Passenger flow forecast of tourist attraction based on MACBL in LBS big data environment

3.1 Overall framework of the MACBL model

This article designs a tourist flow forecast algorithm model based on the fusion of the MHA mechanism, CNN, and BiLSTM, abbreviated as the MACBL model. Figure 1 shows the overall framework of the MACBL model.

Figure 1

The overall framework of the MACBL model.

The MACBL model integrates multiple attention mechanisms to highlight important features and discard irrelevant data. CNN excels in extracting local trends and dimensionality reduction of high-dimensional data, while BiLSTM excels in capturing feature information from bidirectional historical passenger flow sequence data. The MACBL is a deep network composed of multi-layer neural networks, which enables this model to effectively mine deep-level information in high-frequency passenger flow data [23,24]. Moreover, the MACBL model is a many-to-one model that can simultaneously mine multiple feature sequence information to output the predicted feature sequence.

The MACBL model consists of four layers of neural networks. The first three layers are a composite network layer and the last one is a fully connected neural network layer. The composite network layer is composed of the MHA mechanism, CNN, and BiLSTM, while the fully connected neural network layer converts the output of the composite network layer into the final predicted passenger flow sequence. Furthermore, a random deactivation layer (dropout) is also set after each layer of the composite network to avoid overfitting. Dropout reduces the neuron number involved in the calculation at each layer during model training by randomly deactivating a certain percentage of neurons. Thus, dropout simplifies the network structure and prevents the model from overfitting [25].

3.2 CNN network structure

CNN is a multi-layer neural network with supervised learning [26]. The convolution and pool sampling layers of the hidden layer are used to implement the feature extraction function of CNN. The neural network model employs gradient descent to minimize the loss function and modify the weights of the network layer by layer, while repeated iteration training increases the accuracy of the network. The lower hidden layer of CNN consists of alternating convolution and maximum pool sampling layers. The upper layer consists of the traditional multilayer perceptron hidden layer and logical regression classifier.

In Figure 2, the input data are x 1 = { x 1 , 1 , x 1 , 2 , ⋯ , x 1 , T } , x = { x 1 , x 2 , ⋯ , x N } ,

Figure 2

The overall framework of CNN.

where N denotes the number of samples, and x 1 and T denote the data at time T in sample 1. The data window slides and changes in the input data, and each sliding change filter will perform a convolution operation on a local data window. Convolution is the multiplication of local features by their respective weights and then adds them together. Then, further maximum pooling is performed on the output convolutional features to preserve the maximum features. Pooling is the process of capturing features from the lower layer to reduce the size of the network and obtain the constant features of the input data. The proposed framework conducts two convolution and pooling operations, and ultimately outputs F 1 . The calculation formulas for convolution and pooling operations are as follows:

where x is the input data; f _conv is the convolution function; σ represents the activation function; W u and b u represent the weight coefficients and bias function, respectively; and u represents the number of convolution kernels. The discrete convolution operation is ⊗ . The convolution kernels’ size parameters are α and β. w m , n denotes the weight coefficient of the convolution kernel, while m and n denote the row and column indexes of the input data, respectively. The final output feature sequence F 1 of the CNN is shown as follows:

where f 1 , T represents the characterization sequence data of the first sample at time T , F 1 represents the characterization sequence data of the first sample, F 1 represents the CNN characterization output sequence of all samples, and N represents the number of samples.

3.3 BiLSTM network architecture

The BiLSTM is composed of two unidirectional LSTMs stacked in forward and reverse directions [27]. Figure 3 shows the overall framework structure of BiLSTM. Compared to unidirectional LSTM, the BiLSTM can capture data feature information from both the positive and negative directions of the present node, fully considering the impact of past and future feature information on the current node.

Figure 3

The overall framework of BiLSTM.

The obtained feature sequence F = { F 1 , F 2 , … , F N } is used as the input of the BiLSTM model. At time t of sample i , the hidden layer contains information from both the front and back directions. The h i , t l forward LSTM contains the hidden state of historical tourist flow, while the backward LSTM network reads the same historical tourist flow in the reverse direction and outputs it in the given unknown tourist flow h i , t r . Finally, these two vectors are connected to form the final output h i , t . The final feature output of BiLSTM is represented by H i , and are as shown in equations (6) and (7):

3.4 Network architecture of MHA

In the last decade, the attention mechanism has been proposed for handling time series [28]. The MHA mechanism can fuse different types of information, thereby better grasping complex information [29]. The MHA mechanism can not only better understand the meaning of information, but also assist the model in improving learning efficiency, saving both time and resources.

Figure 4 shows the MHA model framework for passenger flow prediction. A query is entered to retrieve the data from the Data Management System. If the query matches the specified key, the associated value is retrieved. Equation (8) illustrates the process of constructing a query Q i , a key K i , and a value V i that are used to predict the tourist flow:

where H i denotes the hidden output of BiLSTM for a data sample i , H i = ( h i , 1 , … , h i , t , … , h i , T ) , and W V ∈ ℝ r × T , W Q ∈ ℝ r × T , and W K ∈ ℝ r × T denote the trainable weight matrices. To obtain the most important information in the V value, the attention vector a is determined and then multiplied by the value matrix to obtain O i = a V i , as shown in equation (9):

Figure 4

The overall framework of MHA.

The attention vector of MHA can be expressed as a h = Soft max ( W Q ( h ) H i ⋅ ( W K ( h ) H ) ′ / T ) , where W K ( h ) , a = S o f t max ( Q i ⋅ K i ′ / T ) , and W V ( h ) represent the trainable weight matrices of the h th attention model and a h represents the attention vector of the h th attention model. The output of the h th attention model is expressed as O i , h = a h ( W V ( h ) H i ) .

Next, all outputs of the attention model are cascaded. Among them, H is the number of attention models that have been layered in the MHA model, as shown in equation (10):

3.5 Loss function

L denotes the loss function that depicts the gap between the predicted and the true values during the training phase of the model. The loss function is also known as the cost function. Through loss function iteration, the classification effect of the model is optimized. The greater the value of L, the lower the classification effect. The optimal value of the loss function for the model is zero. During the training process, certain parameter values are also obtained for the model. The squared loss function is used as the objective function to train each domain, as shown in equation (11):

where y i , j and y ˆ i , j represent the real and predicted passenger flow data of sample i at moment j, respectively; N represents the sample domain; λ represents the regularization parameters that are added to prevent overfitting problems; and Θ represents the model parameters.

3.6 Technical route

The technical route of the proposed framework mainly includes three parts: passenger flow data preprocessing, model construction and prediction, and prediction effect evaluation training and performance testing experiments, as shown in Figure 5.

Figure 5

Technical route for predicting tourist flow in scenic spots.

4.1 Experimental environment

In this experiment, the Windows 10 system was used as the platform. The main configuration information is shown in Table 1.

4.2 Evaluation indexes

To perform an objective analysis of the experimental results, this article adopts MAE, RMSE, and MAPE as experimental performance indexes. The computation formulas for the evaluation indicators are displayed in equations (12)–(14):

Among them, y i and y ˆ i represent the actual and predicted tourist flow data, respectively, while n is the total number of samples. The effectiveness of a prediction increases with decreasing RMSE, MAE, and MAPE values.

4.3 Datasets

The experimental dataset included historical passenger flow data, weather data, and vacation data of scenic spots. The historical passenger flow data of the scenic area were the LBS data of tourists in the Shenzhen Dameisha Coastal Park scenic area, collected from the regional thermal map data on Tencent’s location big data platform. These data recorded real-time longitude and latitude coordinate information of the locations of all mobile terminal devices in a certain area. For the experiments in this study, the data were converted into temporal passenger flow data of the Dameisha Coastal Park scenic area for prediction. The weather data were collected on Worldweatheronline. To enable deep learning models to recognize weather data, research has used dummy variable assignment to assign values to certain types of data such as weather types (nominal data). Since the focus of this study was a short-term high-frequency prediction, there were no extreme weather events in Shenzhen during the time covered by the data. Therefore, the weather types were only divided into two categories for assignment. Among them, the relatively mild weather types such as sunny days and cloudy days were assigned as 0, and the relatively extreme weather types such as rainstorms and moderate rain were assigned as 1. In terms of processing holiday data, the dummy variable assignment method was also used for assignment, where the workday period was assigned 0 and the holiday period was assigned 1. The holiday data included weekends and holidays. Finally, the assigned vacation days and weather variables were combined with the passenger flow sequence to form input features in the form of a sequence composed of 0 and 1, which are imported into the model in the form of a numerical matrix. During the data preparation process, the data were split into training, validation, and testing sets. The data used in this study spanned from 0:00 on May 8, 2020, to 23:00 on May 9, 2020, with a total duration of 792 h and 2,563 samples. The first 720 h of data were divided into training sets for model training and data mining. The data from the last 72 h were selected as a test set to test the prediction effect.

Due to the different data dimensions of passenger flow, weather, and vacation data in the input data, it is important to normalize the data before importing it into the deep learning model. The data were normalized using the Min-max normalization method, which maps three sets of data features uniformly in the interval range of [0, 1] to eliminate data dimensionality and avoid adverse effects on the computational accuracy and convergence speed of the model due to the large differences in dimensionality between data [30].

4.4 Prediction effect evaluation training

To avoid the impact of random factors such as the initialization of random parameters, the MACBL-based passenger flow forecasting method for tourist attractions was trained many times before the hyperparameter was determined. During the training process, the model checkpoint and early halting method in the callback function were used to save and terminate the model to avoid overfitting. Before the model terminates training, its training and validation losses are shown in Figure 6.

Figure 6

Variations in training and validation losses.

It can be seen from Figure 6 that the training and validation losses fluctuated significantly in the first 2,100 iterations but then stabilized, demonstrating good robustness of the model.

The deep learning model needs to be repeatedly debugged to achieve the optimal prediction effect. Hyperparameter debugging is a process of combinatorial optimization. Table 2 displays the final model parameter settings determined after a series of parameter debugging. The hyperparameters of the deep learning model are divided into network parameters, optimization parameters, and regularization and training parameters. The network parameters represent various parameters related to the structure of the deep neural networks. The MACBL network model consists of four layers of networks, including three composite network layers and one fully connected network layer. The network selects the ReLU function as the activation function, which helps to decrease the probability of gradient disappearance. The sparse network generated by this function can alleviate the overfitting phenomenon that may occur in the model.

4.5 Model performance verification experiment

4.5.1 Comparison of model fitting curves at different depths

To verify the effectiveness of the deep neural network model, network models of different depths were set up for curve fitting evaluation and comparison. The proposed MACBL model is a neural network model with three composite network layers. Therefore, experimental designs were made for single-, two-, three-, and four-layer composite networks with the same neural structure as the MACBL model. The selected observation time was 3 days and the time frequency was 1 h. Figure 7 shows that the first actual peak of passenger flow reaches 3,015 persons, while the predicted values are 2,405, 2,547, 2,743, and 2,745 persons when the numbers of floors are 1, 2, 3, and 4, respectively. Thus, when the depth of the network layer gradually grows under the same neural algorithm, the trend of the passenger flow prediction curve becomes increasingly consistent with the actual passenger flow curve, indicating that the deep network model has a satisfactory effect in predicting the passenger flow in scenic areas. Among them, the prediction curves of the three-layer MACBL and the neural network model of the four-layer composite network layers with the same neural structure have the highest overlap with the real passenger flow curve, reflecting the most excellent prediction performance of these two models. Hence, this study selects a three-layer MACBL model as the passenger flow prediction model.

Figure 7

Effect chart comparison of passenger flow prediction by MACBL model with different number of layers: (a) 1 layer, (b) 2 layers, (c) 3 layers, and (d) 4 layers.

4.5.2 Comparison with state-of-the-art models

To demonstrate the superiority of the proposed model, it was compared with the state-of-the-art models including SARIMA + LSTM [18], GCN-RNN [16], and IA-GRU [21].

The comparison result data are shown in Table 3 and Figure 8. It can be seen from Table 3 that compared to the state-of-the-art models, the MACBL model has improved the RMSE index by at least 2.049, 2.926, and 1.338 for prediction steps of 24, 32, and 60 h, respectively. The MAE index has also been improved by at least 1.352, 1.489, and 0.938 for the same prediction steps. The MAPE index has also been improved by at least 0.0447, 0.0345, and 0.0379% for the same prediction steps. The experimental results demonstrate that the MACBL model outperforms SARIMA + LSTM, GCN-RNN, and IA-GRU models in predicting future scenic area passenger flow. Moreover, the prediction effect of the proposed MACBL model on the real scenic area passenger flow dataset is significantly improved, and the prediction error increases slowly with the prediction length. In contrast to the comparative models, the MACBL model is more comprehensive in obtaining the feature information of passenger flow data, achieving local feature extraction, reducing the dimensionality of high-dimensional data, capturing long-term dependencies of two-way time-series data, and deepening the extraction of important feature information. In this study, 24 h is used as the prediction step.

Table 3

Model effects at different time granularities

Model	Time/h	RMSE	MAE	MAPE/%
SARIMA + LSTM	24	18.326	16.146	0.1942
GCN-RNN		17.354	12.354	0.1796
IA-GRU		15.886	10.280	0.1468
MACBL (ours)		13.837	8.928	0.1021
SARIMA + LSTM	32	22.726	19.136	0.2043
GCN-RNN		19.454	14.334	0.1836
IA-GRU		17.615	11.086	0.1508
MACBL (ours)		14.689	9.597	0.1163
SARIMA + LSTM	60	25.536	20.146	0.2167
GCN-RNN		21.354	16.354	0.1896
IA-GRU		18.147	12.275	0.1697
MACBL (ours)		16.809	11.337	0.1318

Figure 8

Comparison of passenger flow prediction by different models at different prediction steps.