不同降水降尺度方法在天山西部区域的适用性评估

An Assessment of Different Precipitation Downscaling Methods in the Western Tianshan Mountains

为能更好地提高外源数据的可靠性与适用性,丰富基础资料匮乏地区可用的数据源,以天山西部区域的伊宁、尼勒克和昭苏站作为典型代表站,基于实测降水数据和NCEP/NCAR再分析数据,采用人工神经网络ANN、极限学习机ELM、长短期记忆神经网络LSTM、支持向量机SVM、组合简单平均CSAM和组合支持向量机CSVM方法建立统计降尺度模型,评估不同降尺度方法在降水降尺度方面的适用性。研究结果表明:①由于降水影响因素的多重性、实测站点降水的差异性,不同降尺度方法的优劣性存在一定的差异,且率定期与验证期最优降尺度方法可能出现不对应的情况;②实测站点降水量越大,降尺度效果越好;降水量年内变化越不均匀,降尺度效果越差;③由于实测站点冬季降雪事件的发生影响了降尺度的效果,总体降雪量越多,降尺度效果越差;④极端降水事件频发,影响到降尺度的效果,降低了降尺度结果的可靠性与精度。总体上,研究区ELM、LSTM和CSVM模型具有一定的优越性,其中CSVM组合模型在季节、年降水方面模拟效果最好,ELM模型在月降水分布方面模拟效果最好。

In order to better improve the reliability and applicability of external data and enrich the data sources available in areas where basic data are scarce.The Yining,Nilek and Zhaosu stations in the western Tianshan region are taken as typical representative stations.Then,the statistical downscaling model is established based on measured precipitation data and NCEP/NCAR reanalysis data by using a series of methods such as artificial neural network(ANN),extreme learning machine(ELM),long and short term memory neural network(LSTM),support vector machine(SVM),combined simple average(CSAM)and combined support vector machine(CSVM).Finally,the applicability of different downscaling methods of precipitation downscaling are evaluated.The results show that:①Due to the multiplicity of precipitation influencing factors and the difference of precipitation at measured stations,there are some differences in the advantages and disadvantages of different downscaling methods.What's more,the optimal downscaling method of the calibration period may not correspond to the optimal downscaling method of the validation period.②The greater the precipitation of the measured site,the better the downscaling effect.The more uneven the precipitation during the year,the worse the downscaling effect.③Because the occurrence of snowfall events in winter affects the effect of downscaling,the more the overall snowfall,the worse the effect of downscaling.④The frequent occurrence of extreme precipitation events affects the effect of downscaling and reduces the reliability and accuracy of downscaling results.Generally speaking,the ELM,LSTM and CSVM models in the study area have certain advantages,among which the CSVM combination model has the best simulation effect on seasonal and annual precipitation,and the ELM model has the best simulation effect on the monthly precipitation distribution.