CN110599236A - Short-time parking demand prediction method based on GRU model - Google Patents

Abstract

The invention discloses a short-term parking demand prediction method based on a GRU model, comprising the following steps: 1) obtaining historical data of parking lot facilities, processing the historical data, and obtaining berth occupancy data at each time point. 2) Use the deep learning Keras framework package, set the GRU neural network structure, and use the GridSearch function in the Keras package to obtain the optimal parameters of the model. 3) Use the training set data to train the GRU model, save the model and predict the berth occupancy rate of the next step. Compared with the prior art, the present invention uses big data processing technology and the latest algorithm of deep learning under the background of obtaining continuous parking data, and proposes a more advanced and accurate parking information induction release method, which can not only consider the parking demand The correlation in the time dimension, and its cell module has a simpler control gate structure, which can greatly improve the training efficiency, thereby improving the utilization rate of parking facilities, improving the satisfaction of users with parking needs, and avoiding users unnecessary traffic, reduce the traffic pressure on the road, and also help the traffic management department to carry out effective traffic control during the peak period of traffic flow.

Description

A Short-term Parking Demand Prediction Method Based on GRU Model

technical field

The invention relates to a parking demand forecasting method, in particular to a short-term parking demand forecasting method based on a GRU model.

Background technique

The rapid development of the automobile industry is in serious conflict with the limited urban space resources, the contradiction between supply and demand is becoming increasingly acute, and the problem of difficult parking has become a challenge for major cities. Parking Guidance System (PGS) is an effective way to alleviate traffic congestion, but the short-term accurate prediction of parking demand as a key technology for the release of vacant parking spaces has not been effectively resolved. During the use of the existing guidance information, there will be a problem that the guidance information display is inaccurate. When the driver sees the guidance information, it shows that there are still parking spaces, but when they arrive at the parking lot, it shows that there are no parking spaces, especially during peak hours. Parking spaces vary greatly. In order to avoid invalid detours for drivers, higher requirements are put forward for the accuracy of guidance information. How to use the parking demand short-term prediction technology to display more accurate vacant parking information on the guidance information screen. Traditional parking demand forecasting methods and technologies are mainly reflected in the planning level. As the key path to effectively improve parking efficiency, refined parking management also puts forward high standards for short-term parking demand forecasting technology.

Research on short-term forecasting of parking demand, the forecasting methods are mainly divided into the following categories: one is the traditional time series forecasting method, which mainly includes ARIMA (Autoregressive Integrated Moving Average Model), multivariate autoregressive forecasting model considering spatio-temporal correlation , Markov model, Kalman filter model. The second is the algorithm in machine learning. Most studies use more simple BP neural network. Because it is difficult to improve the prediction accuracy of berth occupancy by using a single prediction model, many scholars have proposed a method of combining prediction models to improve prediction accuracy, such as fuzzy neural network prediction method, combined with wavelet analysis and Markov chain effective berth occupancy rate is short Time prediction method, compound prediction method based on chaos and BP neural network, improved neural network method based on wavelet transform, etc. According to previous studies, the accuracy of these methods cannot meet the requirements of refined management of parking lots for parking demand prediction. Therefore, it is necessary to explore a machine learning algorithm with higher accuracy. The GRU model is the Gated Recurrent Unit. This model was proposed by Cho, van Merrienboer, Bahdanau and Bengio in 2014. The GRU model is a variant of the LSTM model. Compared with LSTM, there is one less gate, which maintains the effect of LSTM while It also makes the structure simpler.

The update gate (z_t) is used to control the degree to which the state information of the previous moment is brought into the current state. The larger the value of the update gate, the more state information of the previous moment is brought in.

The reset gate (r_t) is used to control how much information in the previous state is written into the current candidate set h ^~ _t. The smaller the value of the reset gate, the less information in the previous state is written.

The GRU forward propagation formula is as follows:

r _t ＝σ(W _r ·[h _t-1 ,x _t ])

z _t =σ(W _z ·[h _t-1 ,x _t ])

y _t = σ(W _o h _t )

Among them, [] indicates that two vectors are connected, and * indicates the product of matrices.

The GRU training formula is as follows:

The parameters to be learned include: W _r , W _z , W _o , the first three parameters are concatenated and need to be separated during learning:

W _r =W _rx +W _rh

W _z ＝ W _z x +W _zh

Input to the output layer:

The output of the output layer:

After the final output is obtained, the loss transmitted by the network can be written. The loss of a single sample at a certain moment is:

Then the loss of a single sample at all times is:

The backward error propagation algorithm is used to learn the network. The partial derivative formula of the loss function for each parameter is as follows. After obtaining the partial derivative of each parameter, the parameters can be continuously updated.

Intermediate parameters:

δ _t = δ _h,t z _t φ′

Contents of the invention

The object of the present invention is to provide a short-term parking demand prediction method based on the GRU model in order to overcome the above-mentioned defects in the prior art.

The purpose of the present invention can be achieved through the following technical solutions:

A partition-based fine-grained parking lot management method mainly includes the following steps:

1. a short-term parking demand prediction method based on GRU model, is characterized in that, comprises the following steps:

1) Obtain historical data of parking facilities, process the historical data, and obtain berth occupancy data at each time point.

2) Use the deep learning Keras framework package to set the GRU neural network structure, and use the GridSearch function in the Keras package to obtain the optimal parameters.

3) Use the training set data to train the GRU model, save the model and predict the berth occupancy rate of the next step.

Described step 1) is specifically:

S1.1, according to the original field information of historical data (the specific time point when a vehicle enters and leaves the parking facility), the number of vehicles entering and leaving the parking facility at each time point is obtained through the Resample function in the Panads package in Python , by adjusting the sampling frequency in the Resample function (60 minutes or 30 minutes or 15 minutes or 5 minutes or 1 minute, etc.), the number of vehicles entering and leaving on different time scales can be obtained.

S1.2, using the Concat function in the Panads package to merge the incoming and outgoing Series according to the index to obtain data in the form of DataFrame. Calculate the number of vehicles on the scene at the initial moment, then the number of vehicles on the scene at the next moment = the number of vehicles on the scene at the previous moment + the number of vehicles entering - the number of vehicles leaving. Create two new columns in the DataFrame, and through loop traversal, one column is the number of vehicles present on each time scale, and the other column is the berth occupancy rate on each time scale.

Described step 2) is specifically:

S2.1, in order to prevent the model from overfitting and underfitting, the obtained historical data set is divided into a training set and a verification set test set. The verification set can be used to adjust the hyperparameters of the model and the ability of the model Make an initial assessment.

S2.2, set the structure of the GRU neural network: the first layer is the GRU layer, the second layer is the Bidirectional-GRU, the third layer is the fully connected layer, and the fourth layer is the output layer.

S2.3, in order to evaluate the prediction accuracy of the model, Root Mean Square Error (RMSE) is used as the optimization objective function; the optimizer is selected as RMSprop; the activation function relu is used; when fitting the training set, shuffling is set to false.

S2.4, through GridSearch grid search, in all candidate parameter selections, through loop traversal, try every possibility, the parameter with the best performance is the final result. The parameters that need to be searched include Lookback, the number of neurons in each layer, and the learning rate.

Described step 3) specifically is:

S3.1, after training the model, use the model.save() method in the keras package to keep the structure of the model training locally, and the file type is .h5.

S3.2, in the actual parking guidance system, to obtain the predicted berth occupancy rate at the next moment, only need to use the model.load() method to read the trained model.

S3.3, when a new week or month of historical data is collected, the GRU model can be retrained to save the latest model training results.

Compared with the prior art, the present invention is based on the prior art, and from the perspective of refined parking management, it aims at the release of parking lot induction information A short-term parking demand prediction method based on the GRU model is proposed. The present invention uses big data and machine learning methods under the background of fully obtaining parking data to propose a more accurate information induction release method, improve the service level for parking users, improve user satisfaction, and avoid unnecessary traffic for users at the same time. Reduce traffic pressure on the road.

The present invention has the following advantages:

1) The GRU model has the ability to learn time series data in the time dimension, so that it can consider the correlation of input samples in the time dimension, thereby improving the prediction accuracy.

2) The GRU model can not only provide higher prediction accuracy, but also its cell module has a simpler control gate structure, which makes it easier to train and greatly improves the training efficiency.

Description of drawings

Fig. 1 is the logic diagram of the short-term parking demand prediction method based on the GRU model of the present invention;

Fig. 2 is GRU cell structure among the present invention;

Fig. 3 is a daily change chart of parking demand on weekdays in one embodiment of the present invention;

Fig. 4 is the iterative loss curve of training model in one embodiment of the present invention;

Fig. 5 is a prediction effect diagram in an embodiment of the present invention.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is carried out on the premise of the technical solution of the present invention, and detailed implementation and specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

Example

As shown in Figure 1, a short-term parking demand forecasting method based on the GRU model includes the following process: First, obtain the historical data of the parking lot facility, process the historical data, and obtain the berth occupancy data at each time point; secondly, Using the deep learning Keras framework package, set the GRU neural network structure, and use the GridSearch function in the Keras package to obtain the optimal parameters; finally, use the training set data to train the GRU model, save the model and predict the berth occupancy rate of the next step. The GRU cell structure in the embodiment is shown in FIG. 2 . The input of the cell structure is the output value h _t-1 of the cell at the previous moment and the observed berth occupancy value x _t at this moment. The z _t and r _t in the figure represent the update gate and the reset gate respectively. The update gate uses It controls the extent to which the state information at the previous moment is brought into the current state. The larger the value of the update gate, the more state information at the previous moment is brought in. The reset gate controls how much information from the previous state is written to the current candidate set On the other hand, the smaller the reset gate is, the less information from the previous state is written.

The concrete method of this embodiment is as follows:

(1) Data preprocessing

Specifically include the following steps:

Step 1: According to the original field information of historical data, count the specific time points when each vehicle enters and leaves the parking facility, and obtain the number of vehicles entering and leaving the parking facility at each time point through the Resample function in the Panads package in Python. Set the sampling frequency in the Resample function to 10 minutes, and count the number of vehicles entering and leaving on a time scale of 10 minutes. The present invention selects a commercial parking facility as an example, the number of parking spaces is 420, and the date of the acquired parking data is 2018.11.1-2018.11.30. The information contained in the specific parking data is shown in Table 1.

Table 1 Parking lot data field information

Step 2: Use the Concat function in the Panads package to merge the incoming and outgoing Series according to the index to obtain data in the form of DataFrame. Calculate the number of vehicles on the scene at the initial moment, then the number of vehicles on the scene at the next moment = the number of vehicles on the scene at the previous moment + the number of vehicles entering - the number of vehicles leaving. Create two new columns in the DataFrame, through loop traversal, one column is the number of vehicles present on each time scale, and the other column is the number of berth occupancy on each time scale.

Step 3: The data selection time period of the number of vehicles in the middle field of the embodiment is 6:00-22:00, a total of 97 data points in one day, and the data of working days and non-working days are reserved. The data format is as shown in Table 2. Working The daily variation of parking demand is shown in Figure 3.

Table 2 Berth occupancy table

(2) GRU neural network construction

Specifically include the following steps:

Step 1: In order to prevent the model from overfitting and underfitting, the obtained historical data set is divided into a training set and a verification set test set. The verification set can be used to adjust the hyperparameters of the model and to test the ability of the model. Preliminary assessment. Take the historical data from November 1, 2018 to November 25, 2018 as the training set, take the data from November 25, 2018 to November 27, 2018 as the verification set, and use the historical data from November 28, 2018 to November 2018 30th is the test set.

Step 2: In order to improve the prediction accuracy, standardize the original berth occupancy data, use the StandardScaler() function in the Sklearn package to standardize the data to the interval from -1 to 1, and use the inverse_transform method to restore the data.

Step 3: Set the structure of the GRU neural network: the first layer is the GRU layer, the second layer is Bidirectional-GRU, the third layer is the fully connected layer, and the fourth layer is the output layer.

Step 4: In order to evaluate the prediction accuracy of the model, Root Mean Square Error (RMSE) is used as the optimization objective function; the optimizer is selected as RMSprop; the activation function Relu is used; when fitting the training set, Shuffling is set to false.

Step 5: Through GridSearch grid search, among all candidate parameter selections, through loop traversal, try every possibility, the parameter with the best performance is the final result, the parameters that need to be searched include Lookback, each layer of neurons number, learning rate, etc.

Step 6: Finally, the optimal parameters of the model are obtained: Lookback=40, the number of GRU cells in the first layer is 60, the number of Bidirectional-GRU cells in the second layer is 60, the number of cells in the third layer of fully connected layer is 60, and the number of cells in the fourth layer is input The number of cells in the layer is 1, the learning rate is 0.01 and iterated 20 times, and the batchsize is 10. The iterative loss curve of the training model in the embodiment is shown in FIG. 4 .

(3) Model preservation and prediction

Step 1: After training the model, use the model.save() method in the keras package to keep the structure of the model training locally, and the file is named ziyangli.h5. In the actual parking guidance system, you only need to use the model.load() method to read the trained model, and then perform real-time prediction.

Step 2: When a new week or month of historical data is collected, the GRU model can be retrained. Follow the above steps to save the final training results of the model for prediction.

(6) Forecast results

Taking the prediction result of 2018.11.30 as an example, the comparison between the prediction result and the real value can be seen from Table 3.

Table 3 2018.11.30 Comparison of predicted value and real value

serial number actual value Predictive value serial number actual value Predictive value serial number actual value Predictive value 0 272 270 33 281 280 66 260 263 1 270 272 34 283 280 67 257 258 2 269 269 35 287 283 68 256 255 3 262 267 36 292 289 69 248 255 4 262 259 37 292 294 70 242 246 5 262 261 38 295 292 71 239 240 6 256 262 39 296 295 72 239 239 7 249 254 40 296 296 73 232 240 8 247 246 41 299 295 74 230 231 9 248 246 42 300 300 75 232 230 10 247 249 43 303 301 76 229 234 11 245 247 44 300 304 77 228 230 12 247 244 45 303 299 78 222 229 13 246 248 46 297 303 79 223 222 14 245 246 47 297 295 80 219 225 15 243 244 48 294 295 81 217 220 16 245 242 49 288 293 82 207 218 17 259 246 50 288 285 83 207 204 18 262 265 51 285 286 84 205 212 19 266 265 52 286 284 85 204 205 20 261 265 53 288 285 86 198 208 twenty one 261 259 54 285 289 87 196 195 twenty two 259 260 55 285 284 88 192 201 twenty three 256 258 56 284 283 89 188 189 twenty four 260 254 57 283 283 90 186 189 25 263 261 58 280 282 91 182 186 26 263 265 59 273 278 92 179 180 27 272 2623 60 268 269 93 175 179 28 274 277 61 267 265 94 171 173 29 275 276 62 264 266 95 165 170 30 277 273 63 265 263 96 162 161 31 279 277 64 264 264 97 160 161 32 280 280 65 264 263 102 154 152

The prediction results of 2018.11.28-2018.11.30 in the embodiment are shown in Figure 5. It can be seen that in the short-term prediction of parking demand, GRU neural network can provide higher prediction accuracy, its MAE is only 3.40, and MAPE is only 1.32%.

Claims (4)

1. A short-time parking demand prediction method based on a GRU model is characterized by comprising the following steps:

1) and obtaining historical data of the parking facility, and processing the historical data to obtain parking space occupancy data at each time point.

2) And setting a GRU neural network structure by using a deep learning Keras framework package, and obtaining optimal parameters by using a GridSearch function in the Keras package.

3) And training the GRU model by using the training set data, storing the GRU model and predicting the occupancy of the berthage of the next step length.

2. The method for predicting the demand for short-term parking based on the GRU model as claimed in claim 1, wherein the step 1) is specifically as follows:

s1.1, obtaining the number of vehicles entering and leaving the parking facility at each time point through a repeat function in a Panads package in Python according to original field information (specific time points of a vehicle entering and leaving the parking facility) of historical data, and obtaining the number of vehicles entering and leaving the parking facility at different time scales by adjusting sampling frequency (60 minutes, 30 minutes, 15 minutes, 5 minutes, 1 minute and the like) in the repeat function.

S1.2, merging the entering and leaving Series according to the index by using the Concat function in the Panads package to obtain data in a DataFrame form. And calculating the number of the vehicles present at the initial moment, wherein the number of the vehicles present at the later moment is equal to the number of the vehicles present at the last moment, the number of the vehicles entering the train and the number of the vehicles leaving the train. And newly building two columns in the DataFrame, wherein one column is the number of the vehicles in the field on each time scale through cyclic traversal, and the other column is the occupancy rate of the berth on each time scale.

S1.3, in order to improve the prediction accuracy, the original berthage-occupying capacity data is standardized, and the data is standardized to a range from-1 to 1 by using a StandardScaler () function in a Sklearn packet. And finally restoring the data by using an inverse _ transform method.

3. The method for predicting the demand for short-term parking based on the GRU model as claimed in claim 1, wherein the step 2) is specifically as follows:

s2.1, in order to prevent the model from generating overfitting and under-fitting, dividing the acquired historical data set into a training set, a verification set and a test set, wherein the verification set can be used for adjusting the hyper-parameters of the model and primarily evaluating the capability of the model, and the test set finally judges the prediction accuracy of the model.

S2.2, setting the structure of the GRU neural network: the first layer is a GRU layer, the second layer is a Bidirectional-GRU layer, the third layer is a full connection layer, and the fourth layer is an output layer.

S2.3, in order to evaluate the prediction accuracy of the model, using Root Mean Square Error (RMSE) as an optimization objective function; the optimizer selects RMSprop; adopting an activation function Relu; when fitting the training set, Shuffling is set to false.

S2.4, through GridSearch grid search, in all candidate parameter selections, each possibility is tried through loop traversal, the parameter which shows the best performance is the final result, and the well-trained GRU model is obtained.

The hyper-parameters needing to be searched mainly include a Lookback, the number of neurons in each layer, a learning rate and the like.

4. The method for predicting the demand for short-term parking based on the GRU model as claimed in claim 1, wherein the step 3) is specifically as follows:

s3.1, after the GRU model is trained, a model save () method in a keras package is used for keeping the model training structure in the local, and the file type is h5 (an efficient file storage format).

And S3.2, in an actual parking guidance system, when the predicted parking space occupancy at the next moment is to be obtained, only a model () method is adopted to read the trained model, and then prediction is carried out.

And S3.3, when new historical data of a week or a month is collected, the data can be imported again, the GRU model can be retrained again, and the latest model training result is stored.