Prediction of soil temperature using regression and artificial neural network models

In this study, monthly soil temperature was modeled by linear regression (LR), nonlinear regression (NLR) and artificial neural network (ANN) methods. The soil temperature and other meteorological parameters, which have been taken from Adana meteorological station, were observed between the years of 2000 and 2007 by the Turkish State Meteorological Service (TSMS). The soil temperatures were measured at depths of 5, 10, 20, 50 and 100 cm below the ground level. A three-layer feed-forward ANN structure was constructed and a back-propagation algorithm was used for the training of ANNs. In order to get a successful simulation, the correlation coefficients between all of the meteorological variables (soil temperature, atmospheric temperature, atmospheric pressure, relative humidity, wind speed, rainfall, global solar radiation and sunshine duration) were calculated taking them two by two. First, all independent variables were split into two time periods such as cold and warm seasons. They were added to the enter regression model. Then, the method of stepwise multiple regression was applied for the selection of the “best” regression equation (model). Thus, the best independent variables were selected for the LR and NLR models and they were also used in the input layer of the ANN method. Results of these methods were compared to each other. Finally, the ANN method was found to provide better performance than the LR and NLR methods.     

人工神经网络法和线性回归法对降水相态的预报效果对比

本文主要对相同条件下线性回归法(LR)和人工神经网络法(ANN)对降雨、雨夹雪和降雪3种降水相态的预报效果进行了对比检验.选取降水发生时和发生前6h的地面2 m温度、露点温度作为预报因子,对降雨、雨夹雪和降雪进行预报.应用国家气象中心2001-2011年我国地面756站实况观测资料,其中应用2001-2010年资料对方法进行训练,2011年资料用来对比检验预报效果.结果显示,(1)两种方法对3种相态降水都有一定的预报能力,对降雪预报最好,其次是降雨和雨夹雪;(2)两种方法对北方的雨雪分界线预报比对南方的好;(3)无论是对全国还是长江中下游流域,在相同条件下,ANN法的预报效果大都优于LR法,当温度和露点温度预报准确时,ANN法对北方的雨雪分界线能进行较准确的预报.     

基于ECMWF产品的站点气温预报集成学习误差订正

提出一种基于数值模式预报产品的气温预报集成学习误差订正方法,通过人工神经网络、长短期记忆网络和线性回归模型组合出新的集成学习模型（ALS模型）,采用2013—2017年的欧洲中期天气预报中心数值天气预报模式2 m气温预报产品和中国部分气象站点数据,利用气象站点气温、风速、气压、相对湿度4个观测要素,挖掘观测数据的时序特征并结合模式2 m气温预报结果训练机器学习模型,对2018年模式2 m气温6~168 h格点预报产品插值到站点后的预报结果进行偏差订正。结果表明:ALS模型可将站点气温预报整体均方根误差由3.11℃降至2.50℃,降幅达0.61℃（19.6%）,而传统的线性回归模型降幅为0.23℃（8.4%）。ALS模型对站点气温预报误差较大的区域和气温峰值预报的订正效果尤为显著,因此,集成学习方法在数值模式预报结果订正中具有较大的应用潜力。     

Estimation of dew point temperature using neuro-fuzzy and neural network techniques

This study investigates the ability of two different artificial neural network (ANN) models, generalized regression neural networks model (GRNNM) and Kohonen self-organizing feature maps neural networks model (KSOFM), and two different adaptive neural fuzzy inference system (ANFIS) models, ANFIS model with sub-clustering identification (ANFIS-SC) and ANFIS model with grid partitioning identification (ANFIS-GP), for estimating daily dew point temperature. The climatic data that consisted of 8 years of daily records of air temperature, sunshine hours, wind speed, saturation vapor pressure, relative humidity, and dew point temperature from three weather stations, Daego, Pohang, and Ulsan, in South Korea were used in the study. The estimates of ANN and ANFIS models were compared according to the three different statistics, root mean square errors, mean absolute errors, and determination coefficient. Comparison results revealed that the ANFIS-SC, ANFIS-GP, and GRNNM models showed almost the same accuracy and they performed better than the KSOFM model. Results also indicated that the sunshine hours, wind speed, and saturation vapor pressure have little effect on dew point temperature. It was found that the dew point temperature could be successfully estimated by using T _mean and R _H variables.     

利用神经网络方法建立热带气旋强度预报模型

以神经网络方法为基础,建立西北太平洋热带气旋强度预测模型,模型首先进行历史相似热带气旋选择。从选择的样本出发,计算得到一组气候持续因子、天气学经验因子和动力学因子, 对这些因子采用逐步回归方法进行筛选,将筛选得到的因子同对应时效的热带气旋强度输入神经网络训练模块,从而得到优化的预测模型。从2004-2005年西北太平洋26个热带气旋过程对12,24,36,48,72h等不同预报时效分别进行的634,582,530,478,426次预测试验结果的统计来看,相对于线性回归模型预测水平,该模型显著降低了各时段的预测误差。从几个热带气旋个例的预测结果来看, 该模型对超强台风, 以及具有强度迅速加强、再次加强等特征的热带气旋过程均有很好的描述能力。     

人工神经网络预报模型的过拟合研究

针对神经网络方法在预报建模中存在的“过拟合”(overfitting)现象和提高泛化性能 (generalizationcapability)问题 ,提出了采用主成分分析构造神经网络低维学习矩阵的预报建模方法。研究结果表明 ,这种新的神经网络预报建模方法 ,通过浓缩预报信息 ,降维去噪 ,使得神经网络的预报建模不需要进行适宜隐节点数的最优网络结构试验 ,没有“过拟合”现象 ,并且与传统的神经网络预报建模方法及逐步回归预报模型相比泛化能力有显著提高     

Relative importance of parameters affecting wind speed prediction using artificial neural networks

In traditional artificial neural networks (ANN) models, the relative importance of the individual meteorological input variables is often overlooked. A case study is presented in this paper to model monthly wind speed values using meteorological data (air pressure, air temperature, relative humidity, and precipitation), where the study also includes an estimate of the relative importance of these variables. Recorded monthly mean data are available at a gauging site in Tabriz, Azerbaijan, Iran, for the period from 2000 to 2005, gauged in the city at the outskirt of alluvial funneling mountains with an established microclimatic conditions and a diurnal wind regime. This provides a sufficiently severe test for the ANN model with a good predictive capability of 1 year of lead time but without any direct approach to refer the predicted results to local microclimatic conditions. A method is used in this paper to calculate the relative importance of each meteorological input parameters affecting wind speed, showing that air pressure and precipitation are the most and least influential parameters with approximate values of 40 and 10 %, respectively. This gained knowledge corresponds to the local knowledge of the microclimatic and geomorphologic conditions surrounding Tabriz.     

Hourly predictive Levenberg–Marquardt ANN and multi linear regression models for predicting of dew point temperature

In this study, the ability of two models of multi linear regression (MLR) and Levenberg–Marquardt (LM) feed-forward neural network was examined to estimate the hourly dew point temperature. Dew point temperature is the temperature at which water vapor in the air condenses into liquid. This temperature can be useful in estimating meteorological variables such as fog, rain, snow, dew, and evapotranspiration and in investigating agronomical issues as stomatal closure in plants. The availability of hourly records of climatic data (air temperature, relative humidity and pressure) which could be used to predict dew point temperature initiated the practice of modeling. Additionally, the wind vector (wind speed magnitude and direction) and conceptual input of weather condition were employed as other input variables. The three quantitative standard statistical performance evaluation measures, i.e. the root mean squared error, mean absolute error, and absolute logarithmic Nash–Sutcliffe efficiency coefficient $ \left( {\left| {{\text{Log}}({\text{NS}})} \right|} \right) $ were employed to evaluate the performances of the developed models. The results showed that applying wind vector and weather condition as input vectors along with meteorological variables could slightly increase the ANN and MLR predictive accuracy. The results also revealed that LM-NN was superior to MLR model and the best performance was obtained by considering all potential input variables in terms of different evaluation criteria.     

Performance comparison of three predictor selection methods for statistical downscaling of daily precipitation

Predictor selection is a critical factor affecting the statistical downscaling of daily precipitation. This study provides a general comparison between uncertainties in downscaled results from three commonly used predictor selection methods (correlation analysis, partial correlation analysis, and stepwise regression analysis). Uncertainty is analyzed by comparing statistical indices, including the mean, variance, and the distribution of monthly mean daily precipitation, wet spell length, and the number of wet days. The downscaled results are produced by the artificial neural network (ANN) statistical downscaling model and 50 years (1961–2010) of observed daily precipitation together with reanalysis predictors. Although results show little difference between downscaling methods, stepwise regression analysis is generally the best method for selecting predictors for the ANN statistical downscaling model of daily precipitation, followed by partial correlation analysis and then correlation analysis.     

Least squares support vector machine for short-term prediction of meteorological time series

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.     

利用人工神经网络模型预测西北太平洋热带气旋生成频数

通过对60年(1950~2009年)北半球夏、秋季(6~10月)热带气旋(TC)频数与春季(3~5月)大尺度环境变量的相关分析,挑选出8个相关性较高的前期预报因子建立人工神经网络(ANN)模型,对2010~2017年8年夏、秋季TC频数进行回报,并将回报结果与传统多元线性回归(MLR)方法所得结果进行对比分析。结果表明,ANN模型对60年历史数据的拟合精度高,相关系数高达0.99,平均绝对误差低至0.77。在8年回报中,ANN模型相关系数为0.80,平均绝对误差为1.97;而MLR模型相关系数仅为0.46,平均绝对误差为3.30。ANN模型在历史数据拟合和回报中的表现都明显优于MLR模型,未来可考虑应用于实际的业务预测中。     

Statistical Downscaling of Wind Variability from Meteorological Fields

Measurements show that on numerous occasions the low-level wind is highly variable across a large portion of south-eastern Australia. Under such conditions the risk of a large rapid change in total wind power is increased. While variability tends to increase with mean wind speed, a large component of wind variability is not explained by wind speed alone. In this work, reanalysis fields from the US National Centers for Environmental Prediction (NCEP) are statistically downscaled to model wind variability at a coastal location in Victoria, Australia. In order to reduce the dimensionality of the problem, the NCEP fields are each decomposed using empirical orthogonal function (EOF) techniques. The downscaling technique is applied to two periods in the seasonal cycle, namely (i) winter to early spring, and (ii) summer. In each case, data representing 2 years are used to form a model that is then validated using independent data from another year. The EOFs that best predict wind variability are examined. To allow for non-linearity and complex interaction between variables, all empirical models are built using random forests. Quantitatively, the model compares favourably with a simple regression of wind variability against wind speed, as well as multiple linear regression models.     

南海热带气旋大风的遗传－神经网络集合预报

利用1980-2012年的南海热带气旋实况资料和NCEP/NCAR再分析资料,将热带气旋定位中心周边6×6格点上的地面风速作为预报对象,以气候持续预报因子和前期风速预报因子作为模型输入,采用遗传—神经网络集合预报方法,进行热带气旋定位中心周边36个格点上的风速预报模型的预报建模研究.分别对2008-2012年7-9月共368个独立预报样本进行遗传-神经网络集合方法的分月预报结果表明,南海热带气旋中心周边风速24h的预报平均绝对误差为2.35m.s-1.另外,本文还进一步将该预报方法与国内外普遍采用的逐步回归预报模型进行对比分析,在相同的预报量和预报因子的条件下的对比分析表明,新预报模型对≥10m.s-1的强风预报结果较逐步回归方法的优势明显,预报性能较好,可为沿海热带气旋大风预报提供新的参考.     

Estimation of relative humidity based on artificial neural network approach in the Aegean Region of Turkey

The aim of this study is to estimate the monthly mean relative humidity (MRH) values in the Aegean Region of Turkey with the help of the topographical and meteorological parameters based on artificial neural network (ANN) approach. The monthly MRH values were calculated from the measurement in the meteorological observing stations established in Izmir, Mugla, Aydin, Denizli, Usak, Manisa, Kutahya and Afyonkarahisar provinces between 2000 and 2006. Latitude, longitude, altitude, precipitation and months of the year were used in the input layer of the ANN network, while the MRH was used in output layer of the network. The ANN model was developed using MATLAB software, and then actual values were compared with those obtained by ANN and multi-linear regression methods. It seemed that the obtained values were in the acceptable error limits. It is concluded that the determination of relative humidity values is possible at any target point of the region where the measurement cannot be performed.     

A neural network atmospheric model for hybrid coupled modelling

 The possibility of using a nonlinear empirical atmospheric model for hybrid coupled atmosphere-ocean modelling has been examined by using a neural network (NN) model for predicting the contemporaneous wind stress field from the upper ocean state. Upper ocean heat content (HC) from a 6-layer ocean model was a better predictor of the wind stress than the (observed or modelled) sea surface temperature (SST). Our results showed that the NN model generally had slightly better skills in predicting the contemporaneous wind stress than the linear regression (LR) model in the off-equatorial tropical Pacific and in the eastern equatorial Pacific. When the wind stresses from the NN and LR models were used to drive the ocean model, slightly better SST skills were found in the off-equatorial tropical Pacific and in the eastern equatorial Pacific when the NN winds were used instead of the LR winds. Better skills for the model HC were found in the western and central equatorial Pacific when the NN winds were used instead of the LR winds. Why NN failed to show more significant improvement over LR in the equatorial Pacific for the wind stress and SST is probably because the relationship between the surface ocean and the atmosphere in the equatorial Pacific over the seasonal time scale is almost linear. Received: 2 March 1999 / Accepted: 13 July 2000     

Annual precipitation forecast for west, southwest, and south provinces of Iran using artificial neural networks

Rainfed agriculture plays an important role in the agricultural production of the southern and western provinces of Iran. In rainfed agriculture, the adequacy of annual precipitation is considered as an important factor for dryland field and supplemental irrigation management. Different methods can be used for predicting the annual precipitation based on climatic and non-climatic inputs. Among which artificial neural networks (ANN) is one of these methods. The purpose of this research was to predict the annual precipitation amount (millimeters) in the west, southwest, and south of Islamic Republic of Iran with the total area of 394,259?km², by applying non-climatic inputs according to the long-time average precipitation in each station (millimeters), 47.5?mm precipitation since the first of autumn (day), t _47.5, and other effective parameters like coordinate and altitude of the stations, by using the artificial neural networks. In order to intelligently estimate the annual amount of precipitation in the study regions (ten provinces), feedforward backpropagation artificial neural network model has been used (method I). To predict the annual precipitation amount more accurately, the region under study was divided into three sub-regions, according to the precipitation mapping, and for each sub-region, the neural networks were developed using t _47.5 and long-time average annual precipitation in each station (method II). It is concluded that neural networks did not significantly increase the prediction accuracy in the study area compared with multiple regression model proposed by other investigators. However, in case of ANN, it is better to use a structure of 2–6–6–10–1 and Levenberg–Marquardt learning algorithm and sigmoid logistic activation function for prediction of annual precipitation.     

GANN models for reference evapotranspiration estimation developed with weather data from different climatic regions

Accurate estimation of reference evapotranspiration (ET₀) becomes imperative for better managing the more and more limited agricultural water resources. This study examined the feasibility of developing generalized artificial neural network (GANN) models for ET₀ estimation using weather data from four locations representing different climatic patterns. Four GANN models with different combinations of meteorological variables as inputs were examined. The developed models were directly tested with climatic data from other four distinct stations. The results showed that the GANN model with five inputs, maximum temperature, minimum temperature, relative humidity, solar radiation, and wind speed, performed the best, while that considering only maximum temperature and minimum temperature resulted in the lowest accuracy. All the GANN models exhibited high accuracy under both arid and humid conditions. The GANN models were also compared with multivariate linear regression (MLR) models and three conventional methods: Hargreaves, Priestley–Taylor, and Penman equations. All the GANN models showed better performance than the corresponding MLR models. Although Hargreaves and Priestley–Taylor equations performed slightly better than the GANN models considering the same inputs at arid and semiarid stations, they showed worse performance at humid and subhumid stations, and GANN models performed better on average. The results of this study demonstrated the great generalization potential of artificial neural techniques in ET₀ modeling.     

Study on the Overfitting of the Artificial Neural Network Forecasting Model

Because of overfitting and the improvement of generalization capability (GC) available in the construction of forecasting models using artificial neural network (ANN), a new method is proposed for model establishment by means of making a low-dimension ANN learning matrix through principal component analysis (PCA). The results show that the PCA is able to construct an ANN model without the need of finding an optimal structure with the appropriate number of hidden-layer nodes, thus avoids overfitting by condensing forecasting information, reducing dimension and removing noise, and GC is greatly raised compared to the traditional ANN and stepwise regression techniques for model establishment.     

北方公路交通气象环境识别及安全管理策略研究

用沈山高速公路沿线10个交通气象站的逐10 min气温、能见度、路面温度、降水、湿度和风等资料,采用多元逐步线性回归、最小二乘曲线拟合、MLP神经网络建立数学模型对北方公路路面温度、能见度以及冰雪路面等交通气象环境进行识别。通过抽取随机样本对模型进行检验,结果表明:所建立的模型对交通气象环境具有较好的模拟和识别作用。基于交通工程理论建立的不同路况条件下车速预估和交通安全指数模型对公路交通安全管理具有很好的应用价值。     

模块化模糊神经网络的数值预报产品释用预报研究

综合应用预报量自身时间序列的拓展,数值预报产品和模块化模糊神经网络方法,进行了一种新的数值预报产品释用预报研究。将这种新方法与常规的数值预报产品完全预报(PP)方法进行了对比试验。结果表明,这种模块化模糊神经网络数值预报产品释用预报方法比PP预报方法的预报精度显著提高。并且,通过对预报模型“过拟合”现象的研究发现,这种模块化模糊神经网络的数值预报产品释用预报模型具有很好的泛化性能。