Power grid dynamic scheduling decision-making method and device based on multi-modal prediction, electronic equipment and storage medium

Abstract

The invention discloses a power grid dynamic scheduling decision-making method and device based on multi-modal prediction, electronic equipment and a storage medium, and belongs to the field of power system regulation and control operation, and the method comprises the steps: obtaining internal state data and external working condition data of each target power grid device, and a future load change curve of a related power transmission and distribution line, and an equipment feature matrix is constructed through space-time alignment. And inputting the feature matrix into a multi-modal neural network, and outputting the health index, the remaining service life and the fault probability. When the equipment health index is lower than a threshold value, a multi-objective optimization model is constructed, a preventive scheduling strategy is generated, and scheduling is executed; and when the equipment fault probability exceeds a set threshold value, updating the power grid line weight based on load prediction, generating a topology reconstruction scheme of the minimum power failure range, and scheduling according to the topology reconstruction scheme. By implementing the method and the device, the problem that the long-term degradation trend and the short-term sudden risk of the equipment cannot be accurately predicted due to single data dimension in the prior art can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system regulation and operation, in particular to a power grid dynamic scheduling decision method and device based on multi-modal prediction, electronic equipment and storage medium. BACKGROUND

[0002] Power grid dynamic scheduling decision is the core link to ensure the safe, stable and economic operation of modern power systems. With the large-scale integration of new energy and the increasing complexity of user load, the operating conditions of power grids change rapidly, which puts unprecedented high requirements on the real-time, foresight and accuracy of scheduling decisions. An excellent dynamic scheduling decision system can achieve optimal allocation of power resources, reduce network loss, and respond quickly to potential grid risks while ensuring power supply reliability, which is self-evident in importance.

[0003] However, the existing power grid dynamic scheduling decision method gradually exposes its inherent technical bottlenecks when dealing with complex operating scenarios. On the one hand, existing scheduling decisions mainly rely on single, macroscopic electrical quantity data (such as current and voltage), lacking comprehensive consideration and deep integration of multi-dimensional information such as the physical state of each grid element (such as temperature and vibration) and the external environment (such as weather and load pressure). This limitation at the data level leads to delays and deviations in the system's assessment of equipment health status, making it difficult to accurately predict its long-term degradation trend and short-term sudden failure risk. On the other hand, existing scheduling methods mostly use decision-making modes based on static rule libraries or fixed thresholds, with relatively single response strategies. When facing two scenarios with completely different natures, i.e. slow aging of equipment requiring optimization of maintenance and sudden failure requiring emergency isolation, it is difficult to generate differentiated and optimal response strategies that fully match the current operating conditions, thereby affecting the effective utilization of assets and the rapid recovery from failures. SUMMARY

[0004] The embodiments of the present application provide a power grid dynamic scheduling decision method and device based on multi-modal prediction, electronic equipment and storage medium, which can solve the problem of inaccurate prediction of long-term degradation trend and short-term sudden risk of equipment due to single data dimension in the prior art.

[0005] An embodiment of the present application provides a power grid dynamic scheduling decision method based on multi-modal prediction, comprising:

[0006] Obtaining internal state data of each target grid device, external operating condition data of each target grid device, and a future load change curve of a target power transmission and distribution line;

[0007] Performing a spatio-temporal alignment operation on the internal state data and external operating condition data of each target grid device to generate a feature matrix of each target grid device;

[0008] input the respective target power grid equipment feature matrix to a preset multi-modal neural network in sequence, so that the multi-modal neural network respectively generates a corresponding health index, a remaining service life and a failure probability according to the respective target power grid equipment feature matrix;

[0009] When the health index of the target power grid equipment is lower than a first preset threshold, a multi-objective optimization model and related constraints are constructed according to the remaining service life of the current equipment and the future load change curve of the target power transmission and distribution line, with the objective of minimizing the maintenance delay risk and economic cost; the multi-objective optimization model is solved to generate a preventive scheduling strategy; and power grid scheduling is performed according to the preventive scheduling strategy;

[0010] When the failure probability of the target power grid equipment is higher than a second preset threshold, the line weight in the power grid topology is dynamically updated according to the future load change curve of each target power transmission and distribution line, and a power grid topology reconfiguration scheme with a minimum power outage range is generated according to the line weight; power grid scheduling is performed according to the power grid topology reconfiguration scheme.

[0011] Further, the spatiotemporal alignment operation is performed on the internal state data and the external working condition data of each target power grid equipment to generate a target power grid equipment feature matrix, including:

[0012] For each target power grid equipment, the internal state data and the external working condition data of the current target power grid equipment are subjected to noise suppression processing and missing compensation processing to generate denoised internal state data and denoised external working condition data of the current target power grid equipment;

[0013] The denoised internal state data and the denoised external working condition data of the current target power grid equipment are subjected to alignment processing and spatial correlation processing to generate a current target power grid equipment feature matrix.

[0014] Further, the multi-objective optimization model is specifically:

[0015]

[0016] In the formula, f1 is an operation risk and economic cost function; f2 is a power grid stability penalty function; t is the maintenance time; P transfer is the total load migration amount;

[0017] The operation risk and economic cost function is specifically:

[0018]

[0019] In the formula, w1 is the first weight coefficient of the operation risk and economic cost function; w2 is the second weight coefficient of the operation risk and economic cost function; k is a risk aversion coefficient; RUL current is the current device remaining useful life; C loss is the unit power economic loss coefficient;

[0020] The power grid stability penalty function is specifically:

[0021]

[0022] In the formula, j is the index of the adjacent bus; L j is the current load of the jth adjacent bus; ΔP j is the load amount allocated to the jth adjacent bus; L max,j is the maximum allowable load of the jth adjacent bus.

[0023] Further, the related constraints include time constraints, capacity constraints and budget constraints:

[0024] The time constraint is specifically: t≤RUL current -RUL min ;

[0025] In the formula, RUL min is the minimum safety margin;

[0026] The capacity constraint is specifically: L j +ΔP j ≤L max,j ;

[0027] The budget constraint is specifically: C loss ·P transfer ≤C budget ;

[0028] In the formula, C budget is the total economic budget of the scheduling operation.

[0029] Further, the training of the multi-modal neural network comprises:

[0030] obtaining a historical operation data set of a target power grid device; wherein the historical operation data set comprises a plurality of historical data samples, and each historical data sample comprises a device feature matrix corresponding to the target power grid device and a corresponding label set; the label set comprises a real health index label, a real remaining useful life label and a fault label for indicating whether a fault exists;

[0031] randomly dividing the historical operation data set into a plurality of batches of training samples according to a preset batch size;

[0032] The device feature matrix of each training sample is sequentially input into the multi-modal neural network for iterative training until a preset number of training rounds is reached; wherein in each iteration, the multi-modal neural network outputs a prediction result including a predicted health index, a predicted remaining useful life, and a failure prediction probability according to the device feature matrix in the current training sample; and a multi-task combined loss function is used to calculate a loss function value according to the prediction result and a label set corresponding to the training sample; and a preset optimizer is used to update the learnable network parameters in the multi-modal neural network according to the loss function value.

[0033] Further, the multi-task combined loss function is specifically:

[0034] L total =λ1·L RUL +λ2·L Fault +λ3·L HI

[0035] In the formula, L total is a multi-task combined loss function; L RUL is a remaining useful life loss; I Fault is a failure loss; L HI is a health index loss; λ1 is a weight hyperparameter of the remaining useful life loss; λ2 is a weight hyperparameter of the failure loss; and λ3 is a weight hyperparameter of the health index loss.

[0036] The remaining useful life loss is specifically:

[0037]

[0038] In the formula, N is the number of samples in a batch size; i is a sample index; is a predicted remaining useful life value of the i-th sample; is a true remaining useful life value of the i-th sample;

[0039] The failure loss is specifically:

[0040]

[0041] In the formula, y i is a true binary failure label of the i-th sample; p i is a predicted failure probability of the i-th sample by the neural network;

[0042] The health index loss is specifically:

[0043]

[0044] In the formula, This is the predicted health index value for the i-th sample. Let be the true value of the health index for the i-th sample.

[0045] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0046] One embodiment of the present invention provides a power grid dynamic dispatch decision-making device based on multimodal prediction, comprising: a multi-source data sensing module, a multimodal feature fusion module, an intelligent state assessment module, a preventive optimization dispatch module, and an emergency fault reconstruction module;

[0047] The multi-source data sensing module is used to acquire the internal status data of each target power grid device, the external operating condition data of each target power grid device, and the future load change curve of the target transmission and distribution lines.

[0048] The multimodal feature fusion module is used to perform spatiotemporal alignment operations on the internal state data and external operating condition data of each target power grid device to generate a feature matrix of each target power grid device.

[0049] The intelligent condition assessment module is used to input the feature matrices of each target power grid device sequentially into a preset multimodal neural network, so that the multimodal neural network generates corresponding health indices, remaining service life and failure probabilities according to the feature matrices of each target power grid device.

[0050] The preventive optimization scheduling module is used to construct a multi-objective optimization model and related constraints based on the remaining service life of the current equipment and the future load change curve of the target transmission and distribution lines, with the goal of minimizing maintenance delay risk and economic cost, when the health index of a target power grid device is lower than a first preset threshold; solve the multi-objective optimization model to generate a preventive scheduling strategy; and perform power grid scheduling according to the preventive scheduling strategy.

[0051] The emergency fault reconstruction module is used to dynamically update the line weights in the power grid topology based on the future load change curves of each target transmission and distribution line when the fault probability of a target power grid device is higher than a second preset threshold, and generate a power grid topology reconstruction scheme with the minimum power outage range based on the line weights; and perform power grid dispatching based on the power grid topology reconstruction scheme.

[0052] Furthermore, the multimodal feature fusion module includes: a data denoising unit and a data spatiotemporal alignment unit;

[0053] The data denoising unit is configured to, for each target power grid device, perform noise suppression processing and missing compensation processing on internal state data and external working condition data of the current target power grid device, and generate denoised internal state data and denoised external working condition data of the current target power grid device.

[0054] The data space-time alignment unit is configured to perform alignment processing and space correlation processing on the denoised internal state data and the denoised external working condition data of the current target power grid device, and generate a feature matrix of the current target power grid device.

[0055] On the basis of the above-mentioned method embodiment, the present application correspondingly provides an electronic device embodiment.

[0056] An embodiment of the present application provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and the processor implements the power grid dynamic scheduling decision method based on multi-modal prediction in any of the above-mentioned method embodiments when executing the computer program.

[0057] On the basis of the above-mentioned method embodiment, the present application correspondingly provides a storage medium embodiment.

[0058] An embodiment of the present application provides a storage medium having a computer program stored thereon, wherein the storage medium controls a device where the storage medium is located to execute the power grid dynamic scheduling decision method based on multi-modal prediction in any of the above-mentioned method embodiments when the computer program runs.

[0059] Compared with the prior art, the present application has the following beneficial effects:

[0060] The embodiment of the present application provides a power grid dynamic scheduling decision method, device, electronic device and storage medium based on multi-modal prediction. The method first acquires internal state data, external working condition data and future load change curve of related transmission and distribution lines of each target power grid device, and constructs a device feature matrix through space-time alignment. Then, the feature matrix is input into a multi-modal neural network, and a health index, a remaining useful life and a failure probability are output. When the device health index is lower than a threshold value, a multi-objective optimization model is constructed to minimize the delay risk and economic cost of maintenance, a preventive scheduling strategy is generated and executed; when the device failure probability exceeds a set threshold value, the power grid line weight is updated based on load prediction, a topology reconstruction scheme with the minimum power outage range is generated and scheduled accordingly, and the dynamic scheduling decision of the power grid is realized.

[0061] The present application solves the problem that the prior art cannot accurately predict the long-term degradation trend and short-term sudden risk of equipment due to single data dimension by comprehensively collecting and fusing the internal state and external working condition data of each power grid equipment, and using a multi-modal neural network for deep analysis. Further, based on the accurate health index and fault probability prediction, the present application constructs a differentiated preventive optimization scheduling and emergency fault reconstruction decision path, overcomes the defects of single response strategy of traditional methods using static rules, and significantly improves the intelligentization and self-adaptation ability of power grid scheduling. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 is a flowchart of a power grid dynamic scheduling decision method based on multi-modal prediction provided by an embodiment of the present application.

[0063] Figure 2 is a structural diagram of a power grid dynamic scheduling decision device based on multi-modal prediction provided by an embodiment of the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0065] As shown in Figure 1 To solve the problem that the prior art cannot accurately predict the long-term degradation trend and short-term sudden risk of equipment due to single data dimension, an embodiment of the present application provides a power grid dynamic scheduling decision method based on multi-modal prediction, which at least includes the following steps:

[0066] Step S1, obtaining internal state data of each target power grid equipment, external working condition data of each target power grid equipment, and a future load change curve of a target power transmission and distribution line;

[0067] Specifically, the present method starts with comprehensive data acquisition of one or more target power grid equipment as monitoring objects. The target power grid equipment is a key physical entity in the power grid. In the present embodiment, the target power grid equipment is a transformer, a circuit breaker, and a cable joint as a key node of the power transmission and distribution line. For each target power grid equipment, the present method synchronously acquires three types of core data to construct a data basis required for subsequent analysis.

[0068] The acquired data first includes internal state data reflecting the physical health of the target power grid equipment itself, specifically, the internal state data includes: infrared thermal imaging maps collected by an infrared thermal imager to detect local hot spots; vibration spectrum data collected by a three-axis vibration sensor to represent mechanical structure abnormalities; and current and voltage waveform data collected by a current transformer to indicate electrical performance. Second, the acquired data also includes external working condition data reflecting the stress of the equipment operating environment, specifically, the external working condition data includes environmental temperature and humidity affecting the heat dissipation efficiency of the equipment and weather conditions such as wind speed that can cause load changes.

[0069] Finally, in order to prospectively assess the work stress that the target power grid equipment will bear and provide a basis for subsequent scheduling decisions, the method also acquires the future load change curve of the target power transmission and distribution line associated with the target power grid equipment. Here, the target power transmission and distribution line specifically includes lines that have a direct electrical connection with the target power grid equipment and adjacent lines that are preset as potential load transfer paths for the target power grid equipment in the power grid topology. In this embodiment, the future load change curve is not directly generated by the multi-modal neural network in the method, but is acquired as a known external working condition data. The acquisition of this future load change curve can come from a dedicated load prediction system in the power grid based on existing technologies, including linear regression models, support vector machines (SVM) or independent neural network models.

[0070] By comprehensively and synchronously collecting these three types of data with different sources and different physical meanings, a perception system is constructed that can provide rich and complete feature inputs for the subsequent multi-modal neural network, thereby laying a solid data foundation for precise state assessment and dynamic decision-making.

[0071] Step S2, performing spatio-temporal alignment operation on the internal state data and external working condition data of each target power grid equipment to generate a feature matrix of each target power grid equipment;

[0072] In a preferred embodiment, the spatio-temporal alignment operation on the internal state data and external working condition data of each target power grid equipment to generate a feature matrix of each target power grid equipment includes:

[0073] For each target power grid equipment, performing noise suppression processing and missing compensation processing on the internal state data and external working condition data of the current target power grid equipment to generate denoised internal state data and denoised external working condition data of the current target power grid equipment;

[0074] The internal state data and the external working condition data of the current target power grid equipment after denoising are aligned and spatially correlated to generate a current target power grid equipment feature matrix.

[0075] Specifically, in one specific embodiment of the present application, after obtaining the internal state data and the external working condition data of each target power grid equipment, a series of preprocessing and spatio-temporal alignment operations are required on these original multi-modal data with different sources and formats. The purpose of this step is to convert the original data into a structured input format that can be effectively processed by the subsequent multi-modal neural network, i.e., a device feature matrix for each target power grid equipment.

[0076] To achieve this purpose, for each target power grid equipment, the related internal state data and external working condition data are first cleaned in the present embodiment. Specifically, the present embodiment adopts appropriate processing methods for different modal data: for high-frequency time series signals such as vibration, wavelet threshold denoising method is used to filter out electromagnetic and other high-frequency interference; for electrical quantity data such as current and voltage, Kalman filter is used for smoothing and correction; and for two-dimensional image data such as infrared thermal imaging spectrum, median filter algorithm is used to eliminate random noise points and improve image quality. Moreover, for the missing points or null values that may occur in data collection, the present embodiment uses linear interpolation or autoregressive difference model to fill in, thereby generating a higher-quality, denoised and compensated internal state data and external working condition data.

[0077] Then, the denoised and compensated data is unified in time and space. In the time dimension, through the GPS synchronization module or the hardware clock synchronization mechanism based on the Network Time Protocol (NTP), all sources of data are given a unified high-precision timestamp, achieving time alignment. In the spatial dimension, the present embodiment pre-constructs a three-dimensional coordinate model of the target power grid equipment, and through coordinate conversion algorithm, the readings of sensors at different positions (such as two-dimensional pixel points of infrared thermal imager) are accurately mapped and correlated to specific physical coordinates on the three-dimensional model, achieving spatial correlation.

[0078] Finally, in order for the subsequent neural network to be able to distinguish the source of the data it is processing, the present embodiment adds an identity feature to the data after the above processing, which uniquely identifies the current target power grid equipment. After a series of operations such as noise suppression, missing compensation, spatio-temporal alignment and identity labeling, the originally heterogeneous and chaotic multi-modal data is finally integrated into a structured and highly condensed information target power grid equipment feature matrix, providing high-quality input for subsequent intelligent state assessment and prediction.

[0079] Step S3, inputting the respective target power grid equipment feature matrices into a preset multi-modal neural network in sequence, so that the multi-modal neural network respectively generates a corresponding health index, a remaining service life, and a failure probability according to the respective target power grid equipment feature matrices;

[0080] Specifically, in one specific embodiment of the present application, after generating the exclusive equipment feature matrix for each target power grid equipment, the method inputs these feature matrices into a pre-trained multi-modal neural network in sequence to perform deep and quantitative evaluation and prediction on the health condition of the equipment. The internal processing procedure of the neural network aims to mine deep patterns indicating performance degradation or potential failure of the equipment from the highly complex feature matrix.

[0081] The multi-modal neural network internally includes multiple parallel feature encoders. For data of different modalities in the equipment feature matrix, the neural network adopts a corresponding encoder to perform primary feature extraction: for time-series electrical quantity data, a bidirectional long short-term memory network (Bi-LSTM) is adopted to capture its dependence relationship and evolution trend in the time dimension; for vibration spectrum and other data, a one-dimensional convolutional neural network (1D-CNN) is adopted to extract its local key features in the frequency domain; and for infrared thermal imaging spectrum and other two-dimensional image data, a classical two-dimensional convolutional neural network (2D-CNN) such as ResNet is adopted to extract its spatial thermal distribution features.

[0082] The multiple primary features extracted by each encoder are then sent to an attention fusion layer. The fusion layer dynamically calculates the correlation strength and importance between different modal features through a cross-attention mechanism and performs weighted fusion, thereby generating a high-dimensional fusion feature vector that can comprehensively represent the current comprehensive state of the equipment. Finally, this fusion feature vector is input into a multi-task decoder. The decoder includes multiple parallel output layers, each of which is responsible for a specific prediction task, and finally respectively decodes to generate the three core indicators of the health index, the remaining service life, and the failure probability corresponding to the current target power grid equipment.

[0083] Through this deep learning processing procedure of modality-specific encoding, attention fusion, and multi-task decoding, the method can mine deep degradation patterns and failure signs from complex and heterogeneous data, and realize high-precision, multi-dimensional, and forward-looking quantitative evaluation of the health condition of the equipment.

[0084] In one preferred embodiment, the training of the multi-modal neural network includes:

[0085] obtain a historical operation data set of a target power grid equipment; wherein the historical operation data set comprises a plurality of historical data samples, and each historical data sample comprises an equipment feature matrix of the corresponding target power grid equipment and a corresponding label set; the label set comprises a real health index label, a real remaining useful life label, and a fault label for indicating whether a fault exists;

[0086] randomly divide the historical operation data set into a plurality of batch training samples according to a preset batch size;

[0087] input the equipment feature matrix of each training sample into the multi-modal neural network in sequence, and perform iterative training until a preset number of training rounds is reached; wherein in each iteration, the multi-modal neural network outputs a prediction result comprising a predicted health index, a predicted remaining useful life, and a fault prediction probability according to the equipment feature matrix in the current training sample; and a multi-task combined loss function is used to calculate a loss function value according to the prediction result and the label set corresponding to the training sample; and a preset optimizer is used to update the learnable network parameters in the multi-modal neural network according to the loss function value.

[0088] In a preferred embodiment, the multi-task combined loss function is specifically:

[0089] L total =λ1·L RUL +λ2·L Fault +λ3·L HI

[0090] wherein L total is the multi-task combined loss function; L RUL is the remaining useful life loss; L Fault is the fault loss; L HI is the health index loss; λ1 is a weight hyperparameter of the remaining useful life loss; λ2 is a weight hyperparameter of the fault loss; and λ3 is a weight hyperparameter of the health index loss.

[0091] wherein the remaining useful life loss is specifically:

[0092]

[0093] wherein N is the number of samples in the batch size; i is the sample index; is the predicted remaining useful life value of the i-th sample; is the real remaining useful life value of the i-th sample;

[0094] the fault loss is specifically:

[0095]

[0096] wherein y i is the true binary fault label of the i-th sample; p i is the predicted failure probability of the i-th sample by the neural network;

[0097] the health index loss, in particular:

[0098]

[0099] wherein, is the predicted health index value of the i-th sample; is the true health index value of the i-th sample.

[0100] Specifically, in one specific embodiment of the present application, in order to enable the multi-modal neural network to have accurate prediction capability, supervised training needs to be performed on the multi-modal neural network. The training process is usually completed in a cloud environment with sufficient computing resources, and the core is to enable the neural network to learn the complex mapping relationship from the input device feature matrix to the future state evaluation result of the device.

[0101] The training process starts with obtaining a historical operation data set of a target power grid device. The historical operation data set contains a large amount of historical data samples, wherein each historical data sample is composed of two parts: one part is a device feature matrix describing the state of the device at a preset time window, and the other part is a label set corresponding to the device feature matrix. In the present embodiment, the label set specifically includes: a true health index label, a true remaining useful life label, and a binary fault label used to represent whether there is a fault.

[0102] In order to generate the above-mentioned labels, the present embodiment systematically processes original records such as historical operation and maintenance work orders, fault reports, and device account books. Among them, the true remaining useful life label is generated by using the time backtracking method; specifically, the time point at which the device occurs a specific failure event is determined from the historical records, and the past is traced back from this end point, so as to calculate the true remaining time of any historical time point from the failure event. The true health index label is calculated by normalization mapping based on the generated remaining useful life label, which quantifies the long-term degradation process of the device into a unified interval of 0 to 1. The binary fault label uses a direct calibration method, according to the historical fault records, the time when the fault occurs and the samples in a small time window before the fault are assigned as 1, and the remaining normal samples are assigned as 0.

[0103] After obtaining and generating the complete labeled historical running dataset, the system randomly divides the historical running dataset into several batches of training samples according to a preset batch size. After the start of iterative training, the device feature matrix of each training sample is input into the multi-modal neural network in turn. In each iteration, the multi-modal neural network will synchronously output a prediction result containing the predicted health index, the predicted remaining useful life and the failure prediction probability according to the currently input device feature matrix. The system uses a preset multi-task combination loss function to measure the overall difference between the prediction result and the true label set corresponding to the training sample. The total loss value calculated is then used by a preset optimizer, such as the Adam optimizer, to adjust all learnable network parameters in the network through the backpropagation algorithm, with the goal of minimizing the total loss function value over the entire training set.

[0104] Through such a rigorous multi-task, supervised training process, a multi-modal neural network is finally obtained, which has the comprehensive ability to accurately and synchronously predict the future health status, life trend and sudden failure risk of the device from the complex device feature matrix, providing reliable and quantitative input basis for subsequent intelligent scheduling decisions.

[0105] Step S4, when the health index of the target power grid device is lower than the first preset threshold, a multi-objective optimization model and related constraints are constructed according to the current device remaining useful life and the future load change curve of the target power transmission and distribution line, with the goal of minimizing the maintenance delay risk and economic cost; the multi-objective optimization model is solved to generate a preventive scheduling strategy; and power grid scheduling is performed according to the preventive scheduling strategy;

[0106] In a preferred embodiment, the multi-objective optimization model is specifically:

[0107]

[0108] In the formula, f1 is an operation risk and economic cost function; f2 is a power grid stability penalty function; t is the maintenance time; P transfer is the total load migration amount;

[0109] The operation risk and economic cost function is specifically:

[0110]

[0111] In the formula, w1 is the first weight coefficient of the operation risk and economic cost function; w2 is the second weight coefficient of the operation risk and economic cost function; k is the risk aversion coefficient; RUL current is the current device remaining useful life; C lossa unit power economic loss coefficient;

[0112] the power grid stability penalty function, in particular:

[0113]

[0114] where j is the index of the adjacent bus; L j is the current load of the jth adjacent bus; ΔP j is the load amount allocated to the jth adjacent bus; L max，j is the maximum allowable load of the jth adjacent bus.

[0115] In a preferred embodiment, the correlation constraints include time constraints, capacity constraints, and budget constraints:

[0116] the time constraint, in particular: t≤RUL current -RUL min ;

[0117] where RUL min is the minimum safety margin;

[0118] the capacity constraint, in particular: L j +ΔP j ≤L max，j ;

[0119] the budget constraint, in particular: C loss ·P transfer ≤C budget ;

[0120] where C budget is the total economic budget of the dispatch operation.

[0121] Specifically, in one specific embodiment of the present application, when the output result of the intelligent state assessment and prediction module shows that the health index of a certain target power grid device is lower than the first preset threshold value, the system will trigger the preventive optimization dispatch process. This process aims to actively and prospectively manage the risks of devices that have been identified as having long-term degradation risks. The core purpose is no longer to simply perform maintenance, but to calculate an optimal dispatch scheme after comprehensively considering multiple conflicting goals such as device failure risks, economic costs, and power grid stability. To achieve this purpose, the present method constructs and solves a multi-objective optimization model to generate a specific preventive dispatch strategy.

[0122] In the embodiment, the multi-objective optimization model includes at least two objective functions that need to be minimized simultaneously. The first objective function is an operation risk and economic cost function. The function aims to quantify the comprehensive cost of different scheduling schemes, which is a combination of the increased equipment failure risk due to postponing maintenance and the direct economic loss caused by performing load migration operations. The second objective function is a grid stability penalty function. The function is used to evaluate the impact of the load migration scheme on the rest of the grid, mainly penalizing schemes that may cause overload or load imbalance of adjacent buses or lines receiving the load, to ensure that the scheduling operation does not trigger new grid stability problems. When solving the above objective functions to find the optimal solution, the model must be within the boundaries of the following three core constraints. One is the time constraint, which is determined by the remaining service life prediction value of the equipment and the preset minimum safety margin, ensuring that any maintenance plan must be completed safely before the predicted failure time of the equipment. Two is the capacity constraint, which is calculated based on the obtained future load variation curve of the target power transmission and distribution line, ensuring that the final load of any adjacent line or bus selected as the load migration path does not exceed its physical or operational capacity limit. Three is the budget constraint, which is limited by the preset total economic budget of the scheduling operation, ensuring that the cost of the entire preventive scheduling scheme is within the acceptable financial range. By solving this complex model containing multiple optimization objectives and multiple constraints, the system can generate a specific preventive scheduling strategy that not only gives the optimal equipment maintenance time window but also plans the optimal load migration path. Finally, the method performs grid scheduling according to this optimal strategy, thereby ensuring equipment safety and avoiding unplanned outages while minimizing operational costs and disturbances to the grid, achieving a shift from passive response to active, fine-grained risk management.

[0123] Step S5, when the failure probability of the target grid equipment is higher than the second preset threshold, dynamically updating the line weight in the grid topology according to the future load variation curve of each target power transmission and distribution line, and generating a grid topology reconfiguration scheme of the minimum power outage range according to the line weight; performing grid scheduling according to the grid topology reconfiguration scheme.

[0124] Specifically, in one specific embodiment of the present application, when the output result of the intelligent state assessment and prediction module shows that the failure probability of a certain target grid equipment is higher than the second preset threshold, the system will trigger an emergency fault reconstruction process. This process aims to quickly and automatically respond and handle the imminent sudden and high-risk equipment failure. The process first accurately locates the target grid equipment at risk according to the failure probability information and immediately sends a trip command to the circuit breaker associated with the equipment to perform emergency isolation, thereby preventing the failure from spreading and forming an initial power outage area.

[0125] After the initial isolation is completed, in order to restore power supply to the outage area with minimal impact, the method needs to dynamically evaluate and select the optimal backup power supply path. To this end, the system dynamically updates the weight of all available lines in the current power grid topology according to the obtained future load variation curve of each target transmission and distribution line. Specifically, the dynamic updating of the line weight first includes calculating the dynamic available capacity of each line. The dynamic available capacity C avail (u,v) of a line (u,v) is calculated by the following formula:

[0126] C avail (u,v) = L max (u,v) - max(L curve (u,v))

[0127] In the formula, C avail (u,v) is the dynamic available capacity of the line (u,v); L max (u,v) is the upper limit of the physical capacity of the line (u,v), which is a system preset value; L curve (u,v) is the future load variation curve of the line (u,v) obtained; max(·) is a function of taking the maximum value, which here refers to the peak load in the obtained future load variation curve.

[0128] After obtaining the dynamic available capacity, the system assigns a weight w(u,v) to each line (u,v) according to the following rules. For the line connected to the faulty device that has been isolated, its weight is:

[0129] w(u,v) = ∞

[0130] For all other available lines, their weights are:

[0131]

[0132] In the formula, w(u,v) is the dynamic weight of the line (u,v). ɑ is a preset basic weight coefficient for scaling.

[0133] After calculating the dynamic weight of all available lines in the power grid topology, the method runs the Dijkstra algorithm in the power grid topology after the fault isolation is performed. The algorithm aims to find the path with the smallest total weight while ensuring that the capacity of the selected path can meet the load demand of the area to be restored, thereby finding an optimal power supply path. This optimal path is finally converted into a power grid topology reconstruction scheme with a specific sequence of switch actions and a minimum outage range.

[0134] Finally, the method performs power grid scheduling according to the generated power grid topology reconstruction scheme, and sends the switch action sequence to the field intelligent switch device for execution through an industrial communication protocol. Through this emergency processing mechanism based on real-time available capacity dynamic routing, the method can generate and execute the optimal fault recovery strategy at the fastest speed under the premise of ensuring the safety of the power grid, minimize the power outage range and impact caused by sudden faults, and significantly improve the resilience and power supply reliability of the power grid.

[0135] On the basis of the above-mentioned method embodiment, the application provides a device embodiment.

[0136] As shown in the accompanying drawings, Figure 2 An embodiment of the application provides a power grid dynamic scheduling decision device based on multi-modal prediction, which comprises a multi-source data sensing module, a multi-modal feature fusion module, an intelligent state evaluation module, a preventive optimization scheduling module and an emergency fault reconstruction module.

[0137] The multi-source data sensing module is used for acquiring internal state data of each target power grid device, external working condition data of each target power grid device and a future load change curve of a target power transmission and distribution line.

[0138] The multi-modal feature fusion module is used for performing a space-time alignment operation on the internal state data and the external working condition data of each target power grid device to generate a feature matrix of each target power grid device.

[0139] The intelligent state evaluation module is used for sequentially inputting the feature matrix of each target power grid device into a preset multi-modal neural network, so that the multi-modal neural network generates a corresponding health index, residual service life and fault probability according to the feature matrix of each target power grid device.

[0140] The preventive optimization scheduling module is used for, when the health index of a target power grid device is lower than a first preset threshold, constructing a multi-objective optimization model and related constraints according to the residual service life of the current device and the future load change curve of the target power transmission and distribution line, with the objective of minimizing the maintenance delay risk and economic cost, solving the multi-objective optimization model to generate a preventive scheduling strategy, and performing power grid scheduling according to the preventive scheduling strategy.

[0141] The emergency fault reconstruction module is used for, when the fault probability of a target power grid device is higher than a second preset threshold, dynamically updating the line weight in the power grid topology according to the future load change curve of each target power transmission and distribution line, generating a power grid topology reconstruction scheme with the minimum power outage range according to the line weight, and performing power grid scheduling according to the power grid topology reconstruction scheme.

[0142] In a preferred embodiment, the multi-modal feature fusion module comprises a data denoising unit and a data space-time alignment unit.

[0143] The data denoising unit is configured to, for each target power grid device, perform noise suppression processing and missing compensation processing on the internal state data and the external working condition data of the current target power grid device to generate denoised internal state data and denoised external working condition data of the current target power grid device.

[0144] The data space-time alignment unit is configured to perform alignment processing and spatial correlation processing on the denoised internal state data and the denoised external working condition data of the current target power grid device to generate a feature matrix of the current target power grid device.

[0145] It should be noted that the above-described embodiments of the device correspond to the above-described embodiments of the application, and can implement any one of the above-described power grid dynamic scheduling decision methods based on multi-modal prediction. In addition, the above-described embodiments of the device are only illustrative, and the modules described as separate components can or can not be physically separated, and the components displayed as modules can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. In addition, in the device embodiment provided by the application, the connection relationship between the modules indicates that there is a communication connection between them, which can be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement it without creative labor.

[0146] On the basis of the above-mentioned method embodiments of the application, an electronic device embodiment is correspondingly provided.

[0147] An embodiment of the application provides an electronic device, which comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the power grid dynamic scheduling decision method based on multi-modal prediction is implemented, or when the processor executes the computer program, the functions of the modules in the above-mentioned device embodiments are implemented.

[0148] For example, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the application. The one or more modules can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the terminal device.

[0149] The terminal device can be a desktop computer, a notebook computer, a palm computer, a cloud server, and the like. The terminal device can include, but is not limited to, a processor and a memory.

[0150] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, and the like. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The processor is a control center of the terminal device, and connects all parts of the terminal device through various interfaces and lines.

[0151] The memory can be used to store the computer program and / or modules, and the processor realizes various functions of the terminal device by running or executing the computer program and / or modules stored in the memory, and calling data stored in the memory. The memory can mainly include a program storage area and a data storage area. The program storage area can store an operating system, at least one application required by a function, and the like; and the data storage area can store data created according to use of the terminal device, and the like. In addition, the memory can include a high-speed random access memory, and can also include a nonvolatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0152] On the basis of the above-mentioned method embodiment, the application provides a storage medium embodiment;

[0153] Another embodiment of the application provides a storage medium, which includes a stored computer program. When the computer program runs, the device where the storage medium is located performs any one of the above-mentioned power grid dynamic scheduling decision methods based on multi-modal prediction.

[0154] The storage medium is a computer readable storage medium, and the computer program includes computer program code in the form of source code, object code, an executable file, or some intermediate form, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a U disk, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, a software distribution medium, etc.

[0155] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials or characteristics described can be combined in any suitable manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0156] The above is the preferred embodiment of the present application. It should be pointed out that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which are also considered within the scope of protection of the present application.

Claims

1. A power grid dynamic dispatch decision method based on multi-modal prediction, characterized in that, The method comprises the following steps: obtaining internal state data of each target power grid device, external working condition data of each target power grid device, and a future load change curve of a target power transmission and distribution line; performing space-time alignment operation on the internal state data and the external working condition data of each target power grid device to generate a feature matrix of each target power grid device; inputting the feature matrix of each target power grid device into a preset multi-modal neural network in sequence, so that the multi-modal neural network generates a corresponding health index, a remaining service life and a failure probability according to the feature matrix of each target power grid device; when the health index of a target power grid device is lower than a first preset threshold, constructing a multi-objective optimization model and related constraints according to the remaining service life of the current device and the future load change curve of the target power transmission and distribution line, with the objective of minimizing the risk of maintenance delay and economic cost; solving the multi-objective optimization model to generate a preventive scheduling strategy; and performing power grid scheduling according to the preventive scheduling strategy; when the failure probability of a target power grid device is higher than a second preset threshold, dynamically updating the line weight in the power grid topology according to the future load change curve of each target power transmission and distribution line, and generating a power grid topology reconstruction scheme with the minimum power outage range according to the line weight; and performing power grid scheduling according to the power grid topology reconstruction scheme.

2. The multi-modal prediction based dynamic dispatch decision method for power grid according to claim 1, characterized in that, The space-time alignment operation on the internal state data and the external working condition data of each target power grid device to generate a feature matrix of each target power grid device comprises: for each target power grid device, performing noise suppression processing and missing compensation processing on the internal state data and the external working condition data of the current target power grid device to generate denoised internal state data and denoised external working condition data of the current target power grid device; performing alignment processing and spatial correlation processing on the denoised internal state data and the denoised external working condition data of the current target power grid device to generate a feature matrix of the current target power grid device.

3. The multi-modal prediction based dynamic dispatch decision method for power grid according to claim 2, characterized in that, The multi-objective optimization model is specifically: In the formula, f1 is an operation risk and economic cost function; f2 is a power grid stability penalty function; t is the maintenance time; P transfer is the total load migration amount; The operation risk and economic cost function is specifically: In the formula, w1 is a first weight coefficient of the operation risk and economic cost function; w2 is a second weight coefficient of the operation risk and economic cost function; k is a risk aversion coefficient; RUL current is the current equipment remaining useful life; C loss is a unit power economic loss coefficient; The power grid stability penalty function is specifically: where j is the index of the adjacent bus; L j is the current load of the jth adjacent bus; ΔP j is the amount of load allocated to the jth adjacent bus; L max,j is the maximum allowable load of the jth adjacent bus.

4. The multi-modal prediction based dynamic dispatch decision method for power grid according to claim 3, characterized in that, The related constraints include time constraints, capacity constraints and budget constraints: The time constraint, in particular: t≤RUL current -RUL min ; where RUL min is the minimum safety margin; The capacity constraint, in particular: L j + ΔP j ≤ L max,j ; The budget constraint, in particular, is: C loss · P transfer ≤ C budget ; In the formula, C budget is the total economic budget for the scheduling operation.

5. The multi-modal prediction based dynamic dispatch decision method for power grid according to claim 4, characterized in that, The training of the multi-modal neural network comprises: obtaining a historical operation data set of the target power grid device; wherein the historical operation data set comprises a plurality of historical data samples, and each historical data sample comprises a device feature matrix of the corresponding target power grid device and a corresponding label set; the label set comprises a real health index label, a real remaining service life label and a failure label for indicating whether there is a failure; randomly dividing the historical operation data set into a plurality of batches of training samples according to a preset batch size; The device feature matrix of each training sample is sequentially input into the multi-modal neural network for iterative training until a preset number of training rounds is reached; wherein in each iteration, the multi-modal neural network outputs a prediction result including a predicted health index, a predicted remaining useful life, and a failure prediction probability according to the device feature matrix in the current training sample; and a multi-task combined loss function is used to calculate a loss function value according to the prediction result and a label set corresponding to the training sample; and a preset optimizer is used to update the learnable network parameters in the multi-modal neural network according to the loss function value.

6. The multi-modal prediction based dynamic dispatch decision method for power grid as claimed in claim 5, wherein, The multi-task combined loss function is specifically: L total = λ1·L RUL + λ2·L Fault + λ3·L HI In the formula, L total is a multi-task combined loss function; L RUL is a remaining useful life loss; L Fault is a failure loss; L HI is a health index loss; λ1 is a weight hyperparameter of the remaining useful life loss; λ2 is a weight hyperparameter of the failure loss; and λ3 is a weight hyperparameter of the health index loss. The remaining useful life loss is specifically: In the formula, N is the number of samples within the batch size; i is a sample index; is the remaining useful life prediction value of the ith sample; is the true value of the remaining useful life of the ith sample; The failure loss is specifically: In the formula, y i is the true binary failure label of the i-th sample; p i is the failure probability predicted by the neural network for the i-th sample; The health index loss is specifically: In the formula, is the health index prediction value of the i-th sample; is the health index true value of the i-th sample.

7. A power grid dynamic dispatch decision device based on multi-modal prediction, characterized in that, It comprises: A multi-source data perception module, a multi-modal feature fusion module, an intelligent state evaluation module, a preventive optimization scheduling module, and an emergency fault reconstruction module; The multi-source data perception module is configured to obtain internal state data of each target power grid device, external working condition data of each target power grid device, and a future load variation curve of a target power transmission and distribution line; The multi-modal feature fusion module is configured to perform a spatio-temporal alignment operation on the internal state data and the external working condition data of each target power grid device to generate a target power grid device feature matrix; The intelligent state evaluation module is configured to sequentially input the target power grid device feature matrix into a preset multi-modal neural network, so that the multi-modal neural network generates a corresponding health index, remaining useful life, and failure probability according to the target power grid device feature matrix; The preventive optimization scheduling module is configured to, when the health index of a target power grid device is lower than a first preset threshold, construct a multi-objective optimization model and related constraints according to the remaining useful life of the current device and the future load variation curve of the target power transmission and distribution line, with the objective of minimizing the risk of maintenance delay and economic cost; solve the multi-objective optimization model to generate a preventive scheduling strategy; According to the preventive scheduling strategy, the power grid is scheduled; The emergency fault reconstruction module is configured to, when the failure probability of a target power grid device is higher than a second preset threshold, dynamically update the line weight in the power grid topology according to the future load variation curve of each target power transmission and distribution line, and generate a power grid topology reconstruction scheme with the minimum power outage range according to the line weight; and the power grid is scheduled according to the power grid topology reconstruction scheme.

8. The multi-modal prediction based power grid dynamic dispatch decision apparatus as claimed in claim 7, wherein, The multi-modal feature fusion module comprises a data denoising unit and a data spatio-temporal alignment unit; The data denoising unit is configured to, for each target power grid device, perform noise suppression processing and missing compensation processing on the internal state data and the external working condition data of the current target power grid device to generate denoised internal state data and denoised external working condition data of the current target power grid device. The data space-time alignment unit is configured to perform alignment processing and space correlation processing on the denoised internal state data and the denoised external working condition data of the current target power grid device, and generate a current target power grid device feature matrix.

9. An electronic device, comprising: The computer program is stored in the memory and configured to be executed by the processor, and the processor implements the power grid dynamic scheduling decision method based on multi-modal prediction according to any one of claims 1 to 6 when executing the computer program.

10. A storage medium, characterized by The storage medium comprises a stored computer program, wherein the storage medium controls the device where the storage medium is located to execute the power grid dynamic scheduling decision method based on multi-modal prediction according to any one of claims 1 to 6 when the computer program is running.

Images