Elevator passenger flow prediction method based on multi-dimensional time sequence memory weight network

Abstract

An elevator passenger flow prediction method based on a multi-dimensional time sequence memory weight network comprises the following steps that firstly, a multi-dimensional passenger flow time sequence data set is constructed, and the process comprises the following steps that (1.1) the data set is constructed; (1.2) collecting feature information; (1.3) preprocessing the feature data; 2, dividing an original data set by using a dynamic noise reduction clustering algorithm, wherein the process is as follows: step (2.1), carrying out dimensionality reduction on a two-dimensional daily average passenger flow curve; (2.2) creating an initial passenger flow clustering center point; (2.3) creating a feature cluster set and distributing data points; (2.4) a central point updating mechanism; and a third step of using a multi-dimensional time sequence memory weight network to realize feature learning extraction and passenger flow prediction of the reconstructed feature data set, wherein the process is as follows: step (3.1), constructing a multi-dimensional time sequence convolutional layer; and step (3.2), constructing a memory weight network. According to the invention, the passenger flow prediction precision is obviously improved.

Description

Technical Field

[0001] The present invention belongs to the fields of intelligent building systems, computer science, and traffic flow prediction, and relates to an elevator passenger flow prediction method based on a multi-dimensional time series memory weight network, which is applied to short-term passenger flow prediction for elevator group control scheduling in high-rise buildings. Background Art

[0002] In recent years, with the rapid development of intelligent building and Internet of Things technologies, the elevator group control system (EGCS) of high-rise buildings has become the mainstream vertical traffic solution. The operating efficiency and user experience of EGCS highly depend on elevator passenger flow prediction. This technology can help the elevator system achieve efficient resource allocation, reduce energy consumption, and improve the response speed during peak hours by predicting the changing trend of passenger flow, ensuring smooth travel for users.

[0003] To improve the accuracy of passenger flow prediction, researchers have proposed various data-driven modeling methods. Currently, the technologies for elevator passenger flow prediction can be roughly divided into three categories: parametric models, non-parametric models, and hybrid models. Parametric models such as ARIMA (Autoregressive Integrated Moving Average) model, SARIMA (Seasonal Autoregressive Integrated Moving Average) model, etc., have good fitting capabilities and perform excellently in short-term trend prediction with their simple linear assumptions, but it is difficult to capture the complex time series characteristics of passenger flow; non-parametric models such as Bidirectional Long Short-Term Memory Network (BiLSTM), Temporal Convolutional Network (TCN), etc., can automatically learn data features without data distribution assumptions, but it is difficult to handle the complex correlations of multi-dimensional data. Hybrid models attempt to combine different methods, such as CNN-GRU, STAFN, etc., which combine multiple network structures to extract complex feature information and can improve the prediction effect, and are currently the research hotspots. For high-rise building scenarios such as office buildings, passenger flow data has significant periodicity and time series dependence, and is significantly affected by various factors such as weather, time, and floor usage. Existing models are difficult to fully exploit such complex features, resulting in limited accuracy of passenger flow prediction, thus affecting the scheduling performance of the elevator group control system and user experience.

[0004] In response to the challenges of elevator passenger flow prediction, many scholars have tried to improve by introducing a method that combines deep learning and clustering algorithms. Yuxin He et al. proposed the MGC-RNN network based on multi-graph encoding and recurrent neural network (RNN), which can effectively perform short-term prediction of passenger flow at station distributions. Yun Jing et al. proposed the LSTM-LGB-DRS network based on LSTM and GBDT decision tree method, which improves the prediction accuracy of passenger flow and the fitting accuracy of passenger flow peaks while reducing costs. Zhang et al. proposed a model combining convolutional neural network and long short-term memory network to improve the prediction accuracy of the elevator scheduling system. Li Feng et al. designed a prediction model based on temporal convolutional network to further explore temporal information.

[0005] However, the existing methods still have deficiencies in terms of processing efficiency and prediction accuracy when dealing with multi-dimensional feature passenger flow data, especially in predicting the dynamic characteristics and complex temporal patterns of elevator traffic flow. Currently, elevator passenger flow prediction faces the following main challenges: (1) How to accurately extract and analyze complex multi-dimensional temporal features to improve the timeliness and accuracy of passenger flow prediction while ensuring high computational efficiency is the core issue in practice; (2) For high-rise building passenger flow data that is significantly affected by fluctuations in external environments (such as weather, time), the existing methods lack effective clustering strategies and prediction model structures with strong adaptability. Therefore, there is an urgent need for an innovative model that can efficiently process multi-dimensional temporal data and significantly improve prediction accuracy. Summary of the Invention

[0006] In order to overcome the shortcomings of the existing technology, the existing models are unable to fully explore the complex features in multidimensional time series data, have low efficiency in processing heterogeneous data and insufficient prediction accuracy when processing elevator passenger flow prediction. This paper proposes an elevator passenger flow prediction method based on a multidimensional temporal memory weight network, which combines the multidimensional time series prediction method NRDC-MTMWN (Noise Reduction Dynamic Clustering-Multi-dimensional Temporal Memory Weight Network) with dynamic noise reduction clustering. First, the elevator data set is reconstructed, and the elevator passenger flow data is reduced in dimension through feature abstraction and clustered by a noise reduction dynamic clustering (NRDC) algorithm; secondly, the divided passenger flow data set is processed into a format suitable for the prediction model input through a sliding window algorithm and input into a multidimensional temporal convolution layer (MTC). The MTC layer can extract local temporal passenger flow features and reduce the number of model parameters, while also increasing the field of view without adding additional parameters or increasing the burden of model training; it is then input into a bidirectional long short-term memory network (BiLSTM) for training, and the output is then weighted by a multi-head adaptive attention mechanism (MHAAM) layer, and finally the passenger flow prediction result is obtained by a fully connected layer. The present invention specifically learns and extracts complex passenger flow feature information in high-rise buildings, and ultimately achieves accurate passenger flow prediction, thereby providing a reliable input basis for the EGCS algorithm and ensuring the operating efficiency of the overall system.

[0007] The technical solution adopted by the present invention to solve its technical problem is:

[0008] A method for predicting elevator passenger flow based on a multi-dimensional temporal memory weight network comprises the following steps:

[0009] The first step is to build a multi-dimensional passenger flow time series dataset. The process is as follows:

[0010] Step (1.1) Dataset construction;

[0011] Step (1.2) feature information collection;

[0012] Step (1.3) feature data preprocessing;

[0013] Step 2: Divide the original dataset using the dynamic noise reduction clustering algorithm. The process is as follows:

[0014] Step (2.1) Reduce the dimension of the two-dimensional daily average passenger flow curve;

[0015] Step (2.2) Create the initial passenger flow clustering center points;

[0016] Step (2.3) Create the feature cluster set and allocate data points;

[0017] Step (2.4) Center point update mechanism;

[0018] Step 3: Use the multi-dimensional time series memory weight network to achieve feature learning extraction and passenger flow prediction for the reconstructed feature dataset. The process is as follows:

[0019] Step (3.1) Construct the multi-dimensional time series convolutional layer;

[0020] Step (3.2) Construct the memory weight network.

[0021] Furthermore, in the said step (1.1), the multi-dimensional passenger flow time series dataset matrix X is defined as follows,

[0022]

[0023] where, x t,d is the j-th feature information value corresponding to the i-th time node, T s is the total number of time steps of the dataset, and D is the total feature dimension of the dataset.

[0024] Still further, the process of the said step (1.2) is:

[0025] Step (1.2.1) Collect the real-time operation parameters of the elevator system, including the waiting passenger flow quantities in the hall and between each floor, the running time of each elevator during the sampling time, the load condition during elevator operation, the number of passengers in the car, and the elevator energy consumption data;

[0026] The elevator control system records the start and stop times of each operation. A weighing sensor is installed in the elevator car, and the sensor data communicates with the central server through the RS485 and Modbus protocols to realize the real-time update of the database; An energy consumption monitoring module is installed in the equipment control cabinet, and cameras are installed in the waiting area outside the car and inside the car. The collected data can update the central server database in real time through the TCP / IP protocol;

[0027] After authentication, data can be directly accessed by calling through the remote data interface to obtain the data;

[0028] Among them, the authentication protocol (OAuth) is an open standard protocol that allows users to securely access protected resources through authorization tokens, ensuring data access rights and security; the interface call (API) is an interface that allows data interaction between different systems, enabling external applications to remotely access databases or services through predefined request and response formats.

[0029] Step (1.2.2) collects environmental information and integrates auxiliary feature information, including encoding the timing information at the current moment, outdoor temperature, rain and snow conditions, special weather annotations, and floor use information.

[0030] Initiate an HTTP request through the interface provided by the meteorological software to obtain meteorological data in JSON format, and incorporate the parsed temperature, rain and snow conditions, and special weather annotation information into the feature dataset; at the same time, encode the system timestamp information obtained during data sampling and integrate it into the feature dataset.

[0031] HTTP is the HyperText Transfer Protocol, which is used to transfer data between the client and the server and is the basic protocol for Internet data exchange; through an HTTP request, the client can obtain resources from the server, such as web pages or API data; JSON is a lightweight data exchange format that uses a key-value pair structure, facilitating data storage and transmission, and is widely used in API data responses and is easy to be parsed and used by programming languages.

[0032] Furthermore, the process of step (1.3) is as follows:

[0033] Step (1.3.1) performs data cleaning, removes duplicate records in the dataset, and replaces invalid or obviously incorrect data points with missing values.

[0034] Step (1.3.2) performs interpolation processing, and fills the missing values generated due to the replacement of invalid values or the lost sampling values in the dataset using the linear interpolation method.

[0035] The linear interpolation method is a method that uses the numerical values of adjacent known data points to estimate the numerical value of an unknown point in a linear manner to smoothly fill the gaps in the dataset.

[0036] Step (1.3.3) performs data standardization, standardizes the numerical data, and scales the data to a fixed range to eliminate the influence of the dimension.

[0037] Performing data standardization means normalizing all numerical features in the multi-dimensional passenger flow time series dataset X constructed in (1.1), so that each feature value x t,d is within a fixed range of [0, 1] to enhance the generalization ability of subsequent model training.

[0038] The process of step (2.1) is as follows:

[0039] In step (2.1.1), the passenger flow data is sorted and organized daily. The passenger flow data for one day can be expressed in the form P x ={x 1,D , x 2,D ,..., x T1,D}, where x 1,D is a vector of one row and D columns, T 1 is the last time step in a day, and D represents the feature dimension;

[0040] In step (2.1.2), a new coordinate mapping is established by analyzing features: the horizontal axis is the time when the first effective maximum passenger flow peak exceeding the minimum peak during the historical morning rush hour appears, and the vertical axis is the corresponding passenger flow peak intensity to construct a coordinate system;

[0041] In step (2.1.3), the daily average curve P x is traversed in a loop and points are selected using an edge detection algorithm. The outer loop traverses N times for the total number of days, and the inner loop traverses T times in total 1 ;

[0042] After the loop in step (2.1.4) ends, a set D of discrete expressions of passenger flow corresponding to N days will be obtained s , and the Noise Factor Recognition (NFR) method is added: calculate the local noise value of each point and divide it by the average distance of its k nearest neighbors. If the result NFR is greater than 1, the point is considered a noise point and is removed.

[0043] The process of step (2.2) is as follows:

[0044] In step (2.2.1), a set C of center points is created, and the local density ρ i ,

[0045]

[0046] where σ i is the dynamic value for point i, determined by the average distance from neighboring points to it, k is the number of neighboring points participating in the operation, and d ij is the distance between point i and point j;

[0047] In step (2.2.2), the distance δ i of each data point is calculated. The distance δ i is defined as the distance between point i and the nearest point with a density higher than it,

[0048]

[0049] Step (2.2.3) Select the first k points with the largest ρ×δ values as the initial cluster centers C = C 1 , C 2 ,..., C k . Each center point represents a typical passenger flow pattern in a day. Let the iteration number iter = 0.

[0050] The process of the said step (2.3) is as follows:

[0051] Step (2.3.1) Create k empty cluster sets S = S 1 , S 2 ,..., S k . Each cluster will contain passenger flow patterns similar to a certain center point;

[0052] Step (2.3.2) Take each passenger flow point x′ s in the total elevator passenger flow data set D j , and calculate the distance from x' j to each passenger flow center point C i . Find the nearest center point,

[0053]

[0054] where, represents finding the i that minimizes the objective function, and dist(x′ j , C i ) represents the Euclidean distance between each passenger flow point x′ j and the passenger flow center point C i . Through the measurement calculation of this formula, the passenger flow point x' j can be assigned to the cluster where the nearest center point is located

[0055] Step (2.3.3) Select the similar passenger flow clusters and select the point closest to the center point C i as the new center point, that is, select the point with the average number of passengers closest to the calculated center value in the cluster as the new center point, and increase the iteration number by one.

[0056] The process of the said step (2.4) is as follows:

[0057] Step (2.4.1) In each iteration process, an adaptive detection mechanism is introduced. According to the average distance from the points in each cluster to the center point, the tightness of the cluster is judged,

[0058]

[0059] where, d(x i , C i ) is the distance from the point xi to the center point C i the distance of, denotes the number of data points in the cluster ;

[0060] In step (2.4.2), if the compactness of a certain cluster is low, indicating that the points in the cluster are more dispersed, the center point of the cluster is reselected through a dynamic adjustment mechanism;

[0061] In step (2.4.3), when the change in the position of the center point is less than the set threshold or the maximum number of iterations is reached, the iteration stops.

[0062] The process of the said step (3.1) is as follows:

[0063] In step (3.1.1), the MTC convolutional layer first performs convolution on the multi-dimensional features along the time dimension by using one-dimensional convolution, so as to extract all local features without omission,

[0064]

[0065] where is the weight vector of the f-th filter at time lag k, b f is the bias term of the f-th filter, t is the current time step, and the final output is a matrix composed of the outputs of all filters at each time step, where F is the number of filters;

[0066] In step (3.1.2), the one-dimensional convolution kernel is enlarged, and the corresponding calculation process after adding the properties of causal convolution is expressed as,

[0067]

[0068] where is the output of the first layer at time t, f is the ReLU activation function, W (1) is the convolution kernel weight of the first layer, x [t,t-1,t-2] is the slice of the input data at time t and the two time steps before it, b (1) is the bias term of the first layer, and "·" represents the dot product of vectors;

[0069] In step (3.1.3), the output is passed to the field-of-view expansion module (TAtrous_Block). After the modification, the convolution kernel is no longer continuous, and the corresponding convolution process is expressed as,

[0070]

[0071] where is the output of the i-th layer at time t, f is the ReLU activation function, W (i) is the convolution kernel weight of the i-th layer, Data at time t and two time steps before it output from the (i - 1)-th layer, b (i) is the bias term of the i-th layer;

[0072] TAtrous_Block consists of dilated convolutional layers with different numbers of kernels and dilation factors in the i-th layer, which can increase the receptive field while reducing the number of model parameters and without adding a heavy burden to model training, and fully extract local temporal passenger flow features.

[0073] The process of step (3.2) is as follows:

[0074] Step (3.2.1) introduces BiLSTM to capture the forward and backward dependencies of the time series. For the input sequence Y, the output of BiLSTM is where L is the number of LSTM units. The calculation formula is expressed as,

[0075]

[0076]

[0077] In the formula and H t are the forward hidden state, backward hidden state, and combined hidden state in BiLSTM at time t respectively, f is the activation function, W x represents the weight matrix from the input to the hidden state, Y t represents the input data at time t, W h is the weight matrix from the hidden state at the previous time step to the current hidden state, and b is the bias term;

[0078] BiLSTM is a recurrent neural network that can capture both the forward and backward dependencies of the input. H represents the output matrix of this layer over the entire time series, and each row corresponds to the combined hidden state H t .

[0079] Step (3.2.2) uses the multi-head adaptive attention mechanism to enhance the model's attention to key time steps. First, three different linear transformations are used to generate the corresponding query Q, key K, and value V matrices,

[0080] Q = H · W Q ;

[0081] K = H · W K ;

[0082] V = H · W V ;

[0083] In the formula, W Q , W K , WV is a learnable weight matrix;

[0084] Step (3.2.3) introduces an adaptive factor α based on the extracted features, multi-dimensional passenger flow feature information, and the feature of the current time step t , which is used to adjust the attention weights at each time step,

[0085] state t = [passenger t , load t ,..., weather t ;

[0086] α t = σ(W α · [H, state t + b α );

[0087] In the formula, W α and b α are learnable parameters, state t is the elevator passenger flow feature set, where passenger t is the number of passengers waiting for the elevator at time t, load t is the elevator operation load condition at time t, weather t is the special weather annotation at time t, and σ is the activation function;

[0088] Step (3.2.4) calculates the attention weights for each head i,

[0089]

[0090] In the formula, d k is the dimension of the key vector, QK T is the dot product of the query and the key, and the softmax function is used to normalize the similarity to a probability distribution;

[0091] Step (3.2.5) concatenates the outputs of all heads and generates the final multi-head adaptive attention output through a linear transformation,

[0092] Context = Concat(Att 1 , Att 2 ,..., Att h ) · W O ;

[0093] In the formula, W O is the linear transformation weight matrix;

[0094] Step (3.2.6) maps the multi-head adaptive attention output to the prediction result. The prediction layer adopts a fully connected layer structure, with Context as the input of the prediction layer. By performing a linear transformation on the features contained in Context, the high-dimensional features in Context are compressed into scalar prediction values, thus achieving the function of mapping the attention output to the prediction dimension of the passenger flow volume.

[0095] The beneficial effects of the present invention are mainly manifested in:

[0096] (1) By reconstructing the feature dataset, the multi-dimensional information of the elevator passenger flow is fully utilized, enabling the model to flexibly adapt to the elevator system requirements of different building types and passenger flow patterns. This feature reconstruction method increases the expressive ability of the features, is applicable to various scenarios, and improves the generalization of the model.

[0097] (2) The dynamic noise reduction clustering algorithm is used to preprocess the original data, effectively removing noise data and outliers, and ensuring the high quality of the input data. Combined with the structural design of the multi-dimensional time series convolutional layer (MTC) and dilated convolution, the model has stronger robustness and adaptability when dealing with complex passenger flow data, and can cope with passenger flow changes under various environments and conditions.

[0098] (3) Through the design of the MTC convolutional layer, not only the receptive field of the model is expanded, but also the number of parameters of the model is reduced, and the computational overhead is lowered. At the same time, the multi-head adaptive attention mechanism (MHAAM) can adaptively focus on important moments, reducing the computational requirements for irrelevant information and improving the computational efficiency of the model.

[0099] (4) The multi-dimensional time series memory weight network comprehensively extracts and analyzes the complex time series features and multi-dimensional features in the elevator passenger flow data. By combining BiLSTM and the multi-head adaptive attention mechanism, the model can capture the dynamic changes and critical moments of the elevator passenger flow, thus significantly improving the prediction accuracy. Brief Description of the Drawings

[0100] Figure 1 It is a graph of the intra-week passenger flow changes after the original passenger flow data is sorted by day.

[0101] Figure 2 It is an architecture diagram of the multi-dimensional time series memory weight network defined by the present invention.

[0102] Figure 3 It is a comparison graph of the prediction effect of model hyperparameter experiments.

[0103] Figure 4 It is a comparison graph of the prediction effects of eight models on the passenger flow of the same day. Detailed Embodiments

[0104] The present invention will be further described below in conjunction with the accompanying drawings.

[0105] Referring to Figures 1 to 4 , an elevator passenger flow prediction method based on a multi-dimensional time series memory weight network, the method comprising the following steps:

[0106] The first step is to construct a multi-dimensional passenger flow time series data set, and the process is as follows:

[0107] Step (1.1) Data set construction, the multi-dimensional passenger flow time series data set matrix X is defined as follows,

[0108]

[0109] where x t,d is the j-th feature information value corresponding to the i-th time node, T s is the total number of time steps of the data set, and D is the total feature dimension of the data set;

[0110] Step (1.2) Feature information collection, the process is:

[0111] Step (1.2.1) Collect the real-time operation parameters of the elevator system, including the waiting passenger flow numbers in the hall and between each floor, the running time of each elevator during the sampling time, the load condition during elevator operation, the number of passengers in the car, and the elevator energy consumption data;

[0112] The elevator control system records the start and stop times of each operation. A weighing sensor is installed in the elevator car, and the sensor data communicates with the central server through the RS485 and Modbus protocols to realize real-time update of the database. An energy consumption monitoring module is installed in the equipment control cabinet, and cameras are installed in the waiting area outside the car and inside the car. The collected data can be used to real-time update the central server database through the TCP / IP protocol.

[0113] After authentication, data acquisition can be directly accessed through the remote data interface call to access the database.

[0114] Among them, the authentication protocol (OAuth) is an open standard protocol that allows users to securely access protected resources through authorization tokens, ensuring data access rights and security. The interface call (API) is an interface that allows data interaction between different systems, enabling external applications to remotely access databases or services through predefined request and response formats.

[0115] Step (1.2.2) Collect environmental information and integrate auxiliary feature information, including encoding the time series information at the current moment, the day of the week information, the outdoor temperature, the rain and snow conditions, the special weather annotation, and the floor use information;

[0116] Initiate an HTTP request through the interface provided by the meteorological software to obtain meteorological data in JSON format, and incorporate the parsed temperature, rain / snow conditions, and special weather annotation information into the feature dataset. At the same time, encode the system timestamp obtained during data sampling and the day-of-week information of the current day and integrate them into the feature dataset.

[0117] HTTP is the HyperText Transfer Protocol, which is used to transfer data between the client and the server and is the basic protocol for Internet data exchange. Through an HTTP request, the client can obtain resources from the server, such as web pages or API data. JSON is a lightweight data interchange format that adopts a key-value pair structure, facilitating data storage and transmission. It is widely used in API data responses and is easy to be parsed and used by programming languages.

[0118] Step (1.3) Feature data preprocessing, the process is as follows:

[0119] Step (1.3.1) Perform data cleaning, remove duplicate records in the dataset, and replace invalid or obviously incorrect data points (such as negative values in a feature set that should be all positive, unreasonable values due to fluctuations during the acquisition process, and unreasonable encoded values) with missing values;

[0120] Step (1.3.2) Perform interpolation processing, and use the linear interpolation method to fill in the missing values generated due to the replacement of invalid values or the lost values during sampling in the dataset;

[0121] The linear interpolation method is a method that uses the numerical values of adjacent known data points to estimate the numerical value of an unknown point in a linear manner to smoothly fill the gaps in the dataset.

[0122] Step (1.3.3) Perform data standardization, standardize the numerical data, and scale the data to a fixed range to eliminate the influence of the dimension;

[0123] Performing data standardization means normalizing all numerical features in the multi-dimensional passenger flow time series dataset X constructed in (1.1), so that each feature value x t,d is within a fixed range of [0, 1] to enhance the generalization ability of subsequent model training.

[0124] The second step: Use the dynamic noise reduction clustering algorithm to divide the original dataset, and the process is as follows:

[0125] Step (2.1) Reduce the dimension of the two-dimensional daily average passenger flow curve, and the process is as follows:

[0126] Step (2.1.1) Organize the passenger flow data on a daily basis, as Figure 1 shown. The passenger flow data for one day can be represented in such a form where x1,D is a vector of one row and D columns, T 1 is the last time step of a day, and D represents the feature dimension;

[0127] Step (2.1.2) Analyze the features to establish a new coordinate mapping: construct a coordinate system with the horizontal axis being the time when the first effective maximum passenger flow peak exceeding the minimum peak during the historical morning rush hour appears, and the vertical axis being the corresponding passenger flow peak intensity;

[0128] Step (2.1.3) Loop through the daily average curve P x and select points using the edge detection algorithm. The outer loop iterates N times for the total number of days, and the inner loop iterates T times in total; 1 ;

[0129] After the loop in Step (2.1.4) ends, a set D of discrete passenger flow expressions corresponding to N days will be obtained s . Add the Noise Factor Recognition (NFR) method: calculate the local noise value of each point and divide it by the average distance of its k nearest neighbors. If the result NFR is greater than 1, then the point is considered a noise point and is removed;

[0130] Step (2.2) Create the initial passenger flow clustering center points, and the process is as follows:

[0131] Step (2.2.1) Create a set of center points C, and calculate the local density ρ of each data point i ,

[0132]

[0133] where σ i is the dynamic value for point i, determined by the average distance from its neighboring points, k is the number of neighboring points participating in the operation, and d ij is the distance between point i and point j;

[0134] Step (2.2.2) Calculate the distance δ of each data point i , and the distance δ i is defined as the distance between point i and the nearest point with a density higher than it,

[0135]

[0136] Step (2.2.3) Select the top k points with the largest ρ×δ value as the initial clustering centers C = C 1 , C 2 ,..., C k , and each center point represents a typical passenger flow pattern in a day. Let the iteration count iter = 0;

[0137] Step (2.3) Create a feature cluster set and assign data points. The process is as follows:

[0138] Step (2.3.1) Create k empty cluster sets S = S 1 , S 2 ,..., S k , where each cluster will contain a passenger flow pattern similar to a certain center point;

[0139] Step (2.3.2) Take each passenger flow point x′ s in the total elevator passenger flow data set D j , calculate the distance from x' j to each passenger flow center point C i , and find the nearest center point.

[0140]

[0141] Among them, represents finding the i that minimizes the objective function, and dist(x′ j , C i ) represents the Euclidean distance between each passenger flow point x′ j and the passenger flow center point C i . Through the metric calculation of this formula, the passenger flow point x' j can be assigned to the cluster where the nearest center point is located.

[0142] Step (2.3.3) Select the similar passenger flow clusters and select the point closest to the center point C i as the new center point, that is, select the point with the average number of passengers closest to the calculated center value in the cluster as the new center point, and increment the iteration count by one;

[0143] Step (2.4) Center point update mechanism. The process is as follows:

[0144] Step (2.4.1) In each iteration process, an adaptive detection mechanism is introduced. According to the average distance from the points in each cluster to the center point, the tightness of the cluster is judged.

[0145]

[0146] Among them, d(x i , C i ) is the distance from the point x i to the center point C i , represents the number of data points in the cluster ;;

[0147] In step (2.4.2), if the compactness of a certain cluster is low, indicating that the points within the cluster are more dispersed, the center point of the cluster is reselected through a dynamic adjustment mechanism;

[0148] In step (2.4.3), when the change in the position of the center point is less than the set threshold or the maximum number of iterations is reached, the iteration stops.

[0149] Thirdly, use a multi-dimensional time-series memory weight network to achieve feature learning extraction and passenger flow prediction for the reconstructed feature data set. The overall structure is as Figure 2 shown, and the process is as follows:

[0150] In step (3.1), construct a multi-dimensional time-series convolutional layer. The process is as follows:

[0151] In step (3.1.1), the MTC convolutional layer first performs convolution on the multi-dimensional features along the time dimension using one-dimensional convolution, so as to extract all local features without omission.

[0152]

[0153] In the formula is the weight vector of the f-th filter at time lag k, b f is the bias term of the f-th filter, t is the current time step, and the final output is a matrix composed of the outputs of all filters at each time step, where F is the number of filters;

[0154] In step (3.1.2), the one-dimensional convolution kernel is enlarged, and the corresponding calculation process after adding the properties of causal convolution is expressed as

[0155]

[0156] In the formula is the output of the first layer at time t, f is the ReLU activation function, W (1) is the convolutional kernel weight of the first layer, x [t,t-1,t-2] is the slice of the input data at time t and the two time steps before it, b (1) is the bias term of the first layer, and "·" represents the dot product of vectors;

[0157] In step (3.1.3), the output is passed to the field-of-view expansion module (TAtrous_Block). After the modification, the convolutional kernel is no longer continuous, and the corresponding convolutional process is expressed as

[0158]

[0159] In the formula is the output of the i-th layer at time t, f is the ReLU activation function, W (i)is the convolution kernel weight of the i-th layer, is the data at time t and two time steps before the output of the (i - 1)-th layer, b (i) is the bias term of the i-th layer;

[0160] TAtrous_Block is composed of i dilated convolutional layers with different kernel sizes and dilation factors, which can increase the receptive field while reducing the number of model parameters and without adding a heavy burden to model training, and fully extract local temporal passenger flow features;

[0161] Step (3.2) constructs a memory weight network, and the process is as follows:

[0162] Step (3.2.1) introduces BiLSTM to capture the forward and backward dependencies of the time series. For the input sequence Y, the output of BiLSTM is where L is the number of LSTM cells. The calculation formula is expressed as:

[0163]

[0164] In the formula and H t are the forward hidden state, backward hidden state, and combined hidden state in BiLSTM at time t respectively, f is the activation function, W x represents the weight matrix from the input to the hidden state, Y t represents the input data at time t, W h is the weight matrix from the hidden state at the previous time step to the current hidden state, and b is the bias term;

[0165] BiLSTM is a recurrent neural network that can capture both the forward and backward dependencies of the input. H represents the output matrix of this layer over the entire time series, and each row corresponds to the combined hidden state H t ;

[0166] Step (3.2.2) uses the multi-head adaptive attention mechanism to enhance the model's attention to key time steps. First, three different linear transformations are used to generate the corresponding query Q, key K, and value V matrices,

[0167] Q = H · W Q ;

[0168] K = H · W K ;

[0169] V = H · W V ;

[0170] In the formula, W Q , W K , W V are learnable weight matrices;

[0171] Step (3.2.3) introduces an adaptive factor α based on the extracted features, multi-dimensional passenger flow feature information, and the feature of the current time step t , which is used to adjust the attention weights at each time step.

[0172] state t = [passenger t , load t ,..., weather t ;

[0173] α t = σ(W α · [H, state t + b α );

[0174] Where W α and b α are learnable parameters, state t is the elevator passenger flow feature set, where passenger t is the number of passengers waiting for the elevator at time t, load t is the elevator operation load condition at time t, weather t is the special weather annotation at time t, and σ is the activation function;

[0175] Step (3.2.4) calculates the attention weights for each head i,

[0176]

[0177] Where d k is the dimension of the key vector, QK T is the dot product of the query and the key, and the softmax function is used to normalize the similarity into a probability distribution;

[0178] Step (3.2.5) concatenates the outputs of all heads and generates the final multi-head adaptive attention output through a linear transformation,

[0179] Context = Concat(Att 1 , Att 2 ,..., Att h ) · W O ;

[0180] Where W O is the linear transformation weight matrix;

[0181] Step (3.2.6) maps the multi-head adaptive attention output to the prediction result. The prediction layer adopts a fully connected layer structure. Taking Context as the input of the prediction layer, through linear transformation of the features contained in Context, the high-dimensional features in Context are compressed into scalar prediction values, thus achieving the function of mapping the attention output to the prediction dimension of the passenger flow volume.

[0182] The application process of this embodiment is as follows:

[0183] Step 1. Composition of the experimental dataset: The experimental dataset consists of the results obtained by preprocessing the original dataset of the Internet of Things platform based on the multi-dimensional data reconstruction method proposed in the present invention. The data sampling interval is 15 minutes, and a total of 6048 records are included, forming a multi-dimensional time-series passenger flow dataset containing 10 features. The input data format is a matrix of size 6048×10, and the output is a one-dimensional prediction result of the passenger flow characteristics. The dataset is divided into a training set and a test set according to a ratio of 8:2 to ensure the effectiveness of model training and evaluation.

[0184] Step 2. Experimental environment settings: The environment configuration used in the experiment is as follows: 64-bit Windows 10 operating system, 32G of running memory, Intel Core i7-12700K@3.60GHz CPU, Nvidia RTX3080 GPU, Python = 3.6 for the program running environment, tensorflow = 1.9.0, and keras = 2.0.8 for the deep learning framework.

[0185] Step 3. Define evaluation metrics: In this experiment, to evaluate the accuracy and robustness of the passenger flow prediction model, the following two commonly used error evaluation metrics are adopted:

[0186] Mean Absolute Error (MAE): This metric is used to measure the average absolute deviation between the predicted value and the actual value. Its calculation formula is as follows,

[0187]

[0188] where y i is the actual value, is the predicted value, and T s is the total number of time steps in the dataset. The smaller the MAE value, the higher the prediction accuracy of the model;

[0189] The output at the first layer at time t, f is the ReLU activation function, and W (1) is the convolutional kernel weight of the first layer

[0190] Root Mean Square Error (RMSE): This metric reflects the average of the squared deviations between the predicted values and the actual values, and is used to evaluate the overall level of prediction error. Its calculation formula is as follows,

[0191]

[0192] RMSE is more sensitive than MAE in emphasizing larger errors, so it can reveal abnormal deviations in prediction errors. The smaller the RMSE value, the better the prediction performance of the model;

[0193] Step 4. Comparative analysis of experimental results: The predicted elevator passenger flow at the current moment is affected by the size of the sliding window (timesteps) used to construct the network input. To determine a reasonable timesteps value, different timesteps are set for testing. For the passenger flow prediction value at time t 1 The model comprehensively considers the eigenvalue characteristics of the time period [t 1 -timesteps, t 1 -1]. The prediction effect of the passenger flow in the next 15 minutes is as Figure 3 shown. The MAE of the model when timesteps takes different values is shown in Table 1.

[0194] Table 1 shows the mean absolute error results for different timesteps;

[0195]

[0196] It can be seen from Table 1 that when timesteps = 17, the average loss is the smallest, indicating that the 17 characteristic data closest to the current moment have the greatest impact on the current moment. Since the passenger flow change during the weekend is more stable and the passenger flow change pattern is more fixed than on weekdays, when timesteps = 15, there is already a good prediction performance. Therefore, when predicting the passenger flow on weekdays, timesteps = 17, and when predicting the passenger flow on weekends, timesteps = 15.

[0197] To verify the effectiveness of the NRDC-MTMWN model in the elevator passenger flow prediction task, it is necessary to compare the performance of the method proposed in this specification with other typical passenger flow prediction models. Select the Temporal Convolutional Network (TCN) and Bidirectional Long Short-Term Memory Network (Bi-LSTM) from non-parametric models; select CNN-GRU, ConvBiLstm-Am, STAFN, Ceemdan-Convla, VMD-BiTCN-Psformer from hybrid models. The prediction results of different models are as Figure 4The prediction performance comparison of different models is shown in Table 2.

[0198] Table 2 shows the comparison of prediction performance of different models;

[0199]

[0200]

[0201] It can be seen from Table 2 that the evaluation indicators of the NRDC-MTMWN model proposed in the present invention under different conditions are significantly better than those of the other 7 models. This shows that compared with the other models, the model can more keenly capture the spatial characteristics of multidimensional data, can effectively improve the accuracy of passenger flow prediction, and also has good performance in long-term and short-term prediction experiments. VMD-BiTCN-Psformer retains the extracted time series feature information well through the improved ProbSparse self-attention mechanism, and can also have a good response speed and prediction accuracy for complex passenger flow conditions. Ceemdan-Convla can match the inflection point of passenger flow fluctuations well due to the use of the Ceemdan algorithm, and obtains a good ability to capture up and down time dependencies by stacking LSTM. STAFN applies the transformation-attention block to extract the time dependency of passenger flow from a global and local perspective, and also has a good prediction effect, but because its convolution mechanism attention network is not fully used here due to the different data forms, its prediction accuracy is not as good as the Ceemdan-Convla model. The ConvBiLstm-Am model uses a combination of convolutional layers and bidirectional LSTM layers, which can also extract time series features well, but its prediction accuracy is still not as good as the STAFN model, which shows that it has certain deficiencies in feature refinement and fusion. The TCN model allows parallel computing due to its convolutional structure. It can also well mine the time features in the data while having high computational efficiency, but it is not as good as the hybrid model in processing data spatial features, so its prediction effect is inferior to the ConvBiLstm-Am model. Compared with ConvBiLstm-Am, CNN-GRU has a simpler structure and is far inferior to the ConvBiLstm-Am model in capturing the long-term and short-term dependencies of complex data and mining data spatial relationships. The remaining Bi-LSTM model lacks the ability to extract spatial features of data, and there may be complex correlations and important interactive information between different dimensions in multi-dimensional time series data, so it can be found that the prediction effect of BiLSTM is not ideal.

[0202] In summary, the NRDC-MTMWN model significantly improves the accuracy of passenger flow prediction by refining and comprehensively extracting the multi-dimensional time series features of elevators, which is superior to other comparison models. This indicates that the NRDC-MTMWN model has strong robustness and reliability in processing complex time series data, providing an effective solution for short-term passenger flow prediction.

[0203] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept and is only for illustrative purposes. The protection scope of the present invention should not be regarded as limited to the specific forms stated in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be conceived by those of ordinary skill in the art based on the inventive concept of the present invention.

Claims

1. A method for predicting elevator passenger flow based on a multi-dimensional temporal memory weight network, characterized in that: The method comprises the following steps: The first step is to build a multi-dimensional passenger flow time series dataset. The process is as follows: Step (1.1) Dataset construction; Step (1.2) feature information collection; Step (1.3) feature data preprocessing; The second step is to use the dynamic noise reduction clustering algorithm to divide the original data set. The process is as follows: Step (2.1) reduces the dimension of the two-dimensional daily average passenger flow curve; Step (2.2) creates the initial passenger flow cluster center point; Step (2.3) creates a set of feature clusters and assigns data points; Step (2.4) center point update mechanism; The third step is to use a multi-dimensional temporal memory weight network to achieve feature learning extraction and passenger flow prediction of the reconstructed feature data set. The process is as follows: Step (3.1) constructs a multi-dimensional temporal convolutional layer; Step (3.2) constructs a memory weight network.

2. The elevator passenger flow prediction method based on a multi-dimensional temporal memory weight network according to claim 1, characterized in that: In the step (1.1), the multi-dimensional passenger flow time series data set matrix X is defined as follows: Among them, x t,d is the jth feature information value corresponding to the i-th time node, T s is the total number of time steps in the dataset, and D is the total feature dimension of the dataset.

3. The elevator passenger flow prediction method based on a multi-dimensional temporal memory weight network according to claim 2, characterized in that: The process of step (1.2) is: Step (1.2.1) collects the real-time operating parameters of the elevator system, including the number of passengers waiting for the elevator in the lobby and between each floor, the running time of each elevator within the sampling time, the load of the elevator during operation, the number of passengers in the car and the energy consumption data of the elevator; Step (1.2.2) collects environmental information and integrates auxiliary feature information, including the encoding of the current time series information, outdoor temperature, rain and snow conditions, special weather annotations and floor usage information.

4. The elevator passenger flow prediction method based on a multi-dimensional temporal memory weight network according to claim 3, characterized in that: The process of step (1.3) is: Step (1.3.1) performs data cleaning, removes duplicate records in the data set, and replaces invalid or obviously erroneous data points with missing values; Step (1.3.2) performs interpolation processing, and uses linear interpolation method to fill in missing values ​​or sampling missing values ​​in the data set due to invalid value replacement; Step (1.3.3) performs data standardization, standardizes numerical data, and scales the data to a fixed range to eliminate the impact of dimension.

5. The elevator passenger flow prediction method based on a multi-dimensional temporal memory weight network according to any one of claims 1 to 4, characterized in that: The process of step (2.1) is: Step (2.1.1) divides and organizes the passenger flow data on a daily basis. The passenger flow data for one day can be expressed in the following form: where x 1,D is a vector with one row and D columns, T1 is the last time step of the day, and D represents the feature dimension; Step (2.1.2) Analyze the characteristics and establish a new coordinate mapping: the horizontal axis is the time when the first effective maximum passenger flow peak value exceeds the minimum peak value of the historical morning peak time period, and the vertical axis is the corresponding passenger flow peak intensity to construct a coordinate system; Step (2.1.3) loops through the daily average curve P x And use the edge detection algorithm to select points, the outer loop traverses the total number of days N, and the total number of traversals of the inner loop is T1; At the end of the loop of step (2.1.4), the discrete expression set D of passenger flow corresponding to N days will be obtained. s , add the noise factor recognition method NFR: calculate the local noise value of each point and divide it with the average distance of k neighboring points. If the result NFR is greater than 1, the point is considered to be a noise point and is removed.

6. The elevator passenger flow prediction method based on multi-dimensional temporal memory weight network according to claim 5, characterized in that: The process of step (2.2) is: Step (2.2.1) creates a set of central points C and calculates the local density ρ of each data point i , where σ i is the dynamic value for point i, which is determined by the average distance from its neighboring points, k is the number of neighboring points involved in the operation, and d ij is the distance between point i and point j; Step (2.2.2) calculates the distance δ of each data point i , distance δ i It is defined as the distance between point i and the nearest point with higher density than it, Step (2.2.3) takes the first k points with the largest ρ×δ value as the initial cluster centers C=C1,C2,...,C k , each center point represents a typical passenger flow pattern in a day, and the number of iterations iter=0.

7. The elevator passenger flow prediction method based on multi-dimensional temporal memory weight network according to claim 6, characterized in that: The process of step (2.3) is: Step (2.3.1) creates k empty cluster sets S = S1, S2, ..., S k , each cluster will contain passenger flow patterns similar to a certain center point; Step (2.3.2) Take the total elevator passenger flow data set D s Each passenger flow point x′ j , calculate x' j To each passenger flow center point C i distance, find the nearest center point, in, It means to find the i that minimizes the objective function, dist(x′ j ,C i ) represents each passenger flow point x′ j and passenger flow center point C i The Euclidean distance of the passenger flow point x' can be calculated by this measurement formula. j Assigned to the cluster with the closest center point Step (2.3.3) Select similar passenger flow clusters Middle distance center point C i The nearest point is taken as the new center point, that is, the point closest to the calculated center value of the average number of passengers in the cluster is selected as the new center point, and the number of iterations is increased by one.

8. The elevator passenger flow prediction method based on a multi-dimensional temporal memory weight network according to claim 7, characterized in that: The process of step (2.4) is: In each iteration of step (2.4.1), an adaptive detection mechanism is introduced to judge the compactness of the cluster by calculating the average distance from each point in the cluster to the center point. Among them, d(x i ,C i ) is point x i To center point C i The distance Representation Cluster The number of data points in ; In step (2.4.2), if the compactness of a cluster is low, it means that the points in the cluster are more dispersed, then the center point of the cluster is reselected through the dynamic adjustment mechanism; In step (2.4.3), the iteration is stopped when the position change of the center point is less than the set threshold or the maximum number of iterations is reached.

9. The elevator passenger flow prediction method based on a multi-dimensional temporal memory weight network according to any one of claims 1 to 4, characterized in that: The process of step (3.1) is: Step (3.1.1) The MTC convolution layer first uses a one-dimensional convolution method to convolve the multi-dimensional features along the time dimension, so as to extract all local features without omission. In the formula is the weight vector of the fth filter at time lag k, b f is the bias term of the fth filter, t is the current time step, and the final output is a matrix consisting of the outputs of all filters at each time step, where F is the number of filters; Step (3.1.2) enlarges the one-dimensional convolution kernel and adds the properties of causal convolution. The corresponding calculation process is expressed as: In the formula is the output of the first layer at time t, f is the ReLU activation function, W (1) is the convolution kernel weight of the first layer, x [t,t-1,t-2] is a slice of the input data at time t and the two time steps before it, b (1) is the bias term of the first layer, "·" represents the vector dot product; Step (3.1.3) passes the output to the field of view expansion module. After the change, the convolution kernel is no longer continuous. The corresponding convolution process is expressed as: In the formula is the output of the i-th layer at time t, f is the ReLU activation function, W (i) is the convolution kernel weight of the i-th layer, is the data output by the i-1th layer at time t and two time steps before it, b (i) is the bias term of the i-th layer.

10. The elevator passenger flow prediction method based on multi-dimensional temporal memory weight network according to claim 9, characterized in that: The process of step (3.2) is: Step (3.2.1) introduces BiLSTM to capture the forward and backward dependencies of the time series. For the input sequence Y, the output of BiLSTM is Where L is the number of LSTM units, and the calculation formula is expressed as, In the formula and H t are the forward hidden state, backward hidden state and combined hidden state in BiLSTM at time t, f is the activation function, W x represents the weight matrix input to the hidden state, Y t represents the input data at time t, W h is the weight matrix from the hidden state of the previous time step to the current hidden state, and b is the bias term; Step (3.2.2) uses a multi-head adaptive attention mechanism to enhance the model's attention to key time steps. First, three different linear transformations are used to generate the corresponding query Q, key K, and value V matrices. Q=H·W Q ; K=H·W K ; V=H·W V ; Where W Q , W K , W V is a learnable weight matrix; Step (3.2.3) introduces an adaptive factor α based on the extracted features, multi-dimensional feature information of passenger flow and the current time step features t , used to adjust the attention weight at each time step, state t =[passenger t ,load t ,...,weather t ]; a t =σ(W α ·[H,state t ]+b α ); Where W α and b α is a learnable parameter, state t is the elevator passenger flow feature set, where passenger t is the number of passengers waiting for the elevator at time t, load t is the elevator load condition at time t, weather t is the special weather label at time t, σ is the activation function; Step (3.2.4) For each head i, calculate the attention weight, Where d k is the dimension of the key vector, QK T is the dot product of the query and the key, and the softmax function is used to normalize the similarity into a probability distribution; Step (3.2.5) concatenates the outputs of all heads and generates the final multi-head adaptive attention output through a linear transformation. Context=Concat(Att1,Att2,...,Att h )·W O ; Where W O is the linear transformation weight matrix; Step (3.2.6) maps the multi-head adaptive attention output to the prediction result. The prediction layer adopts a fully connected layer structure and takes Context as the input of the prediction layer. By linearly transforming the features contained in Context, the high-dimensional features in Context are compressed into scalar prediction values, thereby achieving the effect of mapping the attention output to the prediction dimension of passenger flow quantity.

Images