WO2011060730A1 - 预测交通流的方法和装置 - Google Patents

预测交通流的方法和装置 Download PDF

Info

Publication number
WO2011060730A1
WO2011060730A1 PCT/CN2010/078883 CN2010078883W WO2011060730A1 WO 2011060730 A1 WO2011060730 A1 WO 2011060730A1 CN 2010078883 W CN2010078883 W CN 2010078883W WO 2011060730 A1 WO2011060730 A1 WO 2011060730A1
Authority
WO
WIPO (PCT)
Prior art keywords
traffic flow
frequency signal
prediction
low frequency
predicting
Prior art date
Application number
PCT/CN2010/078883
Other languages
English (en)
French (fr)
Inventor
昝艳
付新刚
贾学力
李建军
Original Assignee
北京世纪高通科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京世纪高通科技有限公司 filed Critical 北京世纪高通科技有限公司
Publication of WO2011060730A1 publication Critical patent/WO2011060730A1/zh

Links

Classifications

    • GPHYSICS
    • G08SIGNALLING
    • G08GTRAFFIC CONTROL SYSTEMS
    • G08G1/00Traffic control systems for road vehicles
    • G08G1/01Detecting movement of traffic to be counted or controlled
    • G08G1/0104Measuring and analyzing of parameters relative to traffic conditions

Definitions

  • the present invention relates to the field of intelligent transportation, and in particular to a method and apparatus for predicting traffic flow.
  • the forecast of traffic flow is an important part of the intelligent transportation system. It plays an important role in traffic travel, urban or regional traffic planning, and supporting traffic control systems.
  • the traffic flow prediction technology can reasonably calculate the traffic conditions in the next 15 minutes through the currently known traffic flow data, thereby intelligently and dynamically planning the route and guiding people's travel routes.
  • the inventors have found that since the traffic flow data is complex and random, and changes frequently, when the above method is used for prediction, the prediction results are inaccurate and the errors are large under some traffic conditions.
  • Embodiments of the present invention provide a method and apparatus for predicting traffic flow that can more accurately predict traffic flow.
  • a method of predicting traffic flow including:
  • a device for predicting traffic flow comprising:
  • a decomposition unit configured to decompose the pre-acquired traffic flow data into a set of low frequency signals and a set of higher frequency signals
  • a first prediction unit configured to predict a development trend of the low frequency signal obtained by the decomposition unit, and obtain a prediction result of the low frequency signal
  • a second prediction unit configured to predict a development trend of the high frequency signal obtained by the decomposition unit, and obtain a prediction result of the high frequency signal
  • a third prediction unit configured to perform traffic flow prediction based on a prediction result of the low frequency signal acquired by the first prediction unit and a prediction result of the high frequency signal acquired by the second prediction unit.
  • the method and device for predicting traffic flow convert a set of complex and random traffic flow data into several groups by decomposing the traffic flow data into a set of low frequency signals and a set of high frequency signals. a signal of regularity and characteristic; predicting the low frequency signal and the high frequency signal separately, performing traffic flow prediction according to the prediction result, and solving the prediction result caused by directly predicting the traffic flow data in the prior art Inaccurate question.
  • the method and apparatus for predicting traffic flow provided by embodiments of the present invention can predict traffic flow more accurately.
  • FIG. 1 is a flowchart of a method for predicting traffic flow according to an embodiment of the present invention
  • FIG. 2 is a flowchart of a method for predicting a development trend of a low frequency signal according to an embodiment of the present invention
  • FIG. 3 is a schematic structural diagram 1 of a device for predicting traffic flow according to an embodiment of the present invention
  • FIG. 4 is a schematic diagram of predicting traffic flow according to an embodiment of the present invention
  • Schematic diagram of the decomposition unit 301 in the device
  • FIG. 5 is a schematic structural diagram of a third prediction unit 304 in an apparatus for predicting traffic flow according to an embodiment of the present invention.
  • embodiments of the present invention provide a method and apparatus for predicting traffic flow.
  • the method and apparatus provided by the embodiments of the present invention are described in detail below with reference to the accompanying drawings.
  • a method for predicting traffic flow includes: Step 101: Decompose pre-acquired traffic flow data into a set of low frequency signals and a set of high frequency signals;
  • the traffic flow data is one-day traffic flow data of a designated road link, and is a set of one-dimensional discrete signals.
  • the acquired traffic flow data components are complex and frequently changing, and contain many signals of different natures. These different nature signals play different roles in the process of traffic flow prediction. Only the signal components of these different properties are treated differently. In order to improve the prediction accuracy.
  • the traffic flow data is decomposed using a wavelet decomposition method, and the traffic flow data is decomposed into a set of low frequency signals and a set of higher frequency signals.
  • the low frequency signal is a basic signal expressing an essential trend of the original traffic flow data; the plurality of high frequency signals are mutually uncorrelated wave signals.
  • the wavelet function needs to be selected. Since the same engineering problem is decomposed by different wavelet functions, the obtained results sometimes differ greatly. Therefore, in this embodiment, it may be adopted. Experience values, or use a continuous experiment to select wavelet functions.
  • Step 102 predict a development trend of the low frequency signal, and obtain a prediction result of the low frequency signal
  • the time series prediction method can be used to predict the development trend of the low-frequency signal.
  • other methods can also be used for prediction, and no longer— - enumeration.
  • the steps of predicting by time series prediction include:
  • Step 201 Perform a time series smoothing test on the low frequency signal.
  • the object of the prediction of the development trend of the low-frequency signal should be a low-frequency signal with a stable time series. Therefore, it is first necessary to check whether the low-frequency signal obtained after the wavelet decomposition is stable.
  • the intuitive meaning of the smoothness of the time series is that there is no trend and periodicity in the time series.
  • the time series is said to be a strictly stationary time series; if the first and second moments of a time series exist, and at any time, the mean of the time series is constant, and The covariance is a function of the time interval, and the time series is said to be a wide stationary time series.
  • the time series is a low frequency signal obtained after wavelet decomposition, and the low frequency signal is considered to be stable as long as the low frequency signal is a wide stationary time sequence, and may be predicted, ie Going to step 203; if the low frequency signal does not satisfy the stationary characteristic, it is subjected to smoothing processing before the prediction, that is, the process proceeds to step 202 to make it a smooth low frequency signal.
  • determining whether the low-frequency signal is stable can be tested according to the statistical significance of the low-frequency signal; or can be tested from the intuitive meaning of the stationarity by drawing, that is, whether the data graph has a trend or periodicity, if not obvious The trend or periodicity, the low-frequency signal is considered to be stationary; or it can be tested by the method of insight, such as autocorrelation function test, partial autocorrelation function test, eigenvalue test, Parameter test, and run test.
  • Step 202 If the low frequency signal is not stable, perform a smoothing process on the low frequency signal.
  • the smooth processing includes methods for eliminating trend items and periodic items of the original sequence, difference methods, and fitting of mean functions.
  • the smoothing test is performed again. If it passes, go to step 203. If not, continue this step until it passes the smoothing test.
  • Step 203 Perform model identification on a low frequency signal that satisfies a smooth condition
  • the purpose of model recognition is to select a model that matches the low frequency signal from the stationary time series model.
  • the stationary time series model is the following three: Autoregressive model (Aut. Regression, AR) , Moving average ( ⁇ ), self-returning slip
  • ARM A Auto Regression-Moving Average
  • is a white noise sequence
  • AR(p) the cylinder is denoted as AR(p).
  • ⁇ ⁇ ⁇ ⁇ + ⁇ 1 ⁇ ⁇ _ 1 + ⁇ + ⁇ ( ⁇ ⁇ ⁇ _ ⁇
  • the moving average coefficient is the white noise order ⁇ ij
  • the above formula is called the step moving average model.
  • the cylinder is recorded as MA(g).
  • autoregressive coefficient satisfies the condition of stationarity, called the moving average coefficient, and is the white noise sequence.
  • the above formula is called the p-order autoregressive _ q-order moving average model, and the cylinder is ARMA (p, q). ).
  • the AR model and the MA model are special cases of the ARMA model.
  • the tailing in Table 1 means that the model autocorrelation function or partial autocorrelation function exhibits exponential decay and tends to zero with increasing time lag, while truncation refers to the model's autocorrelation function or partial autocorrelation function at a certain step. After all, it is zero.
  • Table 1 if the autocorrelation function of the low frequency signal is smear, and the partial autocorrelation function is truncated, an AR model may be selected; if the autocorrelation function of the low frequency signal is truncated The partial autocorrelation function is smeared, and the MA model can be selected. If the autocorrelation function and the partial autocorrelation function of the low frequency signal are both smeared, the ARMA model can be selected.
  • Step 204 Perform a ranking on the selected model.
  • the specific ordering methods include residual variance pattern fixed-order method, F-test fixed-order method, optimal criterion fixed-order method, and autocorrelation function and partial autocorrelation function fixed-order method.
  • the specific method of ordering using autocorrelation and partial autocorrelation function is: If the AR model is selected, and the partial autocorrelation function of the low frequency signal is truncated after p step, the order of the AR model is p; if selected The MA model, and the autocorrelation function of the low frequency signal is truncated after the q step, then the order of the MA model is q.
  • Other fixed-order methods are not described here.
  • Step 205 Perform parameter estimation on the determined model.
  • the parameter estimation can be performed.
  • moment estimation maximum likelihood estimation
  • least squares estimation Least squares estimates and maximum likelihood estimates have higher precision and are therefore generally referred to as fine estimates of model parameters.
  • the maximum likelihood estimation method is more complicated.
  • the last solved equations are all nonlinear equations, which are difficult to solve, so numerical algorithms can generally be used.
  • the method is to arbitrarily give a set of values of the parameters, and initially obtain the result, and calculate a likelihood function value; then, according to a certain rule, a set of values of the parameter is given, and a likelihood function value is calculated. And so on, compare the likelihood function values, and select the set of parameters that maximizes the likelihood function value.
  • the moment estimation method is used.
  • the AR model is taken as an example to illustrate the method:
  • the parameter that needs to be estimated is ⁇ ... .
  • y t y t -j y t — P y t —j + s t y t _
  • ⁇ + ⁇ 2 ⁇ ]-2 + ⁇ + ⁇ ⁇ ⁇ , ⁇ / ⁇ > 0
  • the autocorrelation function can be replaced by the autocorrelation function of the sample (ie, the low frequency signal in this embodiment), so the Yule-Walker equation at this time has only p unknowns, and the solution equation can be obtained.
  • Estimated value expressed in matrix
  • the prediction model used by the time series prediction method establishes step 206 to verify the rationality of the prediction model; For the prediction model established in the above steps, it is necessary to test to see if it is reasonable.
  • the specific test method can use the substitution method to compare the output result with the known result. If the comparison result is within the specified error, the prediction model is reasonable and can be used for prediction; if the comparison result exceeds the specified The error, the prediction model is unreasonable, and the prediction model needs to be re-established until the established prediction model meets the requirements.
  • Step 207 Prediction of a development trend of the low frequency signal according to the established prediction model.
  • the orthogonal prediction method or the conditional expectation method may be used to predict the prediction model of the established low frequency signal. Since the low frequency signal is a low frequency signal after smoothing, in the smoothing step, the trend term and the period term are eliminated, so the prediction result is only a prediction result of a part of the original low frequency signal. In this step, the trend item and the period item are also added to obtain the final prediction result of the original low frequency signal.
  • Step 103 predict a development trend of the high frequency signal, and obtain a prediction result of the high frequency signal;
  • a neural network prediction method can be selected to predict the development trend of the high frequency signal.
  • Neural network is an emerging mathematical modeling method, which has the characteristics of identifying complex nonlinear systems.
  • the Back Propagation (BP) algorithm in the neural network model is selected, which is generally called BP neural network.
  • the network model which needs to work through learning and training, consists of two processes: forward propagation of information and back propagation of errors.
  • the specific learning and training process of the BP model is as follows:
  • the input layer neurons are responsible for receiving input information from the outside world and transmitting them to the middle layer neurons;
  • the middle layer is the internal information processing layer, responsible for information transformation, according to the needs of information transformation capabilities,
  • the middle layer can be designed as a single hidden layer or multiple hidden layer structure; the last hidden layer is passed to the information of each neuron in the output layer, and after further processing, a forward propagation process of learning is completed, and the output layer outputs information to the outside world. process result.
  • the backpropagation phase of the error is performed.
  • the error is corrected according to the error gradient, and the hidden layer and the input layer are back-transported layer by layer.
  • the process of information forward propagation and error back propagation is the process of constant adjustment of weights at all levels, and also the process of neural network learning and training. This process continues until the error in the network output is reduced to an acceptable level.
  • the BP neural network model with acceptable error can be used.
  • To predict the development trend of the high-frequency signal obtained after wavelet decomposition that is, input the high-frequency signal as input data into the BP neural network model, and obtain an output result after the model is calculated, and the result is a high frequency. The predicted value of the signal.
  • Step 104 Perform traffic flow prediction according to the prediction result of the low frequency signal and the prediction result of the high frequency signal.
  • weight values of prediction results of the low frequency signal and the high frequency signal are respectively set, and a weighted average value of prediction results of the low frequency signal and the high frequency signal is calculated according to the set weight value,
  • the resulting weighted average is the predicted value that ultimately contains the original traffic flow information.
  • the method for predicting traffic flow converts a set of complex and random traffic flow data into groups with certain regularity by decomposing the traffic flow data into a set of low frequency signals and a set of high frequency signals. a characteristic signal; predicting the low frequency signal and the high frequency signal separately, performing traffic flow prediction according to the prediction result, and solving the inaccurate prediction result caused by directly predicting the traffic flow data in the prior art The problem.
  • the method and apparatus for predicting traffic flow provided by embodiments of the present invention are capable of predicting traffic flow more accurately.
  • an embodiment of the present invention further provides an apparatus for predicting traffic flow, where the apparatus includes:
  • the decomposition unit 301 is configured to decompose the pre-acquired traffic flow data into a set of low frequency signals and a set of high frequency signals;
  • the traffic flow data is decomposed using a wavelet decomposition method, and the traffic flow data is decomposed into a set of low frequency signals and a set of higher frequency signals.
  • the low frequency signal is a basic signal expressing an essential change trend of the original traffic flow data; and the set of the high frequency signals are mutually unrelated fluctuation signals.
  • the first prediction unit 302 is configured to predict a development trend of the low frequency signal obtained by the decomposition unit 301, and obtain a prediction result of the low frequency signal;
  • the time series prediction method can be used to predict the development trend of the low frequency signal.
  • a second prediction unit 303 configured to predict a development trend of the high frequency signal obtained by the decomposition unit 301, and obtain a prediction result of the high frequency signal;
  • a neural network prediction method can be selected to predict the development trend of the high frequency signal.
  • the third prediction unit 304 is configured to perform traffic flow prediction according to the prediction result of the low frequency signal acquired by the first prediction unit 302 and the prediction result of the high frequency signal acquired by the second prediction unit 303.
  • weight values of prediction results of the low frequency signal and the high frequency signal are respectively set, and a weighted average value of prediction results of the low frequency signal and the high frequency signal is calculated according to the set weight value,
  • the resulting weighted average is the predicted value that ultimately contains the original traffic flow information.
  • the decomposition unit 301 includes:
  • the selecting unit 3011 is configured to select a wavelet function
  • the decomposing subunit 3012 is configured to decompose the pre-acquired traffic flow data according to the wavelet function selected by the selecting unit 3011.
  • the third prediction unit 304 includes:
  • a setting unit 3041 configured to separately set weights of prediction results of the low frequency signal and the high frequency signal acquired by the first prediction unit 302 and the second prediction unit 303;
  • the obtaining unit 3042 is configured to obtain a weighted average value of the prediction result of the low frequency signal and the high frequency signal according to the weight set by the setting unit 3041;
  • the third prediction subunit 3043 is configured to predict the traffic flow based on the weighted average obtained by the acquisition unit 3042.
  • the device for predicting traffic flow converts a set of complex and random traffic flow data into groups with certain regularity by decomposing the traffic flow data into a set of low frequency signals and a set of high frequency signals. a characteristic signal; predicting the low frequency signal and the high frequency signal separately, performing traffic flow prediction according to the prediction result, and solving the inaccurate prediction result caused by directly predicting the traffic flow data in the prior art The problem.
  • the method and apparatus for predicting traffic flow provided by embodiments of the present invention are capable of predicting traffic flow more accurately.
  • the technical solution provided by the embodiment of the invention is applicable to the traffic flow prediction system for urban road crossing Forecasting through the flow.

Landscapes

  • Chemical & Material Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Traffic Control Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Description

预测交通流的方法和装置 本申请要求于 2 009 年 1 1 月 1 9 日提交中国专利局、 申请号为 2 009 1 02 379 39. 9、 发明名称为 "预测交通流的方法和装置" 的中国 专利申请的优先权, 其全部内容通过引用结合在本申请中。
技术领域
本发明涉及智能交通领域, 尤其涉及一种预测交通流的方法和 装置。
背景技术
交通流的预测是智能交通***的重要组成部分, 它对交通出行、 城市或区域交通规划、 支撑交通控制***等都有重要的作用。 交通 流预测技术能够通过当前已知的交通流数据, 合理地推算出未来 1 5 分钟内的交通状况, 从而智能动态地进行路径规划, 指导人们的出行 路线。
现有技术中用于预测交通流的方法很多, 常用的方法有: 回归 分析法、 时间序列法、 神经网络法和卡尔曼滤波法等。
在实现本发明的过程中, 发明人发现, 由于交通流数据具有复 杂性和随机性, 并且变化频繁, 当采用上述方法进行预测时, 在有 些交通条件下预测结果不准确、 误差较大。
发明内容
本发明的实施例提供一种预测交通流的方法和装置,能够更加准确地 预测交通流。
为达到上述目的, 本发明的实施例采用如下技术方案:
一种预测交通流的方法, 包括:
将预先获取的交通流数据分解为一组低频信号和一组以上高频信号; 预测所述低频信号的发展趋势, 获取所述低频信号的预测结果; 预测所述高频信号的发展趋势, 获取所述高频信号的预测结果; 根据所述低频信号的预测结果和所述高频信号的预测结果进行交通 流预测。
一种预测交通流的装置, 包括:
分解单元,用于将预先获取的交通流数据分解为一组低频信号和一组 以上高频信号;
第一预测单元, 用于预测由所述分解单元获得的低频信号的发展趋 势, 获取所述低频信号的预测结果;
第二预测单元, 用于预测由所述分解单元获得的高频信号的发展趋 势, 获取所述高频信号的预测结果;
第三预测单元,用于根据由所述第一预测单元获取的低频信号的预测 结果和由所述第二预测单元获取的高频信号的预测结果进行交通流预测。
本发明实施例提供的预测交通流的方法和装置,通过将交通流数据分 解为一组低频信号和一组以上高频信号,从而将一组复杂、 随机的交通流 数据转变为几组具有一定规律和特性的信号;并对所述低频信号和高频信 号分别进行预测, 根据所述预测结果进行交通流预测, 解决了现有技术中 直接对所述交通流数据进行预测而造成的预测结果不准确的问题。本发明 的实施例提供的预测交通流的方法和装置, 能够更加准确地预测交通流。 附图说明
图 1为本发明实施例提供的预测交通流的方法流程图;
图 2为本发明实施例提供的预测低频信号发展趋势的方法流程图; 图 3为本发明实施例提供的预测交通流的装置结构示意图一; 图 4为本发明实施例提供的预测交通流的装置中分解单元 301的结构 示意图;
图 5 为本发明实施例提供的预测交通流的装置中第三预测单元 304 的结构示意图。
具体实施方式
为使本发明实施例的目的、技术方案和优点更加清楚, 下面将结合本 发明实施例中的附图, 对本发明实施例中的技术方案进行清楚、 完整地描 述,显然,所描述的实施例是本发明一部分实施例,而不是全部的实施例。 基于本发明中的实施例,本领域普通技术人员在没有付出创造性劳动前提 下所获得的所有其他实施例, 都属于本发明保护的范围。
为了解决现有技术中直接对所获取的交通流数据进行预测而造成的 预测结果不准确的问题, 本发明实施例提供一种预测交通流的方法和装 置。 下面结合附图对本发明实施例提供的方法和装置进行详细描述。
如图 1所示, 本发明实施例提供的预测交通流的方法, 包括: 步骤 101 , 将预先获取的交通流数据分解为一组低频信号和一组以上 高频信号;
在本实施例中, 所述交通流数据是指定路链的一天的交通流数据, 是 一组一维的离散信号。 所获取的交通流数据成分复杂、 变化频繁, 并且其 中包含有许多不同性质的信号,这些不同性质的信号在交通流的预测过程 中所起的作用不同, 只有区别对待这些不同性质的信号成分, 才能提高预 测精度。
大量仿真交通流和现实交通流的非线性特征的分析表明,交通流数据 在不同的时间频率尺度上具有自相似性和多尺度特性,因此可以将一组包 含不同性质信号的交通流数据分解为几组具有特定性质的信号。本实施例 中, 采用小波分解的方法分解所述交通流数据, 将所述交通流数据分解为 一组低频信号和一组以上高频信号。 其中, 所述低频信号为表达原始交通 流数据本质变化趋势的基本信号;所述一组以上高频信号为互不相关的波 动信号。此外,在对交通流数据进行小波分解之前,还需要选择小波函数, 由于同一个工程问题用不同的小波函数进行分解时,所得到的结果有时相 差甚远, 因此在本实施例中, 可以采用经验值, 或者采用不断试验的方法 来选择小波函数。
步骤 102 , 预测所述低频信号的发展趋势, 获取所述低频信号的预测 结果;
在本实施例中, 由于经过小波分解后得到的低频信号变化稳定, 因此 可以采用时间序列预测法预测所述低频信号的发展趋势, 当然, 也可以采 用其它的方法进行预测, 此处不再——列举。 如图 2所示, 采用时间序列 预测法预测的步骤包括:
步骤 201 , 对所述低频信号进行时间序列平稳化检验; 本实施例中,预测所述低频信号的发展趋势的作用对象应为时间序列 平稳的低频信号,因此首先要检验经过小波分解后得到的低频信号是否是 平稳的。时间序列平稳的直观含义就是时间序列中不存在任何的趋势性和 周期性。按时间序列的统计特性划分, 有平稳时间序列和非平稳时间序列 两类。 其中, 所述平稳时间序列又分为严格平稳时间序列和宽平稳时间序 列。 如果一个时间序列的概率分布与时间无关, 则称该时间序列为严格平 稳时间序列; 如果一个时间序列的一、 二阶矩存在, 而且在任意时刻, 所 述时间序列的均值为常数, 并且其协方差为时间间隔的函数, 则称所述时 间序列为宽平稳时间序列。 具体到本实施例中, 所述时间序列即为经过小 波分解后得到的低频信号, 只要所述低频信号为宽平稳时间序列, 则认为 所述低频信号是平稳的, 可以对其进行预测, 即转入步骤 203 ; 如果所述 低频信号不满足平稳特性, 则在预测之前, 还要对其进行平稳化处理, 即 转入步骤 202 , 使其成为平稳的低频信号。
具体地, 判断所述低频信号是否平稳, 可以按照低频信号的统计意义 来检验; 也可以通过绘图从平稳性的直观意义来检验, 即观察其数据图是 否存在趋势性或周期性, 如果没有明显的趋势性或周期性, 就认为所述低 频信号是平稳的; 或者, 也可以通过见解反推的方法来检验, 例如, 自相 关函数检验法, 偏自相关函数检验法, 特征根检验法, 参数检验法, 和游 程检验法等。
步骤 202 , 若所述低频信号不平稳, 则对其进行平稳化处理; 在本实施例中, 当所述低频信号经过步骤 201的检验后, 若不符合平 稳的条件, 则还需要对其进行平稳化处理。 具体的平稳化处理的方法有, 消除原始序列的趋势项和周期项的方法、 差分法、 均值函数拟合等。 对于 完成平稳化处理的低频信号, 再次进行平稳化检验, 若通过, 转入步骤 203 , 若不通过, 继续此步骤, 直到通过平稳化检验。
步骤 203 , 对满足平稳化条件的低频信号进行模型识别;
本实施例中,进行模型识别的目的就是从平稳时间序列模型中选择一 个与所述低频信号相吻合的模型, 所述平稳时间序列模型为以下三个: 自 回归模型(Aut。 Regression ,AR) , 滑动平均模型(Moving Average ,ΜΑ) , 自回归滑 动平均模型 (Auto Regression-Moving Average, ARM A) , 下面筒单介绍所述三个时 间序列模型:
( 1 ) 自回归(AR)模型的一般 AR模型的数学形式为:
yt = {-ι + <kyt—i +··· + p + st
其中, … 称为自回归系数,满足平稳性条件, ^为白噪声序列, 上式称为是 阶自回归模型, 筒记为 AR(p).
( 2) 滑动平均(MA)模型的一般 MA模型的数学形式为:
γίί1εί_1 + ··· + φ(}εί_ί} 其中, 称为滑动平均系数, 为白噪声序歹 ij ,上式称为是 阶滑动平均模型, 筒记为 MA(g).
( 3 )自回归滑动平均(ARMA)模型的一般 ARMA模型的数学形式为:
Figure imgf000007_0001
+… + t-q
其中, , … 称为自回归系数, 满足平稳性条件, 称为 滑动平均系数, 为白噪声序列, 上式称为是 p阶自回归 _ q阶滑动平均 模型, 筒记为 ARMA(p,q).
从以上定义中可以看出, AR模型和 MA模型即为 ARMA模型的特 例, 当 p=0时, p阶自回归 - q阶滑动平均模型 ARMA(p, q)转化为 p阶自 回归模型 AR(p); 当 q=0时, p阶自回归 - q阶滑动平均模型 ARMA(p, q) 转化为 q阶滑动平均模型 MA(q)。
对于三类模型 AR, MA, ARMA, 它们各自的自相关函数和偏自相关 函数特点如表一所示, 采用这些特点可以进行模型识别。
表一:
Figure imgf000007_0002
表一中的拖尾是指模型自相关函数或偏自相关函数随着时滞的增加 呈现指数衰减并趋于零,而截尾则是指模型的自相关函数或偏自相关函数 在某步之后全部为零。 具体地, 如表一所示, 如果所述低频信号的自相关函数是拖尾的, 偏 自相关函数是截尾的, 可选用 AR模型; 如果所述低频信号的自相关函数 是截尾的, 偏自相关函数是拖尾的, 可选用 MA 模型; 如果所述低频信 号的自相关函数和偏自相关函数都是拖尾的, 可选用 ARMA模型。
步骤 204, 对所选择的模型进行定阶;
在步骤 203中确定了模型类型之后, 为了建立具体的模型, 还需要知 道模型的阶数。 具体定阶的方法有残差方差图定阶法, F检验定阶法, 最 佳准则定阶法, 以及自相关函数和偏自相关函数定阶法等。 例如, 采用自 相关和偏自相关函数定阶的具体方法为: 如果选用的是 AR模型, 并且低 频信号的偏自相关函数在 p步以后截尾,则 AR模型的阶数为 p;如果选用 的是 MA模型, 并且低频信号的自相关函数在 q步以后截尾, 则 MA模 型的阶数为 q。 其它定阶方法在此不再赘述。
步骤 205 , 对所述定阶后的模型进行参数估计;
模型的阶数确定之后,就可以进行参数估计了。主要有三种估计方法: 矩估计, 极大似然估计和最小二乘估计。 最小二乘估计和极大似然估计的 精度较高, 因而一般称之为模型参数的精估计。 而极大似然估计计算方法 较为复杂, 最后求解的方程皆为非线性方程, 很难求解, 所以一般可以采 用数值算法。 方法是, 任意给出参数的一组数值, 初步估计得到的结果, 计算出一个似然函数值; 然后,根据一定的法则,再给出参数的一组数值, 又计算出一个似然函数值; 依此类推, 比较似然函数值, 选择使似然函数 值最大的那组参数。 本实施例中采用矩估计法, 下面以 AR模型为例说明 方法:
下式是一零均值的 AR(p)模型
yt = y + ΑΛ— 2 +… + ΦΡ yt-p + ^
需要估计的参数是 ^… 。
在模型两边同乘以 > 0 , 可得
yt yt-j =
Figure imgf000008_0001
ytP yt—j + st yt_
两边取期望, 得
Eyt yt-j =
Figure imgf000008_0002
yt-j + Est yt-j 由于 与 _ _ >0)不相关, 所以 ^.=0, 因此
η = Φ + Φ2 Γ]-2 +···+ Ρη~Ρ , ■/· > 0
其中 ^是序列 { }的自协方差函数, 易知序列的自相关函数 ^也满足 上述关系式, 即
Pj = ΦιΡ + Φ2Ρ]-2 + ·" + ΦρΡ]-ρ , J' = 1, 2, 3…
把自相关函数展开成 p个方程
Ρι = ΦιΡο + 2Ρι + ··· + ρΡΡ
Pi = ΦιΡι + Φ2Ρο + ··· + ΡΡΡ-2
ΡΡ = ΦιΡρ-ι + ΦιΡρ-ι +··· + ρΡο 上述 ρ 个方程, 表示了平稳序列的自相关函数与模型未知参数的关 系, 被称为 Yule- Walker方程。
自相关函数可以用样本(即本实施例中的低频信号)自相关函数代替, 所以此时的 Yule-Walker方程只有 p个未知数, 解方程可以得到
Figure imgf000009_0001
的估计值, 用矩阵表示
Figure imgf000009_0002
对于二阶自回归模型 AR(1), 根据上述结果可知 对于二阶自回归模型 AR(2), 根据上述结果可知
Figure imgf000009_0003
获得模型的参数后, 时间序列预测法所采用的预测模型就建立起来 步骤 206, 检验所述预测模型的合理性; 对于上述步骤建立的预测模型, 还需要进行检验, 看其是否合理。 具 体的检验方法可以采用代入法, 将其输出结果与已知的结果相比较, 如果 比较的结果在指定的误差内, 则所述预测模型合理, 可以用其进行预测; 如果比较的结果超出指定误差, 则所述预测模型不合理, 需要重新建立预 测模型, 直到建立的预测模型符合要求为止。
步骤 207 , 根据所建立的预测模型预测低频信号的发展趋势。
在本实施例中, 可以采用正交投影法, 或者条件期望法对所建立的低 频信号的预测模型进行预测。由于所述低频信号是经过平稳化以后的低频 信号, 在平稳化步骤中, 消除了趋势项和周期项, 所以预测结果只是原低 频信号一部分的预测结果。 本步骤中, 还需要加上所述趋势项和周期项, 以获得原低频信号的最终预测结果。
步骤 103 , 预测所述高频信号的发展趋势, 获取所述高频信号的预测 结果;
本实施例中, 由于所述高频信号是非线性的, 所以可以选用神经网络 预测法来预测高频信号的发展趋势。 神经网络是一种新兴的数学建模方 法, 它具有识别复杂非线性***的特性, 本实施例中选取神经网络模型中 的误差反向传播算法( Back Propagation, BP ),—般称为 BP神经网络模型, 此模型需要通过学习训练才能工作,由信息的正向传播和误差的反向传播 两个过程组成。 BP模型具体的学习训练过程如下: 输入层各神经元负责 接收来自外界的输入信息, 并传递给中间层各神经元; 中间层是内部信息 处理层, 负责信息变换, 根据信息变换能力的需求, 中间层可以设计为单 隐层或者多隐层结构; 最后一个隐层传递到输出层各神经元的信息, 经过 进一步处理后, 完成一次学习的正向传播处理过程, 由输出层向外界输出 信息处理结果。当实际输出与期望输出不符时,进行误差的反向传播阶段。 误差通过输出层, 按照误差梯度下降的方式修正各层权值, 向隐层、 输入 层逐层反传。 周而复始的信息正向传播和误差反向传播的过程, 是各层权 值不断调整的过程, 也是神经网络学习训练的过程。 此过程一直进行到网 络输出的误差减少到可以接受的程度。
对于经过学习训练的、 误差可以被接受的 BP神经网络模型, 可以用 来预测经过小波分解后得到的高频信号的发展趋势,即将所述高频信号作 为输入数据输入到所述 BP神经网络模型中, 经所述模型计算后获得输出 结果, 此结果即为高频信号的预测值。
步骤 104 , 根据所述低频信号的预测结果和所述高频信号的预测结果 进行交通流预测。
本实施例中,首先分别设置所述低频信号和所述高频信号的预测结果 的权值,根据所设置的权值计算所述低频信号和所述高频信号的预测结果 的加权平均值, 得到的加权平均值即为最终包含原始交通流信息的预测 值。
本发明实施例提供的预测交通流的方法,通过将交通流数据分解为一 组低频信号和一组以上高频信号, 从而将一组复杂、 随机的交通流数据转 变为几组具有一定规律和特性的信号;并对所述低频信号和高频信号分别 进行预测,根据所述预测结果进行交通流预测, 解决了现有技术中直接对 所述交通流数据进行预测而造成的预测结果不准确的问题。本发明的实施 例提供的预测交通流的方法和装置, 能够更加准确地预测交通流。
如图 3所示, 本发明实施例还提供一种预测交通流的装置, 所述装置 包括:
分解单元 301 , 用于将预先获取的交通流数据分解为一组低频信号和 一组以上高频信号;
本实施例中, 采用小波分解的方法分解所述交通流数据, 将所述交通 流数据分解为一组低频信号和一组以上高频信号。 其中, 所述低频信号为 表达原始交通流数据本质变化趋势的基本信号;所述一组以上高频信号为 互不相关的波动信号。
第一预测单元 302 , 用于预测由所述分解单元 301获得的低频信号的 发展趋势, 获取所述低频信号的预测结果;
在本实施例中, 由于经过小波分解后得到的低频信号变化稳定, 因此 可以采用时间序列预测法预测所述低频信号的发展趋势。
第二预测单元 303 , 用于预测由所述分解单元 301获得的高频信号的 发展趋势, 获取所述高频信号的预测结果; 本实施例中, 由于所述高频信号是非线性的, 所以可以选用神经网络 预测法来预测高频信号的发展趋势。
第三预测单元 304 , 用于根据由所述第一预测单元 302获取的低频信 号的预测结果和由所述第二预测单元 303 获取的高频信号的预测结果进 行交通流预测。
本实施例中,首先分别设置所述低频信号和所述高频信号的预测结果 的权值,根据所设置的权值计算所述低频信号和所述高频信号的预测结果 的加权平均值, 得到的加权平均值即为最终包含原始交通流信息的预测 值。
进一步地, 如图 4所示, 所述分解单元 301 包括:
选取单元 3011 , 用于选取小波函数;
分解子单元 3012 , 用于根据由所述选取单元 3011选取的小波函数分 解所述预先获取的交通流数据。
进一步地, 如图 5所示, 所述第三预测单元 304包括:
设置单元 3041 , 用于分别设置由所述第一预测单元 302 和所述第二 预测单元 303获取的低频信号和高频信号的预测结果的权值;
获取单元 3042 , 用于根据由所述设置单元 3041设置的权值获取所述 低频信号和所述高频信号的预测结果的加权平均值;
第三预测子单元 3043 , 用于根据由所述获取单元 3042获取的加权平 均值预测交通流。
以上装置的具体实现方法可以参见如图 1和图 2所示的方法流程图, 此处不再赘述。
本发明实施例提供的预测交通流的装置,通过将交通流数据分解为一 组低频信号和一组以上高频信号, 从而将一组复杂、 随机的交通流数据转 变为几组具有一定规律和特性的信号;并对所述低频信号和高频信号分别 进行预测,根据所述预测结果进行交通流预测, 解决了现有技术中直接对 所述交通流数据进行预测而造成的预测结果不准确的问题。本发明的实施 例提供的预测交通流的方法和装置, 能够更加准确地预测交通流。
本发明实施例提供的技术方案,适用于交通流预测***对城市道路交 通流进行预测。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分 步骤是可以通过程序来指令相关的硬件完成,所述的程序可以存储于一计 算机可读存储介质中, 如 ROM/RAM、 磁碟或光盘等。
以上所述, 仅为本发明的具体实施方式, 但本发明的保护范围并不局 限于此, 任何熟悉本技术领域的技术人员在本发明揭露的技术范围内, 可 轻易想到变化或替换, 都应涵盖在本发明的保护范围之内。 因此, 本发明 的保护范围应所述以权利要求的保护范围为准

Claims

权 利 要 求 书
1、 一种预测交通流的方法, 其特征在于, 包括:
将预先获取的交通流数据分解为一组低频信号和一组以上高频信号; 预测所述低频信号的发展趋势, 获取所述低频信号的预测结果; 预测所述高频信号的发展趋势, 获取所述高频信号的预测结果; 根据所述低频信号的预测结果和所述高频信号的预测结果进行交通流 预测。
2、 根据权利要求 1所述的预测交通流的方法, 其特征在于, 所述交通 流数据为一维离散数据。
3、 根据权利要求 1所述的预测交通流的方法, 其特征在于, 所述将预 先获取的交通流数据分解为一组低频信号和一组以上高频信号包括:
选取小波函数;
根据所述小波函数分解所述预先获取的交通流数据。
4、 根据权利要求 1所述的预测交通流的方法, 其特征在于, 所述预测 所述低频信号的发展趋势采用时间序列预测法; 所述预测所述高频信号的 发展趋势采用神经网络预测法。
5、 根据权利要求 1所述的预测交通流的方法, 其特征在于, 所述根据 所述低频信号的预测结果和所述高频信号的预测结果进行交通流预测包 括:
分别设置所述低频信号和所述高频信号的预测结果的权值;
根据所述权值获取所述低频信号和所述高频信号的预测结果的加权平 均值;
根据所述加权平均值预测交通流。
6、 一种预测交通流的装置, 其特征在于, 包括:
分解单元, 用于将预先获取的交通流数据分解为一组低频信号和一组 以上高频信号;
第一预测单元,用于预测由所述分解单元获得的低频信号的发展趋势, 获取所述低频信号的预测结果;
第二预测单元,用于预测由所述分解单元获得的高频信号的发展趋势, 获取所述高频信号的预测结果;
第三预测单元, 用于根据由所述第一预测单元获取的低频信号的预测 结果和由所述第二预测单元获取的高频信号的预测结果进行交通流预测。
7、 根据权利要求 6所述的预测交通流的装置, 其特征在于, 所述分解 单元包括:
选取单元, 用于选取小波函数;
分解子单元, 用于根据由所述选取单元选取的小波函数分解所述预先 获取的交通流数据。
8、 根据权利要求 6所述的预测交通流的装置, 其特征在于, 所述第三 预测单元包括:
设置单元, 用于分别设置由所述第一预测单元和所述第二预测单元获 取的低频信号和高频信号的预测结果的权值;
获取单元, 用于根据由所述设置单元设置的权值获取所述低频信号和 所述高频信号的预测结果的加权平均值;
第三预测子单元, 用于根据由所述获取单元获取的加权平均值预测交 通流。
PCT/CN2010/078883 2009-11-19 2010-11-18 预测交通流的方法和装置 WO2011060730A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN200910237939A CN101739819A (zh) 2009-11-19 2009-11-19 预测交通流的方法和装置
CN200910237939.9 2009-11-19

Publications (1)

Publication Number Publication Date
WO2011060730A1 true WO2011060730A1 (zh) 2011-05-26

Family

ID=42463248

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2010/078883 WO2011060730A1 (zh) 2009-11-19 2010-11-18 预测交通流的方法和装置

Country Status (2)

Country Link
CN (1) CN101739819A (zh)
WO (1) WO2011060730A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106571882A (zh) * 2016-09-06 2017-04-19 北京航空航天大学 一种超外差接收机性能退化评估与预测方法及***
EP3358542A1 (en) * 2017-02-01 2018-08-08 Kapsch TrafficCom AG A method of predicting a traffic behaviour in a road system
CN113496314A (zh) * 2021-09-07 2021-10-12 南京感动科技有限公司 一种神经网络模型预测道路交通流量的方法

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101739819A (zh) * 2009-11-19 2010-06-16 北京世纪高通科技有限公司 预测交通流的方法和装置
CN102034351B (zh) * 2010-09-30 2012-07-25 天津大学 一种交通流短时预测***
CN102184335B (zh) * 2011-05-20 2013-04-10 公安部上海消防研究所 一种基于集合经验模式分解和相空间重构的火灾时间序列预测方法
CN102568205B (zh) * 2012-01-10 2013-12-04 吉林大学 非常态下基于经验模态分解和分类组合预测的交通参数短时预测方法
CN102629418B (zh) * 2012-04-09 2014-10-29 浙江工业大学 基于模糊卡尔曼滤波的交通流预测方法
CN102855773A (zh) * 2012-09-13 2013-01-02 东南大学 一种停车场有效泊位占有率短时预测方法
CN103870890B (zh) * 2014-03-19 2016-08-24 四川大学 高速公路网交通流量分布的预测方法
CN108111353B (zh) * 2017-12-26 2021-10-15 深圳广联赛讯股份有限公司 预付卡剩余流量预测方法、网络终端和存储介质
CN110021161B (zh) * 2018-01-09 2021-12-21 株式会社日立制作所 一种交通流向的预测方法及***
CN109061544B (zh) * 2018-08-23 2020-11-06 广东工业大学 一种电能计量误差估计方法
CN109377754B (zh) * 2018-10-29 2021-07-02 东南大学 一种车联网环境下的短时交通流速度预测方法
CN110853375B (zh) * 2019-11-21 2020-12-01 东南大学 考虑重叠路径的随机用户均衡逐日动态交通流预测方法
CN111369794B (zh) * 2020-02-28 2022-01-25 腾讯科技(深圳)有限公司 交通参与信息的确定方法、装置、设备及存储介质
CN111862592B (zh) * 2020-05-27 2021-12-17 浙江工业大学 一种基于rgcn的交通流预测方法
CN114239948B (zh) * 2021-12-10 2023-07-21 浙江省交通投资集团有限公司智慧交通研究分公司 基于时序分解单元的深度交通流预测方法、介质及其设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004178518A (ja) * 2002-11-29 2004-06-24 Nippon Telegr & Teleph Corp <Ntt> 旅行時間予測方法,旅行時間予測装置,旅行時間予測プログラムおよびそのプログラムを記録したコンピュータ読み取り可能な記録媒体
US20050091176A1 (en) * 2003-10-24 2005-04-28 Mitsubishi Denki Kabushiki Kaisha Forecasting apparatus
CN1304987C (zh) * 2004-03-09 2007-03-14 北京交通大学 一种智能交通数据处理方法
CN101188002A (zh) * 2007-12-24 2008-05-28 北京大学 一种具有实时和连续特性的城市交通状态预测***及方法
CN100456335C (zh) * 2006-10-12 2009-01-28 华南理工大学 基于交通流相特征的城市交通***状态可视化评价方法及其应用
CN101739819A (zh) * 2009-11-19 2010-06-16 北京世纪高通科技有限公司 预测交通流的方法和装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004178518A (ja) * 2002-11-29 2004-06-24 Nippon Telegr & Teleph Corp <Ntt> 旅行時間予測方法,旅行時間予測装置,旅行時間予測プログラムおよびそのプログラムを記録したコンピュータ読み取り可能な記録媒体
US20050091176A1 (en) * 2003-10-24 2005-04-28 Mitsubishi Denki Kabushiki Kaisha Forecasting apparatus
CN1304987C (zh) * 2004-03-09 2007-03-14 北京交通大学 一种智能交通数据处理方法
CN100456335C (zh) * 2006-10-12 2009-01-28 华南理工大学 基于交通流相特征的城市交通***状态可视化评价方法及其应用
CN101188002A (zh) * 2007-12-24 2008-05-28 北京大学 一种具有实时和连续特性的城市交通状态预测***及方法
CN101739819A (zh) * 2009-11-19 2010-06-16 北京世纪高通科技有限公司 预测交通流的方法和装置

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106571882A (zh) * 2016-09-06 2017-04-19 北京航空航天大学 一种超外差接收机性能退化评估与预测方法及***
EP3358542A1 (en) * 2017-02-01 2018-08-08 Kapsch TrafficCom AG A method of predicting a traffic behaviour in a road system
WO2018141582A1 (en) * 2017-02-01 2018-08-09 Kapsch Trafficcom Ag A method of predicting a traffic behaviour in a road system
US11670163B2 (en) 2017-02-01 2023-06-06 Kapsch Trafficcom Ag Method of predicting a traffic behaviour in a road system
CN113496314A (zh) * 2021-09-07 2021-10-12 南京感动科技有限公司 一种神经网络模型预测道路交通流量的方法

Also Published As

Publication number Publication date
CN101739819A (zh) 2010-06-16

Similar Documents

Publication Publication Date Title
WO2011060730A1 (zh) 预测交通流的方法和装置
CN109492808B (zh) 一种室内停车场剩余车位预测方法
CN104392390B (zh) 一种基于topsis模型的智能变电站二次设备评估方法
CN102469103B (zh) 基于bp神经网络的木马事件预测方法
Reggiani et al. Neural networks and logit models applied to commuters’ mobility in the metropolitan area of Milan
CN104536412A (zh) 基于指标预报和解相似度分析的光刻工序动态调度方法
Claveria et al. Data pre-processing for neural network-based forecasting: does it really matter?
CN110633859B (zh) 一种两阶段分解集成的水文序列预测方法
CN110858973A (zh) 小区网络流量预测方法及装置
CN108090788A (zh) 基于时间信息集成模型的广告转化率预估方法
CN106940407A (zh) 一种配电网***电能质量扰动定位与识别方法
WO2017071369A1 (zh) 一种预测用户离网的方法和设备
US20220004959A1 (en) Method and system for prefiltering in ride-hailing platforms
CN113988441A (zh) 电力无线网络链路质量预测、模型训练方法及装置
CN102034111A (zh) 一种对飞行器结构损伤状态进行多元化识别与检测的方法
CN110852522A (zh) 一种短期电力负荷预测方法及***
CN108537322A (zh) 神经网络层间激活值量化方法及装置
CN110740063B (zh) 基于信号分解和周期特性的网络流量特征指标预测方法
CN107622301A (zh) 一种停车场空余停车泊位数的预测方法
CN110222840B (zh) 一种基于注意力机制的集群资源预测方法和装置
CN103607219A (zh) 一种电力线通信***的噪声预测方法
CN116192665A (zh) 数据处理方法、装置、计算机设备及存储介质
CN109711555A (zh) 一种预测深度学习模型单轮迭代时间的方法和***
WO2019119319A1 (zh) 停车场泊位占用率预测方法、装置、设备及存储介质
CN114239948A (zh) 基于时序分解单元的深度交通流预测方法、介质及其设备

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: 10831145

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: 10831145

Country of ref document: EP

Kind code of ref document: A1