WO2020024319A1 - 用于交通流量预测的卷积神经网络多点回归预测模型 - Google Patents
用于交通流量预测的卷积神经网络多点回归预测模型 Download PDFInfo
- Publication number
- WO2020024319A1 WO2020024319A1 PCT/CN2018/099498 CN2018099498W WO2020024319A1 WO 2020024319 A1 WO2020024319 A1 WO 2020024319A1 CN 2018099498 W CN2018099498 W CN 2018099498W WO 2020024319 A1 WO2020024319 A1 WO 2020024319A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- layer
- output
- regression
- neural network
- convolution
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Definitions
- the present invention relates to a multi-point regression prediction model of a convolutional neural network, and more particularly to a multi-point regression prediction model of a convolutional neural network for traffic flow prediction.
- the regression analysis and prediction method is based on analyzing the correlation between the independent variables and the dependent variables of various phenomena, establishing a regression equation between the variables, and using the regression equation as a prediction model to predict based on the number of independent variables during the forecast period. Most of the dependent variable relationships are related. Therefore, the regression analysis and forecasting method is an important forecasting method. When we predict the future development status and level of the phenomenon of the research object, if it can affect the main prediction object of the research, Factors can be found and their quantity data can be obtained, and then regression analysis and prediction can be used to make predictions. It is a specific, effective, and commonly used forecasting method with high practical value.
- Neural network is a complex model with multi-layer structure, which can fit complex non-linear systems, and has been applied in regression prediction models [1] [6] [8] .
- Taylor [2] first proposed a neural network regression model in 2000. In the application of financial asset rate of return analysis, the assumption of conditional distribution of financial asset rate of return was avoided. On the other hand, a neural network structure was used. Estimate potential non-linear models. Taylor chose the daily log rate of return of the German mark against the US dollar and the yen against the US dollar as research objects, and empirically compared the performance of the neural network regression model and the GARCH model in the multi-period confidence level risk measurement. The results show that the neural network regression model improves the accuracy of multi-period confidence level risk measurement; Feng [12] and other researchers applied neural network regression models to credit portfolio investment decision-making problems.
- Cannon [13] and others introduced the software package qrnn based on a neural network regression model, and pointed out that the neural network regression model provides a hybrid for discrete continuous variables such as rainfall, wind speed, and pollutant concentration. This kind of non-linear and non-parametric regression method, and applied the neural network regression model to predict the rainfall. The research shows that the neural network regression model performs better than traditional regression in rainfall prediction.
- He Yaoyao [11] proposed a method of probability density prediction based on neural network regression model to realize the prediction of the complete probability distribution of future power load and the probability density of the actual data of power load in a city in China
- the prediction shows that the probability density prediction method based on the neural network regression model can obtain the complete probability density function result of short-term load.
- They also applied the neural network regression model to the medium-term power load probability density prediction, studied the influence of temperature and historical load on the medium-term power load distribution at different quantiles, and compared the temperature factors with and without temperature factors.
- the conditional probability density prediction curve and the point prediction value corresponding to the peak value show that the temperature of the prediction day has a more important impact on the medium-term load forecast, which provides more decision-making information and prediction results for reducing the uncertain factors of the medium-term power load forecast.
- Yeh et al Used a neural network regression model to estimate the compressive strength distribution on high-performance sea coagulation, and pointed out that the ability to estimate the compressive strength distribution of high-performance coagulation king is an important advantage of the neural network regression model. Studies show that neural network regression The model can establish an accurate estimation model, which can estimate the distribution of compressive strength of high-performance concrete. In addition, the log-normal distribution is more suitable for fitting the compressive strength distribution of high-performance concrete than the normal distribution.
- Convolutional neural network [3] [4] [5] is a deep neural network with feature extraction capabilities, which has achieved great success in image recognition, speech recognition and other aspects.
- the application of convolutional neural network in NLP problem, Zeng [14] and others carried out the extraction of associations by convolutional neural network; Chen [15] and others carried out the extraction of event information.
- the question-answering system He [16] et al. Regarded the question-and-answer matching sub-task of the question-answering system as sentence similarity matching; and generally used convolutional neural networks to identify displacement, scaling, and other forms of distortion-invariant 2D graphics.
- CNN's feature detection layer learns from training data, it avoids explicit feature extraction and learns implicitly from training data. Due to the above characteristics, convolutional networks are mainly used in classification tasks. This patent proposes a six-layer unpooled convolutional neural network regression model, which can be used for regression modeling and multi-point prediction tasks of complex systems.
- the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art. Compared with traditional statistical regression models, it has the feature of data space correlation feature extraction and the advantages of local perceptual field and weight sharing. There is a better balance in feature selection.
- a convolutional neural network multi-point regression prediction model for traffic flow prediction including the following steps:
- the first perceptual input layer the input of training data, which usually needs to be converted into a matrix form
- the second convolution layer convolves the input layer data and outputs it after activating the function
- Multi-layer convolutional layer Convolve the output of the previous layer as an input, and output it after activating the function
- the fourth fully-linked layer The output of the previous layer is the input.
- the "fully-linked layer” implements the regression calculation of the feature vector.
- Q nodes are set in this layer, and the matrices obtained by the convolution layer are all stitched into a unique vector. Map it to Q nodes and combine them with weights;
- the fifth discarding layer discards some redundant neurons, and retains 40% -70% of the fully-linked nodes in the upper layer,
- the sixth output layer The effective node output of the discarding layer is subjected to regression calculation, and the obtained regression value is the output of the entire network. A total of m output nodes are set, and the fifth discarding layer is mapped to the output layer for weight combination.
- a further improvement scheme of the present invention is that the multi-layer convolution layer is a third convolution layer, and the convolution network has a six-layer structure.
- a further improvement scheme of the present invention is: the sixth output layer, that is, the output cascade: the training output of the previous output node is taken as the input to the next output node, which reflects the multi-point output. Time series relationship.
- a further improvement scheme of the present invention is: the output of the convolution layer is calculated by convolution of the input of the upper layer, x i, j is the i-th row and j-th column traffic flow data after matrixization,
- This article uses To represent the i-th row and j-th column of the feature map of the k-th layer convolution, To input the corresponding convolution weights, Is the convolution bias term, t is the size of the convolution kernel, f () is the activation function, and Relu function is selected as the activation function.
- the function of each node of the convolution layer is as follows:
- the sixth output layer includes m nodes, w im represents the connection weight from the i-th node to the m-th output node, and b im represents the i-th node to the i-th node.
- m output node connection bias w m-1 represents the connection weight of the m-1th output node to the mth output node
- b m-1 represents the connection bias of the m-1th output node to the mth output node
- the function O m of each node in the sixth output layer is as follows:
- O m Relu (w m-1 O m-1 + b m-1 + ⁇ w im x + b im )
- a further improvement scheme of the present invention is: the full link layer is a feedforward network, and the regression information is integrated through the feature information extraction after convolution:
- a further improvement scheme of the present invention is: the realization of the discard layer is to make the activation value of neurons become 0 with probability p, so that these neurons can be shielded and their activation values should be 0.
- the first perceptual input layer refers to a one-dimensional feature data and a single-dimensional feature information into a two-dimensional matrix of m rows and n columns and k depth, where k is the number of channels , And the size of the product of m and n should be equal to the original feature size.
- a further improvement scheme of the present invention is: the Q is selected between 100-200.
- the beneficial effects of the present invention are: compared with the traditional statistical regression model, the multi-layer convolutional layer has the feature of data space correlation feature extraction, and has the advantages of local perceptual field and weight sharing, which makes time complexity and features The selection has a better balance; in the present invention, the step of using the pooling layer is not used after the convolution layer. This step can well retain the features that need to be extracted, and avoid the loss of spatial information during the pooling process. .
- the six-layer structure has stronger feature extraction ability than the three-layer shallow network, and its training complexity is greatly reduced compared to the 20-layer deep convolutional network, which saves computing resources and improves training efficiency.
- the neural network regression model structure of the patent has an output cascade structure, which has the function of multi-point prediction on a time series, and can simultaneously output prediction values of multiple consecutive time points.
- FIG. 1 is a structural diagram of a multi-point regression prediction model of a six-layer unpooled convolutional neural network according to the present invention
- FIG. 2 is a feature extraction process diagram of a six-layer unpooled convolutional neural network according to the present invention
- FIG. 3 is a result diagram of the predicted values obtained by using a six-layer unpooled convolutional neural network multi-point regression prediction model in the first time period of the present invention
- FIG. 4 is a result diagram of the predicted values obtained by using a six-layer unpooled convolutional neural network multi-point regression prediction model in the second time period of the present invention
- FIG. 5 is a result diagram of a predicted value obtained by using a six-layer unpooled convolutional neural network multi-point regression prediction model in the third time period of the present invention.
- a convolutional neural network multi-point regression prediction model for traffic flow prediction includes the following steps:
- the first perception input layer the input of training data, which usually needs to be converted into a matrix form
- the second convolution layer convolves the input layer data and outputs it after activating the function
- the third convolution layer Convolve the output of the previous layer as an input and output it after activating the function; the number of convolution layers is determined according to the actual effect, and more convolution layers cannot guarantee the performance of the network model. Improved. Three layers are the best results after our experiments. For this convolution layer, there can also be a fourth convolution layer, a fifth convolution layer, and multiple convolution layers.
- the fourth fully-linked layer The output of the third layer is an input.
- the "fully-linked layer” implements the regression calculation of the feature vector.
- This layer is provided with Q nodes.
- the preferred solution of the Q nodes is: the selected value is 100-200.
- the "random discarding layer” discards some redundant neurons, retaining 40% -70% of the fully-linked nodes in the upper layer,
- the sixth output layer the effective node output of the discarding layer is subjected to regression calculation, and the obtained regression value is the output of the entire network. A total of m output nodes are set, and the fifth discarding layer is mapped to the output layer for weight combination.
- the sixth output layer that is, the output cascade, takes the training output of the previous output node as the input to the next output node, and reflects the time series relationship between the multi-point outputs.
- one-dimensional feature data (non-image common data format) is used to convert single-dimensional feature information into a two-dimensional matrix with m rows and n columns and k depths, where k is the number of channels, and m and n
- the product size should be equal to the original feature size.
- the output of the convolution layer is calculated by convolution of the input of the upper layer, x i, j are the i-th row and j-th column traffic flow data after matrixing.
- x i, j are the i-th row and j-th column traffic flow data after matrixing.
- the function of each node of the convolution layer is as follows:
- the activation function of the convolutional layer can also be sigmoid / tanh.
- the convergence rate of SGD (stochastic gradient descent) obtained by ReLU will be much faster than sigmoid / tanh. Compared with sigmoid / tanh, it needs to calculate the index, etc., and the calculation complexity is high. ReLU Only a threshold is needed to get the activation value.
- multi-layer convolution is often used, and then fully-linked layers are used for training.
- the purpose of multi-layer convolution is that the features learned by one layer of convolution are often local. The higher the number of layers, the learned features The more global.
- the full link layer is a feedforward network, which integrates regression through feature information extraction after convolution:
- the fully-linked layer maps the feature map generated by the convolutional layer into a fixed-length (typically the number of image categories in the input image data set) feature vector.
- This feature vector contains the combined information of all features of the input image. Although the position information of the image is lost, this vector retains the most characteristic image features in the image to complete the image classification task.
- the output layer contains m nodes, w im represents the connection weight of the i-th node to the m-th output node, and b im represents the connection bias of the i-th node to the m-th output node.
- w m-1 represents the connection weight of the m-1th output node to the m-th output node
- b m-1 represents the connection bias of the m-1th output node to the m-th output node
- each node function of the output layer is O m as follows:
- O m Relu (w m-1 O m-1 + b m-1 + ⁇ w im x + b im )
- the invention relates to a feature extraction process based on association information of a six-layer convolutional neural network.
- the purpose of the invention is to process convolutions with spatially associated feature data.
- RMSE RMSE
- MAPE MAPE
- MSE MSE indicator
- Figure 2 shows the feature extraction process of the present invention.
- Data preprocessing matrix the existing data, and convert the single-dimensional feature information into a two-dimensional matrix with m rows and n columns and k depths, where k represents the number of data channels (single channel in Figure 3), Convenient for convolutional networks;
- the first layer of convolution input the matrix data into the first layer of convolution layer.
- the convolution kernel is a window of 3 * 3.
- the size of the convolution kernel cannot be larger than m-1 and n-1. For different For prediction objects, you can choose different convolution kernel sizes.
- the number of convolution kernels is 20, low-level features (basic feature information) are extracted, and the opposite edges of the output matrix are set to 0; (that is, the diagonal value of the output matrix is set to 0)
- Second layer convolution take the output low-order features as input to the second layer convolution layer for convolution processing, the convolution kernel is a 3 * 3 window, the number of convolution kernels is 20, and then The second layer of convolution performs partial weight combination of information, extracts high-order features (complex combined feature information), and sets the output matrix to 0 for edges;
- the convolved matrix is output through the activation function ReLu to obtain a two-dimensional matrix with m rows and n columns and a k depth;
- h represents the actual vehicle speed
- f represents the free flow speed
- TSI identifies the congested nodes
- MAPE is the average error percentage
- RMSE is the root mean square error
- MIN represents the amount of data in which the model results are closest to the true values in the three models.
- CNN stands for Convolutional Neural Network. The calculation formulas of MAPE and RMSE are the same for all sections.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Operations Research (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Image Analysis (AREA)
Abstract
一种用于交通流量预测的卷积神经网络多点回归预测模型,包括如下步骤:第一感知输入层,第二卷积层:对输入层数据进行卷积,通过激活函数后输出;多层卷积层:对上一层的输出作为输入进行卷积处理,通过激活函数后输出;第四全链接层,第五丢弃层:"随机丢弃层"舍弃一些冗余的神经元,保留上层全链接节点的40%-70%;第六输出层:丢弃层的有效节点输出进行回归计算,得到的回归数值就是整个网络的输出,共设置m个输出节点,即将全链接层映射到输出层,作权重组合。相比传统的统计回归模型,具有数据空间关联的特征提取能力,具有局部感知野和权值共享的优势,使得在时间复杂度和特征选择上具有更好的平衡。
Description
本发明涉及卷积神经网络的多点回归预测模型,尤其涉及用于交通流量预测的卷积神经网络多点回归预测模型。
回归分析预测法是在分析各类现象自变量和因变量之间相关关系的基础上,建立变量之间的回归方程,并将回归方程作为预测模型,根据自变量在预测期的数量变化来预测因变量关系大多表现为相关关系,因此,回归分析预测法是一种重要的预测方法,当我们在对研究对象的现象未来发展状况和水平进行预测时,如果能将影响研究的预测对象的主要因素找到,并且能够取得其数量资料,就可以采用回归分析预测法进行预测。它是一种具体的、行之有效的、实用价值很高的常用预测方法。研究人员针对回归预测的模型,一般分为线性回归,逻辑回归,多项式回归,逐步回归,岭回归,套索回归,ElasticNet回归。神经网络是一种多层结构的复杂模型,可以拟合复杂的非线性***,在回归预测模型
[1][6][8]中得到了应用。
在金融领域研究方面,Taylor
[2]于2000年首次提出了神经网络回归模型,在金融资产收益率分析的应用中,避免了对金融资产收益率条件分布的假设,另一方面使用神经网络结构估计潜在的非线性模型。Taylor选取了德国马克兑美元汇率和日元兑美元汇率的日对数收益率作为研究对象,实证比较了神经网络回归模型与GARCH模型在多期 置信水平风险测度中的表现。结果表明,神经网络回归模型提高了多期置信水平风险测度的精度;Feng
[12]等将神经网络回归模型应用于信用组合投资决策问题,蒙特卡罗数值模拟和信用组合投资数据的实证分析表明,神经网络回归模型在异常值数据的拟合方面比局部线性回归和样条回归更具有稳健性。许启发
[7]等使用神经网络回归模型测度上证综合指数的置信水平风险,并与传统的置信水平风险测度方法进行比较,实证结果表明,基于神经网络回归模型的置信水平风险测度方法,在样本内与样本外都取得了较好的效果。
在非金融领域研究方面,Cannon
[13]等介绍了实施了基于神经网络回归模型的软件包qrnn,指出神经网络回归模型为混合离散连续变量,如降雨量、风速、污染物浓度等提供了一种非线性、非参数的回归方法,并应用神经网络回归模型对降雨量进行了预测,研究表明,神经网络回归模型在降雨量预测中的表现优于传统回归。何耀耀
[11]等针对电力***短期负荷预测问题,提出了基于神经网络回归模型的概率密度预测方法,实现对未来电力负荷完整概率分布的预测,并对中国某市的电力负荷实际数据进行概率密度预测,结果表明,基于神经网络回归模型的概率密度预测方法能够获得短期负荷完整的概率密度函数结果。他们还将神经网络回归模型应用于中期电力负荷概率密度预测,研究在不同分位点上温度和历史负荷对中期电力负荷分布规律的影响,比较了在考虑温度因素下和不考虑温度因素下的条件概率密度预测曲线及峰值对应的点预测值,结果表明,预测当天温度对中期负荷预测有较重要的影响,为降低中期电力负荷预测的不确定因 素提供了更多的决策信息和预测结果。Yeh等人使用神经网络回归模型估计高性能海凝上抗压强度的分布,并指出估计高性能混凝王抗压强度分布的能力是神经网络回归模型的一个重要优势,研究表明,神经网络回归模型可建立准确的估计模型,可对高性能混凝土抗压强度的分布进行估计,此外,对数正态分布比正态分布更适合拟合高性能混凝土抗压强度分布。
卷积神经网络
[3][4][5]是一种具有特征提取能力的深层神经网络,在图像识别、语音识别等方面取得了巨大的成功。首先卷积神经网络在NLP问题中的应用,Zeng
[14]等人进行了卷积神经网络对关联性的提取;Chen
[15]等人进行了事件信息的抽取。在问答***中,He
[16]等人对问答***的问答匹配子任务看作句子相似度匹配;而一般通过卷积神经来识别位移、缩放及其他形式扭曲不变性的二维图形。由于CNN的特征检测层通过训练数据进行学习,所以避免了显式的特征抽取,而隐式地从训练数据中进行学习。由于以上特点,卷积网络主要用于分类任务中。本专利则提出一种六层无池化的卷积神经网络的回归模型,可用于复杂***的回归建模和多点预测任务。
参考文献
[1]Chun-Hsin Wu,Jan-Ming Ho,D.T.Lee.Travel-Time Prediction With Support Vector Regression[J].IEEE Transactions on Intelligent Transportation Systems,2004,5(4):276-281.
[2]Taylor J W,A quantile regression neural network approach to estimating the condition density of multiperiod returns[J].Journal of Forecasting,2000,19(4):299-311.
[3]Xiang-Xue Wang,Lun-Hui Xu.Short-term Traffic Flow Prediction Based on Deep Learning[J].Journal of Transportation Systems Engineering and Information Technology,2018,18(1):81-88.
[4]Nicholas G.Polson,Vadim O.Sokolov.Deep learning for short-term traffic flow prediction[J].Transportation Research Part C,2017(79):1–17.
[5]罗文慧,董宝田,王泽胜.基于CNN-SVR混合深度学习模型的短时交通流预测[J].交通运输***工程与信息,2017,17(5):68-74.
[6]Lee Kiryung,Kim Dong Sik,Kim Taejeong,Regression-based prediction for blocking artifact reduction in JPEG-compressed images.[j]IEEE Transactions on Image Processing.2005,Vol.14(1).pp.36-48 PubMed;
[7]许启发,徐金菊,蒋翠侠,刘晓华,Financial risk measure of VaR based on quantile regression neural network.[J]Journal of Hefei University of Technology(Natural Science).
[8]罗小妹,陈晓宏,蔡斯龙Mid-Long-Term Runoff Forecasting Based on an Improved Normalized Periodic Regression Model—A Case Study in the Xingfeng River Basin基于改进正规化周期回归模型的中长期径流预测—以新丰江流域为例[J].Journal of Water Resources Research,2013,Vol.02(01),pp.27-32汉斯出版社
[9]于达仁;万杰;胡清华;刘金福;郭钰锋;苏鹏宇;基于深度学习方法 的短期风速多步预测方法(专利),[P].CN201310389450.X 2013-11-27 CNKI
[10]史世雄;杨夙,一种基于时空关联数据挖掘的交通流预测方法(专利),[P].CN201310749807.0 2014-04-02 CNKI
[11]陆宁,周建中,何耀耀,Particle swarm optimization-based neural network model for short-term load forecasting,[J].POWER SYSTEM PROTECTIONAND CONTROL
[12]Feng Y,Li R,Sudjianto A,et al.Robust neural network with applications to credit portfolio data analysis[J].Statistics and Its Interface,2010,3(4):437.
[13]CannonA J.Quantile regression neural networks:implementation in R and application to precipitation downscaling[J].Computers&Geosciences,2010,37(9):1277-1284
[14]Zeng D,Liu K,Lai S,et al.Relation Classification via Convolutional Deep NeuralNetwork[C]//COLING.2014:2335-2344.
[15]Chen Y,Xu L,Liu K,et al.Event Extraction via Dynamic Multipooling Convolution Neural Networks[C]//ACL(1).2015:167-176.
[16]He H,Gimpel K,Lin J J,Multi-Persperctive Sentence Similarity Modeling with Convolutional Neural Networks[C]//EMNLP.2015:1576-1586
发明内容
本发明所要解决的技术问题是,克服现有技术的缺点,相比传统 的统计回归模型,具有数据空间关联的特征提取能力,具有局部感知野和权值共享的优势,使得在时间复杂度和特征选择上具有更好的平衡。
本发明解决以上技术问题的技术方案是:用于交通流量预测的卷积神经网络多点回归预测模型,包括如下步骤:
(1)第一感知输入层:训练数据的输入,通常需要转换成矩阵形式;
(2)第二卷积层:对输入层数据进行卷积,通过激活函数后输出;
(3)多层卷积层:对上一层的输出作为输入进行卷积处理,通过激活函数后输出;
(4)第四全链接层:上一层的输出为输入,“全链接层”实现特征向量的回归计算,在该层设置Q个节点,把卷积层得到的矩阵全部拼接成一唯向量,再映射成Q个节点上,作权重组合;
(5)第五丢弃层:“随机丢弃层”舍弃一些冗余的神经元,保留上层全链接节点的40%-70%,
(6)第六输出层:丢弃层的有效节点输出进行回归计算,得到的回归数值就是整个网络的输出,共设置m个输出节点,将第五丢弃层映射到输出层,作权重组合。
基于以上技术问题,本发明进一步的改进方案是:所述的多层卷积层为第三卷积层,所述的卷积网络为六层结构。
基于以上技术问题,本发明进一步的改进方案是:所述的第六输出层,即输出级联:将上一个输出节点的训练输出作为输入到下一个输出节点,反映了多点输出之间的时间序列关系。
基于以上技术问题,本发明进一步的改进方案是:卷积层的输出是通过上层的输入经过卷积来计算的,x
i,j是矩阵化后的第i行,第j列交通流数据,本文用
来表示第k层卷积的特征图的第i行,第j列输出,
为输入对应卷积权重,
为卷积偏置项,t为卷积核的大小,用f()表示激活函数,选择Relu函数作为的激活函数,卷积层各节点函数如下:
基于以上技术问题,本发明进一步的改进方案是:所述的第六输出层包含m个节点,w
im表示第i个节点到第m输出节点的连接权重,b
im表示第i个节点到第m输出节点的连接偏置,w
m-1表示第m-1个输出节点到第m输出节点的连接权重,b
m-1表示第m-1个输出节点到第m输出节点的连接偏置,第六输出层各节点函数O
m如下:
O
1=Relu(∑w
i1x+b
i1)
O
2=Relu(w
1O
1+b
1+∑w
i2x+b
i2)
...
O
m=Relu(w
m-1O
m-1+b
m-1+∑w
imx+b
im)
基于以上技术问题,本发明进一步的改进方案是:全链接层是一个前馈网络,通过卷积后的特征信息提取,进行回归的集成:
h
3=Relu(h
2*ω
3+b
3)。 (2)
基于以上技术问题,本发明进一步的改进方案是:丢弃层的实现就是将让神经元的激活值以概率p变为0,这样就可以屏蔽这些神经元,使其激活值为0以后,需要对神经元向量进行重构:
w
k=p×w
k
基于以上技术问题,本发明进一步的改进方案是:第一感知输入层是指将一维特征数据,将单维度的特征信息转化为m行n列k深度的二维矩阵,其中k是通道数,而m与n的乘积大小应等于原始特征大小。
基于以上技术问题,本发明进一步的改进方案是:所述的Q选取100-200之间。
本发明的有益效果是:,相比传统的统计回归模型,采用多层卷积层,具有数据空间关联的特征提取能力,具有局部感知野和权值共享的优势,使得在时间复杂度和特征选择上具有更好的平衡;本发明中在卷积层后没有使用池化层的步骤,这一步很好的能够充分保留所需提取的特征,避免了在池化过程中造成空间信息的消失。
六层结构相比于三层的浅层网络具有更强的特征提取能力,而相比于20层深度卷积网络,其训练复杂度大大降低,节省了计算资源,提高了训练效率。
本专利的神经网络回归模型结构中具有输出级联结构,具有在时间序列上多点预测的功能,可同时输出多个连续时间点的预测数值。
图1是本发明六层无池化卷积神经网络多点回归预测模型结构图;
图2是本发明六层无池化卷积神经网络的特征提取过程图;
图3是本发明在第一时间周期内使用六层无池化卷积神经网络多点回归预测模型后的所得到的预测值的结果图;
图4是本发明在第二时间周期内使用六层无池化卷积神经网络多 点回归预测模型后的所得到的预测值的结果图;
图5是本发明在第三时间周期内使用六层无池化卷积神经网络多点回归预测模型后的所得到的预测值的结果图;
实施例1
如图1和图2所示,一种用于交通流量预测的卷积神经网络多点回归预测模型,包括如下步骤:
第一感知输入层:训练数据的输入,通常需要转换成矩阵形式;
第二卷积层:对输入层数据进行卷积,通过激活函数后输出;
第三卷积层:对上一层的输出作为输入进行卷积处理,通过激活函数后输出;卷积层的数量是根据实际效果确定的,更多的卷积层不能保证网络模型的性能得到提高.三层是我们实验后最好的结果.对于该卷积层也可以还包括第四卷积层、第五卷积层,等多个卷积层。
第四全链接层:第三层的输出为输入,“全链接层”实现特征向量的回归计算,该层设置Q个节点,所述Q个节点的优选方案为:选取值为100-200之间;具体的讲就是把卷积层得到的矩阵全部拼接成一唯向量,再映射成Q个节点上,作权重组合;
第五丢弃层:“随机丢弃层”舍弃一些冗余的神经元,保留上层全链接节点的40%-70%,
第六输出层:丢弃层的有效节点输出进行回归计算,得到的回归数值就是整个网络的输出,共设置m个输出节点,将第五丢弃层映射到输出层,作权重组合。所述的第六输出层,即输出级联,将上一个输出 节点的训练输出作为输入到下一个输出节点,反映了多点输出之间的时间序列关系。
上述技术方案涉及卷积神经网络的各层输出函数如下:
(1)第一感知输入层
如图3所示,将一维特征数据(非图像的普通数据格式),将单维度的特征信息转化为m行n列k深度的二维矩阵,其中k是通道数,而m与n的乘积大小应等于原始特征大小。
(2)第二卷积层
卷积层的输出是通过上层的输入经过卷积来计算的,x
i,j是矩阵化后的第i行,第j列交通流数据,本文用
来表示第k层卷积的特征图的第i行,第j列输出,
为输入对应卷积权重,
为卷积偏置项,t为卷积核的大小,用f()表示激活函数,选择Relu函数作为的激活函数,卷积层各节点函数如下:
卷积层的激活函数还可以是sigmoid/tanh,ReLU得到的SGD(随机梯度下降)的收敛速度会比sigmoid/tanh快很多,相比于sigmoid/tanh需要计算指数等,计算复杂度高,ReLU只需要一个阈值就可以得到激活值。在实际应用中,往往使用多层卷积,然后再使用全链接层进行训练,多层卷积的目的是一层卷积学到的特征往往是局部的,层数越高,学到的特征就越全局化。
(3)全链接层是一个前馈网络,通过卷积后的特征信息提取,进行回归的集成:
h
3=Relu(h
2*ω
3+b
3)
全链接层将卷积层产生的特征图映射成一个固定长度(一般为输入图像数据集中的图像类别数)的特征向量。这个特征向量包含了输入图像所有特征的组合信息,虽然丢失了图像的位置信息,但是该向量将图像中含有最具有特点的图像特征保留了下来以此完成图像分类任务。
(4)丢弃层或稀疏层
Drop out层(
丢弃层)的实现就是将让神经元的激活值以概率p变为0,这样就可以屏蔽这些神经元。使其激活值为0以后,需要对神经元向量进行重构:
w
k=p×w
k
(5)输出层
输出层包含m个节点,w
im表示第i个节点到第m输出节点的连接权重,b
im表示第i个节点到第m输出节点的连接偏置。w
m-1表示第m-1个输出节点到第m输出节点的连接权重,b
m-1表示第m-1个输出节点到第m输出节点的连接偏置,输出层各节点函数O
m如下:
O
1=Relu(∑w
i1x+b
i1)
O
2=Relu(w
1O
1+b
1+∑w
i2x+b
i2)
...
O
m=Relu(w
m-1O
m-1+b
m-1+∑w
imx+b
im)
本发明涉及到基于六层卷积神经网络的关联信息的特征提取过程,其目的是对具有空间关联特征数据的卷积处理,在训练网络并利用RMSE,MAPE,MSE指标对卷积核参数进行优化。图2为本发 明的特征提取流程。
为了实现该目标,具体步骤如下:
(1)数据预处理:将已有的数据进行矩阵化,将单维度的特征信息转化为m行n列k深度的二维矩阵,k代表数据的通道数(图3中为单通道),便于卷积网络的卷积处理;
(2)第一层卷积:将矩阵化的数据输入第一层卷积层,卷积核为3*3的窗口,选取卷积核尺寸不能大于m-1和n-1,对于不同的预测对象,可以选取不同的卷积核大小。卷积核的数量为20,提取低阶特征(基本特征信息),输出矩阵对边置0;(即输出矩阵对角线的值置0)
(3)卷积后的矩阵经过激活函数ReLu输出,得到m行n列k深度的二维矩阵;
(4)第二层卷积:将输出后的低阶特征作为输入到第二层卷积层进行卷积处理,卷积核为3*3的窗口,卷积核的数量为20,再由第二层卷积进行信息的部分权重组合,提取高阶特征(复杂的组合特征信息),输出矩阵对边置0;
(5)卷积后的矩阵经过激活函数ReLu输出,得到m行n列k深度的二维矩阵;
(6)在训练卷积神经网络时,卷积核参数朝向损失函数不断减小的方向变化。(也就是梯度下降的方向)
实施例2
对上海市快速路进行交通状态预测的实施步骤如下:
1)通过上海市快速路地感线圈的数据,通过对线圈的时空位置关系分成不同的断面类型:普通断面,匝道断面,分流断面,交织区断面,合流断面;
2)通过上海市的TSI指数,对拥堵节点进行标定:
其中,h代表实际车速,f代表自由流速度;TSI识别出拥堵的节点;
表1不同指数区间对应的道路交通状态
3)进行特征的敏感性分析,拿目标点上游不同节点数,下游不同节点数不同的特征样本对最后的目标点进行预测;启用六层无池化卷积神经网络的多点回归预测模型,得出如下结论:
表2选取不同模型特征的模型结果比较
故本处,选取目标节点和上游及下游各五个节点的五个周期作为特征,进行训练;
4)用选择的特征点,对位置关系不同的断面类型,使用本文的模型进行预测:
5)对测试集中的数据预测出的结果与真实值之间的结果使用RMSE,MAPE,MIN值进行误差分析,来判定模型的优劣;所述的RMSE,MAPE,MIN公式分别如下:
min=min{h
i}
下表中MAPE是平均误差百分比,RMSE为均方根误差,MIN代表在三个模型中模型结果与真实值最接近的数据量。CNN表示的是卷积神经网络,所有截面的MAPE和RMSE计算公式是相同的。
表3不同断面类型模型结果比较
断面类型 | 模型选择 | MAPE | RMSE | MIN |
正常断面 | CNN | 0.047 | 2.56 | 232 |
合流节点断面 | CNN | 0.036 | 6.02 | 259 |
分流节点断面 | CNN | 0.086 | 5.05 | 284 |
交织节点断面 | CNN | 0.110 | 6.53 | 291 |
匝道断面 | CNN | 0.031 | 12.06 | 425 |
使用我们的模型最后得到的预测值的结果图如图3至图5所示。
Claims (9)
- 用于交通流量预测的卷积神经网络多点回归预测模型,其特征在于,包括如下步骤:(1)第一感知输入层:训练数据的输入,通常需要转换成矩阵形式;(2)第二卷积层:对输入层数据进行卷积,通过激活函数后输出;(3)多层卷积层:对上一层的输出作为输入进行卷积处理,通过激活函数后输出;(4)第四全链接层:上一层的输出为输入,“全链接层”实现特征向量的回归计算,在该层设置Q个节点,把卷积层得到的矩阵全部拼接成一唯向量,再映射成Q个节点上,作权重组合;(5)第五丢弃层:“随机丢弃层”舍弃一些冗余的神经元,保留上层全链接节点的40%-70%,(6)第六输出层:丢弃层的有效节点输出进行回归计算,得到的回归数值就是整个网络的输出,共设置m个输出节点,将第五丢弃层映射到输出层,作权重组合。
- 如权利要求1所述的用于交通流量预测的卷积神经网络多点回归预测模型,其特征在于:所述的多层卷积层为第三卷积层,所述的卷积网络为六层结构。
- 如权利要求1或2所述的用于交通流量预测的卷积神经网络多点回归预测模型,其特征在于:所述的第六输出层,即输出级联:将上一个输出节点的训练输出作为输入到下一个输出节点,反映了多点输出之间的时间序列关系。
- 如权利要求1或2所述的用于交通流量预测的卷积神经网络多点回归预测模型,其特征在于:所述的第六输出层包含m个节点,w im表示第i个节点到第m输出节点的连接权重,b im表示第i个节点到第m输出节点的连接偏置,w m-1表示第m-1个输出节点到第m输出节点的连接权重,b m-1表示第m-1个输出节点到第m输出节点的连接偏置,第六输出层各节点函数O m如下:O 1=Relu(∑w i1x+b i1)O 2=Relu(w 1O 1+b 1+∑w i2x+b i2)...O m=Relu(w m-1O m-1+b m-1+∑w imx+b im)
- 如权利要求1或2所述的用于交通流量预测的卷积神经网络多点回归预测模型,其特征在于:全链接层是一个前馈网络,通过卷积后的特征信息提取,进行回归的集成:h 3=Relu(h 2*ω 3+b 3)。 (2)
- 如权利要求1或2所述的用于交通流量预测的卷积神经网络多点回归预测模型,其特征在于:丢弃层的实现就是将让神经元的激活值以概率p变为0,这样就可以屏蔽这些神经元,使其激活值为0以后, 需要对神经元向量进行重构:w k=p×w k
- 如权利要求1或2所述的用于交通流量预测的卷积神经网络多点回归预测模型,其特征在于:第一感知输入层是指将一维特征数据,将单维度的特征信息转化为m行n列k深度的二维矩阵,其中k是通道数,而m与n的乘积大小应等于原始特征大小。
- 如权利要求1或2所述的用于交通流量预测的卷积神经网络多点回归预测模型,其特征在于:所述的Q选取100-200之间。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810866657.4 | 2018-08-01 | ||
CN201810866657.4A CN108830430A (zh) | 2018-08-01 | 2018-08-01 | 用于交通流量预测的卷积神经网络多点回归预测模型 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2020024319A1 true WO2020024319A1 (zh) | 2020-02-06 |
Family
ID=64153462
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2018/099498 WO2020024319A1 (zh) | 2018-08-01 | 2018-08-09 | 用于交通流量预测的卷积神经网络多点回归预测模型 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN108830430A (zh) |
WO (1) | WO2020024319A1 (zh) |
Cited By (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111414482A (zh) * | 2020-03-20 | 2020-07-14 | 北京百度网讯科技有限公司 | 一种事件论元抽取方法、装置以及电子设备 |
CN111627101A (zh) * | 2020-05-22 | 2020-09-04 | 北京工业大学 | 一种基于图卷积的三维人体重构方法 |
CN111680781A (zh) * | 2020-04-20 | 2020-09-18 | 北京迈格威科技有限公司 | 神经网络处理方法、装置、电子设备及存储介质 |
CN111737578A (zh) * | 2020-06-22 | 2020-10-02 | 陕西师范大学 | 一种推荐方法及*** |
CN111898311A (zh) * | 2020-06-24 | 2020-11-06 | 浙江大学 | 一种基于卷积神经网络的多阶段半导体过程虚拟计量方法 |
CN111931948A (zh) * | 2020-04-07 | 2020-11-13 | 北京工业大学 | 用于测量混凝土抗压强度的深度集成森林回归建模方法 |
CN112216101A (zh) * | 2020-09-08 | 2021-01-12 | 吉林大学 | 一种基于弹性学习框架的交通预测方法及*** |
CN112418547A (zh) * | 2020-12-03 | 2021-02-26 | 北京工业大学 | 一种基于gcn-lstm组合模型的公交车站点客流量预测方法 |
CN112508173A (zh) * | 2020-12-02 | 2021-03-16 | 中南大学 | 交通时空序列多步预测方法、***及存储介质 |
CN113421252A (zh) * | 2021-07-07 | 2021-09-21 | 南京思飞捷软件科技有限公司 | 基于改进的卷积神经网络的车身焊接缺陷实际检测方法 |
CN113657582A (zh) * | 2021-08-20 | 2021-11-16 | 西安科技大学 | 一种基于时空数据的时空图神经网络瓦斯浓度预测方法 |
CN113792929A (zh) * | 2021-04-26 | 2021-12-14 | 青岛大学 | 交通流量预测方法、电子设备及存储介质 |
CN113807613A (zh) * | 2021-10-13 | 2021-12-17 | 国网安徽省电力有限公司 | 一种电网工程全周期预算优化配置工具 |
CN113903170A (zh) * | 2021-08-30 | 2022-01-07 | 航天科工广信智能技术有限公司 | 基于结构自适应自演化深度学习的大规模路网交通流预测方法 |
CN114282431A (zh) * | 2021-12-09 | 2022-04-05 | 淮阴工学院 | 一种基于改进sca和qrgru的径流区间预测方法及*** |
CN114418606A (zh) * | 2021-12-01 | 2022-04-29 | 武汉大学 | 基于时空卷积网络的网约车订单需求预测方法 |
CN114707598A (zh) * | 2022-03-31 | 2022-07-05 | 西安交通大学 | 一种混合气体识别方法、***、终端设备及可读存储介质 |
CN115019504A (zh) * | 2022-05-17 | 2022-09-06 | 汕头大学 | 基于新深空时自适应融合图网络的短期交通流预测方法 |
CN115134816A (zh) * | 2021-03-18 | 2022-09-30 | 中国电信股份有限公司 | 基于时空卷积和多时间尺度的基站流量预测方法 |
CN115578851A (zh) * | 2022-07-14 | 2023-01-06 | 西北师范大学 | 一种基于mgcn的交通预测方法 |
CN115951584A (zh) * | 2023-02-09 | 2023-04-11 | 浙江上洋机械股份有限公司 | 用于滚筒杀青机的温度控制***及方法 |
CN113689035B (zh) * | 2021-08-23 | 2023-06-20 | 安徽大学 | 一种基于卷积神经网络的max-doas光谱预测对流层no2廓线的方法 |
CN116596170A (zh) * | 2023-07-18 | 2023-08-15 | 合肥城市云数据中心股份有限公司 | 一种基于时空注意力机制的送达时间智能预测方法 |
Families Citing this family (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109886347A (zh) * | 2019-02-28 | 2019-06-14 | 泉州师范学院 | 基于rbf和lstm模型的多因素网络的血压预测方法 |
CN109859477B (zh) * | 2019-03-15 | 2021-03-19 | 同盾控股有限公司 | 一种拥堵数据的确定方法和装置 |
CN110210644A (zh) * | 2019-04-17 | 2019-09-06 | 浙江大学 | 基于深度神经网络集成的交通流量预测方法 |
CN110231820B (zh) * | 2019-05-31 | 2022-08-05 | 江苏亿科达科技发展有限公司 | 一种基于物联网的车辆行驶控制方法 |
CN110222899B (zh) * | 2019-06-12 | 2022-04-22 | 哈尔滨工业大学 | 一种越野滑雪赛道风速场预测方法 |
CN110517494A (zh) * | 2019-09-03 | 2019-11-29 | 中国科学院自动化研究所 | 基于集成学习的交通流预测模型、预测方法、***、装置 |
CN111833605B (zh) * | 2020-07-10 | 2022-04-26 | 北京嘀嘀无限科技发展有限公司 | 路况预测方法、路况预测模型训练方法、装置及存储介质 |
CN111950810B (zh) * | 2020-08-27 | 2023-12-15 | 南京大学 | 一种基于自演化预训练的多变量时间序列预测方法和设备 |
CN112183846B (zh) * | 2020-09-25 | 2022-04-19 | 合肥工业大学 | 基于模糊c均值聚类的tvf-emd-mcqrnn负荷概率预测方法 |
CN114648880B (zh) * | 2022-05-24 | 2022-09-06 | 阿里巴巴达摩院(杭州)科技有限公司 | 预测交通流量的方法、车辆和可读存储介质 |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9286524B1 (en) * | 2015-04-15 | 2016-03-15 | Toyota Motor Engineering & Manufacturing North America, Inc. | Multi-task deep convolutional neural networks for efficient and robust traffic lane detection |
CN105654729A (zh) * | 2016-03-28 | 2016-06-08 | 南京邮电大学 | 一种基于卷积神经网络的短时交通流量预测方法 |
CN106447119A (zh) * | 2016-10-11 | 2017-02-22 | 济南观澜数据技术有限公司 | 一种基于卷积神经网络的短期交通流量预测方法及*** |
CN107103754A (zh) * | 2017-05-10 | 2017-08-29 | 华南师范大学 | 一种道路交通状况预测方法及*** |
CN107529651A (zh) * | 2017-08-18 | 2018-01-02 | 北京航空航天大学 | 一种基于深度学习的城市交通客流预测方法和设备 |
-
2018
- 2018-08-01 CN CN201810866657.4A patent/CN108830430A/zh active Pending
- 2018-08-09 WO PCT/CN2018/099498 patent/WO2020024319A1/zh active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9286524B1 (en) * | 2015-04-15 | 2016-03-15 | Toyota Motor Engineering & Manufacturing North America, Inc. | Multi-task deep convolutional neural networks for efficient and robust traffic lane detection |
CN105654729A (zh) * | 2016-03-28 | 2016-06-08 | 南京邮电大学 | 一种基于卷积神经网络的短时交通流量预测方法 |
CN106447119A (zh) * | 2016-10-11 | 2017-02-22 | 济南观澜数据技术有限公司 | 一种基于卷积神经网络的短期交通流量预测方法及*** |
CN107103754A (zh) * | 2017-05-10 | 2017-08-29 | 华南师范大学 | 一种道路交通状况预测方法及*** |
CN107529651A (zh) * | 2017-08-18 | 2018-01-02 | 北京航空航天大学 | 一种基于深度学习的城市交通客流预测方法和设备 |
Cited By (36)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111414482A (zh) * | 2020-03-20 | 2020-07-14 | 北京百度网讯科技有限公司 | 一种事件论元抽取方法、装置以及电子设备 |
CN111414482B (zh) * | 2020-03-20 | 2024-02-20 | 北京百度网讯科技有限公司 | 一种事件论元抽取方法、装置以及电子设备 |
CN111931948A (zh) * | 2020-04-07 | 2020-11-13 | 北京工业大学 | 用于测量混凝土抗压强度的深度集成森林回归建模方法 |
CN111680781A (zh) * | 2020-04-20 | 2020-09-18 | 北京迈格威科技有限公司 | 神经网络处理方法、装置、电子设备及存储介质 |
CN111627101A (zh) * | 2020-05-22 | 2020-09-04 | 北京工业大学 | 一种基于图卷积的三维人体重构方法 |
CN111737578A (zh) * | 2020-06-22 | 2020-10-02 | 陕西师范大学 | 一种推荐方法及*** |
CN111737578B (zh) * | 2020-06-22 | 2024-04-02 | 陕西师范大学 | 一种推荐方法及*** |
CN111898311A (zh) * | 2020-06-24 | 2020-11-06 | 浙江大学 | 一种基于卷积神经网络的多阶段半导体过程虚拟计量方法 |
CN112216101A (zh) * | 2020-09-08 | 2021-01-12 | 吉林大学 | 一种基于弹性学习框架的交通预测方法及*** |
CN112508173A (zh) * | 2020-12-02 | 2021-03-16 | 中南大学 | 交通时空序列多步预测方法、***及存储介质 |
CN112418547A (zh) * | 2020-12-03 | 2021-02-26 | 北京工业大学 | 一种基于gcn-lstm组合模型的公交车站点客流量预测方法 |
CN112418547B (zh) * | 2020-12-03 | 2024-04-09 | 北京工业大学 | 一种基于gcn-lstm组合模型的公交车站点客流量预测方法 |
CN115134816A (zh) * | 2021-03-18 | 2022-09-30 | 中国电信股份有限公司 | 基于时空卷积和多时间尺度的基站流量预测方法 |
CN115134816B (zh) * | 2021-03-18 | 2024-03-08 | 中国电信股份有限公司 | 基于时空卷积和多时间尺度的基站流量预测方法 |
CN113792929B (zh) * | 2021-04-26 | 2023-09-22 | 青岛大学 | 交通流量预测方法、电子设备及存储介质 |
CN113792929A (zh) * | 2021-04-26 | 2021-12-14 | 青岛大学 | 交通流量预测方法、电子设备及存储介质 |
CN113421252B (zh) * | 2021-07-07 | 2024-04-19 | 南京思飞捷软件科技有限公司 | 一种基于改进的卷积神经网络的车身焊接缺陷的检测方法 |
CN113421252A (zh) * | 2021-07-07 | 2021-09-21 | 南京思飞捷软件科技有限公司 | 基于改进的卷积神经网络的车身焊接缺陷实际检测方法 |
CN113657582B (zh) * | 2021-08-20 | 2024-04-09 | 西安科技大学 | 一种基于时空数据的时空图神经网络瓦斯浓度预测方法 |
CN113657582A (zh) * | 2021-08-20 | 2021-11-16 | 西安科技大学 | 一种基于时空数据的时空图神经网络瓦斯浓度预测方法 |
CN113689035B (zh) * | 2021-08-23 | 2023-06-20 | 安徽大学 | 一种基于卷积神经网络的max-doas光谱预测对流层no2廓线的方法 |
CN113903170A (zh) * | 2021-08-30 | 2022-01-07 | 航天科工广信智能技术有限公司 | 基于结构自适应自演化深度学习的大规模路网交通流预测方法 |
CN113903170B (zh) * | 2021-08-30 | 2023-07-14 | 航天科工广信智能技术有限公司 | 基于结构自适应自演化深度学习的大规模路网交通流预测方法 |
CN113807613A (zh) * | 2021-10-13 | 2021-12-17 | 国网安徽省电力有限公司 | 一种电网工程全周期预算优化配置工具 |
CN114418606A (zh) * | 2021-12-01 | 2022-04-29 | 武汉大学 | 基于时空卷积网络的网约车订单需求预测方法 |
CN114418606B (zh) * | 2021-12-01 | 2024-05-28 | 武汉大学 | 基于时空卷积网络的网约车订单需求预测方法 |
CN114282431B (zh) * | 2021-12-09 | 2023-08-18 | 淮阴工学院 | 一种基于改进sca和qrgru的径流区间预测方法及*** |
CN114282431A (zh) * | 2021-12-09 | 2022-04-05 | 淮阴工学院 | 一种基于改进sca和qrgru的径流区间预测方法及*** |
CN114707598A (zh) * | 2022-03-31 | 2022-07-05 | 西安交通大学 | 一种混合气体识别方法、***、终端设备及可读存储介质 |
CN115019504A (zh) * | 2022-05-17 | 2022-09-06 | 汕头大学 | 基于新深空时自适应融合图网络的短期交通流预测方法 |
CN115578851A (zh) * | 2022-07-14 | 2023-01-06 | 西北师范大学 | 一种基于mgcn的交通预测方法 |
CN115578851B (zh) * | 2022-07-14 | 2024-06-07 | 西北师范大学 | 一种基于mgcn的交通预测方法 |
CN115951584A (zh) * | 2023-02-09 | 2023-04-11 | 浙江上洋机械股份有限公司 | 用于滚筒杀青机的温度控制***及方法 |
CN115951584B (zh) * | 2023-02-09 | 2024-03-15 | 浙江上洋机械股份有限公司 | 用于滚筒杀青机的温度控制***及方法 |
CN116596170A (zh) * | 2023-07-18 | 2023-08-15 | 合肥城市云数据中心股份有限公司 | 一种基于时空注意力机制的送达时间智能预测方法 |
CN116596170B (zh) * | 2023-07-18 | 2023-09-22 | 合肥城市云数据中心股份有限公司 | 一种基于时空注意力机制的送达时间智能预测方法 |
Also Published As
Publication number | Publication date |
---|---|
CN108830430A (zh) | 2018-11-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2020024319A1 (zh) | 用于交通流量预测的卷积神经网络多点回归预测模型 | |
Qin et al. | A novel combined prediction scheme based on CNN and LSTM for urban PM 2.5 concentration | |
Bui et al. | Spatial-temporal graph neural network for traffic forecasting: An overview and open research issues | |
CN111161535B (zh) | 基于注意力机制的图神经网络交通流量预测方法及*** | |
Lv et al. | Deep learning in the COVID-19 epidemic: A deep model for urban traffic revitalization index | |
CN107529651B (zh) | 一种基于深度学习的城市交通客流预测方法和设备 | |
CN111612243B (zh) | 交通速度预测方法、***及存储介质 | |
Bi et al. | Large-scale water quality prediction with integrated deep neural network | |
CN107909206B (zh) | 一种基于深层结构循环神经网络的pm2.5预测方法 | |
CN112071065A (zh) | 一种基于全局扩散卷积残差网络的交通流预测方法 | |
CN110648014B (zh) | 一种基于时空分位数回归的区域风电预测方法及*** | |
CN114220271A (zh) | 基于动态时空图卷积循环网络的交通流预测方法、设备及存储介质 | |
Deng et al. | Deep insight into daily runoff forecasting based on a CNN-LSTM model | |
CN111047078A (zh) | 交通特征预测方法、***及存储介质 | |
Zhou et al. | Numerical solution for ruin probability of continuous time model based on neural network algorithm | |
Lv et al. | An improved long short-term memory neural network for stock forecast | |
Wang et al. | Spatial–temporal multi-feature fusion network for long short-term traffic prediction | |
Zhang et al. | IGCRRN: Improved Graph Convolution Res-Recurrent Network for spatio-temporal dependence capturing and traffic flow prediction | |
Lv et al. | A new approach to COVID-19 data mining: A deep spatial–temporal prediction model based on tree structure for traffic revitalization index | |
Wang et al. | Mid-long term forecasting of reservoir inflow using the coupling of time-varying filter-based empirical mode decomposition and gated recurrent unit | |
Feng et al. | Hydrological uncertainty processor (HUP) with estimation of the marginal distribution by a Gaussian mixture model | |
Liao et al. | MFGCN: a multimodal fusion graph convolutional network for online car-hailing demand prediction | |
CN111160600A (zh) | 行为预测方法、装置、计算机设备和存储介质 | |
CN112668711B (zh) | 基于深度学习的洪水流量预测方法、装置和电子设备 | |
Lin et al. | Exploring a similarity search-based data-driven framework for multi-step-ahead flood forecasting |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 18928190 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 18928190 Country of ref document: EP Kind code of ref document: A1 |