本课题在国内外学者关于BP算法的研究基础上，通过理论研究与分析神经网络，对BP算法进行了改进，并通过对分体式空调器室内机的换热性能模拟计算和实验，验证了模拟的改进效果，得出以下结论：

1．利用直接式蒸发器的实验数据，构成神经网络的训练样本集，并进行了适当的数据预处理，进而对该换热器神经网络模型在不同学习速率、不同中间层单元数以及不同数值训练算法方面的效果进行比较，经过讨论与分析得出：

（1）神经网络的中间层单元数强烈地影响着训练的收敛结果，将中间层单元数设置较大，可以使网络模型有效地避开局部极小，最终收敛于全局最小值；而学习速率则较强烈地影响着网络训练的稳定性，易导致局部极小的产生，即当学习速率设置为0.4以上时，网络将得不到正确的输出；

（2）在提高网络训练速度方面，共轭梯度算法和Levenberg-Marquardt算法可以获得比梯度下降算法更快的训练速度。

2．通过对实测的蒸发器实验数据进行神经网络计算，可知经过以上对BP算法的改进，本文提出的改进后的神经网络BP算法可以更加快速更加可靠地学习到隐含在大量数据中的内在规律性，为空调系统的仿真计算提供了可靠的方法和手段。

3．归纳与汇集了数据归一化处理的计算公式，有利于编程上的数据预处理和程序的实现。

4．提出了确定学习样本集的原则，为神经网络的正确学习提供了一定的参考。

5．本文通过分析Matlab数值计算软件中神经网络工具箱，提出了运用其进行换热器仿真计算的新途径，实践证明，Matlab神经网络工具可以成功地应用于换热器神经网络模型的建立和计算，并得出了有用的结论，为今后在空调换热器研究领域运用此先进数学工具进行科学研究提供了借鉴。

Based on the researches to BP algorithm inside and outside, the study improves BP algorithm, and applies it to the simulating experiments of the transfering performances for the indoor part of parted Air-conditioner, proves the improved effects of the simulating experiments, and acquires the following conclusions:

(1) Using the datas of the experiments for Direct Evaporator, the study establishes the Training samples of NN, makes properly the data pre-resolving, calculates the NN model of Direct Evaporator in different Learning Rates, different Unit numbers of the middle layer and different Training algorithms, and concludes by discussing and analysing:

1) The Unit numbers of Neural network intermediate level influence strongly the convergence results of Training. When it is counted to about 24, it will make the NN model avoid effectively the partly minimums and finally reach the general minimum; Learning Rate affects strongly the Training stability, and leads to the partly minimum happening, that is, as it is above 0.4, the correct results of the NN model can not be gained.

2) As for increasing the Training speed, Gradient algorithm of Conjugation and Levenberg-Marquardt algorithm can provide the faster descending speed than BP algorithm does.

(2) The article uses Matlab Numerical Calculating software, and makes a new path

to study the performances of Transfers with using computer tools. It not only accomplishes the data samples pre-resolving, but constructs the NN model and calculates it, which gives some references for researchers to study in it.

(3) The study analyses and calculates the experimental data of Evaporators, and

know that with the improvements of BP algorithm, the NN model can learn the internal principles hidden in quantity of data more quickly and reliably, which supports the further simulating of Air-condition systems.

(4) It lays calculation formulas on data generalizing, which are served for pro_

gramming, and suggests the principles desiding Learning samples that resure the right learning of NN.