A statistical comparison of observations and forecasts using the Regional Atmospheric Modelling System (RAMS) for operations at the Savannah River Site (SRS) in South Carolina is discussed. Simulated and observed values of wind direction and wind speed, collected twice daily for a two-year period from April 1998 to March 2000, are compared in a variety of ways for five different locations in the southeastern United States. Simulated and observed turbulence intensities are also compared for a one-year period beginning in February 1999 for the SRS. Statistical measures used for comparison are the mean absolute error, mean relative error and standard deviation. Directional errors tend to increase with forecast time while speeds tend to be underestimated over the entire forecast range, especially those exceeding 8 m s−1. Seasonally, winds are predicted with greater accuracy during the winter and spring months, when speeds in the southeastern United States are higher and direction is less variable. Vertical turbulence intensity at SRS is slightly underestimated, while horizontal turbulence intensity is slightly overestimated. In general, the statistical measures show that the model produces results with reasonable skill for the wind speed, direction and turbulence.

A high pressure system between 20°S and 35°S, moving from east to west, passed over the Andes mountains on 27 and 28 February 1997. Some days earlier, on 20 February, another centre of anticyclonic vorticity had moved westward without crossing the Andes. These events developed during a South American Seesaw System (SASS) event that had been displaced southward from its climatological position.

This southward displacement of the local circulation pattern was accompanied by intense convective activity during the positive phase of the event, which favoured the development of strong storms in north and central Argentina. The unusual configuration of the motion field gave rise to atypically high temperatures in the mid-latitude Argentinian region.

Both anticyclonic systems reached the mountain barrier: one from the north and the other from the south. The analysis of the factors involved in this phenomenon reveals that the static stability in each system behaved differently near the mountain barrier. Advection of potential vorticity was responsible for the westward motion of the system, while the radiative processes influenced the production of cyclonic and anticyclonic potential vorticity which balanced out during each day.

A recursive statistical method is given to post-process the 2-metre temperature ECMWF General Circulation Model by using the SAR network of surface stations in Sardinia. The method calculates daily maximum and minimum temperature forecasts up to 5 days ahead for 55 stations. Forecasts over one year for 55 ground stations have been analysed. A comparison with Direct Model Output (DMO) and persistence is then made, giving special attention to the winter season and to freezing conditions. The error analysis shows a BIAS very close to zero (between 0°C and 0.5°C) on all the stations and a marked reduction of mean absolute error compared to the DMO forecast. The error reduction is more marked in the case of freezing conditions than with the DMO forecast. Such situations are difficult to forecast and the very high BIAS of DMO has not been eliminated completely. The advantages of the procedure are: (a) there is no need to store large numbers of measurements and large amounts of model forecast data; (b) the estimated correction is easily adapted to new meteorological conditions; (c) the procedure is transparent to the updating of the model; and (d) the procedure is very easy to implement operationally.

In this paper, an adaptive Kalman filtering procedure is developed and applied to 2-metre temperature and 10-metre wind-speed forecasts in Iceland. The goal is to reduce the systematic bias and improve the accuracy of the local forecasts derived from a numerical weather prediction model and, in addition, to produce reliable prediction intervals. The method consists of adding two algorithms to the traditional Kalman filter procedure that adaptively estimate the noise statistics individually at each forecast location and each time step. The tested data set shows that the method is able to remove the systematic errors and to quantify the prediction uncertainty in a consistent manner.

Neural networks (NNs) have been successfully used in the environmental sciences over the last two decades. However, only a few review papers have been published, most of which cover image processing, classification, prediction and geophysical retrieval in general, while neglecting rainfall estimation issues. This paper reviews, without aiming to be exhaustive, NN approaches to satellite rainfall estimation (SRE) by providing an overview of some of the methodologies proposed. A basic introduction to NNs is provided and the advantages of using NNs in SRE are explained, illustrating how NNs can be used to complement more computational-expensive methods to generate quick and accurate results in near real time. The role of the NNs in statistical-empirical algorithms is also reviewed. The last section aims to generate some discussion through comparing the empirical and deterministic algorithmic approaches and contrasting some of the apparent drawbacks of using NNs with a statistically based view of the satellite geophysical parameter estimation.

In Part I of this study recent versions of the ECWMF Integrated Forecasting System (IFS) are used together with historical observational data to carry out reanalyses and deterministic reforecasts of three major north-west European wind storms of the twentieth century. The storms considered are the Dutch storm of 1 February 1953, the Hamburg storm of 17 February 1962, and the British October storm of 1987 (Great October Storm). Common to all these storms is their severity, which caused large loss of life and widespread damage.

Reanalysis of the storms is based on a 3D-Var and 4D-Var assimilation scheme at a horizontal resolution of T$_L$159 ($\approx$125 km) and T$_L$511 ($\approx$50 km), respectively. Similarly, two different horizontal resolutions (T$_L$159 and T$_L$511) are used to investigate the deterministic predictability of these storms. The lower-resolution system is exactly that used in the ERA-40 reanalysis project. The high-resolution system is a more recent version of the ECMWF IFS.

It is shown that the basic characteristics of the Dutch and Hamburg storms that gave rise to the storm surge are well predicted by the single deterministic forecasts up to about 48 and 84 hours, respectively, in advance.

Our capability to predict the Great October Storm is more difficult to assess. On the one hand, even recent versions of the ECMWF IFS underestimate the severity of the storm in the very short-range (12–24-hour forecasts started at 12 UTC 15 October 1987). On the other hand, the high-resolution version of the ECMWF IFS provides excellent deterministic forecasts of the track and intensity of the storm up to 96 hours in advance. However, there are errors in the timing of the storm (12 hours for the 96 hour forecast).

From the results presented in this study it is concluded – bearing in mind the limited number of cases considered – that with the current ECMWF forecasting system reliable deterministic predictions of some European wind storms are possible several days in advance.

Operational forecasting of mesoscale weather systems in Hong Kong is supported by an automated nowcasting system called SWIRLS (Short-range Warning of Intense Rainstorms in Localised Systems). SWIRLS is based on the extrapolation of radar echoes using the TREC (Tracking Radar Echoes by Correlation) technique. With a suitable choice of pixel array size on the radar reflectivity maps, the TREC vectors derived can be used to monitor and extrapolate echo motion right across the mesoscale spectrum, from individual convective cells, to supercells and clusters, and to groups of rainbands or squall lines.

On the basis of TREC, quantitative precipitation forecast (QPF) algorithms have been developed to produce high resolution forecast rainfall distribution maps over the local area. These maps provide useful objective guidance for forecasters to assess the likely rain scenario in the next few hours and to facilitate decision-making in operating the Rainstorm Warning System. This review, however, will focus on other aspects of TREC applications and some side benefits for operational mesoscale forecasting in Hong Kong. First, extension of the TREC technique to group tracking of echoes (GTrack) have enabled forecasters to make qualitative educated guesses of the likelihood of prolonged heavy rain or the potential of enhanced storm development. Secondly, extension of the QPF algorithms to a forecast range of 3 hours, supplemented by real-time accumulated rainfall data from a high density raingauge network, has enabled forecasters to provide reliable advice and guidance to geotechnical engineers for landslip risk assessment. Thirdly, rain-related applications aside, TREC vectors are also ingested into a data analysis system LAPS (Local Analysis and Prediction System) adapted from Forecast Systems Laboratory (FSL) of NOAA. Through the assimilation of TREC vectors and other wind observations, a three-dimensional wind structure of tropical cyclones can be generated in near real-time for forecasters' reference in assessing landfall impact.

Regional air pollution models are usually off-line coupled with numerical weather prediction models. The present study, however, focuses on on-line coupled modelling, for which the air pollution model is an integral part of the meteorological model. To this purpose, simulations of the first European Tracer Experiment (ETEX), which supplies the best suited verification data for dispersion modelling of a passive tracer, have been performed using versions of the Danish Meteorological Institute High Resolution Limited-Area Model (DMI-HIRLAM). Simulation results have been compared with observed values as well as with results from other models.

The operational version of DMI-HIRLAM using well-known basic advection schemes has been tested, showing the need for improvements for air pollution purposes. The fourth-order Bott algorithm for advection has thus been implemented in the semi-Lagrangian version of DMI-HIRLAM. Due to its practically perfect mass conservation properties and its low computational cost, this method turns out to be efficient for simulations of transport of pollutants in the atmosphere. This represents a first effort towards a fully integrated air pollution model, e.g. for forecasting purposes.

Theories are good for the intellect, but are no more useful than a bit of practical experience. Perhaps the most significant thing about them is that they have been accorded quite unjustified status by engineers.