【作者】 胡冰；

【导师】 王丽学；

【作者基本信息】 沈阳农业大学 ， 农业水土工程， 2007， 硕士

【摘要】 实际作物需水量预测是实时灌溉预报的核心，而实时灌溉预报又是编制与执行灌区动态用水计划的必要条件，是实现灌区节水的核心内容，只有实时灌溉预报可靠、准确，动态用水计划才可能符合实际，才能发挥指导用水，以取得节水、高产、高效益的效果。今后灌溉预报的发展趋势是开展实时灌溉预报，因此研究实时灌溉预报的核心——作物需水量就具有了重要的理论意义与实用价值。在实际作物需水量预测中参考作物需水量的预测是最基础的。故本文结合沈阳市浑蒲灌区实际资料重点研究了逐日参考作物需水量的预测模型，探讨了作物需水量的确定方法，之后对作物需水量进行了预测。文中介绍了BP神经网络和粒子群优化算法。BP网络包含神经网络理论中最为精华的部分，它的优点在于结构简单、可塑性强，数学意义明确，学习算法步骤分明；但有其自身的缺陷，如收敛速度慢、易陷入局部最小值、冗余性大、学习和记忆的不稳定性等。粒子群算法(PSO)是一种群体智能算法，它具有发现能力强、记忆功能强、效率较高、收敛快、操作简单、易于与其他方法结合、易于编程实现等优点。本文将二者相结合，建立了基于粒子群优化的BP神经网络预测模型(PSO-BP)，该模型融和了粒子群算法全局寻优功能和BP网络的映射能力强的优点，克服了BP网络的学习速度慢、易陷入局部极小值、网络稳定性差等缺点，适应性强，预测能力好。在PSO-BP的基础上对沈阳市浑蒲灌区参考作物需水量进行了预测。本研究采用沈阳市浑蒲灌区1995-2004的逐日气象资料作为基础资料。通过对各种气象数据的分析，取日平均气温、日照时数和日序数作为神经网络的输入向量；逐日参考作物需水量作为网络的输出向量；网络的隐含层神经元数通过模型实验确定为4；采用三层反向传播网络，之后运用Matlab7.0编程对网络进行训练。以1997年相关数据作为测试样本为例，分别用未经优化的BP神经网络和PSO-BP网络对参考作物需水量进行预测。未优化的BP网络采用与PSO-BP完全相同的样本数据和训练函数。由未经优化的BP神经网络和PSO-BP网络预测数据与原始数据的比较分析来看，应用PSO-BP网络预测的参考作物需水量的最大相对误差为25.15％，该误差小于未经优化的BP网络预测相对误差。1997年预测数据为137个，经PSO-BP预测的数据相对误差绝对值小于10％的有101个，小于15％的有22个，小于20％的有8个，小于25％的有5个，小于30％的有1个，预测的合格率较高。网络预测结果曲线与原始数据曲线基本重合，证明了PSO-BP预测能力强。从BP神经网络和PSO-BP网络预测结果中选取部分预测数据进行对比，可以看出PSO-BP网络预测结果更接近实际值。在训练的过程中，PSO-BP迭代400次就可达到精度要求，而普通BP网络需经20000次训练才可达到精度要求。综述了作物需水量的测定与计算方法，之后根据浑蒲灌区的实际情况，选用单系数法计算浑蒲灌区作物需水量，其中K_c采用了中国科学院沈阳生态实验站测定的结果；将选取的K_c与PSO-BP预测的的参考作物需水量相结合来预测灌区逐日作物需水量。更多还原

【Abstract】 The crop evapotranspiration forecast is core content of the real-time irrigation forecast. Real-time irrigation forecast is the necessary conditions of establishment and execution irrigation area dynamic irrigation water plan. It is also the core content of realizing irrigation area irrigation water-saving. Only real-time irrigation forecast reliable, accurate, dynamic water plan can be practical and play guiding water, to obtain water-saving, high-yield and high benefit effect. The irrigation forecast development tendency is the development real-time irrigation forecast, therefore the study of crop evapotranspiration has the important theory significance and the practical value.The reference crop evapotranspiration forecast is the most basic in crop evapotranspiration forecast. Therefore, according to Shenyang City Hunpu irrigation practical information this paper focuses on daily reference crop evapotranspiration forecast model, and discuss on the crop evapotranspiration and field soil moisture forecast.The paper introduces the back propagation（BP） network and particle swarm optimization（PSO）. The BP network contains the essence part of the neural network, its simple structure, high plasticity, clear mathematical significance, distinct learning algorithm steps; but it has its own shortcomings, such as slow convergence, easy to fall into a local minimum, redundancy, learning and memory of instability and so on. Particle Swarm Optimization （PSO） is a smart intelligence algorithm, it has many advantages, such as strong discovery ability, strong memory function, higher efficiency, fast convergence, simple operation, easy to integrate with other methods, easy to program realization and so on. A model of PSO-BP is established for predicting ET₀.This model integrates PSO global optimization function and BP network mapping capability advantages, and overcomes the shortcomings such as network slow learning,easy to fall into the local minimum and network poor stability. This model has adaptability and good predictive ability.The reference crop evapotranspiration of Hunpu irrigation area is predicted by PSO-BP. The daily weather information from 1995 to 2004 is as basis information of this study. Through the various meteorological data analysis, the daily average temperature, sunshine hours and the sequence number are selected as the neural network input vector; daily reference crop evapotranspiration is as network output vector. The number of hidden layer is one.The number of hidden layer neuron is four that is determined by model experiments. Matlab7.0 is used for programming and training. To 1997 data as test samples as an example, the reference crop evapotranspiration is forecasted by BP network and PSO-BP. BP network and PSO-BP use the same sample data and training function. Through the analysis of the forecast data and original data, it shows that the maximum relative error of the reference crop evapotranspiration forecasted by PSO-BP is 25.15% and the error is smaller than BP network’s. Forecast data of 1997 is 137. PSO-BP forecasting data relative absolute error less than 10% is 101,less than 15% is 22, less than 20% is 8, less than 25% is 5 and less than 30% is 1.The forecast qualified rate is high. The network forecast result curve and the primary data curve basically superposes; it proves the PSO-BP forecast ability is strong. Selecting some data from BP networks and PSO-BP forecast results and comparing these data,the comparison result shows that PSO-BP forecast results is closer to actual value. In the training process, PSO-BP iteration 400 precision can be achieved, BP network requires 20,000 training to achieve the required precision.The crop evapotranspiration determination method and computation method is summarized. According to Hunpu irrigation area actual situation, this paper chooses single crop coefficient approach to calculate Hunpu irrigation area crop evapotranspiration. K_c adopts the CAS Shenyang Experimental Station results. Irrigation area crop evapotranspiration is forecasted by K_c and PSO-BP forecast reference crop evapotranspiration.更多还原