ANN、ANFIS和AR模型在日径流时间序列预测中的应用比较

水文预测是水文学为经济和社会服务的重要方面。其预报结果不仅能为水库优化调度提供决策支持,而且对水电系统的经济运行、航运以及防洪等方面具有重大意义。自回归模型(AR模型)、人工神经网络(ANN)和自适应神经模糊推理系统(ANFIS)在日径流时间序列中应用广泛。将这三种模型应用于桐子林的日径流时间序列预测中,不仅采用纳什系数(NS系数)、均方根误差(RMSE)和平均相对误差(MARE)为评价指标,对三种模型的综合性能进行了比较。而且,在对三种模型预测结果的平均相对误差的阈值统计基础上,分析了三种模型的预测误差分布。同时,通过研究模型性能指标随预见期的变化过程评价了三种模型不同预见期下的预测能力。结果表明ANFIS相对于ANN和AR模型不仅具有更好的模拟能力、泛化能力,而且在相同的预见期下具有更优的模型性能,可以作为日径流时间序列预测的推荐模型。     

Estimation of Daily Pan Evaporation Using Two Different Adaptive Neuro-Fuzzy Computing Techniques

This paper investigates the ability of two different adaptive neuro-fuzzy inference systems (ANFIS) including grid partitioning (GP) and subtractive clustering (SC), in modeling daily pan evaporation (E_pan). The daily climatic variables, air temperature, wind speed, solar radiation and relative humidity of two automated weather stations, San Francisco and San Diego, in California State are used for pan evaporation estimation. The results of ANFIS-GP and ANFIS-SC models are compared with multivariate non-linear regression (MNLR), artificial neural network (ANN), Stephens-Stewart (SS) and Penman models. Determination coefficient (R²), root mean square error (RMSE) and mean absolute relative error (MARE) are used to evaluate the performance of the applied models. Comparison of results indicates that both ANFIS-GP and ANFIS-SC are superior to the MNLR, ANN, SS and Penman in modeling E_pan. The results also show that the difference between the performances of ANFIS-GP and ANFIS-SC is not significant in evaporation estimation. It is found that two different ANFIS models could be employed successfully in modeling evaporation from available climatic data.     

Development of Nonlinear Model Based on Wavelet-ANFIS for Rainfall Forecasting at Klang Gates Dam

Rainfall is one of the most complicated effective hydrologic processes in runoff prediction and water management. Artificial neural networks (ANN) have been found efficient, particularly in problems where characteristics of the processes are stochastic and difficult to describe using explicit mathematical models. However, time series prediction based on ANN algorithms is fundamentally difficult and faces some other problems. For this purpose, one method that has been identified as a possible alternative for ANN in hydrology and water resources problems is the adaptive neuro-fuzzy inference system (ANFIS). Nevertheless, the data arising from the monitoring stations and experiment might be corrupted by noise signals owing to systematic and non-systematic errors. This noisy data often made the prediction task relatively difficult. Thus, in order to compensate for this augmented noise, the primary objective of this paper is to develop a technique that could enhance the accuracy of rainfall prediction. Therefore, the wavelet decomposition method is proposed to link to ANFIS and ANN models. In this paper, two scenarios are employed; in the first scenario, monthly rainfall value is imposed solely as an input in different time delays from the time (t) to the time (t-4) into ANN and ANFIS, second scenario uses the wavelet transform to eliminate the error and prepares sub-series as inputs in different time delays to the ANN and ANFIS. The four criteria as Root Mean Square Error (RMSE), Correlation Coefficient (R ²), Gamma coefficient (G), and Spearman Correlation Coefficient (ρ) are used to evaluate the proposed models. The results showed that the model based on wavelet decomposition conjoined with ANFIS could perform better than the ANN and ANFIS models individually.     

Modelling Intelligent Water Resources Allocation for Multi-users

This study proposes intelligent water resources allocation strategies for multiple users through hybrid artificial intelligence techniques implemented for reservoir operation optimization and water shortage rate estimation. A two-fold scheme is developed for (1) knowledge acquisition through searching input–output patterns of optimal reservoir operation by optimization methods and (2) the inference system through mapping the current input pattern to estimate the water shortage rate by artificial neural networks (ANNs). The Shihmen Reservoir in northern Taiwan is the study case. We first design nine possible water demand conditions by investigating the changes in historical water supply. With the nine designed conditions and 44-year historical 10-day reservoir inflow data collected during the growth season (3 months) of the first paddy crop, we first conduct the optimization search of reservoir operation by using the non-dominated sorting genetic algorithm-II (NSGA-II) in consideration of agricultural and public water demands simultaneously. The simulation method is used as a comparative model to the NSGA-II. Results demonstrate that the NSGA-II can suitably search the optimal water allocation series and obtain much lower seasonal water shortage rates than those of the simulation method. Then seasonal water shortage rates in response to future water demands for both sectors are estimated by using the adaptive network fuzzy inference system (ANFIS). The back-propagation neural network (BPNN) is adopted as a comparative model to the ANFIS. During model construction, future water demands, predicted monthly inflows (or seasonal inflow) of the reservoir in the next coming quarter and historical initial reservoir storages configure the input patterns while the optimal seasonal water shortage rates obtained from the NSGA-II serve as output targets (training targets) for both neural networks. Results indicate that the ANFIS and the BPNN models produce almost equally good performance in estimating water shortage rates, yet the ANFIS model produces even better stability. The reliability of the proposed scheme is further examined by scenario analysis. The scenario analysis indicates that an increase in public water demand or a decrease in agricultural water demand would bring more impacts of water supply on agricultural sectors than public sectors. Similarly, a bigger decrease in inflow amount would obviously bring more influence on agricultural sectors than public one. Consequently, given predicted inflow, decision makers can pre-experience the possible outcomes in response to competing water demands through the estimation models in order to determine adequate water supply as well as preparedness measures, if needed, for drought mitigation.     

基于GPR、CatBoost、XGBoost三种模型预测江西地区水面蒸发量

水面蒸发量是水资源规划与管理、农业灌溉设计和水文模拟等方面的基础数据,它是水量平衡计算中的关键要素。为了提高水面蒸发量的预测精度,选用了3种经验模型和3种学习机模型预测江西地区水面蒸发量,3种学习机模型包括GPR模型、XGBoost模型和CatBoost模型。依据江西地区2001-2015年16个气象站的逐日气象资料,如最高(低)气温、全球太阳辐射、地外太阳辐射、相对湿度和风速,构建10种不同的输入参数,通过对4种统计指标（R²、RMSE、MBE、MAE）的大小进行评估来评价模型的模拟精度。结果表明：当气象资料充足时,推荐CatBoost 10模型为江西地区水面蒸发量的预测模型,该模型在验证期的R²、RMSE、MBE、MAE值分别为0.744、0.842、0.006、0.633 mm/d;在输入组合相同的条件下,3种学习机模型的模拟精度均优于相应的经验模型。通过研究对比提高了江西地区水面蒸发量模型预测的精度。     

Managing Water Supply and Demand to Achieve Economic and Environmental Objectives: Application of Mathematical Programming and ANFIS Models

The integrated management of water supply and demand has been considered by many policymakers; due to its complexity the decision makers have faced many challenges so far. In this study, we proposed an efficient framework for managing water supply and demand in line with the economic and environmental objectives of the basin. To design this framework, a combination of ANFIS and multi-objective augmented ε-constraint programming models and TOPSIS were used. First, using hydrological data from 2001 to 2017, the rate of water release from the dam reservoir was estimated with the ANFIS model; afterwards, its allocation to agricultural areas was performed by combining multi-objective augmented ε-constraint models and TOPSIS. To prove the reliability of the proposed model, the southern Karkheh basin in Khuzestan province, Iran, was considered as a case study. The results have showed that this model is able to reduce irrigation water consumption and to improve its economic productivity in the basin.

     

A Wavelet-ANFIS Hybrid Model for Groundwater Level Forecasting for Different Prediction Periods

Artificial neural network (ANN) and Adaptive Neuro-Fuzzy Inference System (ANFIS) have an extensive range of applications in water resources management. Wavelet transformation as a preprocessing approach can improve the ability of a forecasting model by capturing useful information on various resolution levels. The objective of this research is to compare several data-driven models for forecasting groundwater level for different prediction periods. In this study, a number of model structures for Artificial Neural Network (ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), Wavelet-ANN and Wavelet-ANFIS models have been compared to evaluate their performances to forecast groundwater level with 1, 2, 3 and 4 months ahead under two case studies in two sub-basins. It was demonstrated that wavelet transform can improve accuracy of groundwater level forecasting. It has been also shown that the forecasts made by Wavelet-ANFIS models are more accurate than those by ANN, ANFIS and Wavelet-ANN models. This study confirms that the optimum number of neurons in the hidden layer cannot be always determined by using a specific formula but trial-and-error method. The decomposition level in wavelet transform should be determined according to the periodicity and seasonality of data series. The prediction of these models is more accurate for 1 and 2 months ahead (for example RMSE?=?0.12, E?=?0.93 and R ²?=?0.99 for wavelet-ANFIS model for 1 month ahead) than for 3 and 4 months ahead (for example RMSE?=?2.07, E?=?0.63 and R ²?=?0.91 for wavelet-ANFIS model for 4 months ahead).     

Estimating Daily Pan Evaporation Using Different Data-Driven Methods and Lag-Time Patterns

This study develops three neural networks models for estimating daily pan evaporation (PE) in South Korea: multilayer perceptron-neural networks model (MLP-NNM), generalized regression neural networks model (GRNNM), and adaptive neuro-fuzzy inference system (ANFIS). Daily PE was estimated at Daegu and Ulsan stations using temperature-based, radiation-based, sunshine duration-based and merged input combinations under lag-time patterns. Daily evaporation values computed by the models using merged inputs agreed with observed values. Comparison was also made between the neural networks models and multiple linear regression model (MLRM), which showed the superiority of MLP-NNM, GRNNM, and ANFIS over MLRM. It is concluded that the applied neural networks models can be successfully employed for estimating daily PE in South Korea.     

Identifying Prominent Explanatory Variables for Water Demand Prediction Using Artificial Neural Networks: A Case Study of Bangkok

The water demand of a city is a complex and non linear function of climatic, socioeconomic, institutional and management variables. Identifying the prominent variables among these is essential in order to adequately predict water demand, and to plan and manage water resources and the supply systems. Further, the need for such identification becomes more pronounced when data constraints arise. The objective of this study was to establish, using correlation and sensitivity analyses, a minimum set of variables required to predict water demand with significant accuracy. Artificial Neural Networks (ANN) models were developed to predict short-term (daily) and medium-term (monthly) demands for Bangkok. Using meteorological and water utility variables for short-term prediction, and different ANN architecture, 16 sets of models with a 1-, 2- and 3-day lead period were developed. Although the best fit models for the three lead periods used different input variables, prediction accuracies over 98% were achieved by using only the historic daily demand (HDD) as the explanatory variable. Similarly, for medium-term prediction, 11 sets of models with lead periods of 1-, 2- and 6-months were developed, using meteorological, water utility and socioeconomic variables. The best fit models for the three lead periods used all explanatory variables but prediction accuracies of more than 98% were obtained by downsizing the variable set. The meteorological variables have a greater influence on medium-term prediction as compared to short-term prediction, suggesting that future water demand in Bangkok could be significantly affected by climate change.     

Soil Water Simulation and Predication Using Stochastic Models Based on LS-SVM for Red Soil Region of China

The seasonal drought and the low available soil moisture affect the agricultural production in red soil region, China. Therefore, it is necessary to simulate and predict the dynamic changes of soil water in the field. Presently, dynamic model has been applied to obtain the soil water information. While the simulation accuracy of dynamic model depends on many complicated parameters, which are difficult to obtain. In this study, the various nonlinear Stochastic Model of soil water simulation systems and chaotic time series analysis methods of prediction systems had been set up. In the nonlinear Stochastic Model of soil water simulation systems, the daily soil water content simulated by Least squares support vector machine (LS-SVM) with the meteorological factors had more stabilities and advantages in soil water simulation performance over the Back Propagation Artificial Neural Network (BP-ANN) and Adaptive Neuro-Fuzzy Inference System (ANFIS). In chaotic time series analysis method of prediction systems, the various signal preprocessing methods including the appropriate de-noising methods and wavelet decomposition methods were applied to preprocess the original chaotic soil water signal. The results of the prediction systems showed that the appropriate de-noising methods and the tendency of wavelet transformation had less effect on the delay time (τ) and embedding dimension (m). The de-noising methods may ignore the detail information of the soil water signal, while the appropriate wavelet transformation to get smaller Maximum Lyapunov Exponent (λ₁) of the chaotic soil water signal detail and tendency information can improve the predicting capacity.     

磁性浮板覆盖下干旱区平原水库节水效率分析

针对干旱区平原水库水深浅、风速大、气温高、空气相对湿度低而导致蒸发损失严重的问题,在吐鲁番市胜金乡胜金沟三期水土保持水库进行了一整年不同覆盖面积的带磁性浮板与非磁性浮板覆盖水面来减少水面蒸发的试验,以此来分析相同覆盖面积下带磁性浮板成片整体分布与非磁性浮板分散分布对水库水面蒸发的节水效率。试验结果表明：在气象因素相同的情况下,带磁性浮板分别覆盖水面25％,50％和75％时,蒸发抑制率分别为27.4％,41.5％和59.1％;非磁性浮板分别覆盖水面25％,50％和75％时,蒸发抑制率分别为22.1％,38.1％和57.3％;带磁性浮板比非磁性浮板蒸发抑制率分别提高了5.3％,3.4％和1.8％。可见,带磁性浮板节水效果更加显著,对提高干旱区平原水库水资源的利用效率具有显著意义。     

Comparison of Recurrent Neural Network,Adaptive Neuro-Fuzzy Inference System and Stochastic Models in Eğirdir Lake Level Forecasting

Accurate prediction of lake-level changes is a very important problem for a wise and sustainable use. In recent years significant lake level fluctuations have occurred and can be related to the climatic change. Such a problem is crucial to the works and decisions related to the water resources and management. This study is aimed to predict future lake levels during hydrometeorological changes and anthropogenic activities taking place in the Lake Eğirdir which is the most important water storage of Lake Region, one of the biggest fresh water lakes of Turkey. For this aim, recurrent neural network (RNN), adaptive network-based fuzzy inference system (ANFIS) as prediction models which have various input structures were constructed and the best fit model was investigated. Also, the classical stochastic models, auto-regressive (AR) and auto-regressive moving average (ARMA) models are generated and compared with RNN and ANFIS models. The performances of the models are examined with the form of numerical and graphical comparisons in addition to some statistic efficiency criteria. The results indicated that the RNN and ANFIS can be applied successfully and provide high accuracy and reliability for lake-level changes than the AR and the ARMA models. Also it was shown that these stochastic models can be used in the lake management policies with the acceptable risk.     

Improving ANFIS Based Model for Long-term Dam Inflow Prediction by Incorporating Monthly Rainfall Forecasts

The necessity of long-term dam inflow forecast has been recognized for many years. Despite numerous studies, the accurate long-term dam inflow prediction is still a challenging task. This paper presents an adaptive neuro-fuzzy inference system (ANFIS) based model and evaluates the applicability of categorical rainfall forecast for improvement of monthly dam inflow prediction. In order to obtain appropriate ANFIS model configuration for dam inflow prediction, several models were trained and tested using various numbers of input variables i.e. monthly observed rainfall, relative humidity, temperature, dam inflow and categorical monthly rainfall forecast. The ANFIS based models were configured and evaluated for six major dams of South Korea i.e. Andong, Chungju, Daecheong, Guesan, Soyang and Sumjin having high, medium and low reservoir capacity. The results showed significant improvement in dam inflow prediction for all the selected dams using the ANFIS based model with categorical rainfall forecast compared to the ANFIS based model with only preceding month’s dam inflow and weather data.     

Forecasting the Level of Reservoirs Using Multiple Input Fuzzification in ANFIS

Estimation the Level of water is one of the crucial subjects in reservoir management influencing on reservoir operation and decision making. One of the most accurate artificial intelligence model used broadly in water resource aspects is adaptive neuro-fuzzy interface system (ANFIS) taking in to account the membership functions (MF) on the basis of the smoothness characteristics and mathematical components each for set of input data. All researches in hydrological estimation used ANFIS, merely a type of MF has been noticed for all sets of inputs without considering the response of each of them. This study is applying a specified certain MFs for each type of input to improve the accuracy of ANFIS model in forecasting the water level in Klang Gates Dam in Malaysia. On the basis of the previous studies, two most popular MFs, Generalized Bell Shape MF and, Gaussian MF, are employed for examine the new pattern in two inputs ANFIS architecture resulted less stress in error performance, and higher accuracy in estimation, compare to the traditional ANFIS model. The aim is achieved by evaluating the performance in and fitness of the model in daily reservoir estimation.     

A Probabilistic Nonlinear Model for Forecasting Daily Water Level in Reservoir

Accurate prediction and monitoring of water level in reservoirs is an important task for the planning, designing, and construction of river-shore structures, and in taking decisions regarding irrigation management and domestic water supply. In this work, a novel probabilistic nonlinear approach based on a hybrid Bayesian network model with exponential residual correction has been proposed for prediction of reservoir water level on daily basis. The proposed approach has been implemented for forecasting daily water levels of Mayurakshi reservoir (Jharkhand, India), using a historic data set of 22 years. A comparative study has also been carried out with linear model (ARIMA) and nonlinear approaches (ANN, standard Bayesian network (BN)) in terms of various performance measures. The proposed approach is comparable with the observed values on every aspect of prediction, and can be applied in case of scarce data, particularly when forcing parameters such as precipitation and other meteorological data are not available.     

A neuro-fuzzy model for inflow forecasting of the Nile river at Aswan high dam

River flow forecasting is an essential procedure that is necessary for proper reservoir operation. Accurate forecasting results in good control of water availability, refined operation of reservoirs and improved hydropower generation. Therefore, it becomes crucial to develop forecasting models for river inflow. Several approaches have been proposed over the past few years based on stochastic modeling or artificial intelligence (AI) techniques. In this article, an adaptive neuro-fuzzy inference system (ANFIS) model is proposed to forecast the inflow for the Nile River at Aswan High Dam (AHD) on monthly basis. A major advantage of the fuzzy system is its ability to deal with imprecision and vagueness in inflow database. The ANFIS model divides the input space into fuzzy sub-spaces and maps the output using a set of linear functions. A historical database of monthly inflows at AHD recorded over the past 130 years is used to train the ANFIS model and test its performance. The performance of the ANFIS model is compared to a recently developed artificial neural networks (ANN) model. The results show that the ANFIS model was capable of providing higher inflow forecasting accuracy specially at extreme inflow events compared with that of the ANN model. It is concluded that the ANFIS model can be quite beneficial in water management of Lake Nasser reservoir at AHD.     

基于 Embedding-GRU 的水库水位预测模型

为解决因水库数据采集设备能力有限、水文数据不全导致预测水库水位时预测精度较低的问题，以四岭水 库每小时水位监测数据为例，提出基于嵌入式-门控循环单元（Embedding-gated?recurrent?unit，Embedding-GRU）的 水库水位预测模型，即利用 Embedding 方法将单维降雨量数据升维至多维数据，扩大降雨的气候特征，结合 GRU 算法进行水库水位预测。将该模型与传统深度学习算法长短期记忆（long?short-term?memory，LSTM）、门控循环单 元（gated?recurrent?unit，GRU）、双向门控循环单元（bidirectional?recurrent?neural?network，BiGRU）这 3 种模型对比， 结果显示：Embedding-GRU 模型的预测效果均优于其他传统模型，平均绝对误差 EMA和均方根误差 ERMS分别平均 下降 19.6% 和 7.7%，并且在预测次日水库水位的应用场景中决定系数 R2能够达到 0.989?37。结果表明：该模型耦 合多种算法，扩大单变量的气候特征，具有较高预测精度和泛化能力。相较传统模型，基于 Embedding-GRU 的水 库水位预测模型能够对缺少温度、气压、风速、蒸发量等监测数据的水库进行可靠度较高的预测，适用水库范围 更广，为水库日常运维、除险加固提供参考。     

利用神经网络-缺省因子法及熵权法分析气象因素对水面蒸发的影响

根据新疆车尔臣流域且末县气象站2007年非结冰期(4-9月)日水面蒸发量及相关常规气象观测资料,利用神经网络-缺省因子法及熵权法分析了各气象要素对水面蒸发的影响程度。结果表明:水面蒸发对温度与风速最为敏感。希望从气象因素角度出发,为区域水资源优化调度提供参考。     

淮北平原基于水文气象多因子的土壤水分动态预测

为研究土壤水分动态变化,利用五道沟水文实验站1989-2015年水文气象和大田土壤水实测资料,采用灰色关联度和线性回归分析,建立了冬小麦各生长阶段不同土层土壤水分预测模型。结果表明:不同土层土壤水分与气象因子的关联度一致;不同生长阶段土壤水分与气温和地下水埋深关联度最强,分别达0. 92和0. 95;分蘖-越冬期,土壤水分与地下水埋深和日照时数关联度最强,其他生长阶段,土壤水分与气温和地下水埋深关联度最强。通过水文气象因子模拟土壤水分拟合度较高,R~2达0. 94。不同生长阶段不同土层,土壤水分计算模型均具有良好的预测能力,R~2达0. 80。成果为实施作物不同生长阶段的灌溉计划提供科学依据。     

Determination of Mean Velocity and Discharge in Natural Streams Using Neuro-Fuzzy and Neural Network Approaches

The applicability of artificial neural networks (ANNs) and the adaptive neuro-fuzzy inference system (ANFIS) for determination of mean velocity and discharge of natural streams is investigated. The 2,184 field data obtained from four different sites on the Sarimsakli and Sosun streams in central Turkey were used in the study. ANNs and ANFIS models use the inputs, water surface velocity and water surface slope, to estimate the mean velocity and discharges of natural streams. The accuracies of both models were compared with the multiple-linear regression (MLR) model. The comparison results showed that the ANFIS model performed better than the ANNs and regression models for estimating mean velocity and discharge. The ANN model also showed better accuracy than the MLR model. The root mean square errors (RMSE) and mean absolute relative errors (MARE) of the MLR model were reduced by 88 and 91 % using the ANFIS model in estimating discharges, respectively. It is found that the optimal ANFIS model with RMSE of 0,063, MARE of 3,47 and determination coefficient (R²) of 0,996 in the test period is superior in estimation of discharge than the MLR model with RMSE of 0,532, MARE of 38,9 and R² of 0,776, respectively. The study reveals that the ANFIS technique can be successfully used for estimating the mean velocity and discharge of natural streams by using only the inputs of water surface velocity and water surface slope.