Post-Processing Air Temperature Weather Forecast Using Artificial Neural Networks with Measurements from Meteorological Stations

Abstract

Human beings attempt to accurately predict the weather based on their knowledge of climate. The Norwegian Meteorological Institute is responsible for climate-related matters in Norway, and among its contributions is the numerical weather forecast, which is presented in a 2.5 km grid. To conduct a post-processing process that improves the resolution of the forecast and reduces its error, the Institute has developed the GRIDPP tool, which reduces the resolution to 1 km and introduces a correction based on altitude and meteorological station measurements. The present work aims to improve the current post-processing approach of the air temperature parameter by employing neural networks, using meteorological station measurements. Two neural network architectures are developed and tested: a multilayer perceptron and a convolutional neural network. Both architectures are able to achieve a smaller error than the original product. These results open doors for the Institute to plan for the practical implementation of this solution on its product for specific scenarios where the traditional numerical methods historically produce large errors. Among the test samples where the GRIDPP error is higher than 3 K, the proposed solution achieves a smaller error in 74.8% of these samples.