A smart park dynamic analysis and early warning platform

Abstract

The application discloses a kind of wisdom park dynamic analysis early warning platform, specifically related to wisdom park dangerous prediction field, including collection and reduction module, feature extraction module, matching mapping module, training classification module, prediction generation module and sorting intervention module, collection and reduction module is by obtaining multiple equipment operation and environmental state data, and according to equipment category and time stamp is arranged into uniform structure multivariate time series collection;By constructing the multivariate time series of fusion multiple equipment operation state and environmental information, extract disturbance feature and construct propagation path, combine risk classification model and dangerous prediction atlas, form a set of dynamic dangerous prediction and hierarchical intervention mechanism for multi-system linkage and nonlinear disturbance scene, to solve the problem that the existing wisdom park energy system lacks dynamic dangerous prediction ability for complex disturbance situation, cannot realize early identification and intervention to nonlinear energy risk.

Description

Technical Field

[0001] This invention relates to the field of smart park hazard prediction technology, and more specifically, to a smart park dynamic analysis and early warning platform. Background Technology

[0002] In the current energy management system of smart parks, the ability to predict dangers relies heavily on static load models and rule-driven mechanisms. Anomaly identification is mainly achieved by comparing preset thresholds with periodic data. However, such mechanisms lack the ability to dynamically model sudden and nonlinear events. Especially in the scenario of highly coupled energy systems in parks, they cannot effectively handle complex situations such as temporary power surges, environmental feedback delays, or concurrent anomalies of multiple systems. In existing technologies, although there are monitoring modules for power fluctuations and operating status, their response mechanisms are mostly post-event compensation and lack feedforward early warning capabilities.

[0003] Furthermore, energy subsystems in smart parks, such as air conditioning, elevators, lighting, and charging piles, often operate asynchronously at multiple points. Under this heterogeneous structure, energy consumption patterns are highly dynamic and uncertain. Current rule-setting methods cannot cover edge scenarios such as cross-load transfer, fault cascading propagation, or short-cycle energy consumption anomalies, resulting in severe delays or false alarms when dealing with energy system operational imbalances. At the same time, weak multi-source data fusion capabilities also make it difficult for the hazard prediction system to distinguish between normal fluctuations and potential threats, ultimately limiting the overall response accuracy and stability of the system.

[0004] Against this backdrop, the core issue is that existing smart park energy systems lack the ability to dynamically predict risks under multi-system linkage and complex disturbance scenarios, and are unable to identify and intervene in nonlinear energy risks in advance. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of existing technologies, embodiments of the present invention provide a smart park dynamic analysis and early warning platform. By constructing a multivariate time series that integrates the operating status and environmental information of multiple devices, it extracts disturbance characteristics and constructs propagation paths. Combined with risk classification models and hazard prediction maps, it forms a dynamic hazard prediction and graded intervention mechanism for multi-system linkage and nonlinear disturbance scenarios. This addresses the problem that existing smart park energy systems lack dynamic hazard prediction capabilities for complex disturbance scenarios and cannot achieve early identification and intervention of nonlinear energy risks.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a smart park dynamic analysis and early warning platform, comprising a data collection and consolidation module, a feature extraction module, a matching and mapping module, a training and classification module, a prediction and generation module, and a ranking and intervention module;

[0007] The data acquisition and processing module obtains operational and environmental status data from multiple devices and organizes them into a unified multivariate time series set according to device category and timestamp;

[0008] The feature extraction module extracts a perturbation feature group consisting of perturbation amplitude, perturbation rate and perturbation frequency by performing Fourier transform with a fixed window length and sliding window segmentation on the multivariate time series set, and generates a standard perturbation vector sequence after normalization.

[0009] The matching and mapping module performs pairing statistics on vector pairs with the same timestamp and an Euclidean distance below a threshold in the standard disturbance vector sequence, and constructs a set of disturbance propagation paths based on device identification and pairing frequency;

[0010] The matching and mapping module includes extracting data with the same timestamp from the standard perturbation vector sequence, constructing a perturbation vector combination group at the corresponding sampling time, and treating any two perturbation vectors in each group as a set of calculation objects;

[0011] Perform Euclidean distance calculation on each set of computational objects to form an Euclidean distance matrix containing each perturbation vector pair. The judgment conditions include that the perturbation amplitude difference in each vector pair is less than a preset perturbation difference threshold, and the calculated Euclidean distance value is lower than a set distance similarity threshold. If both judgment conditions are met, it is identified as a valid perturbation vector pair and output to the perturbation vector pair set; otherwise, reconstruct the perturbation vector combination group under the timestamp and re-execute the Euclidean distance calculation.

[0012] Using the device number and device type of each perturbation vector pair in the perturbation vector pair set as the matching key, perform pairing frequency statistics and timestamp location recording to generate a perturbation linkage statistics record table containing the paired device number, pairing frequency and sampling time index;

[0013] The device number pairs in the disturbance linkage statistics record table that have reached the specified lower limit of pairing frequency are constructed into directed node pairs, and directional labels are established according to the sampling time index to generate a set of disturbance propagation paths containing node number, connection weight and time label fields;

[0014] The training classification module inputs the set of perturbation propagation paths and the standard perturbation vector sequence into the classification calculation process, and establishes the mapping relationship between perturbation feature groups and risk labels by training the perturbation risk classification model;

[0015] The prediction generation module inputs the current standard perturbation vector into the perturbation risk classification model, matches the risk label with the propagation path node number, and generates a hazard prediction map containing node index and risk number.

[0016] The sorting intervention module sorts the nodes in the hazard prediction map by risk number and path connectivity, and generates load unloading instructions, cycle reset instructions and visual alarm signals in sequence, and writes the execution data into a time series set.

[0017] In a preferred embodiment, the data acquisition and aggregation module acquires the unit time energy consumption, voltage change, current response, and load percentage of air conditioning equipment, lighting equipment, elevator drive device, and electric vehicle charging device, as well as the ambient temperature and humidity values ​​corresponding to each sampling, by setting a sampling period. Then, it tags the operating status data of each device and the environmental status data according to the sampling time and outputs a synchronized data set of device data and environmental data.

[0018] Each set of running status data and environmental status data in the synchronized data group is matched at the field level according to the timestamp. An aligned data structure containing sampling time, device type, device number, running status parameters and environmental status parameters is constructed through field correspondence. The output is a device-level bound data block with a unified field order and complete dimensions.

[0019] Perform unified device type encoding, field position regularization, and time axis alignment operations on the device-level bound data blocks. Arrange all data blocks according to the set time granularity and field structure, and output a multivariate time series set with consistent field structure, clear device identification, and continuous time.

[0020] In a preferred embodiment, the feature extraction module groups the multivariate time series set by device number, obtains the time series segment corresponding to each type of device, and then performs a Fourier transform with a fixed window length on each segment to extract the frequency domain signal containing the frequency distribution of energy consumption change, voltage change and current response, and outputs the frequency domain signal set of the corresponding device.

[0021] By setting a time sliding window with a fixed width and a fixed overlap ratio, continuous segments in the frequency domain signal set are obtained. Then, the truncation operation is performed according to the window movement order to form a sequence of frequency domain sub-segments with continuous time indices. The output is a set of frequency domain time distributions containing all window segments.

[0022] In a preferred embodiment, in the feature extraction module, the disturbance amplitude is obtained by calculating the difference between the maximum and minimum values ​​of the signal amplitude of each sub-segment in the frequency domain time distribution set, the disturbance rate is obtained by calculating the slope of amplitude change between adjacent segments, the disturbance frequency is obtained by locating the main frequency point of each sub-segment, and then the three disturbance quantities are combined to form a disturbance feature data set, and the disturbance feature set is output.

[0023] The perturbation amplitude, perturbation rate, and perturbation frequency in the perturbation feature set are linearly normalized according to preset upper and lower limits, respectively. All feature data sets are rearranged by time index to generate a vector format expression with consistent fields and stable order, and a standard perturbation vector sequence is output.

[0024] In a preferred embodiment, the training classification module extracts node numbers, connection weights, and time labels from the set of perturbation propagation paths, and combines them with the perturbation amplitude, perturbation rate, and perturbation frequency of the corresponding timestamps in the standard perturbation vector sequence to construct a joint feature input set containing path features and perturbation features.

[0025] The joint feature input group is subjected to training set partitioning operation, and a classification label set is generated based on the historical running status label or manual annotation result corresponding to each path, and training data pairs between feature group and risk label are constructed.

[0026] The training data is input into the classification calculation process, and feature encoding, weight initialization and iterative optimization operations are performed to train the perturbation risk classification model and form a mapping structure between input features and output labels.

[0027] The trained perturbation risk classification model is validated by calculating the classification accuracy of the joint feature input group on the test data and solving the convergence rate of the loss function based on the risk label prediction bias and the rate of change of training rounds.

[0028] If the classification accuracy is higher than the lower limit of the recognition accuracy and the convergence rate exceeds the threshold of the loss decrease rate, then the model is confirmed to meet the training effectiveness requirements, the perturbation risk classification model is output and input into the prediction generation module to perform the risk map generation operation; if any judgment condition is not met, then the feature encoding and weight initialization are re-executed on the current training data pair to construct a new perturbation risk classification model.

[0029] In a preferred embodiment, the prediction generation module inputs the standard perturbation vector at the current moment into the perturbation risk classification model, performs risk classification calculation, and outputs the risk category number corresponding to the standard perturbation vector.

[0030] Based on the time index and device number of the standard disturbance vector, the corresponding node number and connection path information are extracted from the disturbance propagation path set to construct an index relationship set between device nodes and risk propagation paths;

[0031] Perform a pairing operation between the risk category number and the index relationship set, combine them in the order of node number to form a risk node set, and attach the corresponding risk category number to each device node;

[0032] The device number, node number, time index, and risk category number in the risk node set are combined to construct a data graph with a consistent field structure, and the output is a hazard prediction graph containing node information, time information, and risk level.

[0033] In a preferred embodiment, the sorting intervention module extracts the corresponding risk category number and the connectivity value of the node in the set of disturbance propagation paths for each node in the hazard prediction map, and constructs a sorting feature set containing risk intensity parameters and structural coupling parameters.

[0034] Perform a two-factor sorting operation on the sorting feature set, arranging them from high to low according to the risk category number. If the risk category numbers are the same, they are arranged from high to low according to the connectivity value, generating a sorted high-priority device node queue.

[0035] Each node in the high-priority device node queue is matched sequentially with its device number and current operating status, and combined with the risk category number to generate corresponding load unloading instructions, operating cycle reset instructions and visual alarm signals.

[0036] After executing the load unloading command, cycle reset command, and visual alarm signal generated by each node, extract the command success mark, reset feedback value, and alarm trigger record corresponding to that node, and write them into a multivariate time series set consistent with the time index of that node to construct a data update record that matches the intervention action.

[0037] The technical effects and advantages of this invention are as follows:

[0038] This solution constructs a set of disturbance feature vector sequences and propagation paths to establish a dynamic correlation map among multiple devices, and uses a classification model to identify risks. This enables the system to handle energy anomaly scenarios with multiple devices linked, frequent disturbances, and delayed responses, thus making up for the shortcomings of traditional static threshold-based mechanisms in dealing with sudden nonlinear risks.

[0039] This invention integrates the operating status parameters of various devices and environmental data into a time series set with a consistent structure by setting a unified time granularity and field order. This ensures that the data dimensions are complete and the labels are clear, which facilitates subsequent modeling and analysis and solves the problem of the difficulty in integrating multi-source asynchronous data in existing systems.

[0040] This invention generates a path set containing physical connection and dynamic propagation characteristics by statistically pairing similar disturbance vectors within the same timestamp after extracting disturbance features, and combining device identification and pairing frequency, enabling the system to restore the regularity and scope of risk transmission between devices.

[0041] This scheme uses perturbation vectors and path information as joint inputs to train a perturbation risk classification model, forming a mapping structure between perturbation features and risk labels. This enables the model to perform graded hazard prediction, distinguish different risk forms and evolution paths, and improve classification accuracy and prediction depth.

[0042] This solution combines the risk number and path connectivity of nodes in the risk prediction map to perform ranking analysis and intervention output, thereby realizing a hierarchical control instruction generation mechanism driven by the hazard prediction results. This enables the system to output corresponding intervention strategies for different risk levels. Attached Figure Description

[0043] Figure 1 This is a system module diagram of the present invention.

[0044] Figure 2 This is a flowchart of the data acquisition and feature extraction process of the present invention.

[0045] Figure 3 This is a flowchart of the matching and mapping process of the present invention.

[0046] Figure 4 This is a flowchart of the training and classification process of the present invention.

[0047] Figure 5 This is a flowchart of the prediction generation process of the present invention.

[0048] Figure 6 This is a flowchart of the sorting intervention process of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Refer to the instruction manual appendix Figure 1-6 An embodiment of the present invention provides a smart park dynamic analysis and early warning platform, comprising a data collection and consolidation module, a feature extraction module, a matching and mapping module, a training and classification module, a prediction and generation module, and a ranking and intervention module.

[0051] The data acquisition and processing module obtains operational and environmental status data from multiple devices and organizes them into a unified multivariate time series set according to device category and timestamp;

[0052] The feature extraction module extracts a perturbation feature group consisting of perturbation amplitude, perturbation rate and perturbation frequency by performing Fourier transform with a fixed window length and sliding window segmentation on the multivariate time series set, and generates a standard perturbation vector sequence after normalization.

[0053] The matching and mapping module performs pairing statistics on vector pairs with the same timestamp and an Euclidean distance below a threshold in the standard disturbance vector sequence, and constructs a set of disturbance propagation paths based on device identification and pairing frequency;

[0054] The training classification module inputs the set of perturbation propagation paths and the standard perturbation vector sequence into the classification calculation process, and establishes the mapping relationship between perturbation feature groups and risk labels by training the perturbation risk classification model;

[0055] The prediction generation module inputs the current standard perturbation vector into the perturbation risk classification model, matches the risk label with the propagation path node number, and generates a hazard prediction map containing node index and risk number.

[0056] The sorting intervention module sorts the nodes in the hazard prediction map by risk number and path connectivity, and generates load unloading instructions, cycle reset instructions and visual alarm signals in sequence, and writes the execution data into a time series set.

[0057] The data acquisition and aggregation module obtains the unit time energy consumption, voltage change, current response, and load ratio of air conditioning equipment, lighting equipment, elevator drive devices, and electric vehicle charging devices by setting the sampling period, as well as the ambient temperature and humidity values ​​corresponding to each sampling. Then, it tags the operating status data of each device and the environmental status data according to the sampling time and outputs the synchronized data group of device data and environmental data.

[0058] Each set of running status data and environmental status data in the synchronized data group is matched at the field level according to the timestamp. An aligned data structure containing sampling time, device type, device number, running status parameters and environmental status parameters is constructed through field correspondence. The output is a device-level bound data block with a unified field order and complete dimensions.

[0059] Perform unified device type encoding, field position regularization, and time axis alignment operations on the device-level bound data blocks. Arrange all data blocks according to the set time granularity and field structure, and output a multivariate time series set with consistent field structure, clear device identification, and continuous time.

[0060] The feature extraction module groups the multivariate time series set by device number, obtains the time series segment corresponding to each type of device, and then performs a Fourier transform with a fixed window length on each segment to extract the frequency domain signal containing the frequency distribution of energy consumption change, voltage change, and current response, and outputs the frequency domain signal set of the corresponding device.

[0061] By setting a time sliding window with a fixed width and a fixed overlap ratio, continuous segments in the frequency domain signal set are obtained. Then, the truncation operation is performed according to the window movement order to form a sequence of frequency domain sub-segments with continuous time indices. The output is a set of frequency domain time distributions containing all window segments.

[0062] In the feature extraction module, the disturbance amplitude is obtained by calculating the difference between the maximum and minimum values ​​of the signal amplitude of each sub-segment in the frequency domain time distribution set, the disturbance rate is obtained by calculating the slope of amplitude change between adjacent segments, the disturbance frequency is obtained by locating the main frequency point of each sub-segment, and then the three disturbance quantities are combined to form a disturbance feature data set, and the disturbance feature set is output.

[0063] The perturbation amplitude, perturbation rate, and perturbation frequency in the perturbation feature set are linearly normalized according to preset upper and lower limits, respectively. All feature data sets are rearranged by time index to generate a vector format expression with consistent fields and stable order, and a standard perturbation vector sequence is output.

[0064] The matching and mapping module extracts data with the same timestamp from the standard perturbation vector sequence, constructs a perturbation vector combination group at the corresponding sampling time, and treats any two perturbation vectors in each group as a set of calculation objects;

[0065] Perform Euclidean distance calculation on each set of computational objects to form an Euclidean distance matrix containing each perturbation vector pair. The judgment conditions include that the perturbation amplitude difference in each vector pair is less than a preset perturbation difference threshold, and the calculated Euclidean distance value is lower than a set distance similarity threshold. If both judgment conditions are met, it is identified as a valid perturbation vector pair and output to the perturbation vector pair set; otherwise, reconstruct the perturbation vector combination group under the timestamp and re-execute the Euclidean distance calculation.

[0066] Using the device number and device type of each perturbation vector pair in the perturbation vector pair set as the matching key, perform pairing frequency statistics and timestamp location recording to generate a perturbation linkage statistics record table containing the paired device number, pairing frequency and sampling time index;

[0067] The device number pairs in the disturbance linkage statistics record table that have reached the specified lower limit of pairing frequency are constructed into directed node pairs, and directional labels are established according to the sampling time index to generate a set of disturbance propagation paths containing node number, connection weight and time label fields.

[0068] The training classification module extracts node numbers, connection weights, and time labels from the set of perturbation propagation paths, and combines them with the perturbation amplitude, perturbation rate, and perturbation frequency of the corresponding timestamps in the standard perturbation vector sequence to construct a joint feature input set containing path features and perturbation features.

[0069] The joint feature input group is subjected to training set partitioning operation, and a classification label set is generated based on the historical running status label or manual annotation result corresponding to each path, and training data pairs between feature group and risk label are constructed.

[0070] The training data is input into the classification calculation process, and feature encoding, weight initialization and iterative optimization operations are performed to train the perturbation risk classification model and form a mapping structure between input features and output labels.

[0071] The trained perturbation risk classification model is validated by calculating the classification accuracy of the joint feature input group on the test data and solving the convergence rate of the loss function based on the risk label prediction bias and the rate of change of training rounds.

[0072] If the classification accuracy is higher than the lower limit of the recognition accuracy and the convergence rate exceeds the threshold of the loss decrease rate, then the model is confirmed to meet the training effectiveness requirements, the perturbation risk classification model is output and input into the prediction generation module to perform the risk map generation operation; if any judgment condition is not met, then the feature encoding and weight initialization are re-executed on the current training data pair to construct a new perturbation risk classification model.

[0073] It should be noted that in the formula structure involved in this scheme, dimensionless terms can be used as proportional or structural adjustment factors. When combined with quantities with units, they only play a role in numerical scaling and do not introduce new physical dimensions. Therefore, they will not change or confuse the overall unit system. This combination of "dimensionless terms and terms with units" can be understood as a composite structural expression commonly used in mathematical physics modeling. It conforms to the principle of dimensional consistency and has a clear physical interpretation basis.

[0074] Secondly, in the formula structure of this scheme, if multiple variables with different physical units are involved, including but not limited to time, mass or energy variables, their joint appearance is to express the collaborative modeling relationship of multiple physical mechanisms. Each variable can form a unified structure through function mapping, ratio combination or normalization adjustment, with clear units and clear meaning. The overall expression conforms to the principle of dimensional consistency and the conventional formula of engineering modeling.

[0075] In this solution, constants, weights, adjustment factors, threshold parameters, proportional coefficients, etc., are all adjustable control parameters for different application environments. Their values ​​depend on the target equipment configuration, data input characteristics, and performance optimization goals. During the implementation phase, they are set to converge within a reasonable range through model verification, performance constraints, or engineering calibration. Although these parameters do not have a unique preset value, they have clear adjustment logic and calculation paths. They belong to the deterministic setting process in engineering implementation. The purpose of this setting is to ensure that the solution is both universally adaptable and reproducible and operable, without affecting its technical clarity and feasibility.

[0076] In the training classification module, define The output value of the perturbation risk classification model is the classification mapping function trained under the combined input of all joint perturbation features and graph structure. The output result is the risk category number.

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] This is the risk label response mapping function, which is used to convert perturbation map features and dynamic judgment results into specific classification label outputs. For the perturbation and path graph structure, the feature aggregation function represents the perturbation feature function ( , ) and path propagation structure function ( , Aggregate integral after tensor coupling; The model trains a dynamic decision function, which is used to measure the current time step. Does the classification training process meet the dual criteria of convergence and error stability? For element-wise multiplication, This represents the dynamic coupling operation between the perturbation evolution term and the structural coupling term, meaning that in time... At time , the perturbation characteristic function With the corresponding path propagation structure function , Perform element-level linkage and combination calculations to form a dynamic response product;

[0084] To represent the current time in the standard perturbation vector sequence The disturbance amplitude function is composed of the signal main frequency amplitude extracted after performing Fourier transform on the device operating state dimensions such as energy consumption, voltage, current, and load. For the current moment The disturbance rate, which represents the disturbance amplitude. The first derivative within a local time window reflects the changing trend of the disturbance intensity; For the current moment The perturbation frequency, representing the dominant frequency in the perturbation signal, is obtained through spectral analysis of a standard perturbation vector sequence and is used to reflect the periodicity of system state changes; the perturbation characteristic function. This is the joint rate of change between the disturbance amplitude and the rate of change. The joint rate of change represents the change through the disturbance amplitude. The dynamic response value, obtained by multiplying the result by the sine function of its rate and differentiating it with respect to time, is used to reveal the degree of instantaneous response to the disturbance state; the disturbance characteristic function The second-order response to changes in the frequency and amplitude of the disturbance is given by the disturbance characteristic function. This indicates that the second derivative is calculated based on the logarithmic growth relationship between the perturbation frequency and the perturbation amplitude, and is used to reveal the trend of curvature level changes in the perturbation.

[0085] This represents the number of valid device nodes in the disturbance propagation path graph. For the path graph, the first One device node; In time At time 1, the path in the graph is... A binary indicator function that determines whether a node is active; if the node is part of the current disturbance path, the value is [value]. Otherwise .

[0086] For the first Each node in time The connection weight value represents the connectivity strength between the node and other nodes in the graph structure, and is used to measure the structural importance of the node in the propagation path; For the first The propagation delay between each node and the starting node of the propagation path represents the time span required for the disturbance signal to travel from the starting point of the path to that node. For the first The relative depth of a node in the graph topology, i.e., its logical level position from the starting point of the propagation path; the deeper the level, the larger the value, used to model the structural attenuation characteristics of path propagation; path propagation structure function. This represents the dynamic coupling strength formed by combining the activation state, connection strength, and propagation delay of each node in the current graph structure using a logarithmic function. It is used to characterize the nonlinear propagation capability of disturbances along the structural path; the path propagation structure function... It represents the exponential decay derivative of the connection strength in the structure propagation path as a function of topological depth, and is used to characterize the energy decay rate of disturbances during deep propagation.

[0087] For the current moment The perturbation risk classification loss function is constructed in the form of cross-entropy and is used to measure the degree of deviation between the model's predicted label and the actual label. Let be the derivative of the loss function with respect to time. This represents the rate of error reduction during training and is used to assess whether the model is in a state of continuous optimization. Aggregate feature function The derivative with respect to time, It reflects the stability and degree of change of the perturbation map features during the training process; Describes the Euclidean norm 2. Used to measure the overall change magnitude of the time gradient vector and as a normalized benchmark for constructing dynamic change thresholds;

[0088] Furthermore, in the training classification module, define In time The perturbation risk classification loss value at any given time is used to measure the overall deviation between the model output and the true risk label, and to reflect the dynamic trend during the training process;

[0089]

[0090]

[0091]

[0092] in, The total number of perturbation samples used in training; For the first The true risk label of a perturbation sample This indicates the actual category of the sample in the disturbance risk classification. Its value is a predefined risk level number, which comes from manual annotation or historical operation status labels and belongs to a fixed set of binary or multi-class labels. For the disturbance risk classification model, the first The predicted output of a sample, after being processed by softmax, determines which sample belongs to the label. The probability of; For the disturbance risk classification model, the first The reconstruction result of the joint feature vector of each sample. As an intermediate representation of the model's output layer; The original joint feature vector is input to the model. The original joint feature vector is composed of the set of perturbation propagation paths and the standard perturbation vector sequence.

[0093] The expression for the reconstruction error of the joint features. Used to measure the model's ability to preserve the original joint feature vector during the encoding and decoding process; An adjustment coefficient is set for the perturbation feature preservation loss term. This adjustment coefficient is used to control the weight of the reconstruction error in the total loss function. The value of the adjustment coefficient for the perturbation feature preservation loss term is selected by the cross-validation strategy during training. This is the adjustment coefficient for the structural disturbance coupling loss term, which is used to adjust the proportion of structural disturbance in the overall loss. For the first The amplitude component of each perturbation sample; Its perturbation rate; The perturbation frequency; The number of effective device nodes in the disturbance propagation path graph; For nodes Does it belong to the first Perturbation path diagram of each sample; For the first The propagation delay of each node; For connection weights; This represents the topology depth.

[0094] The prediction generation module inputs the standard disturbance vector at the current moment into the disturbance risk classification model, performs risk classification calculation, and outputs the risk category number corresponding to the standard disturbance vector;

[0095] Based on the time index and device number of the standard disturbance vector, the corresponding node number and connection path information are extracted from the disturbance propagation path set to construct an index relationship set between device nodes and risk propagation paths;

[0096] Perform a pairing operation between the risk category number and the index relationship set, combine them in the order of node number to form a risk node set, and attach the corresponding risk category number to each device node;

[0097] The device number, node number, time index, and risk category number in the risk node set are combined to construct a data graph with a consistent field structure. The output is a hazard prediction graph containing node information, time information, and risk level, which is then used by the sorting and intervention module to generate intervention instructions.

[0098] The ranking intervention module extracts the corresponding risk category number and the connectivity value of the node in the set of disturbance propagation paths for each node in the hazard prediction map, and constructs a ranking feature set containing risk intensity parameters and structural coupling parameters.

[0099] Perform a two-factor sorting operation on the sorting feature set, arranging them from high to low according to the risk category number. If the risk category numbers are the same, they are arranged from high to low according to the connectivity value, generating a sorted high-priority device node queue.

[0100] Each node in the high-priority device node queue is matched sequentially with its device number and current operating status, and combined with the risk category number to generate corresponding load unloading instructions, operating cycle reset instructions and visual alarm signals.

[0101] After executing the load unloading command, cycle reset command, and visual alarm signal generated by each node, extract the command success mark, reset feedback value, and alarm trigger record corresponding to that node, and write them into a multivariate time series set consistent with the time index of that node to construct a data update record that matches the intervention action, which is used as the model iteration input in the training classification module.

[0102] This solution addresses key issues in current smart park energy systems, such as insufficient hazard prediction capabilities, delayed response, and weak multi-source data fusion capabilities. It proposes a systematic solution with "dynamic analysis and early warning" as its core objective. The solution designs and constructs six functional modules: data collection and consolidation, decomposition and feature extraction, matching and mapping, training and classification, prediction generation, and ranking intervention. These modules are linked by a data processing chain, forming a dynamic closed-loop process system from bottom-level data collection to high-level decision-making. This enables precise modeling and real-time response throughout the entire hazard prediction process. Based on the current situation where park energy management systems generally rely on static load models and rule-driven logic, this solution proposes a hazard prediction mechanism with structural propagation modeling capabilities and supports disturbance identification and response control linkage.

[0103] In practical implementation, the solution starts from the operating status of equipment and environment. It collects operating data such as energy consumption, voltage, and current of air conditioners, elevators, lighting, and charging devices through the data collection and integration module, and simultaneously obtains the environmental temperature and humidity information at the corresponding time. Based on a unified field structure, it generates multivariate time series data. The feature extraction module uses Fourier transform with a fixed window length and a sliding window mechanism to extract frequency domain features such as disturbance amplitude, disturbance rate, and disturbance frequency, forming a standard disturbance vector sequence with time continuity and frequency characteristics. This vector sequence provides key feature support for subsequent disturbance path structure modeling and hazard prediction.

[0104] The matching and mapping module pairs perturbation vectors based on the Euclidean distance and perturbation difference threshold, identifies potential interference relationships between devices within the same time period, and generates a set of perturbation propagation paths with pairing strength and directionality. The training and classification module inputs path structure information and perturbation vector features together, constructs a perturbation risk classification model through supervised learning, and performs dynamic model verification and optimization based on classification accuracy and the convergence rate of the loss function. Then, the prediction and generation module generates a hazard prediction map based on the risk judgment result of the current perturbation vector in the model, combined with propagation path information, and marks high-risk device nodes and their associated structures. The sorting and intervention module sorts nodes based on their risk level and path coupling degree, converts the prediction results into executable unloading instructions, cycle reset commands, and visual alarm signals, and writes them into a multivariate time series, completing the complete closed loop of early warning and intervention.

[0105] The modular serial structure design is based on the following considerations: First, to address heterogeneous equipment structures and nonlinear disturbance behaviors, a unified field representation is needed to enhance the fusion computing capabilities between systems. Second, multi-source dynamic features exhibit complex coupling characteristics, requiring the introduction of frequency domain analysis and structural propagation mechanisms for effective modeling. Third, the hazard prediction model must possess training and validation feedback capabilities to ensure classification accuracy and stability. Fourth, early warning outputs must be rapidly converted into intervention commands and written into the original data structure in a traceable closed-loop manner to support the linkage control requirements of the energy management platform. Therefore, this solution possesses good system feasibility and engineering practicality.

[0106] Finally, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart park dynamic analysis and early warning platform, comprising a data collection and organization module, a feature extraction module, a matching and mapping module, a training and classification module, a prediction and generation module, and a ranking and intervention module, characterized in that: The data acquisition and processing module obtains operational and environmental status data from multiple devices and organizes them into a unified multivariate time series set according to device category and timestamp; The feature extraction module extracts a perturbation feature group consisting of perturbation amplitude, perturbation rate and perturbation frequency by performing Fourier transform with a fixed window length and sliding window segmentation on the multivariate time series set, and generates a standard perturbation vector sequence after normalization. The matching and mapping module performs pairing statistics on vector pairs with the same timestamp and an Euclidean distance below a threshold in the standard disturbance vector sequence, and constructs a set of disturbance propagation paths based on device identification and pairing frequency; The matching and mapping module includes extracting data with the same timestamp from the standard perturbation vector sequence, constructing a perturbation vector combination group at the corresponding sampling time, and treating any two perturbation vectors in each group as a set of calculation objects; Perform Euclidean distance calculation on each set of computational objects to form an Euclidean distance matrix containing each perturbation vector pair. The judgment conditions include that the perturbation amplitude difference in each vector pair is less than the preset perturbation difference threshold, and the value of the Euclidean distance calculation is lower than the set distance similarity threshold. If both judgment conditions are met simultaneously, the pair is identified as a valid perturbation vector pair and output to the perturbation vector pair set. Otherwise, reconstruct the perturbation vector combination group under that timestamp and re-perform the Euclidean distance calculation; Using the device number and device type of each perturbation vector pair in the perturbation vector pair set as the matching key, perform pairing frequency statistics and timestamp location recording to generate a perturbation linkage statistics record table containing the paired device number, pairing frequency and sampling time index; The device number pairs in the disturbance linkage statistics record table that have reached the specified lower limit of pairing frequency are constructed into directed node pairs, and directional labels are established according to the sampling time index to generate a set of disturbance propagation paths containing node number, connection weight and time label fields; The training classification module inputs the set of perturbation propagation paths and the standard perturbation vector sequence into the classification calculation process, and establishes the mapping relationship between perturbation feature groups and risk labels by training the perturbation risk classification model; The prediction generation module inputs the current standard perturbation vector into the perturbation risk classification model, matches the risk label with the propagation path node number, and generates a hazard prediction map containing node index and risk number. The sorting intervention module sorts the nodes in the hazard prediction map by risk number and path connectivity, and generates load unloading instructions, cycle reset instructions and visual alarm signals in sequence, and writes the execution data into a time series set.

2. The smart park dynamic analysis and early warning platform according to claim 1, characterized in that: The data acquisition and aggregation module obtains the unit time energy consumption, voltage change, current response, and load ratio of air conditioning equipment, lighting equipment, elevator drive devices, and electric vehicle charging devices, as well as the ambient temperature and humidity values ​​corresponding to each sampling, by setting the sampling period. Then, it tags the operating status data of each device and the environmental status data according to the sampling time and outputs the synchronized data group of device data and environmental data. Each set of running status data and environmental status data in the synchronized data group is matched at the field level according to the timestamp. An aligned data structure containing sampling time, device type, device number, running status parameters and environmental status parameters is constructed through field correspondence. The output is a device-level bound data block with a unified field order and complete dimensions. Perform unified device type encoding, field position regularization, and time axis alignment operations on the device-level bound data blocks. Arrange all data blocks according to the set time granularity and field structure, and output a multivariate time series set with consistent field structure, clear device identification, and continuous time.

3. The smart park dynamic analysis and early warning platform according to claim 2, characterized in that: The feature extraction module groups the multivariate time series set by device number, obtains the time series segment corresponding to each type of device, and then performs a Fourier transform with a fixed window length on each segment to extract the frequency domain signal containing the frequency distribution of energy consumption change, voltage change and current response, and outputs the frequency domain signal set of the corresponding device. By setting a time sliding window with a fixed width and a fixed overlap ratio, continuous segments in the frequency domain signal set are obtained. Then, the truncation operation is performed according to the window movement order to form a sequence of frequency domain sub-segments with continuous time indices. The output is a set of frequency domain time distributions containing all window segments.

4. The smart park dynamic analysis and early warning platform according to claim 3, characterized in that: In the feature extraction module, the disturbance amplitude is obtained by calculating the difference between the maximum and minimum values ​​of the signal amplitude of each sub-segment in the frequency domain time distribution set, the disturbance rate is obtained by calculating the slope of amplitude change between adjacent segments, the disturbance frequency is obtained by locating the main frequency point of each sub-segment, and then the three disturbance quantities are combined to form a disturbance feature data set, and the disturbance feature set is output. The perturbation amplitude, perturbation rate, and perturbation frequency in the perturbation feature set are linearly normalized according to preset upper and lower limits, and all feature data sets are rearranged by time index to generate a vector format expression with consistent fields and stable order, and output a standard perturbation vector sequence.

5. The smart park dynamic analysis and early warning platform according to claim 4, characterized in that: The training classification module extracts node numbers, connection weights, and time labels from the set of perturbation propagation paths, and combines them with the perturbation amplitude, perturbation rate, and perturbation frequency of the corresponding timestamps in the standard perturbation vector sequence to construct a joint feature input set containing path features and perturbation features. The joint feature input group is subjected to training set partitioning operation, and a classification label set is generated based on the historical running status label or manual annotation result corresponding to each path, thus constructing training data pairs between feature groups and risk labels; The training data is input into the classification calculation process, and feature encoding, weight initialization and iterative optimization operations are performed to train the perturbation risk classification model and form a mapping structure between input features and output labels. The trained perturbation risk classification model is validated by calculating the classification accuracy of the joint feature input group on the test data and solving the convergence rate of the loss function based on the risk label prediction bias and the rate of change of training rounds. If the classification accuracy is higher than the lower limit of the recognition accuracy and the convergence rate exceeds the threshold of the loss decrease rate, then the model is confirmed to meet the training effectiveness requirements, the perturbation risk classification model is output and input into the prediction generation module to perform the risk map generation operation; if any judgment condition is not met, then the feature encoding and weight initialization are re-executed on the current training data pair to construct a new perturbation risk classification model.

6. The smart park dynamic analysis and early warning platform according to claim 5, characterized in that: The prediction generation module inputs the standard disturbance vector at the current moment into the disturbance risk classification model, performs risk classification calculation, and outputs the risk category number corresponding to the standard disturbance vector; Based on the time index and device number of the standard disturbance vector, the corresponding node number and connection path information are extracted from the disturbance propagation path set to construct an index relationship set between device nodes and risk propagation paths; Perform a pairing operation between the risk category number and the index relationship set, combine them in the order of node number to form a risk node set, and attach the corresponding risk category number to each device node; The device number, node number, time index, and risk category number in the risk node set are combined to construct a data graph with a consistent field structure, and the output is a hazard prediction graph containing node information, time information, and risk level.

7. The smart park dynamic analysis and early warning platform according to claim 6, characterized in that: The ranking intervention module extracts the corresponding risk category number and the connectivity value of the node in the set of disturbance propagation paths for each node in the hazard prediction map, and constructs a ranking feature set containing risk intensity parameters and structural coupling parameters. Perform a two-factor sorting operation on the sorting feature set, arranging them from high to low according to the risk category number. If the risk category numbers are the same, they are arranged from high to low according to the connectivity value, generating a sorted high-priority device node queue. Each node in the high-priority device node queue is matched sequentially with its device number and current operating status, and combined with the risk category number to generate corresponding load unloading instructions, operating cycle reset instructions and visual alarm signals. After executing the load unloading command, cycle reset command, and visual alarm signal generated by each node, extract the command success mark, reset feedback value, and alarm trigger record corresponding to that node, and write them into a multivariate time series set consistent with the time index of that node to construct a data update record that matches the intervention action.

Images