan - University of Leeds

This project aims to predict and warn dangerous wind around key infrastructure to increase public safety and wellbeing.

Critical transport infrastructure, such as the Clifton Suspension Bridge and Severn Crossings, are essential to the UK and millions of commuters each year. It is, therefore, a key requirement to be able to precisely predict when infrastructure must be closed due to high winds for public safety and economic reasons. This prediction is done using Wind Early Warning Systems (WEWS) from which results are currently inferred using either a limited number of weather stations giving poor accuracy or meteorological forecasts which are too slow to give any warning time.

Project aims

This project aims to find some middle ground between accuracy and computational expense that provides acceptable results as well as the time to act on them.

Explaining the science

The data sourced for this project is from the Met Office's Data Assimilation System (MIDAS) provides hourly mean average wind behaviour readings.  11 sites from around Bristol (Figure 1) are used to predict wind behaviour one hour ahead and compared against a 12th site at Filton nearest to the Clifton Bridge. Wind behaviour will be predicted over storm Frank, which closed the Bridge on the 30 December 2015, as well as over the whole of December for a more general view.

The methodology for this technique is divided into two stages with time-based forecasting using a Recurrent Neural Network and spatial prediction using Radial Basis Function based spatial interpolation. This prediction is compared against the true data at Filton as well as prediction solely in time one hour ahead using the true Filton data. The mean average error and R-squared metrics are used to test efficacy.

Figure 1. Bristol MIDAS sites used in prediction.

 

Results

Prediction using solely a Neural Network at Filton shows effective results over storm Frank and the month of December (mae = 2.86, R^2 = 0.6718 and mae = 2.265,  R^2 = 0.7175) respectively. However, while showing overall strong prediction in figure 2 it struggles to precisely predict maximum wind speeds. On the other hand space-time prediction using the other 11 sites shows weaker models in both cases (mae = 4.13, R^2 = 0.3942 and mae = 3.524, R^2 = 0.3618) but shows in figure 2 much more promise in prediction of wind maxima.

Figure 2. Time and space-time prediction at Filton over storm Frank.

 

Applications

This methodology allows prediction of the wind speed one hour in advance at Filton quickly with acceptable accuracy. Paired with data from the Clifton Bridge this could be translated into an assessment of the Bridges structural health to quantify danger to the public. In theory, advancements upon this methodology could lead to more generalised predictions to anywhere of interest in the UK.