Please note this is a comparison between Version 1 by Feihu Sun and Version 3 by Catherine Yang.

Agricultural price prediction is a hot research topic in the field of agriculture, and accurate prediction of agricultural prices is crucial to realize the sustainable and healthy development of agriculture. Compared with econometric and mathematical-statistical methods, intelligent forecasting methods have fewer restrictions and assumptions in modeling and can effectively model nonlinear relationships in price series.农产品价格预测是农业领域的研究热点，准确预测农产品价格是实现农业持续健康发展的关键。智能预测方法在建模中的限制和假设较少，可以有效地对价格序列中的非线性关系进行建模。传统的机器学习方法，如决策树、支持向量机和普通贝叶斯，具有简单、训练快速和鲁棒性等优点，但处理复杂非线性关系的能力有限，需要人工选择和提取特征，泛化能力不足。深度学习模型凭借其强大的表达和特征提取能力，无需依赖特征工程即可从原始序列中提取有效的特征信息，在监督有效、数据量充足、数据质量高的情况下，对序列中的非线性关系具有较好的处理能力。

- price forecasting
- combined models
- intelligent prediction methods

The support vector machine (SVM) is a machine learning approach rooted in statistical learning theory ^{[1]}[19]. It hinges on VC dimensional theory, the principle of structural risk minimization ^{[2][3]}[20,21], and represents the pioneering algorithm grounded in geometric distance ^{[4]}[22]. Serving as a small-sample learning technique with a robust theoretical foundation, an SVM’s final decision function is influenced by only a handful of support vectors. Its computational complexity hinges on these vectors rather than the sample space’s dimensionality, sidestepping the so-called “dimensional disaster”. Wang et al. ^{[5]}[23] harnessed SVM to predict the nonlinear facet of garlic prices, coupling it with ARIMA for linear price prediction, yielding accurate results. Nevertheless, SVM does have drawbacks, including diminished performance when data features (dimensions) surpass the sample size, sensitivity to parameters and kernel functions. Consequently, approaches like parameter optimization are frequently employed to enhance SVM prediction performance. Duan et al. ^{[6]}[24] employed a genetic algorithm to identify optimal parameter combinations for a support vector regression model. With these optimized parameters, they constructed a support vector regression model for predicting fish prices, yielding precise outcomes with minor errors. SVR’s remarkable ability to manage high-dimensional, nonlinear, and small-sample data positions is a vital technique in agricultural price prediction.

A Bayesian network is essentially a directed acyclic graph that uses probabilistic networks to make uncertainty inferences. The excellence of Bayesian networks in solving agricultural price forecasting as well as other agricultural problems stems mainly from the following key features: (贝叶斯网络本质上是一个有向无环图，它使用概率网络进行不确定性推断。贝叶斯网络在解决农产品价格预测和其他农业问题方面的卓越表现主要源于以下主要特点：（1) Bayesian networks can handle incomplete datasets; (2) Bayesian networks allow one to understand the relationships between variables and quantify the strength of these relationships; (3) the ability to combine quantitative and qualitative data; (4) the ability to combine expert knowledge and data into Bayesian network; and (5) Bayesian methods can relatively easily avoid data overfitting during the learning process. ）贝叶斯网络可以处理不完整的数据集;（2）贝叶斯网络允许人们理解变量之间的关系并量化这些关系的强度;（3）定量和定性数据相结合的能力;（4）将专业知识和数据结合到贝叶斯网络中的能力;（5）贝叶斯方法可以相对容易地避免学习过程中的数据过度拟合。Putri ^{[7]} used Bayesian network algorithms as a data mining classification method to predict pepper commodity prices in Bandung region based on weather information. One disadvantage of Bayesian networks is that they do not support ring networks ^{[8]}, which would weaken the robust inference capability of the network, and this limitation is not friendly to static Bayesian networks. Dynamic Bayesian network ([25]使用贝叶斯网络算法作为数据挖掘分类方法，根据天气信息预测万隆地区的胡椒商品价格。贝叶斯网络的一个缺点是它们不支持环形网络[26]，这会削弱网络的鲁棒推理能力，并且这种限制对静态贝叶斯网络并不友好。动态贝叶斯网络（DBN) is a dynamic model amalgamating probability theory and influence diagram. It combines a time-varying hidden Markov model with a traditional static Bayesian network, capturing benefits from both while sidestepping their limitations through dynamic adaptability over time and the incorporation of new states ^{[9]}. Ma Zaixing ^{[10]} used the ）是一个融合概率论和影响图的动态模型。它将时变隐马尔可夫模型与传统的静态贝叶斯网络相结合，从两者中获益，同时通过随时间推移的动态适应性和新状态的合并来避开它们的局限性[27]。马再兴[28]利用PC algorithm to learn from data, construct according to expert knowledge, and combine expert knowledge and 算法从数据中学习，根据专家知识进行构造，并将专家知识和PC algorithm to perform structural learning. After obtaining the initial structure, he adjusted the obtained initial structure to obtain the network structure of the model, and then used the EM algorithm to perform parameter learning. Moreover, he obtained a complete dynamic Bayesian network model for price prediction, and selected the best model based on the prediction results to predict the price and output of live pigs. The results show that the prediction effect is better than the control group’s ARIMA, SVM, and BP neural network models.算法相结合进行结构学习。得到初始结构后，他调整得到的初始结构得到模型的网络结构，然后用EM算法进行参数学习。此外，他获得了用于价格预测的完整动态贝叶斯网络模型，并根据预测结果选择最佳模型来预测生猪的价格和产量。结果表明，预测效果优于对照组的ARIMA、SVM和BP神经网络模型。

Neural神经网络通常被称为人工神经网络 networks are commonly referred to as artificial neural networks (ANN). They constitute a complex nonlinear network system composed of numerous processing units interconnected in a manner resembling biological neurons. Neural networks exhibit robust nonlinear fitting capabilities, enabling them to map intricate nonlinear relationships. Furthermore, their learning rules are simple, making them easily implementable on computers. They possess strong robustness, memory, nonlinear mapping abilities, and powerful self-learning capabilities, showcasing unique advantages in addressing agricultural commodity price prediction challenges. In 1987, （ANN）。它们构成了一个复杂的非线性网络系统，由许多以类似于生物神经元的方式相互连接的处理单元组成。神经网络表现出强大的非线性拟合能力，使它们能够映射复杂的非线性关系。此外，他们的学习规则很简单，可以在计算机上轻松实现。它们具有较强的鲁棒性、记忆力、非线性映射能力和强大的自学习能力，在应对农产品价格预测挑战方面具有独特的优势。1987年，Lapedes and 和Farber ^{[11]} pioneered the application of neural networks to forecasting, marking the inception of neural network predictions. In [29]率先将神经网络应用于预测，标志着神经网络预测的开始。1993, 年，Kohzadi et al. ^{[12]} were among the first to employ feed-forward neural networks for predicting US wheat and cattle prices. They compared the predictive results with those from 等人[30]是最早使用前馈神经网络来预测美国小麦和牛价格的人之一。他们将预测结果与ARIMA, concluding that neural networks exhibited superior turning point prediction capabilities and achieved more accurate price forecasting.的结果进行了比较，得出的结论是神经网络表现出卓越的转折点预测能力，并实现了更准确的价格预测。

As big data and artificial intelligence technology advance, neural networks find increasingly wide application in the agricultural domain ^{[13]}. In the realm of price prediction, prevalent neural network models are as follows (**Table 1**), along with examples, summarizes the applications of neural networks in agricultural commodity price prediction):随着大数据和人工智能技术的进步，神经网络在农业领域的应用越来越广泛[31]。在价格预测领域，流行的神经网络模型如下（**表2**和示例总结了神经网络在农产品价格预测中的应用）：

- 径向基函数神经网络（Radial Basis Function Neural Networks (RBFNN)
^{[17][18]}: A BFNN）[35，36]：BP network is a global approximation of a nonlinear mapping, whereas an 网络是非线性映射的全局近似，而RBF network is a local approximation of a nonlinear mapping and is faster to train. RBF can handle complex nonlinear relationships and has good generalization ability. However, it is sensitive to the network structure and hyperparameters, and the training and tuning are relatively complicated. When the problem involves complex nonlinear relationships and there is enough training data, you can try to use RBF neural network.网络是非线性映射的局部近似，训练速度更快。RBF可以处理复杂的非线性关系，具有良好的泛化能力。但是，它对网络结构和超参数敏感，训练和调优相对复杂。当问题涉及复杂的非线性关系并且有足够的训练数据时，您可以尝试使用 RBF 神经网络。

^{[} | ^{26} | ^{]} | [ | 44 | ] |

Models |

- 长短期记忆网络（
- Long Short-Term Memory Networks (LSTM)
^{[19][20]}: STM）[37，38]：LSTM neural network is a special kind of recurrent neural network that solves the problems of long-term dependency and gradient vanishing by introducing structures, such as forgetting gates, input gates, and output gates, to control the flow of information through the unit states. 神经网络是一种特殊的递归神经网络，它通过引入遗忘门、输入门和输出门等结构来控制通过单元状态的信息流，从而解决长期依赖性和梯度消失的问题。LSTM neural networks have the ability to memorize and capture long-term dependencies. Therefore, LSTM is a good choice when the prediction problem involves time series data, especially with long-term dependencies.神经网络具有记忆和捕获长期依赖关系的能力。因此，当预测问题涉及时间序列数据时，LSTM 是一个不错的选择，尤其是具有长期依赖关系的数据。

模型 | |||||||

/Algorithms算法 |
BPNN | RBFNN | LSTM | CNN美国有线电视新闻网 | Chaotic neural networks混沌神经网络 | ELM榆树 | WNN |

Characteristic特征 |
Strong nonlinear mapping ability, high self-learning and self-adaptive ability, ability to apply learning outcomes to new knowledge, and certain fault tolerance. Research results show that the 较强的非线性映射能力，较高的自学习和自适应能力，将学习成果应用于新知识的能力，具有一定的容错能力。研究结果表明，BP neural network model has the long-term prediction ability for the futures market.神经网络模型对期货市场具有长期预测能力。 | It has better approximation ability, classification ability, and learning speed than 它比BP neural network, simple structure, concise training, fast learning convergence speed, can approximate any nonlinear function, and overcome the local minimum problem. Research results show that the influencing factors of soybean price are different at different price levels, and the construction of this model is beneficial to the prediction of soybean price.神经网络具有更好的逼近能力、分类能力和学习速度，结构简单，训练简洁，学习收敛速度快，可以逼近任何非线性函数，并克服局部最小问题。研究结果表明，不同价格水平下大豆价格的影响因素不同，该模型的构建有利于大豆价格的预测。 | It effectively overcomes the problem of gradient vanishing caused by the increase in network layers in 它有效地克服了RNN. This model is especially suitable for tasks with very long time intervals and delays, and has excellent performance. Research results show that parameter tuning has a large impact on the prediction effect of LSTM network model, and the main parameters with large impact include iteration times, learning rate, window size, and network layers. Compared with ARIMA model, MLP model and SVR model, LSTM network model has higher accuracy in prediction results.中网络层增加导致的梯度消失问题。该型号特别适用于时间间隔和延迟非常长的任务，并且具有出色的性能。研究结果表明，参数调优对LSTM网络模型的预测效果影响较大，影响较大的主要参数包括迭代次数、学习速率、窗口大小和网络层。与ARIMA模型、MLP模型和SVR模型相比，LSTM网络模型对预测结果的准确性更高。 | The effectiveness of CNN in feature extraction and autonomous learning of nonlinear patterns makes it perform well in image classification and audio recognition tasks. This study reviews the factors that affect crop yield and proposes a 在特征提取和非线性模式自主学习方面的有效性使其在图像分类和音频识别任务中表现良好。本研究综述了影响作物产量的因素，并提出了一个3D CNN model to predict future crop prices. The model helps decision-makers to better predict crop price trends and formulate strategic plans, select trade partners, reduce costs, and solve food insecurity issues.模型来预测未来的作物价格。该模型可帮助决策者更好地预测作物价格趋势并制定战略计划，选择贸易伙伴，降低成本并解决粮食不安全问题。 | The output of the network not only depends on the current input, but also on the past output. 网络的输出不仅取决于当前输入，还取决于过去的输出。经过训练后，网络对非线性数据的适应性会更好，非常适合预测复杂、非平稳、非线性的时间序列。所设计的基于动态混沌神经网络的马铃薯价格时间序列预测模型在预测精度和性能上优于After training, the network will have better adaptability to nonlinear data and is very suitable for predicting complex, non-stationary, and nonlinear time series. The designed potato price time series prediction model based on dynamic chaotic neural network has clear advantages over ARMA model in prediction accuracy and performance.RMA模型。 | The algorithm can randomly generate the input weights and hidden layer thresholds required by the neural network without multiple adjustments. As long as the number of hidden layer nodes is reasonable, a unique optimal solution can be obtained. Its parameter setting process is simple, does not need to be adjusted repeatedly, the training speed is significantly improved, and the prediction results are more accurate. Compared with traditional neural network learning algorithms (such as 该算法可以随机生成神经网络所需的输入权重和隐藏层阈值，而无需多次调整。只要隐藏层节点数量合理，就可以得到唯一的最优解。其参数设置过程简单，不需要反复调整，训练速度明显提高，预测结果更准确。与传统的神经网络学习算法（如BP algorithm), it overcomes the disadvantage of falling into local optimum. This study uses 算法）相比，克服了陷入局部最优的缺点。本研究采用PCA-ELM model to predict grain prices and achieves good prediction results.模型预测粮食价格，取得了较好的预测效果。 | Wavelet neural network combines the advantages of neural network and wavelet function, using 小波神经网络结合了神经网络和小波函数的优点，采用Morlet wavelet as the hidden layer basis function, which can extract local dynamic features, and can build a local approximation feed-forward neural network, reduce the interference between nodes, and improve the prediction accuracy. This study uses wavelet neural network to predict the prices of two kinds of Chinese medicinal materials, Radix Codonopsis and Angelica sinensis, and the results show that the prediction error is very small and the prediction accuracy is very high.小波作为隐层基函数，可以提取局部动态特征，并且可以构建局部近似前馈神经网络，减少节点之间的干扰，提高预测精度。本研究利用小波神经网络对党参和当归两种中药材的价格进行预测，结果表明，预测误差很小，预测准确率很高。 |

Agricultural Product农产品 |
Egg蛋 | Soybeans大豆 | Soybeans大豆 | Five different Crops五种不同的作物 | Potato土豆 | Grain粮食 | Chinese herbal medicine中草药 |

Observed Features观察到的特征 |
Soybean meal price, cull chicken price, corn price, egg seedling price, duck egg price豆粕价格，剔除鸡价格，玉米价格，鸡蛋苗价格，鸭蛋价格 | Domestic国内大豆生产， Soybean Production, Soybean Imports, Global Soybean Production, Domestic Soybean Demand, Consumer Price Index, Consumer Confidence Index, Money Supply, Imported Soybean Port Delivery Prices大豆进口， 全球大豆产量， 国内大豆需求， 居民消费价格指数， 消费者信心指数， 货币供应， 进口大豆港货价格 | Price time series价格时间序列 | Environmental, economic, and commodity trading data环境、经济和商品交易数据 | Price time series价格时间序列 | Total grain production, per capita grain consumption, average grain production price index, per capita disposable income of urban residents, consumer price index, grain sown area粮食总产量、人均粮食消费量、粮食生产平均价格指数、城镇居民人均可支配收入、居民消费价格指数、粮食播种面积 | Planting area, yield, province’s disaster area, hype factor, and market demand种植面积、产量、省灾区、炒作因素、市场需求 |

Evaluation Method评价方法 |
Mean Absolute Percentage Error平均绝对百分比误差 | Mean Absolute Percentage Error, Relative error平均绝对百分比误差、相对误差 | Mean Absolute Error, 平均绝对误差、均方根误差、平均绝对百分比误差、Root Mean Square Error, Mean Absolute Percentage Error, R-Square 平方 | Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error平均绝对误差、均方根误差、平均绝对百分比误差 | Mean Square Error均方误差 | Mean Square Error均方误差 | Relative error相对误差 |

- 卷积神经网络（
- C
- onvolutional Neural Networks (CNN)
^{[21]}: NN）[39]：CNN is a multi-layer feed-forward neural network that extracts local and global features from data through structures, such as convolutional, pooling, and fully connected layers to enable automatic feature learning and abstraction. In price prediction tasks, 是一个多层前馈神经网络，通过卷积层、池化层和全连接层等结构从数据中提取局部和全局特征，以实现自动特征学习和抽象。在价格预测任务中，CNNs can learn and capture important features, such as time series, data trends, periodicity, etc., in the input data. Market prices are usually affected by a combination of several factors, and CNNs can better handle these complex nonlinear relationships. 可以学习和捕获输入数据中的重要特征，例如时间序列、数据趋势、周期性等。市场价格通常受到多种因素的组合影响，CNN可以更好地处理这些复杂的非线性关系。 -
混沌神经网络（Chaos Neural Networks (CNN)
^{[22][23]}: NN）[40，41]：混沌神经网络（Chaos neural network (CNN) is a kind of intelligent information processing system that combines chaos theory and neural network. Chaotic neural networks exploit the sensitivity and unpredictability of chaotic phenomena to enhance the learning and generalization capabilities of neural networks, thus improving the accuracy of prediction and modeling. By introducing methods, such as chaotic noise or logistic maps, chaotic neural networks are able to avoid neural networks from falling into local minima to a certain extent, thus speeding up the training process and increasing the convergence rate.）是一种结合混沌理论和神经网络的智能信息处理系统。混沌神经网络利用混沌现象的敏感性和不可预测性，增强神经网络的学习和泛化能力，从而提高预测和建模的准确性。通过引入混沌噪声或逻辑图等方法，混沌神经网络能够在一定程度上避免神经网络陷入局部极小值，从而加快训练过程并提高收敛率。

- 极限学习机（
- E
- xtreme Learning Machines (ELM)
^{[}^{24]}: The extreme learning machine is a feed-forward neural network that was first proposed by Professor Huang Guangbin of Nanyang Technological University in Singapore in LM）[42]：极限学习机是一种前馈神经网络，由新加坡南洋理工大学的黄广斌教授于2006. 年首次提出。ELM has the advantages of fast training, high generalization ability, and simple implementation.具有训练快、泛化能力强、实现简单等优点。-
小波神经网络（Wavelet Neural Networks (WNN)
^{[25][26]}: NN）[43，44]：小波神经网络（Wavelet neural network (WNN) is a method based on wavelet transform and neural network. By decomposing the original data into wavelet coefficients at different scales, it is able to effectively extract a variety of features in the data, such as trend, cycle, seasonality, etc. WNN combines the powerful fitting ability of neural networks, which is capable of nonlinear mapping, thus achieving accurate prediction of future prices. However, high complexity and high data requirements are the unavoidable drawbacks of this method.）是一种基于小波变换和神经网络的方法。通过将原始数据分解为不同尺度的小波系数，能够有效地提取数据中的各种特征，如趋势、周期、季节性等，WNN结合了神经网络强大的拟合能力，能够进行非线性映射，从而实现对未来价格的准确预测。然而，高复杂性和高数据要求是该方法不可避免的缺点。

基于神经网络的农产品价格预测应用实例.

Reference参考 |