共查询到19条相似文献，搜索用时 62 毫秒
1.
动态学习率神经网络预测气温的尝试 总被引：3，自引：1，他引：3
采用单站气象资料建立动态学习率的神经网络预测模型，进行逐日气温模拟预测，并与逐步回归预测模型和固定学习率的神经网络模型比较，结果表明，神经网络模型预测能力较好，但训练时间长。采用动态学习率的网络模型在不损失预测精度的前提下大大减少了训练时间，为神经网络在气象中的应用提供了一种方法。 相似文献
2.
基于深度学习的强对流高分辨率临近预报试验 总被引：1，自引：0，他引：1
强对流天气临近预报、预警在气象灾害防御中具有极为重要的地位。在气象业务中,因对强对流天气临近预报、预警准确率和时、空分辨率的极高要求,使其成为业务难点和研究热点之一。对于高时、空分辨率强对流临近预报问题,尝试用深度学习方法来解决。首先将强对流临近预报抽象成同时包含时间和空间的序列预测问题;然后基于改进的循环神经网络算法形成的自编码模型,使用京津冀地区长序列、高时空分辨率天气雷达组网拼图数据进行模型训练;最后利用基于历史0.5 h雷达回波拼图数据训练得到的端到端神经网络,预报未来1 h内的逐6 min回波演变特征。通过基于传统外推算法的临近预报方法与深度学习算法的临近预报方法进行对比,发现使用的深度学习方法可以有效“学习”到高时、空分辨率序列雷达数据特征的内在关联,通过多层神经网络构造出抽象的深层特征,能够有效捕捉到雷达回波的演变规律和运动状态。通过计算雷达回波预报的命中率（POD）、虚警率（FAR）、临界成功指数（CSI）等检验表明,相较传统外推临近预报方法,在强对流回波临近预报准确率上有较明显提高。 相似文献
3.
近年来深度学习模型在解决对防灾减灾影响巨大且极具挑战性的临近预报问题的应用中日益增多。本文中,我们把临近预报作为一个时空序列预测的任务,将雷达反射率因子作为试验对象,使用基于对抗神经网络（GAN）优化构建的TAGAN深度学习模型预测未来1小时的雷达回波图像,并且与Rover光流法、基于卷积神经网络的3D U-Net模型进行对比试验。选取2018年全球气象AI挑战赛雷达回波数据集进行训练与测试,检验结果表明TAGAN模型在命中率（POD）,虚警率（FAR）,临界成功指数（CSI）以及相关系数等多种评分上要优于传统的光流法和对比的3D U-Net深度学习模型,TAGAN模型在以上的检验评分表现出色,并且随预测时间的增加较之传统光流模型效果更优,这为拓展和提升深度学习模型在临近天气预报中的应用提供了参考依据。 相似文献
4.
藻类叶绿素a浓度是反映太湖水体富营养化程度的重要参数指标。以太湖2010—2011年5—10月旬平均叶绿素a浓度和气象资料数据作为建模样本, 通过对气象资料进行主成分分析, 得到4种主要气象因子作为输入, 建立时间序列ARMA预测模型与BP神经网络预测模型, 并对2012年数据进行预测。利用两种模型在线性空间和非线性空间的预测优势, 将叶绿素a数据结构分解为线性自相关主体和非线性残差两部分。首先用ARMA模型预测序列的线性主体, 然后用BP模型对其非线性残差进行估计, 最终集成整个序列的预测结果, 建立了ARMA-BP预测模型。3种模型的预测效果为ARMA-BP＞BP＞ARMA。 相似文献
5.
6.
天气雷达探测资料是进行强对流天气临近预报的主要参考数据。针对传统雷达回波外推方法中存在资料信息利用率不足和外推时效有限的问题,文中利用神经网络进行雷达回波的外推、利用预测神经网络模型进行2 h以内的回波变化预报。回波外推问题的关键是回波时、空序列预测问题,该网络具有解决时间记忆问题的长、短时记忆单元（Long Short-Term Memory,LSTM）和提取空间特征的卷积模块。应用福建、江苏和河南多年的雷达探测资料构造训练和测试数据集。为消除降水的不平衡和提高对强回波的预报准确率,网络采用带权重的损失函数进行训练。对光流法和预测神经网络进行测试集检验以及个例分析,结果表明,在相同外推时效和检验反射率阈值的情况下,预测神经网络的临界成功指数、命中率均高于光流法,虚警率低于光流法。不同类型降水预测神经网络的SSIM值（structural similarity）均高于光流法,且层状云降水的SSIM值比对流云降水的大。因此,预测神经网络对强回波的预报能力高于光流法;在预报时效性上,预测神经网络模型具有一定的优越性;预测神经网络对层状云降水预报的准确率比对流云降水的高。 相似文献
7.
王永弟 《南京气象学院学报》2012,4(4):316-320
将模糊时间序列模型引入短期气候预报,利用重庆34个地面气象观测站的逐日观测资料(1971—2007年)和重庆市旱涝灾害监测预警决策服务系统计算的干旱指数和洪涝指数等资料,运用模糊时间序列模型分别对2001—2007年重庆市城口县1月降水、1月平均气温的预报结果(年度预测)和重庆市春旱指数的预报结果(年度预测)进行了模糊时间序列分析,预测了2004—2007年的发展趋势,用2004—2007年实测值与预测结果进行了比较,并与加权集成、人工神经网络集成、数据挖掘集成等模型进行了精度比较和分析.结果表明:模糊时间序列模型各项精度评定指标优良,并且计算简单,具有一定的实用价值. 相似文献
8.
9.
BP神经网络在建模中的参数优化问题研究 总被引：2，自引：0，他引：2
神经网络方法已经在过去很多年中得到了大量研究,特别是基于误差反向传播算法的人工神经网络(简称BP-ANN)在很多天气预报业务上发挥了重要作用.对于BP-ANN训练有这样的一个问题,在同一个样本模型、同样的网络结构和同样的输入参数情况下,每次训练得到的权重和最终的误差结果几乎都是不一样的,有的会很好,有的会较差.在利用BP-ANN建模训练中,希望都寻找到因子模型的局部最优解,使它具有较好的泛化能力.为了提高BP-ANN在业务预报中的建模和预测能力,将对BP-ANN进行改进试验.利用2009-2010年每年5月15日至9月15日的T639模式预测数据和北京地区4个站点的最高温度实况资料作为建模样本数据,对4个站点进行数值模拟试验.通过对4个模型样本的拟合建模试验发现:BP-ANN的随机初始权重场服从高斯分布,或者初始权重场进行多次初始化,或者采用动态的隐层神经元网络结构都能让BP-ANN对样本的拟合命中率有一定的提高.最后选择2011年5月15日至9月15日115天的资料作为预报测试数据,集成3种改进方法于一个BP-ANN中,和未改进前的BP-ANN进行比较,对比后发现优化后的BP-ANN训练出的模型预测得到的验证样本预测命中率要高于未优化的BP-ANN训练出的模型得到的验证样本预测命中率,优化后的BP-ANN具有更好的泛化能力. 相似文献
10.
11.
针对海量气象观测数据间存在大量的物理噪声、与气温无关的冗余特征以及时间相关性,提出了一种将一维卷积神经网络(1DCNN)和长短期记忆神经网络(LSTM)相结合的多信息融合气温预报方法。首先,运用差分法对气象观测数据进行预处理,得到平稳时间序列数据;其次,运用1DCNN提取与气温变化相关的特征变量作为神经网络模型的输入变量;最后,运用1DCNN和LSTM构建多信息融合气温预报模型1DCNN-LSTM,并以云南省昆明市历史气象观测数据为例,与传统的LSTM、1DCNN和反向传播神经网络(BP)对未来24小时的逐时气温预报进行了比较研究。研究结果表明,1DCNN-LSTM的均方根误差(RMSE)相较于LSTM、1DCNN和BP最大降低了5.221%、19.350% 和9.253%,平均绝对误差(MAE)最大降低了4.419%、17.520% 和8.089%。为气温的精准预报提供了参考依据。 相似文献
12.
Abstract We examine Arctic sea‐ice concentration (SIC) and sea‐level pressure (SLP) data using principal oscillation pattern (POP) and neural network methods. The POP method extracts oscillating patterns from multivariate time series, each pattern being characterized by an oscillation period and a decay time. Predictions can be made for patterns whose decay time is comparable with the period. For both the SIC and SLP, however, the decay times are much shorter than the oscillation periods, and therefore the forcast skill is poor. A neural network is a model of the learning behaviour of a living neural system. Presented with training data, a neural network can learn the linear or non‐linear rules embedded in the data. We trained neural networks with sea‐ice and sea‐level pressure data, and estimated the forecast skill using a cross‐validation technique. The neural networks did not exhibit forecast skill significantly better than that of persistence. We contrast the Arctic situation with previous studies in which POP and neural networks were successfully used to forecast El Niño at lead times up to 6 months. Reasons for the lack of skill in both methods are discussed. 相似文献
13.
Deviation exists between measured and simulated microwave radiometer sounding data. The bias results in low-accuracy atmospheric temperature and humidity profiles simulated by Back Propagation artificial neural network models. This paper evaluated a retrieving atmospheric temperature and humidity profiles method by adopting an input data adjustment-based Back Propagation artificial neural networks model. First, the sounding data acquired at a Nanjing meteorological site in June 2014 are inputted into the MonoRTM Radiative transfer model to simulate atmospheric downwelling radiance at the 22 spectral channels from 22.234GHz to 58.8GHz, and we performed a comparison and analysis of the real observed data; an adjustment model for the measured microwave radiometer sounding data was built. Second, we simulated the sounding data of the 22 channels using the sounding data acquired at the site from 2011 to 2013. Based on the simulated rightness temperature data and the sounding data, BP neural network-based models were trained for the retrieval of atmospheric temperature, water vapor density and relative humidity profiles. Finally, we applied the adjustment model to the microwave radiometer sounding data collected in July 2014, generating the corrected data. After that, we inputted the corrected data into the BP neural network regression model to predict the atmospheric temperature, vapor density and relative humidity profile at 58 high levels from 0 to 10 km. We evaluated our model’s effect by comparing its output with the real measured data and the microwave radiometer’s own second-level product. The experiments showed that the inversion model improves atmospheric temperature and humidity profile retrieval accuracy; the atmospheric temperature RMS error is between 1K and 2.0K; the water vapor density’s RMS error is between 0.2 g/m3 and 1.93g/m3; and the relative humidity’s RMS error is between 2.5% and 18.6%. 相似文献
14.
海洋表面温度(Sea Surface Temperature,SST)具有非平稳、非线性的特征,直接将处理平稳数据序列的方法应用到非平稳非线性特征明显的序列上显然是不合适的,预测的误差将会很大。为了提高预测精度,更好地解决非平稳非线性序列预测的问题,本文以东北部太平洋(40°N~50°N、150°W~135°W)区域的月平均海洋表面距平温度为例,首先分别应用集合经验模态分解(EEMD)和互补集合经验模态分解(CEEMD)方法将SST分解为不同尺度的一系列模态分量(IMF),再运用BP(Back Propagation)神经网络模型对每一个模态分量进行分析预测,最后将各IMF预测结果进行重构得到SST的预测值。数值试验的结果表明,CEEMD分解精度比EEMD分解精度高,CEEMD提高了基于BP神经网络的预测精度。系列试验统计分析说明应用这种方法对SST的1年预测是有效的。 相似文献
15.
Several studies have been devoted to dynamic and statistical downscaling for both climate variability and climate change. This paper introduces an application of temporal neural networks for downscaling global climate model output and autocorrelation functions. This method is proposed for downscaling daily precipitation time series for a region in the Amazon Basin. The downscaling models were developed and validated using IPCC AR4 model output and observed daily precipitation. In this paper, five AOGCMs for the twentieth century (20C3M; 1970–1999) and three SRES scenarios (A2, A1B, and B1) were used. The performance in downscaling of the temporal neural network was compared to that of an autocorrelation statistical downscaling model with emphasis on its ability to reproduce the observed climate variability and tendency for the period 1970–1999. The model test results indicate that the neural network model significantly outperforms the statistical models for the downscaling of daily precipitation variability. 相似文献
16.
G. Mihalakakou M. Santamouris D. N. Asimakopoulos 《Theoretical and Applied Climatology》2000,66(3-4):185-197
Summary The present study describes a neural network approach for modeling and making short-term predictions on the total solar radiation
time series.
The future hourly values of total solar radiation for several years are predicted, by extracting knowledge from their past
values, using feedforward backpropagation neural networks. The results are tested using various sets of non training measurements,
the findings are very encouraging and the model is found able to simulate the future values of total solar radiation time
series based on their past values. “Multi-lag” output predictions are performed using the predicted values to the input database
in order to model future total solar radiation values with sufficient accuracy. Furthermore, an autoregressive model is developed
for analysing and representing the total solar radiation time series. The predicted values of solar radiation are compared
with the observed data series and it was found that the neural network approach leads to better predictions than the AR model.
Received November 22, 1999 Revised February 17, 2000 相似文献
17.
人工神经网络算法在红外高光谱资料反演大气温度廓线中的应用 总被引：1，自引：0，他引：1
基于红外高光谱大气探测器AIRS实况观测资料,尝试用人工神经网络算法来实现晴空时大气的温度垂直廓线反演,重点将该算法与目前已经发展比较成熟的特征向量统计反演算法进行比较。结果表明,两种算法在计算时间效率和反演精度上相当。鉴于人工神经网络算法的统计物理本质,误差反向传播BP神经网络可以求解非线性问题,在优化训练样本和继续调试网络种类和网络参数的基础上,有望能进一步提高反演精度。 相似文献
18.
Estimating daily pan evaporation using artificial neural network in a semi-arid environment 总被引：3，自引：1，他引：2
Ali Rahimikhoob 《Theoretical and Applied Climatology》2009,98(1-2):101-105
The objective of this study was to test an artificial neural network (ANN) for estimating the evaporation from pan (E Pan) as a function of air temperature data in the Safiabad Agricultural Research Center (SARC) located in Khuzestan plain in the southwest of Iran. The ANNs (multilayer perceptron type) were trained to estimate E Pan as a function of the maximum and minimum air temperature and extraterrestrial radiation. The data used in the network training were obtained from a historical series (1996–2001) of daily climatic data collected in weather station of SARC. The empirical Hargreaves equation (HG) is also considered for the comparison. The HG equation calibrated for converting grass evapotranspiration to open water evaporation by applying the same data used for neural network training. Two historical series (2002–2003) were utilized to test the network and for comparison between the ANN and calibrated Hargreaves method. The results show that both empirical and neural network methods provided closer agreement with the measured values (R 2?>?0.88 and RMSE?<?1.2 mm day?1), but the ANN method gave better estimates than the calibrated Hargreaves method. 相似文献
19.
Least squares support vector machine for short-term prediction of meteorological time series 总被引：3，自引：1，他引：2
The prediction of meteorological time series plays very important role in several fields. In this paper, an application of least squares support vector machine (LS-SVM) for short-term prediction of meteorological time series (e.g. solar irradiation, air temperature, relative humidity, wind speed, wind direction and pressure) is presented. In order to check the generalization capability of the LS-SVM approach, a K-fold cross-validation and Kolmogorov–Smirnov test have been carried out. A comparison between LS-SVM and different artificial neural network (ANN) architectures (recurrent neural network, multi-layered perceptron, radial basis function and probabilistic neural network) is presented and discussed. The comparison showed that the LS-SVM produced significantly better results than ANN architectures. It also indicates that LS-SVM provides promising results for short-term prediction of meteorological data. 相似文献