关键词: 长短期记忆网络, 广义回归神经网络, 小时气温, 玛纳斯河流域

Abstract:

As an indispensable component of freshwater resources, snow accumulation and melting processes are profoundly affected by temperature fluctuations. Accurate estimation and analysis of temperature data is paramount for ensuring the security of ecosystems in watershed and promoting the sustainable utilization of water resources use. This paper focus on the Manas river basin, which was located in the middle section of Tianshan mountain in Xinjiang. Based on hourly meteorological observation data from 139 national observation stations in the basin from 2012 to 2017, the optimal set of temperature factors was selected using the embedded method. Two kinds of precise temperature estimation models were constructed: the Long Short−Term Memory Network (LSTM) and the Long Short−Term Memory Network−Generalized Regression Neural Network (LSTM-GRNN). These models were used to simulate the distribution of land surface air temperatures in the study region. The results indicated that: (1) the temperature data simulated by the two models exhibited a similar trend to that observed in the measured data, with correlation coefficients R² of 0.89 for the LSTM model and 0.94 for the LSTM-GRNN model, indicating that both models were capable of achieving comparable results. (2) Hourly temperature estimates of both the LSTM model and the LSTM-GRNN model were in close proximity to the actual observations, the root mean square error (RMSE) for each season (spring, summer, autumn and winter) of the LSTM model was 1.93℃, 2.67℃, 2.16℃and 1.71℃, respectively. In comparison, the RMSE values of the LSTM-GRNN model were 1.79℃, 2.42℃, 1.91℃ and 1.46℃, with an overall improvement of 10.4% in accuracy over the former. Both models showed differences in estimation accuracy across seasons, with the highest accuracy in winter and spring, the second in autumn, and the lowest in summer. (3) In contrast to the LSTM model, which was constrained to single−site estimation, the LSTM-GRNN model, which integrated the spatial characteristics of meteorological data, could provide a higher accuracy of the spatial distribution of air temperature, and a more accurate representation of the spatial and temporal distribution of hourly temperatures in the Manas river basin. The research results will facilitate the generation of data for the simulation of snowmelt in the region and the protection against disasters in subsequent studies.

Key words: Long Short-Term Memory, Generalized Regression Neural Network, Hourly air temperature, Manas river basin