Scheduling method and scheduling system for pre-judging idle state of intelligent parking space based on AI

Abstract

The invention discloses a scheduling method and a scheduling system for pre-judging an idle state of an intelligent parking space based on AI, and belongs to the technical field of big data, and the method comprises the following steps: S1, data collection and preprocessing; s2, feature engineering construction; s3, AI model training is carried out; and S4, recommending an operation and maintenance window. According to the method, the prediction accuracy is improved through multi-source data fusion, the time features, the environment features and the social features are extracted from the preprocessed data, and the space-time dependency and external factor relevance of the parking space are comprehensively described. The high-reliability idle windows are screened through the probability threshold value, invalid attendance caused by temporary parking of the vehicle is directly reduced, and the utilization rate of operation and maintenance resources is effectively improved. And an operable time window is generated based on the idle probability of AI prediction, and a path optimization algorithm is automatically matched to generate a task scheduling instruction, so that end-to-end automation from prediction to execution is realized.

Description

Technical Field

[0001] The present invention belongs to the field of big data technology, and specifically relates to a scheduling method and a scheduling system based on AI-based prediction of the idle state of intelligent berths. Background Art

[0002] With the advancement of smart city development, intelligent parking systems have become a crucial component of urban traffic management. Currently, mainstream intelligent parking systems utilize geomagnetic sensors, Internet of Things (IoT) communication modules, and cloud-based management platforms to collect and transmit real-time parking space occupancy data. These systems, supplemented by mobile applications, provide users with parking space query, navigation, and contactless payment functions. These systems significantly improve parking space utilization and the driver experience, and are widely used in commercial areas, residential areas, and public parking lots, forming a complete technology ecosystem encompassing data collection, transmission, analysis, and application.

[0003] However, for parking space operation and maintenance management (such as geomagnetic equipment maintenance, berth number repair, road maintenance, etc.), existing technologies still rely heavily on manual inspections or real-time geomagnetic monitoring data to schedule operation and maintenance tasks. This model has significant flaws: First, real-time geomagnetic data can only reflect the current parking space occupancy status. When operation and maintenance personnel arrive at the site, they may be unable to operate due to temporarily parked vehicles, resulting in frequent invalid attendance; second, frequent temporary scheduling forces the operation and maintenance plan to be repeatedly adjusted, resulting in wasted manpower and time costs and overall low efficiency; third, existing technologies lack the ability to deeply mine and analyze historical parking data, and cannot combine dynamic factors such as time, date, holidays and special events to predict future idle periods for operation. Among the current solutions, the monitoring system based on geomagnetic induction can only provide static occupancy status and has no predictive function; the method of manually recording parking patterns relies on experience and is difficult to adapt to dynamic changes; and although video surveillance technology can obtain multi-dimensional information, it has problems such as high deployment cost, high environmental interference, and privacy leakage risks.

[0004] In summary, existing parking space operation and maintenance scheduling technologies are unable to meet the needs of efficient smart city management. There is an urgent need to develop a method for predicting idle periods based on historical geomagnetic data and artificial intelligence predictions. By accurately predicting future work windows, this method can optimize the scheduling paths and timing of operation and maintenance tasks, thereby reducing ineffective attendance and improving the utilization of urban operation and maintenance resources. Summary of the Invention

[0005] In response to the problems in related technologies, the present invention provides a scheduling method and a scheduling system based on AI to predict the idle status of intelligent parking spaces, so as to solve technical problems such as the difficulty in operation and maintenance management and scheduling of existing parking spaces, and the difficulty in predicting the timing of operations.

[0006] The technical solution of the present invention is implemented as follows: a scheduling method based on AI-based prediction of the idle state of intelligent berths, comprising the following steps:

[0007] S1. Data Collection and Preprocessing: Each smart berth is equipped with a geomagnetic sensor. This sensor acquires historical geomagnetic data for the target area, including occupied / unoccupied status, timestamps, and signal strength. It also collects external environmental data, including weather data, holiday information, and surrounding events. This historical geomagnetic data is then filtered for outliers and subjected to noise reduction.

[0008] S2. Feature Engineering: Extract multidimensional features from the preprocessed data, including temporal, environmental, and social features. Temporal features include date type, time segmentation, and cyclical trends; environmental features include temperature and rainfall; and social features include holiday and event markers.

[0009] S3. AI model training: A time series prediction model is used to learn historical parking patterns. The model includes at least one of LSTM, Prophet, or XGBoost-Time. The model inputs the multi-dimensional features to output the probability of parking availability for each time period within the next N days.

[0010] S4. Operation and maintenance window recommendation: Generate an operational time window based on the idle probability. When the idle probability is greater than a preset threshold, mark the corresponding time period as a high-probability idle window, and generate an operation and maintenance task scheduling plan based on the path optimization algorithm.

[0011] As a further improvement to the above solution, the data preprocessing in step S1 includes:

[0012] Filter abnormal occupancy status. If the occupancy status of the same parking space lasts for more than 24 hours and the signal strength is lower than the preset fault threshold, it is determined to be sensor fault data and discarded.

[0013] The Kalman filter is used to reduce the noise of the signal strength data. The observation noise covariance matrix of the Kalman filter is dynamically adjusted according to the electromagnetic interference intensity to eliminate the influence of the electromagnetic interference.

[0014] As a further improvement to the above solution, the social features in step S2 also include a dynamic adjustment factor. The dynamic adjustment factor is calculated based on the impact radius of the event and the parking pattern offset caused by similar events in historical data. Specifically, it is:

[0015] Δ=γ·R+(1-γ)·δ;

[0016] Δ is a dynamic adjustment factor used to correct the prediction results; R is the impact radius of the active event; δ is the standard deviation of the historical parking pattern offset; γ is the weight coefficient; where 0.6≤γ≤0.8.

[0017] As a further improvement to the above solution, the training of the time series prediction model in step S3 further includes:

[0018] The dataset is divided into training set, validation set and test set, and the sliding window mechanism is used to generate time series samples;

[0019] During the model training process, an early stopping mechanism is introduced to terminate the training when the validation set loss does not decrease for 5 consecutive epochs to prevent overfitting, and the hyperparameters are optimized through cross-validation.

[0020] As a further improvement to the above solution, the LSTM model is used in step S3 to predict the idle probability, and its output formula is:

[0021] P t =σ(W p ·[h t-1 ,x t ]+b p );

[0022] Among them, P t is the idle probability in period t, h t-1 is the state of the LSTM hidden layer at the previous moment, x t is the characteristic vector of period t, W p is the weight matrix, b p is the bias term, and σ is the Sigmoid activation function.

[0023] As a further improvement of the above solution, the feature vector x t The construction includes: time encoding vector T t , environmental feature vector E t , social feature vector S t , and satisfy:

[0024] x t =T t ⊕E t ⊕S t ;

[0025] Among them, ⊕ represents the vector concatenation operation, T t Including the unique hot encoding of the day of the week, hour, and whether it is the peak period, E t Including normalized values ​​of temperature and rainfall, the normalization range is 0-1, S t Includes binary tags for holidays and event events.

[0026] As a further improvement to the above solution, the path optimization algorithm in step S4 adopts a genetic algorithm, and its fitness function is defined as:

[0027]

[0028] Among them, F is the fitness value, P i is the idle probability of the i-th operation and maintenance point, D j is the driving distance of the jth segment of the path, α and β are weight coefficients, and they satisfy α + β = 1; M is the total number of operation and maintenance points to be scheduled, and N is the total number of segmented driving distances in the operation and maintenance path.

[0029] A scheduling system based on AI prediction of the idle state of intelligent berths, which applies the above-mentioned scheduling method based on AI prediction of the idle state of intelligent berths; comprising:

[0030] Data acquisition module, used to obtain historical data of geomagnetic sensors and external environment data, and perform data cleaning and noise reduction;

[0031] A feature extraction module, connected to the data acquisition module, for constructing time, environment and social multi-dimensional features;

[0032] An AI prediction module, connected to the feature extraction module, uses a time series prediction model to output future idle probability;

[0033] The scheduling optimization module is connected to the AI ​​prediction module and generates operation and maintenance task scheduling instructions based on the idle probability and path constraints.

[0034] As a further improvement of the above solution, the AI ​​prediction module further includes a model updating unit, which regularly receives new data and triggers incremental training. The incremental training uses an online learning algorithm to update model parameters, and the update frequency is once a week.

[0035] As a further improvement of the above solution, the scheduling optimization module also includes a visualization interaction unit, which is used to mark high-probability idle windows and recommended paths on the map, and supports manual adjustment of scheduling priorities. The priority adjustment is dynamically configured based on the urgency of the operation and maintenance task and the idle probability threshold.

[0036] Compared with the prior art, the present invention has the following advantages:

[0037] (1) Multi-source data fusion improves prediction accuracy: Geomagnetic sensors are used to collect historical geomagnetic data such as occupancy / idle status, timestamps, and signal strength, and external environmental data such as weather, holidays, and surrounding events are integrated to build a multi-dimensional data input foundation. Combined with outlier filtering and Kalman filter noise reduction processing, it effectively eliminates sensor noise and environmental interference, ensures the integrity and reliability of training data, provides high-precision input for AI models, significantly enhances the confidence of predictions for future idle periods, and avoids misjudgments caused by data bias.

[0038] (2) Multi-dimensional feature engineering adapts to dynamic parking patterns: Temporal, environmental, and social features are extracted from pre-processed data to comprehensively characterize the spatiotemporal dependencies of parking spaces and their correlation with external factors. By combining multi-dimensional features, the AI ​​model can accurately capture parking behavior patterns in different scenarios, such as commuting peaks and holiday traffic fluctuations, dynamically adapt to complex changes in urban operations, and improve the robustness of idle window predictions.

[0039] (3) Time series prediction models drive intelligent O&M decisions: Using time series prediction models such as LSTM, Prophet, or XGBoost-Time, based on historical patterns and multi-dimensional feature learning, output the idle probability for each time period within the next N days. This upgrades traditional passive O&M to an active prediction mode, using probability thresholds to screen high-reliability idle windows, directly reducing invalid attendance caused by temporary parking of vehicles and effectively improving O&M resource utilization.

[0040] (4) Closed-loop scheduling optimization reduces operation and maintenance costs: Based on AI-predicted idle probability, it generates an operational time window and automatically matches the path optimization algorithm to generate task scheduling instructions, achieving end-to-end automation from prediction to execution. This breaks through the inefficient bottleneck of traditional manual experience-based scheduling, reduces the frequency of temporary plan adjustments, effectively shortens the execution time of a single operation and maintenance task, and significantly reduces manpower, fuel, and time costs, meeting the refined needs of smart cities for efficient operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of a scheduling method based on AI prediction of the idle state of intelligent berths provided in an embodiment;

[0042] Figure 2 This is an architectural diagram of a scheduling system based on AI prediction of intelligent berth idle status provided in an embodiment. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0044] Example:

[0045] like Figure 1 As shown, a scheduling method based on AI prediction of the idle state of intelligent berths includes the following steps:

[0046] S1. Data collection and preprocessing: Each smart berth is equipped with a geomagnetic sensor. This embodiment uses a geomagnetic sensor of model MLX90393. The geomagnetic sensor is used to obtain historical geomagnetic data of the target area, including occupied / idle status, such as 0 / 1, 0 represents occupied status, 1 represents idle status, timestamp (accurate to seconds), signal strength (dBm) and sensor temperature (°C). The sampling frequency is configured to be 1Hz. Collect external environmental data, which includes weather data, holiday information and surrounding activities, and perform outlier filtering and noise reduction on the historical geomagnetic data. Specifically, it is possible to access weather data such as rainfall and temperature from the Meteorological Bureau, holiday lists issued by the government, and information on surrounding large-scale events provided by the urban activity management platform, such as concerts and sports events, through open APIs.

[0047] In this embodiment, the data preprocessing in step S1 includes:

[0048] S11. Outlier Filtering: Filter abnormal occupancy status. If the occupancy status of the same parking space lasts for more than 24 hours and the signal strength is lower than the preset fault threshold, it is determined to be sensor fault data and discarded. Specifically, in this embodiment, the signal strength below -90dB is set as the preset fault threshold.

[0049] S12. Kalman filter noise reduction: Kalman filter is used to reduce the noise of signal strength data. The observation noise covariance matrix of the Kalman filter is dynamically adjusted according to the electromagnetic interference intensity to eliminate the influence of electromagnetic interference. Specifically, the state transfer matrix can be set: The observation matrix H = [1 0], and the observation noise covariance matrix R is dynamically adjusted according to the intensity of electromagnetic interference. For example, when the interference is strong, R = 5, and when the interference is weak, R = 1. By fusing geomagnetic data with external environmental data, and using dual conditions to filter outliers and dynamic noise reduction, sensor noise and environmental interference are eliminated, the signal-to-noise ratio of the input data is improved, and a high-precision data foundation is provided for subsequent model training. This embodiment uses geomagnetic sensors to collect historical geomagnetic data such as occupancy / idle status, timestamps, and signal strength, and integrates external environmental data such as weather, holidays, and surrounding events to build a multi-dimensional data input foundation. Combined with outlier filtering and Kalman filter noise reduction processing, it effectively eliminates sensor noise and environmental interference, ensures the integrity and reliability of training data, provides high-precision input for AI models, significantly enhances the confidence level of predictions for future idle periods, and avoids misjudgments caused by data bias.

[0050] S2. Feature Engineering: Extracting multidimensional features from preprocessed data, including temporal, environmental, and social features.

[0051] The time features include date type, time segmentation and periodic trend; specifically, the date type includes weekdays / weekends / holidays, which are encoded in binary; the time segmentation includes morning peak, off-peak and evening peak, which are encoded in one-hot encoding; the periodic trend can be the average occupancy rate in the same time period over the past 7 days.

[0052] The environmental characteristics include temperature and rainfall; specifically, the temperature is normalized to 0-1, and the rainfall is normalized to 0-1.

[0053] The social features include holiday tags and event tags; specifically, holiday tags are set to 1 / 0; for event tags, such as during a concert, the tag is set to 1. In this embodiment, the social features in step S2 also include a dynamic adjustment factor, which is calculated based on the impact radius of the event and the parking mode offset caused by similar events in historical data. Specifically, it is:

[0054] Δ=γ·R+(1-γ)·δ;

[0055] Δ is a dynamic adjustment factor used to correct the prediction results; R is the impact radius of the event, in kilometers; δ is the standard deviation of the historical parking pattern offset; γ is the weight coefficient; where 0.6≤γ≤0.8.

[0056] In this embodiment, the calculation method of δ is:

[0057]

[0058] Among them, x iis the hourly parking rate during similar historical events, μ is the average parking rate, and N is the number of data samples.

[0059] For example, if there is a large-scale concert event in the target area, the parameter settings are:

[0060] R = 2 km, the event’s impact range is provided by the organizer;

[0061] δ = 0.5, that is, the standard deviation of parking rate during similar concerts in history is 0.5;

[0062] γ = 0.7, focusing more on the impact of real-time activities;

[0063] Then the dynamic adjustment factor Δ = 0.7 × 2 + (1 - 0.7) × 0.5 = 1.4 + 0.15 = 1.55; the dynamic adjustment factor Δ = 1.55 will be added to the social characteristics, indicating that the probability of vacant parking spaces within 2 kilometers of the concert needs to be revised downward. Because the concert event attracts a large number of vehicles, the correction range is controlled by the Δ weighting.

[0064] For example, if the target area only has small community events, such as a market, the parameter settings are:

[0065] R = 0.5 km, the activity only affects 500 meters around the area;

[0066] δ = 0.3, which means that the parking fluctuation of similar historical activities is small;

[0067] γ = 0.8, more dependent on the real-time activity range;

[0068] The dynamic adjustment factor Δ = 0.8 × 0.5 + (1-0.8) × 0.3 = 0.4 + 0.06 = 0.46. Δ = 0.46 indicates that the activity has limited impact, and the model only fine-tunes the idle probability of the local area.

[0069] As can be seen from the above example, the dynamic adjustment factor Δ enables the AI ​​model to dynamically correct prediction results by integrating real-time activity range and historical fluctuation data. This is particularly suitable for complex scenarios such as holidays and large-scale events. It takes into account both scientificity and practicality, and significantly improves prediction robustness and operation and maintenance scheduling efficiency.

[0070] S3. AI model training: A time series prediction model is used to learn historical parking patterns. The time series prediction model includes at least one of LSTM, Prophet or XGBoost-Time, and the multi-dimensional features are input to output the idle probability of each time period in the next N days. In this embodiment, the training of the time series prediction model in step S3 further includes: dividing the data set into a training set, a validation set and a test set, and using a sliding window mechanism to generate time series samples. During the model training process, an early stopping mechanism is introduced to terminate the training when the validation set loss has not decreased for 5 consecutive epochs to prevent overfitting, and hyperparameters are optimized through cross-validation.

[0071] This embodiment takes the LSTM model as an example and specifically includes the following process:

[0072] S31. Model selection and training:

[0073] The LSTM model is used, in which the input layer has 64 nodes, the hidden layer has 32 nodes, and the output layer has 1 node. The input feature vector x t =T t ⊕E t ⊕S t ; Among them, T t is the time encoding vector, including the unique hot encoding of the day of the week, hour, and whether it is the peak period; E t is the environmental characteristic vector, including the normalized values ​​of temperature and rainfall, with a normalization range of 0-1, S t is a social feature vector, including binary labels of holidays and events; ⊕ represents a vector concatenation operation.

[0074] The dataset is divided into training set, validation set and test set according to the ratio of 7:2:1, and a sliding window mechanism (window length 7 days, step length 1 hour) is used to generate time series samples.

[0075] An early stopping mechanism is introduced to terminate training when the validation loss does not decrease for five consecutive epochs, and 5-fold cross-validation is used to optimize hyperparameters. It should be noted that the early stopping mechanism is a strategy to prevent model overfitting. During training, by monitoring the changes in validation loss, training is terminated early when the loss value stops decreasing within a certain number of epochs, preventing the model from overfitting the training set and losing generalization ability.

[0076] For example, suppose we train an LSTM model to predict the probability of a parking space being available:

[0077] Epoch 1: validation loss = 0.25;

[0078] Epoch 2: validation loss = 0.20;

[0079] Epoch 3: validation loss = 0.18;

[0080] Epoch 4: Validation loss = 0.17;

[0081] Epoch 5: Validation loss = 0.17;

[0082] Epoch 6: Validation loss = 0.17;

[0083] From Epoch 4 to Epoch 6, the validation set loss did not decrease for 3 consecutive times (but did not reach 5 times), so training continued.

[0084] Epoch 7: Validation set loss = 0.18 (starting to rise);

[0085] Epoch 8: Validation loss = 0.19;

[0086] Epoch 9: Validation loss = 0.20;

[0087] After five consecutive runs from Epoch 5 to Epoch 9, where the validation loss did not decrease (actual loss increased), training was terminated at Epoch 9, retaining the model parameters from Epoch 4. At this point, the validation loss reached a minimum of 0.17. Introducing the early stopping mechanism can avoid overfitting and prevent the model from excessively memorizing noise in the training data. It also saves computing resources and reduces ineffective training rounds.

[0088] Five-fold cross-validation is a hyperparameter optimization method that divides a dataset into five equal parts, alternately using four parts as training sets and one part as validation sets. Training is repeated five times to evaluate the performance of different hyperparameter combinations. Ultimately, the hyperparameters with the best average performance are selected, thus avoiding the randomness of a single validation set. Combined with an early stopping mechanism, this method can efficiently optimize the model and prevent overfitting, significantly improving prediction accuracy.

[0089] The specific implementation steps include the following:

[0090] S311. Data partitioning: Randomly divide the training set (excluding the test set) into 5 parts (Fold 1 to 5).

[0091] S312. Cross Validation:

[0092] Round 1: Training on Folds 1 to 4, and verification on Fold 5.

[0093] Round 2: training on Folds 1 to 3 and Fold 5, and verification on Fold 4.

[0094] Repeat until all 5 copies are used as validation sets.

[0095] S313. Hyperparameter selection: Calculate the average of the five rounds of validation results and select the hyperparameter combination with the best average performance.

[0096] S32. Idle probability output:

[0097] The LSTM model is used to predict the idle probability, and its output formula is:

[0098] P t =σ(W p ·[h t-1 ,x t ]+b p );

[0099] Among them, P t is the idle probability in period t; h t-1 is the state of the LSTM hidden layer at the previous moment; x t is the characteristic vector of period t; W p is the weight matrix, dimension 32x1; b p is the bias term, σ is the Sigmoid activation function, and outputs the idle probability P for each time period in the next 3 days t ∈[0,1].

[0100] S4. Operation and maintenance window recommendation: Generate an operational time window based on the idle probability. When the idle probability is greater than a preset threshold, mark the corresponding time period as a high-probability idle window, and generate an operation and maintenance task scheduling plan based on the path optimization algorithm.

[0101] The path optimization algorithm in this embodiment adopts a genetic algorithm, which specifically includes the following steps:

[0102] S41. Window generation: Set the idle probability threshold (default 90%). t When the value is >0.9, it is marked as a high-probability idle window.

[0103] S42. Path optimization: Genetic algorithm (population size 50, iteration number 100), weight coefficients α and β are 0.7 and 0.3 respectively; fitness function is:

[0104]

[0105] P i is the idle probability of the i-th operation and maintenance point, D j is the driving distance of the jth segment of the path, M is the total number of operation and maintenance points to be scheduled, and N is the total number of segmented driving distances in the operation and maintenance path.

[0106] For example, a commercial district needs to conduct geomagnetic inspections on five parking spaces (A, B, C, D, and E). The AI-predicted vacancy probabilities and path distances are shown in the following table:

[0107] Operation and maintenance points <![CDATA[Idle probability P i > <![CDATA[Adjacent path segment distance D j (km)]]> A 0.95 A→B:1.2 B 0.90 B→C:0.8 C 0.85 C→D:1.5 D 0.92 D→E:0.7 E 0.88 E→A:2.0 (closed loop path)

[0108] The comparison of candidate path solutions is shown in the following table:

[0109]

[0110] From the above comparative analysis, it can be seen that the fitness value of the sequential path and the optimal path is the same (F=3.221), but the optimal path adjusts the access order, reduces the number of turns or complex sections, such as avoiding congested areas, and has higher actual execution efficiency. The optimized path is eliminated by the algorithm because of its longer total distance (5.0 kilometers) and the lowest fitness value (F=3.210). This embodiment generates an operational time window based on the idle probability predicted by AI, and automatically matches the path optimization algorithm to generate task scheduling instructions, realizing end-to-end automation from prediction to execution, ensuring that operation and maintenance personnel can work directly after arriving at the site, and the configuration of the fitness function makes the algorithm naturally prefer shorter paths. At the same time, it also has a certain degree of dynamic adaptability. If the idle probability of a certain point drops suddenly, such as temporary activities causing P i =0.6, the algorithm automatically removes that point and replans the route, reducing manual intervention. The path optimization algorithm utilizes a genetic algorithm to overcome the inefficiency bottleneck of traditional manual scheduling, reducing the frequency of temporary plan adjustments, effectively shortening the execution time of a single operation and maintenance task, and significantly reducing labor, fuel, and time costs, meeting the refined needs of smart cities for efficient operation and maintenance.

[0111] like Figure 2 As shown, this embodiment also provides a scheduling system based on AI prediction of the idle state of intelligent berths, which applies the above-mentioned scheduling method based on AI prediction of the idle state of intelligent berths; including:

[0112] The data acquisition module is used to obtain historical data of geomagnetic sensors and external environment data, and perform data cleaning and noise reduction. Specifically, it can receive sensor data through the MQTT protocol and store it in a MySQL database.

[0113] A feature extraction module, connected to the data acquisition module, for constructing time, environment and social multi-dimensional features;

[0114] The AI ​​prediction module, connected to the feature extraction module, uses a time series prediction model to output future idleness probabilities. The AI ​​prediction module further includes a model update unit that regularly receives new data and triggers incremental training. This incremental training uses an online learning algorithm to update model parameters weekly. Specifically, an LSTM model is deployed based on the TensorFlow framework, with incremental training triggered weekly and a learning rate set to 0.001.

[0115] The scheduling optimization module is connected to the AI ​​prediction module and generates operation and maintenance task scheduling instructions based on idle probabilities and path constraints. The scheduling optimization module also includes a visualization interaction unit, which is used to mark high-probability idle windows and recommended paths on the map and supports manual adjustment of scheduling priorities. The priority adjustment is dynamically configured based on the urgency of the operation and maintenance task and the idle probability threshold. Specifically, the Leaflet.js map library is used to display high-probability windows and recommended paths, and manual adjustment of priorities is supported, such as adjusting the emergency task threshold to 95%.

[0116] Based on the disclosure and teachings of the above description, those skilled in the art may also make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and any modifications and variations of the invention should also fall within the scope of protection of the claims of the present invention. In addition, although certain specific terms are used in this description, these terms are for convenience of description only and do not constitute any limitation to the present invention.

Claims

1. A scheduling method based on AI prediction of intelligent berth idle status, characterized in that: The following steps are involved: S1. Data Collection and Preprocessing: Each smart berth is equipped with a geomagnetic sensor. This sensor acquires historical geomagnetic data for the target area, including occupied / unoccupied status, timestamps, and signal strength. It also collects external environmental data, including weather data, holiday information, and surrounding events. This historical geomagnetic data is then filtered for outliers and subjected to noise reduction. S2. Feature Engineering: Extract multidimensional features from the preprocessed data, including temporal, environmental, and social features. Temporal features include date type, time segmentation, and cyclical trends; environmental features include temperature and rainfall; and social features include holiday and event markers. S3. AI model training: A time series prediction model is used to learn historical parking patterns. The model includes at least one of LSTM, Prophet, or XGBoost-Time. The model inputs the multi-dimensional features to output the probability of parking availability for each time period within the next N days. S4. Operation and maintenance window recommendation: Generate an operational time window based on the idle probability. When the idle probability is greater than a preset threshold, mark the corresponding time period as a high-probability idle window, and generate an operation and maintenance task scheduling plan based on the path optimization algorithm.

2. The scheduling method based on AI prediction of intelligent berth idle status according to claim 1 is characterized in that: The data preprocessing in step S1 includes: Filter abnormal occupancy status. If the occupancy status of the same parking space lasts for more than 24 hours and the signal strength is lower than the preset fault threshold, it is determined to be sensor fault data and discarded. The Kalman filter is used to reduce the noise of the signal strength data. The observation noise covariance matrix of the Kalman filter is dynamically adjusted according to the electromagnetic interference intensity to eliminate the influence of the electromagnetic interference.

3. The scheduling method based on AI prediction of intelligent berth idle status according to claim 1 is characterized in that: The social features in step S2 also include a dynamic adjustment factor, which is calculated based on the impact radius of the event and the parking pattern offset caused by similar events in historical data. Specifically, it is: Δ=γ·R+(1-γ)·δ; Δ is a dynamic adjustment factor used to correct the prediction results; R is the impact radius of the active event; δ is the standard deviation of the historical parking pattern offset; γ is the weight coefficient; where 0.6≤γ≤0.

8.

4. The scheduling method based on AI prediction of intelligent berth idle status according to claim 1 is characterized in that: The training of the time series prediction model in step S3 further includes: The dataset is divided into training set, validation set and test set, and the sliding window mechanism is used to generate time series samples; During the model training process, an early stopping mechanism is introduced to terminate the training when the validation set loss does not decrease for 5 consecutive epochs to prevent overfitting, and the hyperparameters are optimized through cross-validation.

5. The scheduling method based on AI prediction of intelligent berth idle status according to claim 1 is characterized in that: In step S3, the LSTM model is used to predict the idle probability, and its output formula is: P t =σ(W p ·[h t-1 ,x t ]+b p ); Among them, P t is the idle probability in period t, h t-1 is the state of the LSTM hidden layer at the previous moment, x t is the characteristic vector of period t, W p is the weight matrix, b p is the bias term, and σ is the Sigmoid activation function.

6. The scheduling method based on AI prediction of intelligent berth idle status according to claim 5 is characterized in that: The eigenvector x t The construction includes: time encoding vector T t , environmental feature vector E t , social feature vector S t , and satisfy: in, Represents vector concatenation operation, T t Including the unique hot encoding of the day of the week, hour, and whether it is the peak period, E t Including normalized values ​​of temperature and rainfall, the normalization range is 0-1, S t Includes binary tags for holidays and event events.

7. The scheduling method based on AI prediction of intelligent berth idle status according to claim 1 is characterized in that: The path optimization algorithm in step S4 adopts a genetic algorithm, and its fitness function is defined as: Among them, F is the fitness value, P i is the idle probability of the i-th operation and maintenance point, D j is the driving distance of the jth segment of the path, α and β are weight coefficients, and they satisfy α + β = 1; M is the total number of operation and maintenance points to be scheduled, and N is the total number of segmented driving distances in the operation and maintenance path.

8. A scheduling system based on AI prediction of the idle state of intelligent berths, applying a scheduling method based on AI prediction of the idle state of intelligent berths as described in any one of claims 1 to 7; characterized in that: include: Data acquisition module, used to obtain historical data of geomagnetic sensors and external environment data, and perform data cleaning and noise reduction; A feature extraction module, connected to the data acquisition module, for constructing time, environment and social multi-dimensional features; An AI prediction module, connected to the feature extraction module, uses a time series prediction model to output future idle probability; The scheduling optimization module is connected to the AI ​​prediction module and generates operation and maintenance task scheduling instructions based on the idle probability and path constraints.

9. The scheduling method based on AI prediction of intelligent berth idle status according to claim 8 is characterized in that: The AI ​​prediction module further includes a model updating unit, which regularly receives new data and triggers incremental training. The incremental training uses an online learning algorithm to update model parameters with an update frequency of once a week.

10. The scheduling method based on AI prediction of intelligent berth idle status according to claim 9 is characterized in that: The scheduling optimization module also includes a visualization interaction unit, which is used to mark high-probability idle windows and recommended paths on the map and supports manual adjustment of scheduling priorities. The priority adjustment is dynamically configured based on the urgency of the operation and maintenance task and the idle probability threshold.

Images