【作者】 赵新宇；

【导师】 费良军；

【作者基本信息】 西安理工大学 ， 水文学及水资源， 2007， 博士

【摘要】 退水广泛存在于我国的大型灌区中,它主要由田面弃水、引水渠退水、地下水排水、降水、山洪、工业废水及生活污水组成。我国西北地区的灌区年退水量接近200亿m3,一些灌区的退水量可达引水量的40%～60%,这部分水量的重复利用不仅对于灌区,而且对整个区域的水资源管理都具有十分重要的意义。由于退水量的组成和时空分布十分复杂,目前国内外关于退水量预测理论与方法的研究较少,而且不够深入。在查阅国内外相关文献资料的基础上,本文采用理论与方法研究和实例应用相结合的技术路线,以宁夏青铜峡灌区为例,研究了灌区退水的变化规律,确定了退水的组成及主要影响因素,将灌区退水分为年退水、月退水、日退水等3种不同的类型,对灌区退水量预测的理论与方法进行了研究。论文的主要研究成果为:(1)通过分析宁夏青铜峡灌区灌溉、排水和降水等长系列历史观测资料,研究了宁夏青铜峡灌区产生大量退水的主要原因,揭示了灌区水资源的转化规律。(2)研究了灌区退水量的组成,揭示了退水量的年际和年内变化规律,采用灰色关联度和相关分析的方法确定了退水量的主要影响因素为引水量、地下水位和降水量。(3)以年退水量为研究对象,根据灌区退水量与其影响因素相关关系比较明显,年退水规律相对容易描述的特点,采用多元逐步回归方法建立了灌区年退水量预测模型,模型结构简单,物理意义明确,对样本要求不大,预测精度在7%以内。(4)灌区月退水量的变化呈一定的周期性,但随机波动较大,用传统数值分析方法很难进行模拟,本文利用神经网络的强模拟能力和在线学习能力,建立了灌区月退水量预测的神经网络模型。实例应用表明,在样本数量较大的情况,退水量神经网络模型具有较高的模拟和预测精度。另外,针对BP神经网络训练速度慢、易于收敛于局部最小点、过度训练和外推能力差的缺点,本文从建模、样本选取、样本数量、算法选取、网络训练、模型更新等6个方面讨论了克服这些缺点的方法,实例分析表明这些方法是切实有效的。(5)神经网络模型要取得好的外推能力和预测效果,必须有大量的训练样本作保证,但许多灌区的月退水量资料序列较短,训练样本的数量较少,达不到神经网络建模的要求。因此,本文引入了以结构化风险最小为原则,适合小样本建模的支持向量机,建立了灌区月退水量预测的支持向量机模型,并探讨了支持向量模型参数的训练方法。实例应用表明,在样本数量较少的情况下,支持向量机模型的预测能力高于神经网络模型。(6)通过对灌区日退水量时间序列的研究,揭示了宁夏青铜峡灌区日退水量时间序列的非平稳性和1阶差分平稳性,建立了灌区日退水量预测的时间序列模型,探讨了退水量时间序列模型预测的新息修正方法。实例应用表明模型能较好地模拟灌区日退水量的变化规律,平均相对误差为7.88%,预测精度较高。(7)通过对灌区日退水量时间序列进行混沌识别和相空间重构,揭示灌区日退水量具有混沌特征,确定了其重构相空间的时间延迟、嵌入维数和关联维数。在此基础上,将相空间重构理论与神经网络方法相结合,建立了灌区日退水量预测的混沌神经网络模型,并与日退水量预测的时间序列方法进行了对比分析。实例应用表明,混沌神经网络模型的模拟和预测精度方面要高于时间序列模型,但时间序列模型的建模和使用复杂程度要低于混沌神经网络模型。更多还原

【Abstract】 The return water widely exists in the large irrigation area of our country. It is mainly composed of the abandoned water of field, return water of diversion ditch, drainage of ground water, precipitation, flash floods, industrial wastewater and domestic sewage. The return water volume of northwest china is nearly 200 billion m3, and it even accounting for 40% -60% of annual diversion. in some irrigation area. The study on return water has great significance not only to the irrigation area, but also to water resources management of whole district. However there is only little research of theories and methods regarding return water volume, and the research is not enough due to the composition and the spatial and temporal distribution of return water is very complex. Utilizing the approach combining the theoretical methods and application example, taking Qingtongxia irrigation area as example, the paper analyses the variation regularity of return water of irrigation area, ascertains the composition and main influencing factors of return water, classifies the return water into three different types: annual return water, monthly return water and daily return water, and researches the theory and method of predicting return water of irrigation area. Among the key findings:(1)The paper reveals the variation regularity of water and researches the reason of generating a great deal of return water in irrigated area by analyses of long historical observation data of Ningxia Qingtongxia irrigation area,,(2)The paper reveals the variation regularity of return water in multi-years and single-year, and ascertains the main influencing factors of return water using methods of the grey relational degree analysis and correlation analysis; they are diversion, precipitation and groundwater table.(3)The paper establishes the multiple stepwise regression methods of predicting annual return water of irrigated area. There are obvious correlation between return water and itseffective factors, but the modeling samples are few. The multiple regression model is simple in structure and clear in physical meaning, and no more requirement for sample. The predicting precision of annual return water model is less than 7%.(4)The paper finds that the monthly return water of irrigation area changes in a certain periodicity, however, the random variation of the change is large. It is difficult to simulate using traditional numerical analysis methods. The paper advances the neural network method of predicting monthly return water of irrigation area taking advantage of strong simulation capabilities and online learning ability of neural network. In addition, the papers compares the various neural network algorithm, and gives the suggestions for improving training method contraposing the disadvantage of BP neural network: slow training speed, more converging to the local minimum and over-training. The case analysis shows that improved method can overcome those three disadvantages and advance the training speed and precision.(5)It need a large number of training samples to ensure the neural network to achieve good extrapolation capability and predicting results. But in many irrigation area, there are shorter data sequences of return water and few training samples, can not meet the requirement of establishing neural network model. Therefore, this paper introduces the Support Vector Machine (SVM) method which is based on the principle of structure risk minimization and fit to establishing model with few samples, advances the SVM method of predicting monthly return water of irrigation area, discusses the training parameters methods of SVM method, and compares it with the predicting method based on neural network.The case application shows that the prediction ability of SVM model is better than neural network model.when the sample is enough.(6)The paper finds the nonstationarity of time-series and stationarity of first-order difference of return water of Qingtongxia irrigation area by studying the time series of daily return water of the irrigation area. The paper advances the time-series methods of predicting daily return water and discusses the information improvement methods of the time-series predicting methods of daily return water. The case application shows that the modle can simulate the change regularity of diary return water well, and the average relative error of the predicted results is 7.88%.(7)By carrying chaotic identification and phase space reconstruction on daily return water of irrigation area. The papers finds the chaotic character of daily return water of irrigation area, ascertains the time delay, embedment dimension and correlation dimension of phase spacereconstruction, then paper combines the phase space reconstruction theory and neural network methods, advances the chaos-neural network predicting method of daily return water, and compares it with the time-series predicting methods of daily return water.The cases appliation shows that the chaos-neural network modle is more accurate than the times series modle,but it is also more complex in modle model building and application.更多还原