CN111524348A - A long-term and short-term traffic flow prediction model and method - Google Patents

Abstract

The present application belongs to the field of traffic technology, and in particular relates to a long-term and short-term traffic flow prediction model and method. Predicting future traffic flow with external environmental influence data containing random errors will only make this random error magnified again, which will inevitably have a greater impact and fluctuation on the accuracy of model prediction. The present application provides a long-term and short-term traffic flow prediction model, which includes a context factor input layer, a feature learning and pattern recognition layer, and a traffic flow data output layer. Provides a multi-scale prediction model with high accuracy, good prediction effect, short training time, strong robustness, and is not affected by the lack of historical data. Using contextual factors as prediction input improves the accuracy of traffic flow prediction models and plays a vital role in advanced traffic management and traveler route planning.

Description

A long-term and short-term traffic flow prediction model and method

technical field

The present application belongs to the field of traffic technology, and in particular relates to a long-term and short-term traffic flow prediction model and method.

Background technique

Traffic flow refers to the traffic flow formed by the continuous driving of cars on the road. In a broad sense, it also includes the traffic and people flow of other vehicles. For a certain period of time, on the road section not affected by the lateral intersection, the traffic flow is in a continuous flow state; when it encounters traffic light control, it is in an intermittent flow state. Traffic flow prediction plays an important role in predicting road congestion and accidents, as well as traffic signal control. Especially for fixed-time signal control strategies, traffic flow prediction is crucial. The traffic flow prediction mode, suitable historical data, etc. have a great influence on the prediction accuracy.

There are various internal and external influencing factors affecting traffic flow. Some traditional traffic flow prediction only considers the historical traffic flow data itself, and does not consider the internal and external influencing factors of traffic flow data. Some combine traffic flow data with external factors, but it is difficult to obtain data due to external factors, and in the model prediction stage, the data of these external factors are unknown like the traffic flow data to be predicted. Only forecast data for external factors are used. Using the external environmental impact data with random errors obtained from these forecasts to predict future traffic flow will only make this random error magnified again, which will inevitably have a greater impact on the accuracy of the model prediction. fluctuation.

SUMMARY OF THE INVENTION

1. Technical problems to be solved

There are various internal and external factors that affect traffic flow. Some traditional traffic flow prediction only considers the historical traffic flow data itself, and does not consider the internal and external factors of traffic flow data. Some combine traffic flow data with external factors, but it is difficult to obtain data due to external factors, and in the model prediction stage, the data of these external factors are unknown like the traffic flow data to be predicted. Only forecast data for external factors are used. Using the external environmental impact data with random errors obtained from these forecasts to predict future traffic flow will only make this random error magnified again, which will inevitably have a greater impact on the accuracy of the model prediction. To solve the problem of fluctuation, this application provides a long-term and short-term traffic flow prediction model and method based on contextual factors and historical traffic flow data.

2. Technical solutions

In order to achieve the above purpose, the present application provides a long-term and short-term traffic flow prediction model, the model sequentially includes a context factor input layer, a feature learning and pattern recognition layer, and a traffic flow data output layer;

The context factor input layer is used to input the context factor into the neural network after preprocessing;

The learning and pattern recognition layer is used to transform the input context factors layer by layer to extract hidden patterns and features;

The traffic flow data output layer is used for aggregating and summarizing the patterns and features learned by the previous hidden layer, and performing nonlinear weighted transformation to obtain corresponding traffic flow prediction data.

Another embodiment provided by the present application is: the context factor input layer uses future context factors as input.

Another embodiment provided by the present application is that the traffic flow prediction output by the traffic flow data output layer includes long-term traffic flow prediction and short-term traffic flow prediction.

Another embodiment provided by the present application is: the traffic flow prediction duration output by the traffic flow data output layer includes 5 minutes, 10 minutes, 15 minutes, 20 minutes, 30 minutes or 1 hour.

The present application also provides a long-term and short-term traffic flow prediction method, the method includes the following steps:

Step 1): Input the historical traffic flow data and contextual factors into the feature learning and pattern recognition layer for training;

Step 2): Utilize the neural network back-propagation algorithm to continuously carry out the updating iteration and testing of the model;

Step 3): generate a deep belief network model that can express the characteristics of traffic flow contextual factors;

Step 4): Load the trained deep belief network model;

Step 5): Feed the prepared future contextual factors into the prediction model in sequence;

Step 6): output the predicted result through the traffic flow data output layer.

Another embodiment provided by the present application is: the context factors include year, month, day, week, holiday, holiday, hour, minute and daily data time point.

Another embodiment provided by the present application is that the prediction method mines the relationship between traffic flow data and contextual factors within a given time interval from historical data.

Another embodiment provided by the present application is that the relationship between the traffic flow data and contextual factors is mined by using a deep belief network model.

Another embodiment provided by the present application is: the relationship between the traffic flow data and contextual factors is mined by using a multi-layer supervised learning algorithm.

3. Beneficial effects

Compared with the prior art, the beneficial effects of a long-term and short-term traffic flow prediction model and method provided by the present application are:

The long-term and short-term traffic flow prediction model provided by the present application improves the accuracy of the traffic flow prediction model. Provides a multi-scale prediction model with high accuracy, good prediction effect, short training time, strong robustness, and is not affected by the lack of historical data. It plays a vital role in advanced traffic management and traveler route planning.

The long-term and short-term traffic flow prediction model provided by this application studies the training methods of different unit models, optimizes the model structure and parameters, and reduces the training time.

The long-term and short-term traffic flow prediction method provided by the present application is a long-term and short-term traffic prediction algorithm based on contextual factors and historical traffic flow data.

The long-term and short-term traffic flow prediction method provided by the present application proposes a long-term and short-term traffic flow prediction algorithm based on a deep belief network, aiming at the problem that the existing traffic prediction flow has low accuracy in terms of prediction accuracy.

The long-term and short-term traffic flow prediction method provided in this application studies the relationship between the traffic flow and the combination of contextual factors in a given time interval.

The long-term and short-term traffic flow prediction method provided in this application can resist the influence of the problem of missing historical data on the prediction accuracy.

The long-term and short-term traffic flow prediction method provided in this application studies the structure and optimization method of the factor feature extraction network.

The long-term and short-term traffic flow prediction method provided in this application studies the realization method of the long-term and short-term multi-time-scale traffic flow prediction model.

In the long-term and short-term traffic flow prediction method provided by this application, the model algorithm extracts the contextual factors that affect the traffic flow operation mode by analyzing the traffic flow data, establishes a traffic flow prediction model of contextual factors, and mines the historical data. The complex relationship between contextual factors and traffic flow patterns. After the model is trained, future contextual factors are used to input the model to achieve accurate estimation and prediction of future traffic flow without relying on any external data, which resists the impact of missing historical data on the prediction accuracy. Finally, at a given time interval (eg, 5 minutes, 10 minutes, 15 minutes, or 1 hour), the collected traffic flow data is fed into the model to achieve long-term and short-term neural network traffic flow prediction.

Description of drawings

1 is a schematic diagram of a long-term and short-term traffic flow prediction model of the present application;

In the figure: 1-context factor input layer, 2-feature learning and pattern recognition layer, 3-traffic flow data output layer.

Detailed ways

Hereinafter, specific embodiments of the present application will be described in detail with reference to the accompanying drawings, from which those skilled in the art can clearly understand the present application and be able to implement the present application. Without departing from the principles of the present application, the features of the various embodiments may be combined to obtain new embodiments, or instead of certain features of certain embodiments, to obtain other preferred embodiments.

Referring to FIG. 1 , the present application provides a long-term and short-term traffic flow prediction model. The model includes a context factor input layer 1 , a feature learning and pattern recognition layer 2 and a traffic flow data output layer 3 .

The whole model is divided into three parts, the leftmost is the context factor input layer, which is responsible for inputting the contextual factors into the neural network after preprocessing; the middle is the feature transformation and pattern recognition layer, which is the core part of the entire model, also known as Hidden layer, where the input contextual factors are transformed layer by layer to extract hidden patterns and features; the last part is the predictor, also known as the output layer, which is a simple artificial neural network layer learned from the previous hidden layer. The corresponding traffic flow prediction data is obtained after non-linear weighted transformation, which is the output part of the entire model.

Further, the context factor input layer adopts a future context factor input layer. Accurate estimation and prediction of future traffic flow can be achieved without relying on any external data, and the impact of missing historical data on prediction accuracy can be resisted.

Further, the traffic flow prediction output by the traffic flow data output layer includes long-term traffic flow prediction and short-term traffic flow prediction.

Further, the traffic flow prediction duration output by the traffic flow data output layer includes 5 minutes, 10 minutes, 15 minutes, 20 minutes, 30 minutes and 1 hour. Multi-scale forecasting was used. Different scales of forecasting models can be selected for long-term and short-term traffic flow forecasting according to forecasting requirements.

The present application also provides a long-term and short-term traffic flow prediction method, the method includes the following steps:

(1) Model training stage

Step 1): Input historical traffic flow data and contextual factors into feature learning and pattern recognition layer 2 for training;

Step 2): Utilize the neural network back-propagation algorithm to continuously carry out the updating iteration and testing of the model;

Step 3): generate a deep belief network model that can express the characteristics of traffic flow contextual factors;

(2) Model prediction stage

Step 4): Load the trained deep belief network model;

Step 5): Feed the prepared future contextual factors into the prediction model in sequence;

Step 6): output the predicted result through the traffic flow data output layer.

Further, the contextual factors include year, month, day, week, holiday, holiday, hour, minute and daily data time point.

Further, the prediction method mines the relationship between traffic flow data and contextual factors within a given time interval from historical data.

Further, the relationship between the traffic flow data and contextual factors is mined by using a deep belief network model.

Further, the relationship between the traffic flow data and contextual factors is mined by using a multi-layer supervised learning algorithm.

1. High accuracy

The long-term and short-term traffic flow prediction algorithm based on deep belief network considers the influence of contextual factors on the traffic flow value, fully researches the traffic flow data, and uses the deep belief network model to excavate the potential relationship between the traffic flow value and the contextual factors. . The method proposed in this application outperforms traditional traffic flow prediction methods in terms of prediction accuracy.

2. Not affected by missing values in historical data

The daily and long-term traffic flow prediction algorithm based on deep belief network uses the future contextual factors to input the model to achieve accurate estimation and prediction of future traffic flow without relying on any external data.

3. Multi-scale prediction

The long-term and short-term traffic flow prediction algorithm based on deep belief network adopts multi-scale model prediction. Unlike previous prediction models, only one time interval can be used for training and prediction. After the model is trained, the time interval can be selected according to your own needs. , multi-scale prediction of future traffic flow.

This application mainly considers the influence of contextual factors on traffic flow values, and uses a multi-layer supervised learning algorithm to mine the relationship between traffic flow values and contextual factors within a given time interval. The feature of this application is that it is the first time to use the future contextual factor input model to achieve accurate estimation and prediction of future traffic flow without relying on any external data. Combined with the unique advantages of deep belief networks in nonlinear transformation, The features of contextual factors are abstracted at a higher level, and the prediction output of future traffic flow is realized.

On the basis of analyzing the long-term trends, periodic characteristics and external factors of traffic flow data, this application focuses on the research on the correlation between traffic flow time series data and its inherent data attributes, and uses time series analysis methods and nonlinear features to fit models. , establish a multi-time-scale long-term and short-term traffic flow prediction model, and mine the potential relationship between traffic flow and contextual factors from historical traffic data, so as to make accurate and reliable predictions of future traffic operation status and trends.

Use historical traffic flow data and contextual factors to conduct multi-data source, multi-time scale test evaluations to verify the correctness, robustness and adaptability of the proposed model for long-term and short-term traffic flow prediction; simultaneously using multi-scale The multi-fold cross-validation method is compared with the prediction results of common and typical traffic flow prediction methods to verify the accuracy, universality and advancement of the proposed prediction model.

This application proposes a long-term and short-term traffic prediction method based on contextual factors and historical traffic flow data using deep belief networks. The main idea is that by mining the correspondence between traffic flow data and contextual factors, it is superior to traditional traffic forecasting in terms of forecasting accuracy and improves the accuracy of traffic flow forecasting. The overall framework of the traffic flow contextual feature prediction model considering contextual factors proposed in this application is shown in Figure 1.

Deep neural networks are often recommended for mining latent relationships in massive multidimensional historical datasets. Deep Belief Network, a type of Deep Neural Network (DNN), is a deep learning architecture that allows computational models to be composed of multiple processing layers in order to learn data representations with multiple levels of abstraction. In order to mine the complex nonlinear relationship between contextual factors and traffic flow patterns, this application uses the classical deep belief network as a tool to build a model.

或者Z在

上的一个分布。这里，Z是自变量，代表从交通流数据中提取出来的语境因素的集合；

是因变量，是通过函数映射变换后得到的未来的交通流参数，也是模型中的想要输出的预测值。假设z为某一语境因素向量，z∈Z，神经元的网络函数f(z)被定义为一些变换函数g_i(z)的组合，这些组合中的函数还可以进一步被分解为更多的函数，以更加方便地表示复杂的网络结构。公式中的箭头描述函数之间的依赖关系，可以方便的使用组合函数来表示，在神经网络里面广泛使用的组合函数就是非线性加权和如公式1所示:A deep belief network is actually an extension of a single-layer artificial neural network in the number of network layers. The so-called "deep" refers to an artificial neural network with more than one layer. In order to realize the extraction and transformation of complex features, the deep belief network cascades multiple single-layer neural network layers with nonlinear processing functions. The nodes of each layer are trained on a different set of patterns based on the output of the previous layer, each subsequent neural network layer is using the output of the network transformation of the previous layer as input, and the context factor is carried out at each layer. with different nonlinear transformations and reorganizations, and pass them back layer by layer. As the deep belief network continues to advance to a deeper level, the more complex patterns and features can be identified, and the deeper abstract relationships that are hidden can be unearthed. Like a single-layer neural network, a deep belief network is a fully connected feedforward network, and its model can be viewed as a simple mathematical function mapping f:

or Z in

a distribution on . Here, Z is the independent variable, representing the set of contextual factors extracted from the traffic flow data;

is the dependent variable, the future traffic flow parameter obtained after transformation by function mapping, and the predicted value that the model wants to output. Assuming that z is a vector of contextual factors, z∈Z, the network function f(z) of neurons is defined as a combination of some transformation functions _gi (z), and the functions in these combinations can be further decomposed into more function to more conveniently represent complex network structures. The arrows in the formula describe the dependencies between functions, which can be easily represented by the combination function. The combination function widely used in neural networks is the nonlinear weighted sum, as shown in formula 1:

In the formula: wj is the weight; gj points to an element in a function vector set g = (g1, g2, ..., gn); bj represents the bias; σ represents a predetermined transformation function, also known as the activation function. An important feature of an activation function is that it provides a smooth transition when the input value changes. The activation function in this paper is tanh, which can generate a nonlinear value and compress between -1 and 1.

The structural model of the deep belief network prediction model is divided into three parts. The leftmost is the contextual factor input layer, which is responsible for preprocessing the contextual factors into the neural network; the middle is the feature transformation and pattern recognition layer, which is the most important part of the entire model. The core part, also known as the hidden layer, where the input contextual factors are transformed layer by layer to extract hidden patterns and features; the last part is the predictor, also known as the output layer, which is a simple artificial neural network layer, The patterns and features learned by the previous hidden layer are aggregated and summarized, and the corresponding traffic flow prediction data is obtained after nonlinear weighted transformation, which is the output part of the entire model.

如同前面的映射关系，整个深度置信网络模型可以表示成如公式2所示：Each input node of the deep belief network prediction model corresponds to a context factor zi (z _{i ∈} Z) in the context factor vector, and the output node of the predictor corresponds to the traffic flow data

Like the previous mapping relationship, the entire deep belief network model can be expressed as Equation 2:

where z is the context factor vector; W is the weight parameter; b is the bias parameter. In this function, the deep belief network model is regarded as a whole that completes the nonlinear transformation, and W and b are regarded as the weight parameters and bias parameters of the entire model, which greatly simplifies the description and understanding of the model function. Since the deep belief network itself is composed of multiple artificial neural network layers stacked, if W and b are given different meanings and ranges, this function can also be used to represent an artificial neural network layer, or even a neuron in the network . After the context factor vector is fed into the network, the internal states of neurons and layers are changed according to their weight parameters and biases, and feature outputs are generated through activation functions. Deep belief networks do this by connecting the outputs of these neurons to the inputs of other neurons, passing from front to back layer by layer, forming a directed weighted graph. These weight parameters in the neural network will not be determined when the model is established. Due to their complex relationship and large number, these weight parameters need to be continuously updated and revised through learning. This learning process is also called model training. Since the weight parameters in the deep belief network model cannot be determined one by one, the best way is to select a random value first, and then iteratively optimize according to the error between the output result of the model and the expected value until the model outputs The error between the result and the expected value reaches a minimum or no longer changes. When the training error reaches a preset minimum value (such as 10 ^-5 or less), the iterative process can be declared over, the parameters of the model are determined, and the training job is completed. Of course, there may also be cases where the training error stops decreasing and never reaches the preset value. In this case, the cause should be found from both the model structure and the data set, and the model should be corrected or the data set should be retrained. From this point of view, when the structure of a deep belief network is determined, the learning or training of the deep belief network is actually to continuously reduce the error of the model through an iterative algorithm, optimize and determine the value of its weight parameters.

(1) The loss function of the model

In the case of supervised learning, after the context data is input into the deep belief network model, after a series of abstraction, transformation and extraction processes on the hidden layer, a prediction value is finally output from the output layer. This context factor exists depending on the traffic flow data, and there is an inevitable error between the predicted value output by the model and the real traffic flow data. Representing this error as a function is the loss function used in model training.

The loss function is used to indicate the degree of inconsistency between the predicted value of the model and the actual value, which is a reflection of the degree of fit between the prediction model and the real data. , so that it can effectively play a role in the optimization process. The loss function should also be differentiable everywhere in the value range, and there should be a reasonable gradient. When the loss value is large, the gradient is large, and when the loss value is small, the gradient becomes significantly smaller, and when the value is 0, the gradient is also 0 at the same time. . In the application of nonlinear regression class, the most commonly used loss function is the mean square error loss function:

为预测值。有了损失函数，就可以方便的计算出每一个语境因素预测模型上的总体损失，对于模型的学习也就转化成了一个最优化问题，优化的目标就是损失函数最小化。In the formula: y is the true value;

is the predicted value. With the loss function, the overall loss of each context factor prediction model can be easily calculated, and the learning of the model is transformed into an optimization problem. The goal of optimization is to minimize the loss function.

(2) Back propagation of response error and parameter update

The loss is caused by the propagation of the error, and the overall loss of the model can also be called the response error, or simply the error. The error generated by the deep neural belief network model is propagated to the output layer from front to back, and the overall loss of the model propagation can be calculated according to the loss function, but these losses are the errors of the parameters adjusted by multiple neurons after multiple combinations and transformations It is not known how much error is generated by which parameter is passed.

Back Propagation (BP) is a common method used in conjunction with gradient descent to train artificial neural networks. In each iteration process of the network model, the method calculates the loss gradients of all weight parameters sequentially from back to front by performing forward calculation and back propagation on the response error of the model, and uses the calculated gradients to update the corresponding Weight parameters, so as to realize the parameter adjustment and optimization of the model.

The error backpropagation algorithm generally consists of three stages. The first stage is the incentive dissemination stage. This stage is mainly the forward propagation of the excitation response of the neural network model. According to the structure and initial parameters of the model, the weighted sum and the excitation response are calculated layer by layer, and the response error of the model is calculated according to the loss function. The second stage is the gradient calculation stage. By taking the derivation of the loss function, the descending gradient of each parameter is calculated in reverse. The third stage is the weight update stage, and the corresponding model parameters are updated according to the calculated gradient values.

In the gradient calculation stage, the error is generated due to the change of the parameters, so the model parameters are regarded as the independent variables of the loss function to calculate the gradient. Let o _i be the excitation response of the i-th neuron in a certain layer of the network, according to formula 4, we can see

为激活函数；net_i为神经元i的加权和；b_i为神经元i的偏置参数；w_ji为与前层神经元j相连的神经元i的权重参数。公式中i为本层神经元的索引，j为前层神经元的索引，因此o_j表示前层神经元j的激励响应，在整个模型的第一层o_j＝z_j，在整个模型的最后一层

根据链式求导法则(Chain Rule)逐层向前求取导数，权重参数w_ji的求导公式如下：where:

is the activation function; net _i is the weighted sum of neuron i; b _i is the bias parameter of neuron i; w _ji is the weight parameter of neuron i connected to neuron j in the previous layer. In the formula, i is the index of the neuron in the layer, and j is the index of the neuron in the front layer, so o _j represents the excitation response of the neuron in the front layer. In the first layer of the whole model, o _j =z _j , in the whole model last layer

According to the chain derivation rule (Chain Rule), the derivative is obtained layer by layer forward, and the derivation formula of the weight parameter w _ji is as follows:

In the formula: E is the response error of neuron i; the definitions of other symbols are the same as the above formula. in,

为激活函数的导数。对于输出层来说，响应误差就是模型的总体损失，可以使用损失函数直接计算，整个输出层响应误差对于响应激励的导数可以使用下式直接计算：where:

is the derivative of the activation function. For the output layer, the response error is the overall loss of the model, which can be calculated directly using the loss function. The derivative of the response error of the entire output layer to the response excitation can be directly calculated using the following formula:

However, the neurons in the hidden layer cannot directly obtain their response error, which needs to be calculated by passing the response error of the output layer from back to front. Let L be the set of all neurons (subscripts) that accept input from neuron i, then the response error of the current neuron can be back-propagated from these connected neurons, and the value is equal to the sum of these errors. The response o _j takes the total differential, and the derivative of the response error of the neuron i to the excitation response can be written as:

Comparing the two ends of the formula, it can be seen that there is an obvious recursive relationship here. If the derivative of the response excitation of all neurons in the next layer can be obtained, then the derivative of the response excitation of the neurons in this hidden layer can be calculated and substituted into the above The equation can obtain the weight parameter W and the corresponding gradient formula is as follows:

in:

Now, the gradient of each neuron weight parameter can be calculated, and the gradient descent method can be used to update the numerical value of the weight parameter. The update formula of the weight parameter W is as follows:

w _ij (t+1)=w _ij (t)+Δw _ij +ξ(t)

In the formula: t is the number of iterations; η is the learning rate, usually a value less than 1; ξ(t) is a random variable.

There is another parameter that needs to be learned in the model - the bias parameter b. Like the weight parameter W, it also uses the chain derivation rule to obtain the derivative forward layer by layer. The derivation formula of the bias parameter b _ji is as follows:

where:

According to the previous derivation process and formula 10, the total differential is taken for the excitation response o _j , and the derivative of the neuron i response error to the excitation response can be written as:

Substituting into the above equation can obtain the corresponding gradient formula of the model bias b as follows:

in:

Now that the gradient of each neuron's bias parameter has been calculated, the gradient descent method can be used to update the value of the bias parameter. The update formula is as follows:

b _ij (t+1)=b _ij (t)+Δb _ij +ξ(t)

Where: t is the number of iterations; η is the learning rate; ξ(t) is a random variable.

Considering that different activation functions may be selected in practical application, the derivative of activation function φ in the above formula is represented by φ’, which should be analyzed according to specific functions in practical application. For example, the derivative of the hyperbolic tangent function used in the model is tanh(x)'=1-tanh(x)2.

The gradient descent method is an optimization algorithm used to find the optimal model parameters. By calculating the response error of the training set on the model, the model error continues to decrease along the direction of the smallest gradient, until the error no longer decreases or the model is found to be in The parameter with the smallest error on the training set. Mini-Batch GradientDescent is a gradient descent algorithm between online learning and offline learning. Combining the advantages of the batch gradient descent algorithm and the stochastic gradient descent algorithm, the training data set is divided into smaller batches, and the offline learning method is used on each batch, which looks like online learning from a macro perspective.

Using the mini-batch gradient descent method, the model update frequency is high, the model update speed is accelerated, the noise interference of a single sample is reduced, the variance of the gradient is reduced, the model is prevented from falling into a local minimum, and the convergence of the model is stronger. The batch update mini-batch gradient descent algorithm finds a balance between the robustness of stochastic gradient descent and the efficiency of batch gradient descent, providing a more efficient computing process and improving the efficiency of storage space utilization. The most commonly used gradient descent method.

Using the mini-batch stochastic gradient descent algorithm introduces a hyperparameter batch_size during model training, which can be assigned empirically or adjusted during the actual training process. Smaller values give an online learning-like process that converges quickly at the expense of introducing noise during training; larger values give an offline learning-like process that can accurately estimate the error gradient but converges slow. , then formula (3) can be written as the following matrix expression:

In the formula: n is the number of samples for each batch of training. In this paper, 256 is selected as the batch size parameter based on experience; m is the dimension of contextual factors; o is the dimension of predicted value. Although the use of a slightly larger batch size brings higher memory usage, since the data is sent to the model in batches for training, the parallel processing capability of the computer is effectively utilized, and the parallelism of the calculation and the throughput of the machine are improved. It reduces the training time while maintaining good generalization performance.

Although the present application has been described above with reference to specific embodiments, it will be understood by those skilled in the art that many modifications may be made in configuration and detail disclosed herein within the spirit and scope of the present disclosure. The scope of protection of the present application is to be determined by the appended claims, and the claims are intended to cover all modifications encompassed by the literal meaning or scope of equivalents to the technical features in the claims.

Claims (9)

1. A long-short term traffic flow prediction model is characterized in that: the model sequentially comprises a context factor input layer, a feature learning and pattern recognition layer and a traffic flow data output layer;

the context factor input layer is used for inputting the preprocessed context factors into the neural network;

the characteristic learning and pattern recognition layer is used for transforming the input context factors layer by layer and extracting implicit patterns and characteristics;

and the traffic flow data output layer is used for aggregating and summarizing the modes and characteristics obtained by learning of the previous hidden layer, and obtaining corresponding traffic flow prediction data after nonlinear weighted transformation.

2. The long and short term traffic flow prediction model according to claim 1, characterized in that: the contextual factor input layer takes future contextual factors as input.

3. The long and short term traffic flow prediction model according to claim 1, characterized in that: the traffic flow prediction output by the traffic flow data output layer comprises long-time traffic flow prediction and short-time traffic flow prediction.

4. The long-and-short-term traffic flow prediction method according to claim 3, characterized in that: the predicted time of the traffic flow output by the traffic flow data output layer comprises 5 minutes, 10 minutes, 15 minutes, 20 minutes, 30 minutes or 1 hour.

5. A long-term and short-term traffic flow prediction method is characterized in that: the method comprises the following steps:

step 1): inputting historical traffic flow data and context factors into a feature learning and pattern recognition layer for training;

step 2): updating iteration and testing of the model are continuously carried out by utilizing a neural network back propagation algorithm;

step 3): generating a deep belief network model capable of expressing the characteristics of the traffic flow context factors;

step 4): loading a trained deep belief network model;

step 5): sending the prepared context factors into a prediction model in sequence;

step 6): and outputting the predicted result through the traffic flow data output layer.

6. The long-and-short-term traffic flow prediction method according to claim 5, characterized in that: the contextual factors include year, month, day, week, holiday, hour, minute, and daily data time points.

7. The long-and-short-term traffic flow prediction method according to claim 5, characterized in that: the prediction method exploits from historical data the relationship between traffic flow data and contextual factors over a given time interval.

8. The long-and-short-term traffic flow prediction method according to claim 7, characterized in that: and mining the relation between the traffic flow data and the contextual factors by adopting a deep belief network model.

9. The long-and-short-term traffic flow prediction method according to claim 7, characterized in that: and mining the relation between the traffic flow data and the contextual factors by adopting a multi-layer supervised learning algorithm.

Images