Dataset on weather types and largescale climate drivers

This dataset links large-scale drivers and weather types relevant for UK weather, with a particular focus on UK winters. This dataset represents stages 1 and 2 of the proposed COMPASS framework presented in D2.2 (i.e. Figure 1 of the report) for understanding and quantifying the influence of large-scale drivers on the likelihood of a selected type of compound event. It also facilitates the remaining steps of the framework.

This dataset combines two existing datasets, the first of which represent a set of large-scale drivers (Sexton, 2024) and the second of which represents UK weather patterns (WPs) (Neal et al., 2016).

For full details, see Deliverable 2.2 Report on understanding of weather types and large-scale climate drivers and Deliverable 2.3 Dataset on weather types and largescale climate drivers.

Large scale drivers

To represent the historical values of large-scale drivers, which research has previously identified as being relevant for UK circulation, we use a set of variables provided by Sexton (2024), following the definitions given in Table 1. Although Sexton (2024) provide these driver variables for a set of climate model simulations, here we only use the variables calculated from the ERA5 Reanalysis (Hersbach et al., 2020) as this dataset is focused on understanding historical relationships.

Table 1: Summary of climate drivers calculated by Sexton et al. (2025). ERA5 data covers Jan 1959-Dec 2021.

Weather patterns

To represent weather types relevant to the UK, we utilise the well-established MO301 weather patterns (referred to as WP throughout this report) (Neal et al., 2016). The set of 30 WPs were derived objectively through k-means clustering of daily mean sea level pressure data (1850-2003) from European and North Atlantic Daily to Multi-decadal Climate Variability (EMULATE) MSLP (EMSLP) data (Ansell et al., 2006). This provides an assignment of one of the 30 WPs to each day from 1850 onwards, with WPs from 2004 onwards assigned by applying the clustering definitions to mean sea level pressure data from ERA5 (Hersbach et al., 2020). 

A smaller set of 8 WPs were derived from the original set of 30 for use in evaluating long-range and seasonal forecasts. At the seasonal timescale, only a broad indication of atmospheric circulation is possible, so fewer clusters are needed. To enable seamless comparisons with the original set of patterns at intermediate 
timescales, the reduced set of patterns was produced by combining highly correlating combinations from the original set of 30, using the MSLP anomalies.

Method

Pre-processing is applied to the monthly large-scale driver values provided by Sexton (2024), in the form of standardisation and de-trending (to remove long-term trends which can cause spurious correlations between drivers due to issues with collinearity). Dataset D2.3 uses both monthly and annual (seasonally aggregated) 
large-scale driver time series, both of which are standardized using the series’ mean and standard deviation. The season aggregations are those provided in Table 1 and align with the months of the year included for each variable in the causal network of Sexton et al. (2025). The dataset includes versions of the time series which 
both have and have not been de-trended (linearly de-trended, except for BKSIC where instead a 30-year high pass filter is applied).

Two different methods for temporally linking the large-scale driver and WP data are applied; either the values of a large-scale driver on a given month are linked to the assigned WPs for each day within that month, or season-aggregated values of the large-scale driver for a year are linked to the assigned WPs for each day within the late-winter period (i.e. January and February, the target period used by Sexton et al. (2025)). There is also the option to further aggregate the data by calculating monthly/seasonal frequencies and proportions of each MO30 WP, thus retaining a one-to-one relationship between large-scale driver values and MO30 WP assignments.

Contents

The dataset consists of 6 Comma Separated Value (CSV) files, each linking the same set of large-scale drivers to WPs (from both the 30 and 8 classifications) using a different combination of the pre-processing, temporal-linking and aggregation methods described above. Table 2 provides a summary of the methods applied to each of the CSV files. A brief description of the contents of each column is given in Table 3.

Table 2: Information about the methods applied to each of the files comprising D3.3.

Table 3: Information about the data in each column of the files in D3.3.