Aerial passenger cableway cabin dynamic load prediction method based on deep neural network

Abstract

The application discloses a dynamic load prediction method for a hanging passenger cableway based on a deep neural network, which comprises the following steps: step one, collecting and preprocessing the running data of multiple hanging cars in the hanging passenger cableway; step two, inputting the running data into a TimesNet network; step three, calculating the numerical difference between adjacent time points of an initial load prediction value sequence, marking local anchor points, and constructing a cross-hanging car anchor point graph; step four, calculating the disturbance intensity of an anchor point pair based on a two-dimensional hanging car position index matrix, and generating a disturbance embedding tensor; step five, performing a Kronecker product operation on the embedding representation tensor and the disturbance embedding tensor, and inputting them into a prediction head structure; and step six, calculating the numerical error and trend error between the load prediction value sequence and the real load sequence, and updating the bias value. The application improves the load prediction accuracy and robustness of the hanging car through the TimesNet network and disturbance reconstruction.

Description

Technical Field

[0001] This invention relates to the field of cableway load monitoring technology, and in particular to a method for predicting the dynamic load of aerial passenger cableway gondolas based on deep neural networks. Background Technology

[0002] With the increasing demand for urban transportation, aerial ropeways have become a widely used, efficient, and safe mode of public transportation. In actual operation, changes in the load on the gondolas directly affect the system's safety, comfort, and scheduling efficiency. Therefore, how to accurately predict the load on the gondolas, especially dynamic load prediction, has become an important research topic in the field of intelligent transportation. Traditional gondola load prediction methods mostly rely on static calculations and simple historical data regression, which cannot effectively handle complex, dynamically changing load data, resulting in low prediction accuracy and failing to meet the needs of modern intelligent scheduling and safety monitoring.

[0003] Existing load forecasting methods mainly fall into three categories: statistical models, physical models, and machine learning models. Traditional statistical models typically employ linear regression and time series analysis. While these methods can handle simple load forecasting problems, they often exhibit poor prediction accuracy and high computational complexity when dealing with complex, nonlinear, large-scale data. Physical models predict load by modeling the motion state of the gondolas, but these models often ignore the influence of environmental factors (such as weather changes and passenger flow fluctuations) and are computationally intensive, making real-time forecasting difficult. In recent years, machine learning methods, especially deep neural networks, have provided new directions for gondola load forecasting. However, existing deep learning models still face some technical challenges, particularly in effectively handling multi-dimensional, multi-time-series, and multi-source data, and in accurately predicting loads by incorporating the dynamic changes and complex disturbances of the gondolas.

[0004] While some deep learning-based load prediction methods have made progress in the field, these methods typically rely solely on raw sensor data, such as load weight and speed, lacking in-depth modeling of temporal dynamic relationships and disturbance factors. Most methods neglect the interactions between multiple load cells and fail to effectively capture dynamic load changes across load cells.

[0005] Therefore, how to provide a method for predicting the dynamic load of aerial passenger cableway gondolas based on deep neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a dynamic load prediction method for aerial passenger cableway gondolas based on deep neural networks. This invention effectively extracts the temporal characteristics of gondola loads using TimesNet networks and perturbation reconstruction techniques, solving the problem that existing methods cannot handle cross-gondola dynamic influences and complex perturbations. By constructing cross-gondola anchor point graphs and temporal causal relationship graphs, the perturbation relationships between gondolas are accurately captured. Combined with Kronecker product and weighted aggregation, the accuracy and robustness of load prediction are significantly improved, ensuring accurate predictions in highly dynamic environments and enhancing system safety and operational efficiency.

[0007] The dynamic load prediction method for aerial passenger ropeway gondolas based on deep neural networks according to embodiments of the present invention includes the following steps:

[0008] Step 1: Collect and preprocess the operating data of multiple gondolas in the aerial passenger cableway;

[0009] Step 2: Input the operating data of each gondola into the TimesNet network to obtain the initial load prediction sequence for each gondola;

[0010] Step 3: Calculate the numerical difference between adjacent time points in the initial load prediction sequence for each gondola, mark local anchor points, and construct a cross-gond anchor point map;

[0011] Step 4: Construct a two-dimensional gondola location index matrix based on the relative geographical location index of the gondola, and calculate the disturbance intensity of anchor point pairs in the cross-gondola anchor point map to construct the anchor point disturbance matrix and generate the disturbance embedding tensor;

[0012] Step 5: Perform Kronecker product operation on each embedding representation tensor of the hoistway with the perturbation embedding tensor to generate a perturbation reconstruction feature representation tensor sequence, and input it into the prediction head structure to obtain the perturbation reconstruction load prediction value sequence.

[0013] Step 6: Calculate the numerical error between the predicted load sequence and the actual load sequence, obtain the trend error through Soft-DTW alignment, and update the bias value based on the numerical error and the trend error.

[0014] Optionally, the operational data includes real-time load weight data, meteorological data, date and time data, passenger flow data, and relative geographical location index of each gondola, wherein the relative geographical location index of the gondola is a mapping relationship between the gondola number and the gondola's spatial coordinates; the preprocessing includes performing missing value imputation, time synchronization processing, outlier removal, and normalization processing on different types of data to generate operational data with a unified structure.

[0015] Optionally, step two specifically includes:

[0016] The TimesNet network includes an input embedding layer, a convolutional perceptual layer, a periodic perceptual attention layer, a temporal dimension compression layer, and a prediction head structure.

[0017] The input embedding layer is used to receive the running data of each gondola, perform linear mapping on the running data, and compress the feature dimension to obtain the initial embedding tensor;

[0018] The initial embedding tensor is input to the convolutional perceptual layer, which consists of several parallel convolutional channels. Each convolutional channel has a different convolutional kernel size and is used to extract feature patterns at different time scales. The feature patterns are the changes in the load weight of the gondola at different time steps. Each convolutional channel outputs a scale convolutional feature tensor.

[0019] Multiple scale convolutional feature tensors are input into the periodic awareness attention layer. The cosine similarity between any two scale convolutional feature tensors is calculated as the attention intensity, and an attention matrix is ​​constructed.

[0020] The attention matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the obtained eigenvalues ​​are used as weight vectors to weight the convolutional feature tensors of multiple scales to obtain the periodic enhancement feature tensor.

[0021] The periodic augmentation feature tensor is input into the time dimension compression layer, and the subsequences of the periodic augmentation feature tensor in different time periods are extracted segment by segment using the set time sliding window size and step size.

[0022] Perform average pooling on each subsequence, and concatenate the pooling results of each time sliding window in time order to generate a time-dimensional compressed embedding representation tensor sequence;

[0023] The embedded representation tensor sequence is input into the prediction head structure, and the output is the sequence of initial load prediction values ​​for the corresponding hoist within a preset future time range.

[0024] Optionally, the processing steps for the prediction head structure are as follows:

[0025] Constructing a temporal causal relationship graph based on the embedded representation tensor sequence, the construction steps are as follows:

[0026] The embedding representation tensor of each time step is used as a node in the temporal causal relationship graph, and all nodes are arranged in chronological order.

[0027] Calculate the Euclidean distance between any two nodes. If it is less than a preset threshold, it is considered that there is a causal relationship between the corresponding nodes. Establish a directed edge from the previous time step to the next time step, and the weight of the edge is the corresponding Euclidean distance.

[0028] For each node in the target prediction time step, find all historical nodes in the temporal causal relationship graph, where the historical nodes are the nodes corresponding to the incoming edges;

[0029] For the embedding representation tensors represented by all historical nodes, linear weighted aggregation is performed according to the weights of the incoming edges to obtain the aggregated feature representation of the target prediction time step.

[0030] Assign a fixed feature weight to each feature dimension in the aggregated feature representation;

[0031] Multiply each feature dimension by its corresponding feature weight, and sum all the product results to obtain the total weighted value;

[0032] The sum of the weighted values ​​is added to a preset bias value to obtain the initial load prediction value for the target prediction time step. The bias value is used to adjust the overall offset of the initial load prediction value.

[0033] Arrange the initial load forecast values ​​of all target prediction time steps in chronological order to obtain the sequence of initial load forecast values ​​for the corresponding hoistway within the future preset time range.

[0034] Optionally, step three specifically includes:

[0035] Calculate the numerical difference between adjacent time points in the initial load prediction sequence for each hoist, forming the corresponding hoist rate of change sequence;

[0036] Calculate the global mean and standard deviation of the rate of change series, and set a preset threshold coefficient;

[0037] All time points in the rate of change sequence that are greater than the global mean plus a threshold coefficient multiple of the standard deviation are marked as local anchor points, forming the corresponding anchor point time set for the gondola.

[0038] The anchor point time sets of all gondolas are merged to form the total anchor point set;

[0039] Based on the total set of anchor points, a cross-cabin anchor point diagram is constructed. The construction steps are as follows:

[0040] Each local anchor point is represented as an anchor node in the cross-cab anchor point diagram;

[0041] For any two anchor nodes in the cross-cabin anchor point diagram, if the time difference between the corresponding time points is less than the preset time window length, then an undirected edge is established between the two anchor nodes.

[0042] The reciprocal of the time difference between two anchor points is used as the edge weight of the corresponding undirected edge, which represents the degree of time proximity between the corresponding anchor points.

[0043] Optionally, step four specifically includes:

[0044] A two-dimensional gondola location index matrix is ​​constructed based on the gondola's relative geographical location index, and the location index matrix represents the relative coordinate relationship of the gondola in the spatial arrangement;

[0045] For any pair of anchor points connected by an edge in the cross-cabin anchor point diagram, extract the corresponding time difference and the cabin index difference. The cabin index difference is the spatial distance between the corresponding relative coordinates in the two-dimensional cabin position index matrix.

[0046] For all anchor point pairs, the product of the corresponding time difference and the pod index difference is calculated as the disturbance intensity of the corresponding anchor point pair.

[0047] An anchor point perturbation matrix is ​​established using the index of the anchor point node in the total anchor point set as the row and column coordinates. Each element in the anchor point perturbation matrix is ​​the perturbation intensity of the corresponding anchor point pair, and the elements corresponding to unconnected anchor points are set to zero.

[0048] The anchor point perturbation matrix is ​​multiplied by the initial load prediction value corresponding to the pod anchor point, and the dimensions are reconstructed to generate a perturbation embedding tensor, which is used to represent the cross-time perturbation correlation between pods.

[0049] Optionally, step five specifically includes:

[0050] Align the embedding representation tensor of each gondola with the corresponding perturbation embedding tensor in the time dimension;

[0051] Perform the Kronecker product operation on the embedding representation tensor and the perturbation embedding tensor corresponding to each time step to obtain the perturbation reconstructed feature representation tensor for the corresponding time step;

[0052] By concatenating the perturbation reconstruction feature representation tensors corresponding to all time steps in chronological order, a sequence of perturbation reconstruction feature representation tensors is obtained.

[0053] The perturbation-reconstructed feature representation tensor sequence is input into the prediction head structure to construct the perturbation-reconstructed temporal causal relationship graph;

[0054] Based on the time-series causal relationship graph after disturbance reconstruction, historical nodes are weighted and aggregated to obtain the load prediction value after disturbance reconstruction.

[0055] Arrange the perturbation-reconstructed load prediction values ​​of all target prediction time steps in chronological order to generate a corresponding sequence of perturbation-reconstructed load prediction values ​​for the gantry within a preset future time range.

[0056] Optionally, step six specifically includes:

[0057] The load prediction sequence after each gondola disturbance reconstruction is aligned with the corresponding real load sequence point by point on the time step, and the squared difference between the prediction value and the real value at each time step is calculated. The real load sequence is obtained by the weighing sensor installed at the bottom of the gondola during actual operation at time intervals.

[0058] The summation of the squared differences of all time steps yields the numerical error of the corresponding hoist within the preset time range;

[0059] A Soft-DTW alignment operation is performed between the perturbation-reconstructed load prediction sequence and the actual load sequence. The Soft-DTW alignment operation includes the following steps:

[0060] A time-aligned distance matrix is ​​constructed based on the load prediction sequence and the actual load sequence. Each element in the time-aligned distance matrix represents the Manhattan distance between the predicted value and the actual value between two time points.

[0061] Based on the time alignment distance matrix, the path with the minimum total cost is searched through dynamic programming and selected as the optimal time alignment path.

[0062] The sum of costs along the optimal time alignment path is used as the trend error, which represents the degree of deviation of the load prediction sequence from the actual load sequence.

[0063] Based on the set weight ratio of numerical error and trend error, the numerical error and trend error are weighted to obtain the comprehensive error;

[0064] If the overall error is greater than the preset error threshold, the difference between the overall error and the error threshold will be subtracted from the bias value in the prediction head structure until the overall error converges to below the error threshold, and finally the load prediction result with controllable error will be output.

[0065] The beneficial effects of this invention are:

[0066] The proposed method for dynamic load prediction of aerial passenger cableway gondolas based on deep neural networks successfully overcomes the shortcomings of existing technologies in effectively handling multi-dimensional and multi-moment data, particularly in capturing dynamic load changes and complex disturbances across gondolas. By combining TimesNet networks with disturbance reconstruction techniques, this method can extract highly discriminative temporal features from multi-source data, improving prediction accuracy and stability. By introducing cross-gondola anchor point graphs and temporal causal relationship graphs, this invention can identify the mutual influence between gondolas in real time, further enhancing load prediction capabilities in multi-gondola systems and avoiding the problem of neglecting cross-gondola mutual influences in traditional methods. Simultaneously, by utilizing Kronecker product and weighted aggregation operations, this invention enhances the sensitivity of load prediction to disturbance factors, enabling more accurate dynamic load prediction when dealing with complex environmental changes. These innovations not only improve prediction accuracy in practical applications but also enhance system safety and operational efficiency, especially under high dynamic load and variable environmental conditions, better meeting the needs of intelligent scheduling and safety monitoring, demonstrating significant technical advantages. Attached Figure Description

[0067] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0068] Figure 1 This is an overall flowchart of the dynamic load prediction method for aerial passenger cableway gondolas based on deep neural networks proposed in this invention;

[0069] Figure 2 This is a TimesNet network processing flowchart of the dynamic load prediction method for aerial passenger cableway gondolas based on deep neural networks proposed in this invention.

[0070] Figure 3 This diagram illustrates the numerical error and trend error of the dynamic load prediction method for aerial passenger cableway gondolas based on deep neural networks proposed in this invention, and how it updates the bias values. Detailed Implementation

[0071] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0072] refer to Figure 1-3 A method for predicting the dynamic load of gondolas in aerial passenger ropeways based on deep neural networks includes the following steps:

[0073] Step 1: Collect and preprocess the operating data of multiple gondolas in the aerial passenger cableway;

[0074] Step 2: Input the operating data of each gondola into the TimesNet network to obtain the initial load prediction sequence for each gondola;

[0075] Step 3: Calculate the numerical difference between adjacent time points in the initial load prediction sequence for each gondola, mark local anchor points, and construct a cross-gond anchor point map;

[0076] Step 4: Construct a two-dimensional gondola location index matrix based on the relative geographical location index of the gondola, and calculate the disturbance intensity of anchor point pairs in the cross-gondola anchor point map to construct the anchor point disturbance matrix and generate the disturbance embedding tensor;

[0077] Step 5: Perform Kronecker product operation on each embedding representation tensor of the hoistway with the perturbation embedding tensor to generate a perturbation reconstruction feature representation tensor sequence, and input it into the prediction head structure to obtain the perturbation reconstruction load prediction value sequence.

[0078] Step 6: Calculate the numerical error between the predicted load sequence and the actual load sequence, obtain the trend error through Soft-DTW alignment, and update the bias value based on the numerical error and the trend error.

[0079] This step innovatively solves the problem of cabin load prediction in aerial passenger cableways by combining the TimesNet network with cross-cabin anchor point graphs. By preprocessing the cabin operation data, embedding perturbation tensors, and applying Kronecker products, the expressive power of the prediction features is enhanced. Soft-DTW alignment is used to calculate trend errors, accurately comparing the predicted load values ​​with the actual load values, thereby optimizing the prediction results. This effectively improves load prediction accuracy, provides more accurate load monitoring, and enhances the safety and operational efficiency of the cableway system.

[0080] In this embodiment, the operational data includes real-time load weight data, meteorological data, date and time data, passenger flow data, and relative geographical location index of each gondola. The relative geographical location index of the gondola is a mapping relationship between the gondola number and the gondola's spatial coordinates. The preprocessing includes performing missing value imputation, time synchronization processing, outlier removal, and normalization processing on different types of data to generate operational data with a unified structure.

[0081] This step preprocesses multi-source data to ensure high-quality and consistent input data. Techniques such as missing value imputation, time synchronization, outlier removal, and normalization are employed to process different data types, effectively reducing noise interference and providing accurate training data for the deep neural network. This processing step significantly improves the model's prediction accuracy and robustness, ensuring the stability and reliability of gondola load prediction and providing reliable data support for the safety and operational efficiency of the cableway system.

[0082] In this embodiment, step two specifically includes:

[0083] The TimesNet network includes an input embedding layer, a convolutional perceptual layer, a periodic perceptual attention layer, a temporal dimension compression layer, and a prediction head structure.

[0084] The input embedding layer is used to receive the running data of each gondola, perform linear mapping on the running data, and compress the feature dimension to obtain the initial embedding tensor;

[0085] The initial embedding tensor is input to the convolutional perceptual layer, which consists of several parallel convolutional channels. Each convolutional channel has a different convolutional kernel size and is used to extract feature patterns at different time scales. The feature patterns are the changes in the load weight of the gondola at different time steps. Each convolutional channel outputs a scale convolutional feature tensor.

[0086] Multiple scale convolutional feature tensors are input into the periodic awareness attention layer. The cosine similarity between any two scale convolutional feature tensors is calculated as the attention intensity, and an attention matrix is ​​constructed.

[0087] The attention matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the obtained eigenvalues ​​are used as weight vectors to weight the convolutional feature tensors of multiple scales to obtain the periodic enhancement feature tensor.

[0088] The periodic augmentation feature tensor is input into the time dimension compression layer, and the subsequences of the periodic augmentation feature tensor in different time periods are extracted segment by segment using the set time sliding window size and step size.

[0089] Perform average pooling on each subsequence, and concatenate the pooling results of each time sliding window in time order to generate a time-dimensional compressed embedding representation tensor sequence;

[0090] The embedded representation tensor sequence is input into the prediction head structure, and the output is the sequence of initial load prediction values ​​for the corresponding hoist within a preset future time range.

[0091] This step innovatively improves the accuracy of gondola load prediction by introducing the TimesNet network and combining multi-level feature extraction with a periodicity-aware attention mechanism. The input embedding layer maps various operational data into initial embedding tensors, followed by the extraction of load features at different time scales through a convolutional perceptual layer. A periodicity-aware attention layer then weights the features at each scale, enhancing the ability to capture periodic patterns. The time-dimensional compression layer pools features through a sliding window, further compressing the information dimensionality and optimizing computational efficiency. The prediction head structure generates accurate load prediction sequences, significantly improving the prediction accuracy and real-time performance of gondola loads, thus contributing to the safety and stability of the cableway system.

[0092] In this embodiment, the processing steps for the prediction head structure are specifically as follows:

[0093] Constructing a temporal causal relationship graph based on the embedded representation tensor sequence, the construction steps are as follows:

[0094] The embedding representation tensor of each time step is used as a node in the temporal causal relationship graph, and all nodes are arranged in chronological order.

[0095] Calculate the Euclidean distance between any two nodes. If it is less than a preset threshold, it is considered that there is a causal relationship between the corresponding nodes. Establish a directed edge from the previous time step to the next time step, and the weight of the edge is the corresponding Euclidean distance.

[0096] For each node in the target prediction time step, find all historical nodes in the temporal causal relationship graph, where the historical nodes are the nodes corresponding to the incoming edges;

[0097] For the embedding representation tensors represented by all historical nodes, linear weighted aggregation is performed according to the weights of the incoming edges to obtain the aggregated feature representation of the target prediction time step.

[0098] Assign a fixed feature weight to each feature dimension in the aggregated feature representation;

[0099] Multiply each feature dimension by its corresponding feature weight, and sum all the product results to obtain the total weighted value;

[0100] The sum of the weighted values ​​is added to a preset bias value to obtain the initial load prediction value for the target prediction time step. The bias value is used to adjust the overall offset of the initial load prediction value.

[0101] Arrange the initial load forecast values ​​of all target prediction time steps in chronological order to obtain the sequence of initial load forecast values ​​for the corresponding hoistway within the future preset time range.

[0102] This step transforms the gondola load prediction problem into a time-series graph modeling and feature weighting task by constructing a time-series causal relationship graph. By calculating Euclidean distance and determining causal relationships based on preset thresholds, directed edges between nodes are established, capturing the impact of historical nodes on the target prediction time step. By linearly weighting and aggregating historical node features and introducing fixed feature weights and bias adjustments, the adaptability to dynamic changes in the time series is improved. This enhances causal inference capabilities, effectively improving the accuracy and timeliness of load prediction, providing precise data support for the load management of the cableway system, and ensuring the safe and efficient operation of the system.

[0103] In this embodiment, step three specifically includes:

[0104] Calculate the numerical difference between adjacent time points in the initial load prediction sequence for each hoist, forming the corresponding hoist rate of change sequence;

[0105] Calculate the global mean and standard deviation of the rate of change series, and set a preset threshold coefficient;

[0106] All time points in the rate of change sequence that are greater than the global mean plus a threshold coefficient multiple of the standard deviation are marked as local anchor points, forming the corresponding anchor point time set for the gondola.

[0107] The anchor point time sets of all gondolas are merged to form the total anchor point set;

[0108] Based on the total set of anchor points, a cross-cabin anchor point diagram is constructed. The construction steps are as follows:

[0109] Each local anchor point is represented as an anchor node in the cross-cab anchor point diagram;

[0110] For any two anchor nodes in the cross-cabin anchor point diagram, if the time difference between the corresponding time points is less than the preset time window length, then an undirected edge is established between the two anchor nodes.

[0111] The reciprocal of the time difference between two anchor points is used as the edge weight of the corresponding undirected edge, which represents the degree of time proximity between the corresponding anchor points.

[0112] This step accurately identifies key moments affecting load fluctuations by calculating the rate of change sequence of predicted gondola load values ​​and combining the global mean and standard deviation. By setting threshold coefficients, marking local anchor points, and merging them into a total anchor point set, a cross-gondola anchor point graph is constructed, effectively capturing the temporal correlation between gondolas. The construction and edge weight assignment of the cross-gondola anchor point graph reflect a quantitative analysis of the temporal proximity between anchor points, helping to improve the prediction model's understanding and modeling accuracy of the correlation between gondolas. This method provides an efficient feature construction approach for multi-gondola dynamic load prediction, contributing to more accurate and real-time load prediction and ensuring the safe operation and load balance of the cableway system.

[0113] In this embodiment, step four specifically includes:

[0114] A two-dimensional gondola location index matrix is ​​constructed based on the gondola's relative geographical location index, and the location index matrix represents the relative coordinate relationship of the gondola in the spatial arrangement;

[0115] For any pair of anchor points connected by an edge in the cross-cabin anchor point diagram, extract the corresponding time difference and the cabin index difference. The cabin index difference is the spatial distance between the corresponding relative coordinates in the two-dimensional cabin position index matrix.

[0116] For all anchor point pairs, the product of the corresponding time difference and the pod index difference is calculated as the disturbance intensity of the corresponding anchor point pair.

[0117] An anchor point perturbation matrix is ​​established using the index of the anchor point node in the total anchor point set as the row and column coordinates. Each element in the anchor point perturbation matrix is ​​the perturbation intensity of the corresponding anchor point pair, and the elements corresponding to unconnected anchor points are set to zero.

[0118] The anchor point perturbation matrix is ​​multiplied by the initial load prediction value corresponding to the pod anchor point, and the dimensions are reconstructed to generate a perturbation embedding tensor, which is used to represent the cross-time perturbation correlation between pods.

[0119] This step enhances the modeling of spatial relationships between gondolas by constructing a two-dimensional gondola location index matrix and utilizing information about the relative geographical locations of the gondolas. By calculating the product of the time difference and the gondola index difference, a disturbance matrix reflecting the intensity of disturbances between gondolas is obtained. The generation of this disturbance matrix not only considers the spatial relationships between gondolas but also effectively integrates the effects of temporal and spatial disturbances. Multiplying the disturbance matrix with the initial load prediction value generates a disturbance embedding tensor, capturing the cross-temporal disturbance correlation between gondolas and further optimizing prediction accuracy. This process provides a more detailed and comprehensive disturbance analysis method for multi-gond dynamic load prediction, improving the accuracy and robustness of load prediction.

[0120] In this embodiment, step five specifically includes:

[0121] Align the embedding representation tensor of each gondola with the corresponding perturbation embedding tensor in the time dimension;

[0122] Perform the Kronecker product operation on the embedding representation tensor and the perturbation embedding tensor corresponding to each time step to obtain the perturbation reconstructed feature representation tensor for the corresponding time step;

[0123] By concatenating the perturbation reconstruction feature representation tensors corresponding to all time steps in chronological order, a sequence of perturbation reconstruction feature representation tensors is obtained.

[0124] The perturbation-reconstructed feature representation tensor sequence is input into the prediction head structure to construct the perturbation-reconstructed temporal causal relationship graph;

[0125] Based on the time-series causal relationship graph after disturbance reconstruction, historical nodes are weighted and aggregated to obtain the load prediction value after disturbance reconstruction.

[0126] Arrange the perturbation-reconstructed load prediction values ​​of all target prediction time steps in chronological order to generate a corresponding sequence of perturbation-reconstructed load prediction values ​​for the gantry within a preset future time range.

[0127] This step successfully reconstructed cross-temporal features in gondola load prediction by aligning the gondola embedding representation tensor and the perturbation embedding tensor in the time dimension and combining them with the Kronecker product operation. By perturbing and reconstructing the features at each time step, the modeling capability of the temporal causal relationship graph was further enhanced, capturing the temporal relationships and perturbation effects between gondolas. Using weighted aggregation of historical nodes, load changes in future time steps were accurately predicted. This provides a high-precision, low-latency prediction mechanism for gondola load prediction, effectively improving the prediction capability and scheduling efficiency of the cableway system in dynamic environments, ensuring operational safety and load balancing.

[0128] In this embodiment, step six specifically includes:

[0129] The load prediction sequence after each gondola disturbance reconstruction is aligned with the corresponding real load sequence point by point on the time step, and the squared difference between the prediction value and the real value at each time step is calculated. The real load sequence is obtained by the weighing sensor installed at the bottom of the gondola during actual operation at time intervals.

[0130] The summation of the squared differences of all time steps yields the numerical error of the corresponding hoist within the preset time range;

[0131] A Soft-DTW alignment operation is performed between the perturbation-reconstructed load prediction sequence and the actual load sequence. The Soft-DTW alignment operation includes the following steps:

[0132] A time-aligned distance matrix is ​​constructed based on the load prediction sequence and the actual load sequence. Each element in the time-aligned distance matrix represents the Manhattan distance between the predicted value and the actual value between two time points.

[0133] Based on the time alignment distance matrix, the path with the minimum total cost is searched through dynamic programming and selected as the optimal time alignment path.

[0134] The sum of costs along the optimal time alignment path is used as the trend error, which represents the degree of deviation of the load prediction sequence from the actual load sequence.

[0135] The numerical error and trend error are weighted according to the set weight ratio to obtain the comprehensive error.

[0136] If the overall error is greater than the preset error threshold, the difference between the overall error and the error threshold will be subtracted from the bias value in the prediction head structure until the overall error converges to below the error threshold, and finally the load prediction result with controllable error will be output.

[0137] This step reconstructs the load prediction and actual load values ​​by aligning the gondola's disturbances point-by-point, calculating and summing the squared differences, thus accurately assessing the numerical error at each time step. Soft-DTW alignment further optimizes the load prediction accuracy by constructing a time alignment distance matrix and using a dynamic programming algorithm to minimize the time deviation between the predicted and actual load values. Combining the weighted calculation of numerical and trend errors effectively adjusts the model bias, ensuring accurate and controllable load prediction results. This process significantly improves the prediction stability and reliability of the cableway system in complex environments, ensuring the safe and efficient operation of the system under various working conditions.

[0138] Example 1:

[0139] To verify the feasibility of this invention in practice, it was applied to an aerial passenger cableway system in a tourist attraction. The cableway system has 10 gondolas, each capable of carrying a maximum weight of 1000 kg, and approximately 500 tourists use the cableway per hour. To ensure safety and optimize operational efficiency, the attraction's management decided to use the method of this invention to predict the dynamic load of the gondolas in real time and monitor their operating status, especially load fluctuations during peak hours.

[0140] In this practical scenario, load data for each gondola is first collected in real time by weighing sensors installed at the bottom of the gondolas. This data includes the gondola's real-time load weight, weather data, date and time data, passenger flow data, and the gondola's relative geographical location index. This data is transmitted to the backend system at time intervals. The real-time load data for each gondola is used as input data, preprocessed, and then fed into a TimesNet deep neural network for dynamic load prediction.

[0141] Through network training and optimization, this method can calculate the load prediction sequence for each gondola in real time, and further improve the accuracy of the prediction by combining cross-gondola disturbance analysis. During peak hours, gondola loads fluctuate significantly, and excessive load may lead to gondola system malfunctions or safety hazards. Therefore, accurate load prediction can help scenic area management departments adjust their operational strategies in real time. For example, when the predicted load of a gondola exceeds a safety threshold, the system will immediately issue an alarm, reminding staff to inspect the gondola or take adjustment measures.

[0142] During data processing, by utilizing preset thresholds and perturbation reconstruction mechanisms within the model, the system can identify sudden load fluctuations caused by external factors such as surges in passenger traffic and weather changes. Under the influence of severe weather (heavy rain and strong winds), the predicted load value of the gondola fluctuates. Through this method, the system can detect load change trends in advance and provide early warnings of potential safety hazards, ensuring the normal operation of the gondola.

[0143] Through comparative analysis, we were able to find that this method significantly improves the accuracy of predictions in practical applications. The following table shows the differences in performance between this method and traditional load forecasting methods (such as prediction methods based on simple linear regression) in multiple experimental scenarios.

[0144] Table 1. Performance Comparison of the Method of the Present Invention with Different Load Prediction Methods

[0145]

[0146] As shown in Table 1, the method of this invention outperforms other traditional methods in all test scenarios. Especially during peak hours and in severe weather, the error of the method of this invention is significantly lower than that of traditional methods. For example, during peak hours, the error of the traditional linear regression method is 11.5 kg, while the error of the method of this invention is only 3.5 kg, a reduction of 69%. In severe weather, the error of the method of this invention is 3.0 kg, a reduction of 69% compared to the traditional method. During periods of high load fluctuation, the prediction error of the method of this invention is only 4.9 kg, compared to 13.2 kg for the linear regression method, a reduction of 63%. This demonstrates that the deep neural network method can effectively cope with complex dynamic load changes and accurately capture the periodicity and sudden fluctuations of the load on the hoisting vehicle.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the dynamic load of aerial passenger cableway gondolas based on deep neural networks, characterized in that, Includes the following steps: Step 1: Collect and preprocess the operating data of multiple gondolas in the aerial passenger cableway; Step 2: Input the operating data of each gondola into the TimesNet network to obtain the initial load prediction sequence for each gondola; Step 3: Calculate the numerical difference between adjacent time points in the initial load prediction sequence for each gondola, mark local anchor points, and construct a cross-gond anchor point map, specifically: Calculate the numerical difference between adjacent time points in the initial load prediction sequence for each hoist, forming the corresponding hoist rate of change sequence; Calculate the global mean and standard deviation of the rate of change series, and set a preset threshold coefficient; All time points in the rate of change sequence that are greater than the global mean plus a threshold coefficient multiple of the standard deviation are marked as local anchor points, forming the corresponding anchor point time set for the gondola. The anchor point time sets of all gondolas are merged to form the total anchor point set; Based on the total set of anchor points, a cross-cabin anchor point diagram is constructed. The construction steps are as follows: Each local anchor point is represented as an anchor node in the cross-cab anchor point diagram; For any two anchor nodes in the cross-cabin anchor point diagram, if the time difference between the corresponding time points is less than the preset time window length, then an undirected edge is established between the two anchor nodes. The reciprocal of the time difference between two anchor points is expressed as the edge weight of the corresponding undirected edge, and the edge weight is used to represent the degree of time proximity between the corresponding anchor points; Step 4: Construct a two-dimensional gondola location index matrix based on the relative geographical location index of the gondolas, and calculate the disturbance intensity of anchor point pairs in the cross-gondola anchor point map to construct the anchor point disturbance matrix, generating the disturbance embedding tensor, specifically: A two-dimensional gondola location index matrix is ​​constructed based on the gondola's relative geographical location index, and the location index matrix represents the relative coordinate relationship of the gondola in the spatial arrangement; For any pair of anchor points connected by an edge in the cross-cabin anchor point diagram, extract the corresponding time difference and the cabin index difference. The cabin index difference is the spatial distance between the corresponding relative coordinates in the two-dimensional cabin position index matrix. For all anchor point pairs, the product of the corresponding time difference and the pod index difference is calculated as the disturbance intensity of the corresponding anchor point pair; An anchor point perturbation matrix is ​​established using the index of the anchor point node in the total anchor point set as the row and column coordinates. Each element in the anchor point perturbation matrix is ​​the perturbation intensity of the corresponding anchor point pair, and the corresponding elements of unconnected anchor point pairs are set to zero. The anchor point perturbation matrix is ​​multiplied by the initial load prediction value corresponding to the pod anchor point, and the dimension is reconstructed to generate a perturbation embedding tensor. The perturbation embedding tensor is used to represent the cross-time perturbation correlation between pods. Step 5: Perform Kronecker product operation on each embedding representation tensor of the hoistway with the perturbation embedding tensor to generate a perturbation reconstruction feature representation tensor sequence, and input it into the prediction head structure to obtain the perturbation reconstruction load prediction value sequence. Step 6: Calculate the numerical error between the predicted load sequence and the actual load sequence, obtain the trend error through Soft-DTW alignment, and update the bias value based on the numerical error and the trend error.

2. The method for dynamic load prediction of aerial passenger cableway gondolas based on deep neural networks according to claim 1, characterized in that, The operational data includes real-time load weight data, meteorological data, date and time data, passenger flow data, and relative geographical location index of each gondola. The relative geographical location index of the gondola is a mapping relationship between the gondola number and the gondola's spatial coordinates. The preprocessing includes performing missing value imputation, time synchronization processing, outlier removal, and normalization processing on different types of data to generate operational data with a unified structure.

3. The method for dynamic load prediction of aerial passenger cableway gondolas based on deep neural networks according to claim 1, characterized in that, Step two specifically involves: The TimesNet network includes an input embedding layer, a convolutional perceptual layer, a periodic perceptual attention layer, a temporal dimension compression layer, and a prediction head structure. The input embedding layer is used to receive the running data of each gondola, perform linear mapping on the running data, and compress the feature dimension to obtain the initial embedding tensor; The initial embedding tensor is input to the convolutional perceptual layer, which consists of several parallel convolutional channels. Each convolutional channel has a different convolutional kernel size and is used to extract feature patterns at different time scales. The feature patterns are the changes in the load weight of the gondola at different time steps. Each convolutional channel outputs a scale convolutional feature tensor. Multiple scale convolutional feature tensors are input into the periodic awareness attention layer. The cosine similarity between any two scale convolutional feature tensors is calculated as the attention intensity, and an attention matrix is ​​constructed. The attention matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the obtained eigenvalues ​​are used as weight vectors to weight the convolutional feature tensors of multiple scales to obtain the periodic enhancement feature tensor. The periodic augmentation feature tensor is input into the time dimension compression layer, and the subsequences of the periodic augmentation feature tensor in different time periods are extracted segment by segment using the set time sliding window size and step size. Perform average pooling on each subsequence, and concatenate the pooling results of each time sliding window in time order to generate a time-dimensional compressed embedding representation tensor sequence; The embedded representation tensor sequence is input into the prediction head structure, and the output is the sequence of initial load prediction values ​​for the corresponding hoist within a preset future time range.

4. The method for dynamic load prediction of aerial passenger cableway gondolas based on deep neural networks according to claim 3, characterized in that, The specific processing steps for the prediction head structure are as follows: Constructing a temporal causal relationship graph based on the embedded representation tensor sequence, the construction steps are as follows: The embedding representation tensor of each time step is used as a node in the temporal causal relationship graph, and all nodes are arranged in chronological order. Calculate the Euclidean distance between any two nodes. If it is less than a preset threshold, it is considered that there is a causal relationship between the corresponding nodes. Establish a directed edge from the previous time step to the next time step, and the weight of the edge is the corresponding Euclidean distance. For each node in the target prediction time step, find all historical nodes in the temporal causal relationship graph, where the historical nodes are the nodes corresponding to the incoming edges; For the embedding representation tensors represented by all historical nodes, linear weighted aggregation is performed according to the weights of the incoming edges to obtain the aggregated feature representation of the target prediction time step. Assign a fixed feature weight to each feature dimension in the aggregated feature representation; Multiply each feature dimension by its corresponding feature weight, and sum all the product results to obtain the total weighted value; The sum of the weighted values ​​is added to a preset bias value to obtain the initial load prediction value for the target prediction time step. The bias value is used to adjust the overall offset of the initial load prediction value. Arrange the initial load forecast values ​​of all target prediction time steps in chronological order to obtain the sequence of initial load forecast values ​​for the corresponding hoistway within the future preset time range.

5. The method for dynamic load prediction of aerial passenger cableway gondolas based on deep neural networks according to claim 1, characterized in that, Step five specifically involves: Align the embedding representation tensor of each gondola with the corresponding perturbation embedding tensor in the time dimension; Perform the Kronecker product operation on the embedding representation tensor and the perturbation embedding tensor corresponding to each time step to obtain the perturbation reconstructed feature representation tensor for the corresponding time step; By concatenating the perturbation reconstruction feature representation tensors corresponding to all time steps in chronological order, a sequence of perturbation reconstruction feature representation tensors is obtained. The perturbation-reconstructed feature representation tensor sequence is input into the prediction head structure to construct the perturbation-reconstructed temporal causal relationship graph; Based on the time-series causal relationship graph after disturbance reconstruction, historical nodes are weighted and aggregated to obtain the load prediction value after disturbance reconstruction. Arrange the perturbation-reconstructed load prediction values ​​of all target prediction time steps in chronological order to generate a corresponding sequence of perturbation-reconstructed load prediction values ​​for the gantry within a preset future time range.

6. The method for dynamic load prediction of aerial passenger cableway gondolas based on deep neural networks according to claim 1, characterized in that, Step six specifically includes: The load prediction sequence after each gondola disturbance reconstruction is aligned with the corresponding real load sequence point by point on the time step, and the squared difference between the prediction value and the real value at each time step is calculated. The real load sequence is obtained by the weighing sensor installed at the bottom of the gondola during actual operation at time intervals. The summation of the squared differences of all time steps yields the numerical error of the corresponding hoist within the preset time range; A Soft-DTW alignment operation is performed between the perturbation-reconstructed load prediction sequence and the actual load sequence. The Soft-DTW alignment operation includes the following steps: A time-aligned distance matrix is ​​constructed based on the load prediction sequence and the actual load sequence. Each element in the time-aligned distance matrix represents the Manhattan distance between the predicted value and the actual value between two time points. Based on the time alignment distance matrix, the path with the minimum total cost is searched through dynamic programming and selected as the optimal time alignment path. The sum of costs along the optimal time alignment path is used as the trend error, which represents the degree of deviation of the load prediction sequence from the actual load sequence. Based on the set weight ratio of numerical error and trend error, the numerical error and trend error are weighted to obtain the comprehensive error; If the overall error is greater than the preset error threshold, the difference between the overall error and the error threshold will be subtracted from the bias value in the prediction head structure until the overall error converges to below the error threshold, and finally the load prediction result with controllable error will be output.

Images