TMC-Net: A temporal multivariate correction network in temperature forecasting

Abstract

Numerical weather prediction and meteorological grand models have emerged as the predominant methods for modern temperature forecasting, with continuous advancements towards higher resolution and accuracy in recent years. However, as the forecast lead time increases, errors inevitably accumulate, necessitating the application of bias correction techniques to mitigate these inaccuracies. Existing bias correction models, however, exhibit several limitations, including suboptimal correction performance and insufficient utilization of historical information. To address these shortcomings, we propose a novel bias correction model called the Temporal Multivariate Correction Net (TMC-Net). The proposed model is composed of three principal modules: a Temporal Extraction Module, which captures the temporal variation patterns of forecast errors by accounting for factors such as seasonality and forecast lead time, making full use of historical information; a Multi-scale Fusion Module, which integrates multi-scale features from multiple variables and selects the most effective features; and a Transformer-based High-order Feature Fusion Module, which performs a deep fusion of interactive features among multiple variables. Empirical results, derived from applying TMC-Net to correct 2-m temperature forecasts from ECMWF HRES, ECMWF ENS, and Pangu models for lead times ranging from 12 to 240 h, demonstrate that TMC-Net can reduce forecast errors by 0.4 °C, enhance forecast accuracy by 5 %, and increase the anomaly correlation coefficient by 0.2 within the 12 to 240-h forecast range. These findings highlight the efficacy of TMC-Net in mitigating numerical forecast errors and improving forecast accuracy, indicating its potential application in high-resolution temperature forecasting.