Performance Assessment of Neural Network Models for Seasonal Weather Forecast Postprocessing in the Alpine Region

Seasonal weather forecasts are crucial for water-related sectors. However, the presence of systematic biases limits the usefulness of global seasonal weather forecasts produced by numerical weather prediction models. Although statistical postprocessing approaches, such as empirical quantile mapping, are widely used to improve accuracy and reliability, they have limitations in the accuracy of forecast values outside the training period and difficulties in incorporating multiple static and dynamic environmental variables to capture non-linear dependencies. This study seeks to overcome these limitations by implementing a neural network-based distributional regression method as a postprocessing tool. The study investigates the performance of these methods using seasonal forecasts of total precipitation and 2-meter temperatures for a one-month lead time over the Trentino-South Tyrol region in the northeastern Italian Alps. The forecast dataset is the fifth-generation seasonal weather forecast system (SEAS5) generated by the European Centre for Medium-Range Weather Forecasts (ECMWF), which has a 0.125°x 0.125°horizontal grid resolution with 25 ensemble members over the period from 1981 to 2016. The reference dataset is a high-resolution (250 m x 250 m) gridded observational data over the region. The performance of both methods is evaluated with a focus on the effects of forecast lead times, location, and seasonal variability. Results show that the neural network-based approach consistently outperforms empirical quantile mapping, especially during short lead times and at higher elevations.