Collocation-analyzed multi-source ensembled wind speed data in lake district: a case study in Dongting Lake of China

Abstract

The lack of reliable near-surface wind data products in lake district regions significantly limits the efficiency of meteorological models, forecasts, services, and wind resource development. This is primarily due to a shortage of observation data and related studies, as well as insufficient validation and promotion of existing data products, which further hinders the comprehension and monitoring of local wind fields. In this study, comprehensive verification research was conducted on mainstream wind field products, namely, ERA5-Land (EC), GLDAS (GL), and HRCLDAS (HR) in the Dongting Lake area of China. This was achieved by utilizing a large volume of measured data and a triple collocation analysis (TCA) method. Additionally, an exploration into the optimal wind field data fusion method was undertaken. HR products demonstrate superior performance in capturing wind speed at the in situ measured scale, while GL outperforms at the grid scale, and EC products show relatively stable performance with minimal outliers. The long short-term memory (LSTM) neural network model, combined with time-series features, emerges as the most optimal data fusion model. LSTM fusion product is superior to the original product (except for HR products at the in situ measured scale), TCA-based weighted fusion products, and multi-layer fully connected neural network (MFCNN) on various parameters. This study quantifies the performance of mainstream wind products in lake areas and provides a benchmark for further application of these products. Furthermore, the successful implementation of an optimal wind data fusion method can provide valuable insights for related research, and the resulting wind fusion products can offer superior basic data support for local terminal applications.