FastNet learns to keep the detail in weather forecasts

FastNet’s experimental UK forecasts preserve local detail when training checks patterns at several scales. A 27 October preprint compares wavelet, Fourier and pointwise mean-squared-error objectives on the same model architecture, forecasting at 1.5-kilometre spacing over the UK.[1]

Developed by the Met Office and the Alan Turing Institute, FastNet learns how atmospheric conditions evolve from historical data. Its graph neural network advances a weather state through successive predictions. The Met Office currently evaluates it alongside existing forecasting systems, with operational use still subject to assessment.[2]

A wavelet transform examines an image at several scales while retaining information about where features lie. Nick Kingsbury’s dual-tree design, described in 2001, also distinguishes directions and is comparatively stable when a feature shifts slightly. Think, for example, of separating a narrow diagonal rainband from a broad structure aligned north to south: scale, position and orientation each carry information.[3]

Pointwise error penalises a displaced rainband both where it appeared and where it was predicted, encouraging smoothing. The new objective compares features’ strength and spatial arrangement within each scale and direction. Held-out 2023 verification included 114 surface stations.[1]

Wind forecasts improved most clearly under neighbourhood scoring. Rainfall fractions skill scores also improved, particularly for heavier rain. Pointwise training generally retained slightly lower errors at exact locations; all variants underestimated the strongest rainfall.[1]

Neighbourhood verification asks how closely a forecast describes the surrounding area. The fractions skill score, introduced by Nigel Roberts and Humphrey Lean, compares the proportions of forecast and observed grid squares exceeding a rainfall threshold within a chosen window. For illustration, if five of ten nearby squares exceed the threshold and a forecast marks four, it compares those fractions. Repeating this over progressively larger windows reveals how the answer changes with spatial scale.[4]

A rainband displaced slightly from its observed position can therefore receive credit for being in the right vicinity. Enlarging the window makes the location requirement less exacting. The useful question becomes how small a neighbourhood can be before that forecast loses its skill.[4]

Report an error or suggest a correction