Travel planning method and system based on deep learning

Abstract

The invention provides a travel planning method and system based on deep learning, and the method comprises the steps: constructing a map and scenic spot layout model of a park based on a CNN (convolutional neural network), constructing a visitor flow prediction model based on a LSTM (long short term memory) network, calculating the waiting time according to a prediction result, helping tourists to better arrange time, avoiding long-time waiting, and improving the travel planning efficiency. According to the method, tourist demands are analyzed, keywords are extracted and matched with related information in a database, an optimal action is selected according to the current states of tourists in combination with a constructed route planning model based on reinforcement learning, and an optimal tour route is planned. The personalized travel plan is generated, the expectations and requirements of different tourists are met, the formulated route plan is more scientific, accurate and reasonable, and the travel experience of the tourists in an amusement park is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of travel planning, and in particular to a travel planning method and system based on deep learning. Background Art

[0002] As people's living standards improve, travel has become a popular leisure activity, and amusement parks are particularly popular among people of all ages. In recent years, the number of visitors to amusement parks has shown a rapid annual growth trend. During peak tourist season, popular amusement parks can see tens of thousands of visitors per day. This dense flow of people inevitably increases wait times for attractions, significantly impacting the visitor experience.

[0003] Under current technological conditions, there is a lack of systems and methods that can develop optimal routes within amusement parks tailored to visitor needs. Visitors often blindly choose attractions upon entering the park. Faced with numerous attractions and winding roads, they rely on experience or randomness to choose routes. This results in visitors wasting considerable time searching for attractions, especially during high-volume events, leading to long waits in line and unable to fully enjoy the park. For example, many visitors wait in line for more than two hours for popular attractions, while missing out on less popular attractions that suit their interests due to lack of location. Furthermore, different visitors have different preferences, with some preferring thrilling roller coasters and others gravitating towards interactive parent-child activities. Existing simple maps and signage simply cannot meet these diverse needs.

[0004] Therefore, it is urgent to design a system and method that can accurately analyze the needs of tourists with different interests and preferences, and then plan the optimal travel route within the park, which will greatly enhance the tourists' experience in the amusement park. Summary of the Invention

[0005] To solve the above problems, the present invention provides a travel planning method and system based on deep learning, which can accurately analyze the needs of tourists with different interests and preferences, and then plan the optimal travel route within the park, which will greatly enhance the tourists' experience in the amusement park.

[0006] To achieve the above object, the technical solution adopted by the present invention is: The present invention provides a travel planning method based on deep learning, which specifically includes the following steps: Step S1: Collect relevant data of the amusement park to form a database, and pre-process the collected data; Step S2: Build a park map and scenic spot layout model based on the convolutional neural network (CNN), build a crowd flow prediction model based on the long short-term memory (LSTM) network, and calculate the waiting time based on the crowd flow prediction results; Step S3: Obtain the demand information input by the visitor on the user interface, use the keyword search algorithm to analyze the text input by the visitor, extract keywords, match the keywords with the information in the database, and identify key information; Step S4: Construct a path planning model based on reinforcement learning, input the identified key information into the trained path planning model, and use the path planning model to combine the calculated waiting time for each item, select the optimal action based on the tourist's current state, and plan the optimal travel path; Step S5: Using a deep learning-based recommendation algorithm, based on tourist preferences and current park project information, recommend suitable attraction combinations and tour sequences for tourists; Step S6: Collect tourists' feedback and suggestions on the park and continuously optimize the park's service quality.

[0007] In step S1, the collected relevant data include entertainment project information, park and surrounding map topology, crowd flow information of each project, tourist preference information and historical play data.

[0008] Furthermore, the step S2 specifically includes the following steps: S21: Collect high-precision map image data of the park and accurately mark the locations of scenic spots in the map images to clarify their locations and categories; collect the flow of people in each project of the amusement park, as well as the relevant factors affecting the flow of people; S22: preprocess all collected data; S23: Build and train the park map and scenic spot layout model based on the convolutional neural network (CNN); S24: Build and train a crowd flow prediction model based on the long short-term memory network (LSTM); S25: Input the current park map image data to obtain the category and location information of the scenic spot; input the relevant data of the flow of people in the past period of time to obtain the predicted results of the flow of people for each project in the park; S26: Calculate the waiting time based on the simulated annealing algorithm according to the crowd flow prediction result.

[0009] Furthermore, step S24 specifically includes the following steps: S241: Initialize the model weights using normal distribution; S242: Inputting the collected historical pedestrian flow related data into the pedestrian flow prediction model and obtaining a prediction result through a forward propagation algorithm; S243: Compare the predicted result with the actual pedestrian flow data and calculate the loss value using the loss function; S244: Update the model weights through the backpropagation algorithm according to the gradient of the loss function; S245: Repeat the above steps S242, S243, and S244 until the model converges or reaches a preset number of training rounds.

[0010] Furthermore, the step S3 specifically includes the following steps: S31: Preprocessing the text data input by the visitor; S32: extract keywords from the text; S33: performing word segmentation processing on the text data stored in the database, and establishing an index structure to associate each keyword with the database record containing the keyword; S34: Match the extracted keywords with the index established in the database to find the database records corresponding to the same keywords; S35: Determine the type of key information based on the matched database record; S36: Integrate the matched relevant information to form a complete key information set.

[0011] In step S3, when matching the identified keywords, a fuzzy matching algorithm is used to match the keywords in the case of spelling errors or expression differences.

[0012] In step S4, the specific process of planning the optimal path includes the following steps: S41: State definition, clearly defining the state information that can represent the current situation of the tourist, and combining the state information into a vector to represent the tourist state; S42: Action definition, which clarifies the choices that tourists can make from the current state; S43: Environment modeling: Abstracting the amusement park into a graph structure, using nodes to represent locations and edges to represent paths between nodes. Edge weights are also set. The environment receives actions taken by visitors, updates its state according to rules, and returns corresponding rewards. S44: Formulate a reward function and give corresponding rewards; S45: Use the reinforcement learning algorithm DQN to build and train the path planning model; S46: Input the current state of the tourist, select the optimal action through the strategy network, and generate the optimal travel path.

[0013] In step S43, the rules include state transition rules and reward setting rules; State transfer rules include location update rules, time update rules and scenic spot state update rules; The reward setting rules include goal achievement reward rules, time reward rules, path length reward rules, attraction preference reward rules, crowd avoidance reward rules and penalty mechanism rules.

[0014] Furthermore, the step S5 specifically includes the following steps: S51: collecting tourist preference information and detailed information of each attraction in the amusement park, and pre-processing the collected data; S52: Build and train a recommendation model based on deep autoencoder; S53: Input the preference data of the new tourist into the trained recommendation model to obtain the low-dimensional feature representation of the tourist; S54: Calculating the similarity score between the feature representation of the visitor and the feature representations of each project in the park; S55: Based on the similarity scores, select attractions with high similarity to form a recommended attraction combination; at the same time, combine the attraction information to plan a reasonable tour sequence for tourists.

[0015] The present invention provides a travel planning system based on deep learning, comprising a data acquisition and processing module, a model building module, a demand acquisition and analysis module, a route planning module, a scenic spot recommendation module, and a feedback suggestion module, which are sequentially communicatively connected; Data collection and processing module: used to collect relevant data of the amusement park to form a database and pre-process the collected data; Model building module: Builds a park map and attraction layout model based on the convolutional neural network (CNN), builds a crowd flow prediction model based on the long short-term memory (LSTM) network, and calculates waiting time based on the crowd flow prediction results; Demand collection and analysis module: obtains the demand information entered by tourists on the user interface, uses keyword search algorithms to analyze the text entered by tourists, extracts keywords, matches the keywords with the information in the database, and identifies key information; Path planning module: Build a path planning model based on reinforcement learning. Input the identified key information into the trained path planning model. The path planning model selects the optimal action based on the current state of the tourist and plans the optimal travel path. Attraction recommendation module: Using a deep learning-based recommendation algorithm, it recommends suitable attraction combinations and visit sequences for tourists based on their preferences and current park project information. Feedback and suggestion module: used to collect visitors' feedback and suggestions on the park and continuously optimize the park's service quality.

[0016] The beneficial effects of the present invention are: the present invention provides a travel planning method and system based on deep learning, constructs a park map and scenic spot layout model based on the convolutional neural network CNN, constructs a crowd flow prediction model based on the long short-term memory network LSTM, and calculates the waiting time based on the prediction results to help tourists better arrange their time, avoid long waiting times, and improve travel efficiency. It analyzes tourist needs, extracts keywords, and matches them with relevant information in the database, constructs a path planning model based on reinforcement learning, selects the optimal action according to the current state of the tourist, and plans the optimal travel path. The present invention can generate personalized travel plans based on the unique preferences and needs of each tourist, meet the expectations and requirements of different tourists, and the formulated route planning is more scientific, accurate and reasonable, which greatly improves the tourists' travel experience in the amusement park. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a travel planning method based on deep learning in the present invention.

[0018] Figure 2 This is a block diagram of a travel planning system based on deep learning in the present invention. DETAILED DESCRIPTION

[0019] See also Figure 1 and Figure 2 As shown, the present invention provides a travel planning method and system based on deep learning.

[0020] Specifically, such as Figure 1 As shown, the present invention provides a travel planning method based on deep learning, comprising the following steps: Step S1: Collect relevant data of the amusement park to form a database, and pre-process the collected data; In step S1, the collected relevant data include entertainment project information (collecting information such as entertainment project name, type, opening hours, closing hours, content introduction, excitement level, suitable crowd, height and age limit, project cost, etc. from the official website of the amusement park, social media, and online travel platforms, wherein the excitement level can be measured by specific indicators or user evaluations), the topological structure of the park and its surrounding maps (using satellite maps, lidar scanning, field measurements, etc. to obtain the division of each area in the park, road distribution, location information of entertainment projects, restaurants, shops, rest areas, toilets and other facilities, as well as information on transportation hubs, hotels, restaurants, attractions, etc. around the park), each Project crowd flow information (by installing crowd flow sensors at the entrances of various projects in the park or using surveillance camera recognition technology to collect real-time information on the number of people currently queuing to visit the project, and at the same time record changes in project crowd flow over the past period of time, such as in different seasons, holidays, and time periods, to facilitate analysis of crowd flow patterns and prediction of future crowd flow), visitor preference information (collecting tourists' preferences for different types of entertainment projects, visit intervals, visit duration, and other consumption in the park, such as consumption in dining and shopping, etc.), historical play data (recording tourists' actual travel paths in the park, including the projects and order of play, visit duration, and tourist evaluations).

[0021] Preprocess the collected data, including data cleaning, data conversion, data encoding, data integration and association, and data labeling.

[0022] Data cleaning removes duplicates, errors, and missing values. Text data such as visitor reviews and project descriptions is segmented, breaking the text into individual words for subsequent semantic analysis and feature extraction. Parts of speech are labeled for each word, and the text data is converted into numerical vectors for processing by machine learning models. The data is then normalized and standardized. For categorical variables with a limited number of possible values, such as project types (roller coasters, carousels, etc.), one-hot encoding is used to convert them into binary vectors. Each category corresponds to a binary bit, and only the bit corresponding to that category is 1, while the rest are 0. Each category of the categorical variable is mapped to an integer. For example, the excitement level of the project is categorized as low, medium, and high, mapped to 0, 1, and 2, respectively.

[0023] Merge data from different sources and formats. For example, associate entertainment project information, visitor flow data, and visitor reviews based on project name to form a dataset containing multiple dimensions of information. Label the data. For example, precisely mark the location of each project on a map to facilitate the construction of a map topology. For image recognition-based visitor flow statistics, it is necessary to label the tourists in the image so that the training model can accurately identify the number of tourists.

[0024] Step S2: Build a park map and scenic spot layout model based on the convolutional neural network (CNN), build a crowd flow prediction model based on the long short-term memory (LSTM) network, and calculate the waiting time based on the crowd flow prediction results; The step S2 specifically includes the following steps: S21: Collect high-precision map image data of the park and accurately mark the locations of scenic spots in the map images to clarify their locations and categories; collect the flow of people in each project of the amusement park, as well as the relevant factors affecting the flow of people; Map image data can clearly display the overall layout of the park, including the locations of roads, buildings, and attractions, and accurately mark the locations of attractions in the map image to clarify the location and category. The marking method can be rectangular box marking (used to mark the boundaries of the attraction) or key point marking (used to mark the center position of the attraction). At the same time, each attraction is assigned a unique category label, such as "roller coaster", "carousel", etc.

[0025] Data on factors that affect foot traffic, such as time, weather, holidays, special events, etc. The time span of data collection should be long enough to cover a variety of different situations.

[0026] S22: preprocess all collected data; The collected map image data is preprocessed, including image scaling, data augmentation, normalization, and encoding labels. The category labels of the scenic spots are encoded using one-hot encoding, and each category is converted into a binary vector for easy processing.

[0027] The data related to pedestrian flow is preprocessed, including data cleaning, data conversion and data normalization. Through data conversion, the collected data related to pedestrian flow is converted into a format suitable for reception by the LSTM model and organized into a time series. Each time step contains the pedestrian flow of each project and related influencing factors.

[0028] S23: Build and train the park map and scenic spot layout model based on the convolutional neural network (CNN); The model structure is designed to be a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer uses convolution kernels to perform sliding convolution operations on the input image to extract image features. The convolutional layer can set different convolution kernel numbers, sizes, and step sizes to control the ability to extract features. The pooling layer is used to reduce the dimension of the feature map, reduce the number of model parameters, and enhance the robustness of the model. The fully connected layer integrates the features extracted by the convolutional layer and the pooling layer and outputs the final classification result. The number of neurons in the last layer of the fully connected layer should be consistent with the number of scenic spot categories. In the technical solution of the present invention, the output can be the category probability or location coordinates of each scenic spot.

[0029] After the model is built, it is iteratively trained using collected historical map image data. In each iteration, the model receives a batch of input data, calculates predictions, and then calculates a loss value based on the cross-entropy loss function. The model parameters are then updated based on this loss value using an optimizer, such as stochastic gradient descent (SGD), Adam, or RMSprop, to gradually reduce the loss value. During training, the appropriate number of training rounds and batch size must be set. After the model is trained, it is evaluated using appropriate metrics, such as accuracy, recall, F1 score, and mean squared error. The evaluation results are used to determine whether the model performance meets actual requirements.

[0030] S24: Build and train a crowd flow prediction model based on the long short-term memory network (LSTM); The designed model consists of two stacked LSTM layers, followed by a fully connected layer to output the predicted pedestrian flow. Each layer has 64 neurons, and the model's hyperparameters are set, such as the learning rate, batch size, and number of training rounds. The learning rate determines the step size of parameter updates during model training, the batch size affects the stability and efficiency of model training, and the number of training rounds determines the number of times the model is trained.

[0031] Wherein, step S24 specifically includes the following steps: S241: Initialize the model weights using normal distribution; S242: Inputting the collected historical pedestrian flow related data into the pedestrian flow prediction model and obtaining a prediction result through a forward propagation algorithm; At each time step, the LSTM unit updates the cell state and hidden state based on the current input and the state of the previous moment, and finally obtains the predicted traffic flow through the fully connected layer.

[0032] S243: Compare the predicted result with the actual pedestrian flow data and calculate the loss value using the loss function; The loss function is defined as the cross-entropy loss function. The pedestrian flow data is discretized and converted into a classification problem. For example, the pedestrian flow is divided into several different intervals, each interval corresponds to a category, and then the cross-entropy loss between the predicted category and the true category is calculated.

[0033] S244: Update the model weights through the backpropagation algorithm according to the gradient of the loss function; The backpropagation algorithm starts from the output layer, propagates the error back to the input layer layer by layer, calculates the gradient of each parameter, and updates the weights according to the gradient, so that the loss function gradually decreases.

[0034] S245: Repeat the above steps S242, S243, and S244 until the model converges or reaches a preset number of training rounds.

[0035] After the model is trained, it is evaluated using appropriate evaluation indicators, such as mean square error, mean absolute error, mean absolute percentage error (MAPE), etc., to assess the model's prediction accuracy. Based on the evaluation results, the model is further optimized and improved.

[0036] S25: Input the current park map image data to obtain the category and location information of the scenic spot; input the relevant data of the flow of people in the past period of time to obtain the predicted results of the flow of people for each project in the park; S26: Calculate the waiting time based on the simulated annealing algorithm according to the crowd flow prediction result.

[0037] The simulated annealing algorithm is a probabilistic global optimization algorithm that simulates the annealing process in physical systems to find the optimal solution. In the problem of calculating waiting time, different waiting time allocation schemes can be viewed as points in the solution space. By continuously adjusting the scheme and accepting less favorable solutions with a certain probability, we can avoid falling into local optimality and ultimately find a near-optimal waiting time calculation result.

[0038] First, an initial waiting time allocation scheme is randomly generated and the objective function value under this scheme is calculated (for example, the shortest total waiting time or the highest tourist satisfaction, etc.). Then, a new scheme is randomly generated in the neighborhood of the current scheme and its objective function value is calculated. According to the Metropolis criterion, the new scheme is accepted with a certain probability, and the probability formula is ,in is the difference between the objective function values ​​of the new solution and the current solution, and T is the simulated annealing temperature parameter. As the number of iterations increases, the temperature T gradually decreases. This process is repeated until the termination condition is met, such as reaching the maximum number of iterations or the change in the objective function value is less than a certain threshold. The solution obtained at this time is the approximately optimal waiting time calculation result.

[0039] Step S3: Obtain the demand information input by the visitor on the user interface, use the keyword search algorithm to analyze the text input by the visitor, extract keywords, match the keywords with the information in the database, and identify key information; The step S3 specifically includes the following steps: S31: Preprocessing the text data input by the visitor; Remove special characters such as punctuation, emoticons, and unnecessary whitespace from visitor input to standardize the text. Convert all characters in the text to lowercase to avoid keyword matching errors caused by capitalization errors. Use a word segmentation tool to break the text into individual words or phrases.

[0040] S32: extract keywords from the text; Count the frequency of each word in the text. Words with higher frequency are likely keywords. Part-of-speech tagging is performed on the segmented words, such as nouns (project name "Ferris Wheel" and location "Entrance to Amusement Park"), verbs ("play" and "see"), and adjectives. Nouns and verbs are more likely to be key information, so words with these parts of speech can be selected as candidate keywords. A stop word list is used to remove words that are not helpful in expressing key information, further refining the candidate keyword set.

[0041] S33: performing word segmentation processing on the text data stored in the database, and establishing an index structure to associate each keyword with the database record containing the keyword; An index structure is established for text data stored in the database, such as entertainment project information and park introductions, to associate each keyword with the database records containing the keyword. For example, the keyword "roller coaster" corresponds to all project records and visitor review records in the database that contain the word "roller coaster".

[0042] S34: Match the extracted keywords with the index established in the database to find the database records corresponding to the same keywords; For example, if a visitor enters "carousel," the index will find all records containing "carousel." For cases where there may be misspellings or discrepancies in the wording, a fuzzy matching algorithm (such as the edit distance algorithm, which measures similarity by calculating the minimum number of edit operations between two strings) is used for matching. For example, if a visitor enters "ferris wheel," the fuzzy matching algorithm will find a high similarity with "ferris wheel," thus finding database records related to "ferris wheel."

[0043] S35: Determine the type of key information based on the matched database record; For example, if data such as project name and opening hours are matched, it belongs to the category of entertainment project information; if data such as starting point and end point are matched, it belongs to the category of park location information.

[0044] S36: Integrate the matched relevant information to form a complete key information set.

[0045] For example, for the keyword "roller coaster," information such as opening hours, waiting times, and location is integrated for use in subsequent modules such as travel planning. Identified key information is verified to ensure accuracy and completeness. If any information is missing, additional information is sought from other relevant records.

[0046] When a keyword corresponds to multiple different entities or information, such as there may be multiple projects with the same name but different characteristics in an amusement park, it is necessary to further analyze the context or provide more prompt information to eliminate ambiguity and accurately identify the key information that tourists really need.

[0047] Step S4: Construct a path planning model based on reinforcement learning, input the identified key information into the trained path planning model, and use the path planning model to combine the calculated waiting time for each item, select the optimal action based on the tourist's current state, and plan the optimal travel path; S41: State definition, clearly defining the state information that can represent the current situation of the tourist, and combining the state information into a vector to represent the tourist state; In the amusement park scenario, the state information includes the visitor's current location, the items played, the remaining items to be played, the waiting time for each item, the current time, etc. This information is combined into a vector to represent the state, for example, [current location coordinates, list of items played, list of remaining items, array of waiting times for each item, current time].

[0048] S42: Action definition, which clarifies the choices that tourists can make from the current state; In amusement park path planning, an action is moving from the current location to the next feasible item. For example, if a visitor is currently at item A, and item A is adjacent to items B, C, and D, then moving to items B, C, and D are three different actions.

[0049] S43: Environment modeling: Abstracting the amusement park into a graph structure, using nodes to represent locations and edges to represent paths between nodes. Edge weights are also set. The environment receives actions taken by visitors, updates its state according to rules, and returns corresponding rewards. Nodes are used to represent locations, such as the locations of various projects, imports and exports, and other key locations. The weight of the edges takes into account factors such as distance and traffic flow. The environment receives the actions taken by tourists, updates the status according to the rules, and returns corresponding rewards.

[0050] Among them, the rules include state transfer rules and reward setting rules; State transition rules include location update rules (when a visitor chooses to go to a certain attraction or move along a certain road, the environment updates their location information on the park map based on the visitor's current location and the action they have chosen), time update rules (each visitor's action may consume a certain amount of time, and the environment needs to update the time status based on the action type and actual situation), and attraction status update rules (when a visitor arrives at an attraction, the environment will update its status based on the actual situation of the attraction). The reward setting rules include goal achievement reward rules (if tourists reach the target attraction or complete a specific game task, the environment will give a large positive reward to encourage tourists to move towards the goal), time reward rules (rewards can be given according to the remaining time to encourage tourists to reduce their total stay time in the park), path length reward rules (encourage tourists to choose shorter paths to avoid unnecessary detours and waste time), attraction preference reward rules (based on tourists' preference information, when tourists reach the attractions they are interested in, they will be given corresponding rewards), congestion avoidance reward rules (when tourists choose to go to attractions or paths with fewer people, the environment will give congestion avoidance rewards), and penalty mechanism rules (if tourists choose unreasonable actions, such as trying to enter unopened areas or choosing an inaccessible road, the environment will give penalties, that is, return a negative reward. In addition, if tourists stay in the park for too long and exceed the reasonable play time, they can also be given certain penalties).

[0051] S44: Formulate a reward function and give corresponding rewards; For example, considering the following three factors: Number of items played: Encourage the agent to play more items by giving a certain positive reward for each item successfully played. Total time spent: Minimize the total playing time. Negative rewards can be given based on the travel time on the path and the waiting time for items. Waiting time: Excessive waiting time will reduce the playing experience, so a larger negative reward is given for longer waiting times. The formulated reward function is: , where N is the number of recreational items, T toal is the total time consumption, T wait is the total waiting time, α, β, and γ are weight coefficients used to adjust the importance of each factor.

[0052] S45: Use the reinforcement learning algorithm DQN to build and train the path planning model; During training, the experience (state, action, reward, next state) generated by the agent's interaction with the environment is stored in an experience replay buffer. During training, a batch of experiences is randomly sampled from the buffer for learning, improving learning efficiency. Because the target network has the same structure as the main network, the target network's parameters are used when calculating the target Q-value, reducing fluctuations in the Q-value estimate. The agent selects an action based on the current state, interacts with the environment to obtain a reward and the next state, and stores the experience in the replay buffer. Experience is sampled from the buffer, and the error between the target Q-value and the current Q-value is calculated. The main network parameters are updated using gradient descent. Simultaneously, the target network parameters are regularly updated.

[0053] S46: Input the current state of the tourist, select the optimal action through the strategy network, and generate the optimal travel path.

[0054] In practical applications, the state space and action space can be very large, making it difficult to accurately calculate the Q-values ​​for all state-action pairs. Therefore, DQN uses a deep neural network (Q-network) to approximate the Q-value function. The input of this network is the state s, and the output is the Q-value of each possible action in that state. In each state s, the agent will choose the action a that maximizes the Q-value, i.e. , where a * is the optimal action.

[0055] Q value refers to the action value function. Q(s,a) represents the expected value of the cumulative discounted reward that the agent can obtain by taking action a in a specific state s and subsequently following the established strategy. The formula is: , Among them, r t is the immediate reward obtained at time t, γ is the discount factor, and its value range is [0,1]. It reflects the importance of future rewards relative to current rewards. The closer γ is to 1, the more the agent values ​​future rewards. The closer γ is to 0, the more the agent values ​​immediate rewards. represents the expectation operator, meaning that the value is averaged over all possible subsequent state and action sequences.

[0056] In DQN, agents make decisions using Q-values, which evaluate the pros and cons of actions in a given state. A larger Q-value indicates that taking an action in that state will yield more cumulative rewards, making it a better choice.

[0057] After the model is trained, the agent selects the optimal action through the policy network based on the tourist's current initial state, that is, the decision to move from the current location to the next item. The agent continuously selects actions based on the current state until all the items to be played are completed or the preset end conditions are reached, generating a complete play path.

[0058] Step S5: Using a deep learning-based recommendation algorithm, based on tourist preferences and current park project information, recommend suitable attraction combinations and tour sequences for tourists; The step S5 specifically includes the following steps: S51: collecting tourist preference information and detailed information of each attraction in the amusement park, and pre-processing the collected data; Through questionnaires, in-app preference settings, and historical behavior analysis, we collect information on visitor preferences for ride types (e.g., thrilling, family-friendly, leisure), theme preferences (e.g., fairy tale, science fiction), and time of day (e.g., morning, afternoon, evening). We also collect detailed information on each attraction within the park, including type, opening hours, duration, popularity, and location.

[0059] Preprocessing includes data cleaning, data encoding and normalization.

[0060] S52: Build and train a recommendation model based on deep autoencoder; Specifically, step S52 includes the following steps: S521: Initialize the weights and biases of the model; S522: Input the collected data into the model, obtain low-dimensional feature representation through the encoder, and then reconstruct it through the decoder; Design multiple fully connected layers to form the encoder and decoder, respectively. Use the input data to train the model. The encoder maps the input data to a low-dimensional feature space, learning the data's latent feature representation. The decoder reconstructs the encoder's output low-dimensional feature vector into a vector with the same dimension as the input data. ReLU can be used as the activation function in the hidden layers of the encoder and decoder, and the Softmax function can be used as the activation function in the output layer.

[0061] S523: Calculate the loss value between the reconstructed data and the original input data; S524: Use an optimization algorithm (such as stochastic gradient descent SGD, Adam, etc.) to update the model parameters according to the loss value, so that the loss value gradually decreases; S525: Repeat the above steps until the model converges or reaches the preset number of training rounds.

[0062] Use appropriate evaluation metrics to measure the performance of the recommendation system, such as accuracy, recall, F1 score, and Mean Average Precision (MAP). Optimize the model based on the evaluation results.

[0063] S3: Input the preference data of the new tourist into the trained recommendation model to obtain the low-dimensional feature representation of the tourist; S54: Calculating the similarity score between the feature representation of the visitor and the feature representations of each project in the park; For example, priority is given to recommending attractions whose opening hours match tourists' visiting time, whose visiting time is moderate, and whose locations are adjacent to each other.

[0064] S55: Based on the similarity scores, select attractions with high similarity to form a recommended attraction combination. At the same time, combine the attraction information to plan a reasonable tour sequence for tourists.

[0065] Step S6: Collect tourists' feedback and suggestions on the park and continuously optimize the park's service quality.

[0066] Among them, the feedback suggestions include: facility experience (asking tourists about their feelings about the safety measures of amusement facilities, their evaluation of the excitement levels of different amusement facilities, and their acceptance of the waiting time for popular facilities), park environment (the sanitation of public areas, their views on the park's green layout, their satisfaction with the facilities in the rest areas, and their feelings about the noise level and air quality in the park), service quality (collecting tourists' evaluations of the service attitudes of park staff and their evaluations of catering in the park), entertainment programs and activities (evaluations of various performances in the park, their participation and satisfaction with the theme activities held in the park, and their experience of the interactive projects and games in the park), and entry and travel (collecting tourists' evaluations of ticket prices, ticket purchase methods, entry and ticket checking procedures, and their views on the convenience of transportation around the park).

[0067] See also Figure 2 As shown, the present invention provides a travel planning system based on deep learning, including a data acquisition and processing module, a model building module, a demand acquisition and analysis module, a route planning module, a scenic spot recommendation module and a feedback suggestion module that are communicatively connected in sequence; Data collection and processing module: used to collect relevant data of the amusement park to form a database and pre-process the collected data; Model building module: Builds a park map and attraction layout model based on the convolutional neural network (CNN), builds a crowd flow prediction model based on the long short-term memory (LSTM) network, and calculates waiting time based on the crowd flow prediction results; Demand collection and analysis module: obtains the demand information entered by tourists on the user interface, uses keyword search algorithms to analyze the text entered by tourists, extracts keywords, matches the keywords with the information in the database, and identifies key information; Path planning module: Build a path planning model based on reinforcement learning. Input the identified key information into the trained path planning model. The path planning model selects the optimal action based on the current state of the tourist and plans the optimal travel path. Attraction recommendation module: Using a deep learning-based recommendation algorithm, it recommends suitable attraction combinations and visit sequences for tourists based on their preferences and current park project information. Feedback and suggestion module: used to collect visitors' feedback and suggestions on the park and continuously optimize the park's service quality.

[0068] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A travel planning method based on deep learning, characterized in that: The following steps are involved: Step S1: Collect relevant data of the amusement park to form a database, and pre-process the collected data; Step S2: Build a park map and scenic spot layout model based on the convolutional neural network (CNN), build a crowd flow prediction model based on the long short-term memory (LSTM) network, and calculate the waiting time based on the crowd flow prediction results; Step S3: Obtain the demand information input by the visitor on the user interface, use the keyword search algorithm to analyze the text input by the visitor, extract keywords, match the keywords with the information in the database, and identify key information; Step S4: Construct a path planning model based on reinforcement learning, input the identified key information into the trained path planning model, and use the path planning model to combine the calculated waiting time for each item, select the optimal action based on the tourist's current state, and plan the optimal travel path; Step S5: Using a deep learning-based recommendation algorithm, based on tourist preferences and current park project information, recommend suitable attraction combinations and tour sequences for tourists; Step S6: Collect tourists' feedback and suggestions on the park and continuously optimize the park's service quality.

2. A travel planning method based on deep learning according to claim 1, characterized in that: In step S1, the collected relevant data include entertainment project information, park and surrounding map topology, crowd flow information of each project, tourist preference information and historical play data.

3. A travel planning method based on deep learning according to claim 1, characterized in that: The step S2 specifically includes the following steps: S21: Collect high-precision map image data of the park and accurately mark the locations of scenic spots in the map images to clarify their locations and categories; collect the flow of people in each project of the amusement park, as well as the relevant factors affecting the flow of people; S22: preprocess all collected data; S23: Build and train the park map and scenic spot layout model based on the convolutional neural network (CNN); S24: Build and train a crowd flow prediction model based on the long short-term memory network (LSTM); S25: Input the current park map image data to obtain the category and location information of the scenic spot; input the relevant data of the flow of people in the past period of time to obtain the predicted results of the flow of people for each project in the park; S26: Calculate the waiting time based on the simulated annealing algorithm according to the crowd flow prediction result.

4. A travel planning method based on deep learning according to claim 3, characterized in that: Step S24 specifically includes the following steps: S241: Initialize the model weights using normal distribution; S242: Inputting the collected historical pedestrian flow related data into the pedestrian flow prediction model and obtaining a prediction result through a forward propagation algorithm; S243: Compare the predicted result with the actual pedestrian flow data and calculate the loss value using the loss function; S244: Update the model weights through the backpropagation algorithm according to the gradient of the loss function; S245: Repeat the above steps S242, S243, and S244 until the model converges or reaches a preset number of training rounds.

5. The deep learning-based travel planning method according to claim 1, characterized in that: The step S3 specifically includes the following steps: S31: Preprocessing the text data input by the visitor; S32: extract keywords from the text; S33: performing word segmentation processing on the text data stored in the database, and establishing an index structure to associate each keyword with the database record containing the keyword; S34: Match the extracted keywords with the index established in the database to find the database records corresponding to the same keywords; S35: Determine the type of key information based on the matched database record; S36: Integrate the matched relevant information to form a complete key information set.

6. A travel planning method based on deep learning according to claim 5, characterized in that: In step S3, when matching the identified keywords, a fuzzy matching algorithm is used to match the keywords in the case of spelling errors or expression differences.

7. A travel planning method based on deep learning according to claim 1, characterized in that: In step S4, the specific process of planning the optimal path includes the following steps: S41: State definition, clearly defining the state information that can represent the current situation of the tourist, and combining the state information into a vector to represent the tourist state; S42: Action definition, which clarifies the choices that tourists can make from the current state; S43: Environment modeling: Abstracting the amusement park into a graph structure, using nodes to represent locations and edges to represent paths between nodes. Edge weights are also set. The environment receives actions taken by visitors, updates its state according to rules, and returns corresponding rewards. S44: Formulate a reward function and give corresponding rewards; S45: Use the reinforcement learning algorithm DQN to build and train the path planning model; S46: Input the current state of the tourist, select the optimal action through the strategy network, and generate the optimal travel path.

8. A travel planning method based on deep learning according to claim 7, characterized in that: In step S43, the rules include state transition rules and reward setting rules; State transfer rules include location update rules, time update rules and scenic spot state update rules; The reward setting rules include goal achievement reward rules, time reward rules, path length reward rules, attraction preference reward rules, crowd avoidance reward rules and penalty mechanism rules.

9. A travel planning method based on deep learning according to claim 1, characterized in that: The step S5 specifically includes the following steps: S51: collecting tourist preference information and detailed information of each attraction in the amusement park, and pre-processing the collected data; S52: Build and train a recommendation model based on deep autoencoder; S53: Input the preference data of the new tourist into the trained recommendation model to obtain the low-dimensional feature representation of the tourist; S54: Calculating the similarity score between the feature representation of the visitor and the feature representations of each project in the park; S55: Based on the similarity scores, select attractions with high similarity to form a recommended attraction combination; at the same time, combine the attraction information to plan a reasonable tour sequence for tourists.

10. A travel planning system based on deep learning, applying a travel planning method based on deep learning according to any one of claims 1 to 9, characterized in that: It includes a data collection and processing module, a model building module, a demand collection and analysis module, a route planning module, a scenic spot recommendation module and a feedback suggestion module which are sequentially connected in communication; Data collection and processing module: used to collect relevant data of the amusement park to form a database and pre-process the collected data; Model building module: Builds a park map and attraction layout model based on the convolutional neural network (CNN), builds a crowd flow prediction model based on the long short-term memory (LSTM) network, and calculates waiting time based on the crowd flow prediction results; Demand collection and analysis module: obtains the demand information entered by tourists on the user interface, uses keyword search algorithms to analyze the text entered by tourists, extracts keywords, matches the keywords with the information in the database, and identifies key information; Path planning module: Build a path planning model based on reinforcement learning. Input the identified key information into the trained path planning model. The path planning model selects the optimal action based on the current state of the tourist and plans the optimal travel path. Attraction recommendation module: Using a deep learning-based recommendation algorithm, it recommends suitable attraction combinations and visit sequences for tourists based on their preferences and current park project information. Feedback and suggestion module: used to collect visitors' feedback and suggestions on the park and continuously optimize the park's service quality.

Images