Power grid section heavy load on-line early warning method and device considering limit transmission capacity

Abstract

The invention provides a power grid section heavy load on-line early warning method and device considering limit transmission capacity, and the method comprises the steps: obtaining multi-dimensional measurable physical variables of a power grid in a heavy load state in a historical process, and forming a supervised multi-dimensional time series data set; building a first learning model, generating a feature subset through an evolution optimization method, and training the first learning model according to the feature subset and the multi-dimensional time sequence data set; after first learning model training is obtained through training, the fitness value of a first learning model is calculated according to the variable dimensions of the feature subsets and the prediction performance of the first learning model through an evolutionary optimization method, the feature subsets are optimized with the fitness value as an optimization target, and the feature subset with the maximum fitness value serves as a target feature subset; and building a second learning model, and training the second learning model by adopting the target feature subset and the supervised multi-dimensional time sequence data set to obtain a heavy load event prediction model.

Description

Technical Field

[0001] This application belongs to the field of power technology, and in particular relates to an online early warning method and device for overloaded power grid sections considering the ultimate transmission capacity. Background Technology

[0002] In power systems, the uneven development of power source and load locations, coupled with relatively lagging transmission network construction, can lead to modern power systems frequently operating in a "tight balance" state close to safety boundaries. This can induce regional section overload events, posing significant risks to the safe and stable operation of the system. Section overload refers to the transmission power of inter-regional tie lines approaching or exceeding the section's transmission limit, which is a comprehensive dynamic upper limit calculated based on static safety and transient stability rules. Due to insufficient early warning capabilities and low levels of intelligent decision-making in current systems, existing control methods are rendered passive in terms of both resources and time when facing real-time overload events. This makes it difficult to quickly optimize resource allocation to meet section transmission limit constraints, resulting in significant control pressure.

[0003] Determining cross-sectional overload events based on cross-sectional transmission limits involves verifying multiple safety rules (such as static safety and transient stability) under large-scale power flow levels. This requires performing numerous complex simulations and time-domain simulations on the anticipated event set. Consequently, existing day-ahead and intraday two-stage dispatching programs typically only consider offline-tuned cross-sectional transmission limits or directly use line thermal ratings for overload management. Furthermore, offline-tuned cross-sectional transmission limits struggle to track the rapidly changing power grid operating conditions in the context of new power systems, while line thermal ratings fail to effectively consider the system's dynamic safety characteristics. Research indicates that due to the computational complexity of the tuning process, the update frequency of actual power grid cross-sectional limits is currently very low (often measured in months), and is even tuned only according to typical quarterly operating conditions.

[0004] Against this backdrop, with the prediction errors of renewable energy power and the potential for unexpected load increases, heavy load events are still very likely to occur in the real-time stage after routine day-ahead and intraday dispatching procedures. This hinders energy transmission and utilization efficiency and reduces the system's safety and stability margin. If a sudden failure occurs at this time, the system's operational reliability will be severely threatened, potentially triggering unexpected cascading failures and widespread power outages, resulting in severe load losses. Due to the lack of early warning procedures, dispatchers can only take action after observing heavy loads, leading to immense control pressure and a passive situation in terms of both resource adjustment and control time. Furthermore, because online system stability analysis through transient time-domain simulation is computationally challenging, dispatchers often need to rely on experience to determine heavy load conditions, making it difficult to accurately track the complex and volatile system operating conditions under high renewable energy penetration.

[0005] Therefore, it is necessary to develop online automatic rapid early warning technology for cross-sectional heavy load events. Based on the dynamic system operation status and taking into account various safety rules, it can provide early warning of potential unsafe situations in the future, thereby supporting proactive regulation (prevention and control) to achieve more comprehensive event-driven heavy load management. Summary of the Invention

[0006] In view of this, this application provides a method, device, equipment and storage medium for online early warning of power grid section overload, which aims to provide high-precision online early warning of power grid section overload events.

[0007] Firstly, this application provides an online early warning method for overloaded power grid sections considering the ultimate transmission capacity, including: Acquire multidimensional measurable physical variables of the power grid under heavy load conditions in the historical process to form a supervised multidimensional time series dataset; A first learning model is built, and a feature subset is generated using an evolutionary optimization method. The first learning model is then trained based on the feature subset and a multidimensional time series dataset. After the first learning model is trained, the fitness value of the first learning model is calculated using an evolutionary optimization method based on the variable dimension of the feature subset and the prediction performance of the first learning model. The feature subset is then optimized using the fitness value as the optimization objective, and the feature subset with the largest fitness value is selected as the target feature subset. A second learning model was built, and trained using a subset of target features and a supervised multidimensional time series dataset to obtain a heavy-load event prediction model.

[0008] Optionally, the method for determining the overload state includes: Based on the repetitive power flow method, the cross-sectional limit transmission capacity is calculated considering AC power flow security, N-1 static security, and transient stability rules. Whether it is a heavy load state is determined based on the cross-sectional limit transmission capacity and the detected cross-sectional transmission power.

[0009] Optionally, the process of constructing a supervised multidimensional time series dataset includes: According to the data sampling interval By sampling the multidimensional measurable physical variables of the power grid, a historical sequence of multidimensional measurable physical variables that changes over time is obtained; Using observation time windows, with step sizes Multiple spatiotemporal data matrices are obtained by sampling on a historical sequence of multidimensional measurable physical variables; Each spatiotemporal data matrix is ​​assigned a category label to indicate the event state and form a supervised data format; and a prospective labeling strategy is used to assign time stage labels to the spatiotemporal data matrices, thereby forming a supervised multidimensional time series dataset.

[0010] Optionally, the evolutionary optimization method is used to calculate the fitness value of the first learning model based on the variable dimension of the feature subset and the prediction performance of the first learning model, and the feature subset is optimized with the fitness value as the optimization objective. The steps of selecting the feature subset with the largest fitness value as the target feature subset include: Based on the correlation between measurable physical variables and the early warning target of overload state, the priority of each measurable physical variable in the multidimensional time series dataset is determined. The binary grasshopper optimization algorithm is used, combined with the priority of measurable physical variables, to perform evolutionary optimization of the feature subset.

[0011] Optionally, the mMIFS-U criterion is adopted, combined with a sequential forward search strategy, to determine the priority of measurable physical variables.

[0012] Optionally, during the evolutionary optimization of the feature subset using the binary grasshopper optimization algorithm, the initialization and evolution rules of the binary grasshopper optimization algorithm are adjusted according to the priority of the measurable physical variables, so that the measurable physical variables with higher priority are retained first.

[0013] Optionally, the first learning model includes a k-NN classifier based on a multidimensional version of the dynamic time warp distance.

[0014] Secondly, this application provides an online early warning device for overloaded power grid sections considering the ultimate transmission capacity, comprising: The acquisition module is used to acquire multidimensional measurable physical variables of the power grid under heavy load conditions in the historical process, forming a supervised multidimensional time series dataset. The optimization module is used to build the first learning model. It generates a feature subset through an evolutionary optimization method and trains the first learning model based on the feature subset and the multidimensional time series dataset. After the first learning model is trained, the evolutionary optimization method is used to calculate the fitness value of the first learning model based on the variable dimension of the feature subset and the prediction performance of the first learning model. The fitness value is used as the optimization objective to optimize the feature subset, and the feature subset with the largest fitness value is taken as the target feature subset. The early warning module is used to build a second learning model. It uses a subset of target features and a supervised multidimensional time series dataset to train the second learning model and obtain a heavy-load event prediction model.

[0015] Thirdly, this application provides an electronic device, including the above-described online early warning device for overloaded power grid sections considering the ultimate transmission capacity.

[0016] Fourthly, this application provides a computer-readable storage medium storing at least one piece of program code, which is executed by a processor to implement the online early warning method for overloaded power grid sections considering the ultimate transmission capacity, as described above.

[0017] The beneficial effects of the technical solution provided in this application include: This application provides an online early warning method for overloaded power grid sections considering the ultimate transmission capacity. The method first acquires multidimensional measurable physical variables of the power grid under overload conditions during historical data collection, forming a supervised multidimensional time-series dataset. Then, an evolutionary optimization method is used to generate a feature subset, which is formed by selecting one or more of the multidimensional measurable physical variables; that is, the feature subset is a dataset formed by selecting a certain number of multidimensional measurable physical variables. After the feature subset is formed, a first learning model is trained by combining the feature subset and the multidimensional time-series dataset. After the first learning model is trained, the fitness value of the feature subset is comprehensively evaluated based on the model's prediction performance and the dimensionality of the measurable physical variables included in the feature subset. Evolutionary optimization (i.e., iteration) is continuously performed. Finally, after the evolution ends (when termination conditions are met, such as the iteration cycle is satisfied), the optimal feature subset is output as the target feature subset. A second learning model is trained offline by combining the target feature subset and the supervised multidimensional time-series dataset to obtain a prediction model for overloaded events with superior prediction performance. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart of an online early warning method for overloaded power grid sections, considering limit transmission capacity and feature combination optimization, provided as an embodiment of this application; Figure 2 A schematic diagram illustrating the risks to safe operation of a power grid under normal and heavy load conditions, provided as an embodiment of this application; Figure 3 A schematic diagram of a look-ahead marking strategy for early warning of cross-sectional overload events provided in an embodiment of this application; Figure 4 This is a general framework diagram of a high-dimensional combination optimization method for input features of a cross-sectional overload early warning model provided in an embodiment of this application; Figure 5This is a diagram illustrating the optimized architecture for combining wrapper features according to an embodiment of this application. Figure 6 Provided for an embodiment of this application =1.5 and A schematic diagram of the characteristics of the social force intensity function when = 0.5; Figure 7 A modified wiring diagram of an IEEE 39-node test system provided for one embodiment of this application; Figure 8 Different embodiments provided for this application Configure the performance of overloaded event alerts on the validation dataset; Figure 9 This is a histogram of mMIFS-U feature priority scores provided in an embodiment of this application; Figure 10 The fitness average convergence curve of the feature combination optimization method provided in an embodiment of this application; Figure 11 A statistical table of commonly used adjustable resource response times provided in an embodiment of this application; Figure 12 A set of candidate variables provided for an embodiment of this application The feature information description table in the middle (N=426); Figure 13 This application provides an embodiment of the information table of the top 16 feature variables ranked by mMIFS-U priority scores; Figure 14 A parameter setting table for an effectiveness analysis experiment of a feature combination optimization algorithm provided in an embodiment of this application; Figure 15 A comparison of the results provided in one embodiment of this application with the Wrapper feature combination optimization method based on evolutionary computation; Figure 16 This is a performance comparison result of a heavy-load event warning test provided in one embodiment of this application; Figure 17 A table of specific information about the combination of feature variables optimized by the method proposed in this invention, provided as an embodiment of this application; Figure 18 A structural block diagram of an online early warning device for overloaded power grid sections considering the ultimate transmission capacity, provided in an embodiment of this application; Figure 19 This is a structural block diagram of an electronic device provided in an embodiment of this application.

[0020] The attached figures are labeled as follows: 11: Acquisition Module; 12: Optimization Module; 13: Early Warning Module; 21: Processor; 22: Memory. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] Figure 1 A flowchart illustrating an online early warning method for overloaded power grid sections, considering both maximum transmission capacity and feature combination optimization, is provided as an embodiment of this application. See also... Figure 1 ,include: S101. Obtain multidimensional measurable physical variables of the power grid under heavy load conditions in the historical process to form a supervised multidimensional time series dataset.

[0023] In some examples provided in this application, the method for determining the overload state includes: Step 1: Calculate the sectional limit transmission capacity based on the repetitive power flow method, taking into account AC power flow security, N-1 static security, and transient stability rules.

[0024] The sectional limit transmission capacity is fundamental to determining whether a system is experiencing sectional overload. Considering computational convergence and the flexibility of embedding complex security constraints, it can be calculated using the Repeated Power Flow (RPF) method. The calculation process comprehensively considers AC power flow security, N-1 static security, and transient stability rules. Specifically, the calculation of the limit transmission capacity considering the above complex security constraints can be modeled as the following optimization problem:

[0025] In the formula, This is the power generation load growth index, used to adjust the power growth in power generation and receiving areas; These are system control variables, including active power injection at generator nodes, node voltage amplitude, and load power, etc. For system dependency variables; , These are the state variables and algebraic variables of the system during the transient process, respectively. , These are the sets of equality and inequality constraints before the fault, respectively. This represents the set of static equality constraints of the system after the N-1 fault is interrupted. For N-1 static security verification rules, Represents the set of N-1 line disconnection faults; Equation (4) represents the initial values ​​of the system state variables calculated by solving the power flow equations before the fault. Equations (5) and (6) represent the differential equation and algebraic equation in the transient process, respectively. For transient stability verification rules, For the purpose of anticipating accidents, This refers to the period of transient stability research.

[0026] This invention primarily considers transient power angle stability constraints and adopts the following determination rules:

[0027] In the formula, and These represent the generator after the fault was cleared. The angle of work and the angle of work of the center of inertia, To constrain the upper limit, it is often set to 180 degrees. Let be the generator's inertial constant.

[0028] The above-mentioned limit transmission capacity calculation model is based on a certain power generation and load growth pattern, and is adjusted by... To iteratively change the system's operating conditions:

[0029] In the formula, For the collection of generators in the sending-end area, For the load collection of the receiving area, , , , These represent the initial states of generator active power output, voltage, load active power, and load reactive power, respectively. , and , respectively represent the generator's active power output, voltage, and load growth mode. Models (1) to (9) can be solved using the RPF method, which continuously increases the generator output in the sending-end region and the load demand in the receiving-end region through iteration, thereby continuously increasing the electrical energy transmitted across the cross-section. At the same time, various safety constraints are comprehensively checked in each iteration to obtain the cross-section transmission limit:

[0030] In the formula, This represents the optimal growth index obtained through iterative calculation. For the collection of inter-regional connecting lines, For connecting lines In the running scenario The maximum transmission capacity under these conditions To calculate the connecting lines It is a function of transmission power. Therefore, by comparing the tie line power flow and transmission limit under this operating scenario, it can be determined whether there is section overload.

[0031] It should be noted that, since the purpose of this invention is to provide early warning of overload events and perform proactive control, it only needs to determine the overload state category (yes / no). For operating scenarios whose initial operating state does not meet the safety verification rules, they can be directly determined as overloaded without needing to reduce the load level downwards to accurately calculate their maximum transmission capacity. For operating scenarios whose initial operating state meets all safety verification rules, their maximum transmission capacity is accurately calculated, and the overload state is determined based on 95% of the limit. Based on this logical design, the execution time of the overload determination program will be effectively reduced when facing large-scale operating scenarios to be evaluated, without affecting the determination results, and therefore will not affect subsequent model construction.

[0032] Specifically, the specific steps of the cross-section overload determination algorithm based on bisection RPF used in this invention are as follows: (1) First, input the power system operation scenario. It also performs AC power flow verification, N-1 static safety verification, and transient power angle stability verification for the input scenario.

[0033] (2) Determine whether the current scenario passes all security rule checks. If it passes, then set... , , If the scenario is determined to be a section overload state, proceed to step (3); otherwise, proceed to step (7).

[0034] (3) If Greater than or equal to the threshold (at the beginning) =0), then pass Update the load level of the receiving area, reschedule the generators in the sending area, and perform AC power flow verification, N-1 static safety verification, and transient power angle stability verification.

[0035] (4) Determine whether all safety rule checks have passed. If they have, update the calculation accordingly. : , , Otherwise, update according to this calculation. : , , .

[0036] (5) Judgment Is it still greater than or equal to the threshold? If yes, return to step (3); if... If it is less than the threshold, then iterative calculation is obtained. The result is obtained by calculation using equation (10). , .

[0037] (6) Verification Scenario If the power flow of the downstream connection exceeds 95% of the transmission limit, it is determined to be a section overload state; otherwise, it is a normal state.

[0038] (7) The overload state determination ends, and the output is based on the input scenario. The result of the overload state determination.

[0039] Step 2: Determine whether the system is under heavy load based on the cross-sectional limit transmission capacity and the detected cross-sectional transmission power.

[0040] The ultimate transmission capacity of inter-regional transmission sections in a power system is a crucial indicator for assessing tie-line transmission limitations. Ultimate transmission capacity measures the ability of a transmission network to reliably transmit electrical energy from one region to another without compromising system security. When the transmission power at a section approaches or exceeds its ultimate transmission capacity, the system may face significant operational risks, specifically manifested as an inability to maintain N-1 static security or transient stability. Figure 2 As shown, this application considers cross-sectional overload as a situation where the cross-sectional transmission power exceeds 95% of the limit transmission capacity, thereby indicating a situation where the system safety margin is insufficient or faces safety risks.

[0041] Based on the above analysis, calculating the ultimate transmission capacity of a transmission section involves verifying multiple safety rules under large-scale power flow, requiring extensive and complex simulation calculations and time-domain simulations. This presents an extremely heavy computational burden in online scenarios, making practical application difficult. Therefore, this application employs a data-driven overload event early warning method, directly learning overload pattern knowledge from data. This avoids the high requirements for network-wide prediction information and accurate topology information, and shifts the computationally intensive overload assessment process considering ultimate transmission capacity to the offline stage, improving online practicality. Furthermore, to consider the temporal evolution trend of grid overload events and the temporal correlation between multi-dimensional measurement sequences, this invention uses time-series data recorded before the action time to train the overload event early warning (CEP) model, thereby uncovering some valuable temporal data patterns.

[0042] This application treats overload event early warning as a multidimensional time series learning task, using deep neural networks to mine key data patterns from the MTS (Medium-Terminal Scale), thereby helping to identify the early stages of power grid overload events and infer possible future event states. To construct an overload event early warning model based on multidimensional time series learning, a supervised MTS dataset is first required. In some examples provided in this application, the construction process of the supervised multidimensional time series dataset includes: Step 1: Based on the data sampling interval By sampling the multidimensional measurable physical variables of the power grid, a historical sequence of multidimensional measurable physical variables that changes over time is obtained.

[0043] Step 2: Use an observation time window with a step size. Multiple spatiotemporal data matrices are obtained by sampling on a historical sequence of multidimensional measurable physical variables.

[0044] Step 3: Assign a category label to each spatiotemporal data matrix to indicate the event state and form a supervised data format; and adopt a prospective labeling strategy to assign time stage labels to the spatiotemporal data matrix, thereby forming a supervised multidimensional time series dataset.

[0045] More specifically, the process of constructing a supervised multidimensional time series dataset includes: Assume there are N-dimensional measurable physical variables in the system, and the data sampling interval is... Given a length of The observation time window (OTW) is used to... The step size slides over the historical multivariate measurement sequence to extract a series of spatiotemporal data matrices. For the ... The extracted data matrix is ​​obtained by sliding the window twice. :

[0046] In the formula, , It is the length of each sequence segment in the data matrix. , , Representing variables Observation time series segments, Assume the first The observation time window in the next slide begins at time... , Representing variables At sampling time The measured value, This represents physical variables. At sampling time The measured values, and so on.

[0047] Subsequently, each data matrix (i.e., MTS samples) all need to be assigned a category label. To indicate event status and form a supervisory data format, a forward-looking labeling strategy is adopted, which can assign time phase labels to MTS samples based on current and forward-looking overload status.

[0048] According to the overload state determination rule considering the limit of transmission capacity, the operating scenario at any moment in the historical multivariate time series can be assigned an overload or non-overload state label, thus forming an event state sequence, such as... Figure 3 As shown. If the power system under study can obtain the corresponding historical overload operation records (including the location, time and control actions of the overload), an event state sequence can also be formed based on the recorded historical overload information.

[0049] After obtaining the overload state sequence, the subsequent modeling process (feature combination optimization, model training, etc.) is the same. Based on the formed cross-sectional overload state sequence, it is assumed that the first historical overload event occurred at time... arrive During the period (see Figure 3 ), and the CEP model hopes to advance Predict the occurrence of overloaded events based on time. Based on a lookahead labeling strategy, all endpoints are located within... The MTS data matrix corresponding to the observation time window within the time interval will be labeled as the "early stage"; all ends are located in The MTS data matrix corresponding to the observation time window within the time interval will be marked as the "overload phase"; while earlier time... The MTS data matrix will be marked as "normal phase".

[0050] In addition, the end is located arrive The MTS data matrix samples corresponding to the observation time windows within the time interval can be considered as the "recovery phase". Since identifying this phase offers little help for early warning and prevention control, this invention ignores the "recovery" label. Using the above annotation strategy, cross-sectional overload events can be mainly divided into three phases: normal, early, and overload phases. Combining the multi-dimensional time series samples generated by sliding, a supervised MTS dataset for the overload event early warning model can be obtained, as shown below:

[0051] In the formula, It is the total number of paired samples. It corresponds to the sample The category label indicates the stage of the event. , .

[0052] Based on supervised multidimensional time series datasets A multi-class classification model for heavy-load event early warning can be learned offline through MTS. In the real-time phase, once the latest MTS samples are collected... And input it into the trained classification model. You can quickly get the prediction results. This indicates the development stage of the overloaded event:

[0053] In the formula, Representation Model Predicted event phase categories, , .

[0054] Specifically, if a sample is classified as "early stage", that is... This means that cross-sectional overload events are very likely to occur in the following... This occurs within a short period of time. Therefore, preventative control measures can be developed during this period to address potential risks.

[0055] S102. Build the first learning model, generate a feature subset using an evolutionary optimization method, and train the first learning model based on the feature subset and the multidimensional time series dataset. After training the first learning model, use the evolutionary optimization method to calculate the fitness value of the first learning model based on the variable dimension of the feature subset and the prediction performance of the first learning model. Optimize the feature subset with the fitness value as the optimization objective, and take the feature subset with the largest fitness value as the target feature subset.

[0056] This application proposes a hybrid high-dimensional feature combination optimization method based on a feature combination optimization framework embedding multiple MTS learning models and a developed enhanced evolutionary computation structure. This method is used for efficient combination optimization of high-dimensional candidate data variables in power systems. The overall structural framework of the proposed high-dimensional feature combination optimization method for the input features of the overload early warning model is as follows: Figure 4 As shown. In this application, the purpose of step S102 is to find the target feature subset that optimizes the model performance, optimize the dataset used for model training, and thus improve the model's predictive performance.

[0057] It should be noted that the data dimension of the feature subset is different from that of the multidimensional time series dataset. Therefore, when training the first learning model using both the feature subset and the multidimensional time series dataset, only the data in the multidimensional time series dataset with the same dimension as the feature subset will be used for training.

[0058] In some examples, step S102 includes: S1021. Based on the correlation between measurable physical variables and the early warning target of overload state, determine the priority of each measurable physical variable in the multidimensional time series dataset.

[0059] In some examples, step S1021 includes: The priority of measurable physical variables is determined by adopting the mMIFS-U criterion and combining it with a sequential forward search strategy.

[0060] This application uses the mMIFS-U criterion (Modified Mutual Information Feature Selector Under Uniform Information Distribution) to quantify the priority score of each candidate feature variable for building a heavy event early warning model. This information criterion can simultaneously consider the correlation between feature variables and prediction targets as well as the redundancy within feature combinations, thus laying the foundation for the formation of an enhanced evolutionary computation method.

[0061] More specifically, step S1021 includes: For two random variables and Their edge distributions are respectively and Mutual information can be defined as follows:

[0062] Given a set of candidate feature variables from a power system , Let be the dimension of the candidate feature variable set, denoted as . This is a variable for the overloaded event phase label.

[0063] Based on the mMIFS-U criterion and the sequential forward search strategy, the optimal variable in each search iteration round is... It can be found:

[0064] In the formula, Indicates the current iteration number. Indicates the first The set of unselected variables in the next iteration Indicates the first One unselected variable; It is the first The set of selected variables in the next iteration, variable set and The following relationship exists: , , , . This represents the unselected variables calculated according to equation (14). With event label variables Mutual information between them Indicates unselected variables With selected variables Mutual information between them , Representing variables Information entropy.

[0065] By employing a sequential forward search strategy, the optimal variable for each iteration can be obtained by maximizing equation (15). and to Add this feature variable from Remove the variable from the middle, that is , At the same time, Substituting back into equation (15) allows us to calculate the variables. The corresponding maximum value It can be regarded as a characteristic variable. The priority score is then saved. And proceed to the next round. Continue iterating until... All candidate features were added, i.e. , Once the sequential forward search is complete, the round priority and corresponding priority score of all candidate feature variables can be obtained.

[0066] To facilitate subsequent use in the initialization of evolutionary computation (EC) methods and the improvement of iterative evolution rules, the obtained feature priority scores need to be normalized, as shown below:

[0067] In the formula, Indicates the first The standardized priority score of each feature variable indicates that the higher the value, the higher the priority and importance in the heavy-load event warning task.

[0068] like Figure 4 As shown, the input to the feature priority score calculation program is also a multidimensional time series supervised dataset, i.e. For each feature variable, it is possible to extract... A time series segment. For the feature variable The extracted time series segments are , Simultaneously, the event tag set can be extracted. Therefore, it can be seen that a time series segment (containing multiple data points) will correspond to only one event type label.

[0069] Therefore, in order to calculate the variables With event label variables Mutual information between This requires compressing time series segments. For example, for segments derived from variables... Time series segments The compression value can be calculated. For example, the average, median, maximum, and minimum values, among which... Indicates the index of the compressed variable. , This is the total number of compression variables used. For the th One compression variable By performing compression on each time series segment, a compressed value vector can be obtained. Therefore, the compressed value vector With category label vector They will have the same data length (number of elements). Therefore, compressed variables With event label variables Mutual information between It can then be calculated using conventional methods.

[0070] Subsequently, according to One compression variable and mutual information evaluation results Mutual information can then be obtained through comprehensive approximate calculation. For example, a comprehensive approach can be used, such as averaging. Similar to mutual information, the mMIFS-U criterion supports filtered feature combination optimization, and both require compression of the measurement time series to quantify the importance of feature variables.

[0071] S1022. The binary grasshopper optimization algorithm is adopted, and the priority of measurable physical variables is combined to perform evolutionary optimization on the feature subset.

[0072] To select feature variables containing time data patterns beneficial for overload event early warning and to interact with the MTS learning model during feature combination optimization, this invention employs a Wrapper framework embedded with the MTS learning model, treating feature combination optimization as an input variable combination optimization task. Furthermore, since evolutionary computation methods possess global search capabilities based on group agents and are applicable to multi-objective combination optimization, this invention sets evolutionary computation as the basic optimization algorithm within the Wrapper feature combination optimization framework. For example... Figure 4 As shown, the Wrapper feature combination optimization framework based on evolutionary computation typically consists of three main components: an optimization algorithm based on evolutionary computation, a learning model, and a fitness function (evaluation criterion).

[0073] In each iteration of the Wrapper feature combination optimization, the evolutionary computation algorithm generates multiple feature variable combinations (i.e., feature subsets) and trains multiple learning models based on these combinations and the MTS supervised dataset. Then, based on the variable dimensionality of each feature subset and the heavy-load event warning performance of the corresponding learning model, a fitness function is used to comprehensively evaluate each feature combination, and the evaluation results guide the next combination optimization iteration. After repeated evolution, the optimized feature subset, i.e., the preferred feature combination, can be output.

[0074] In some embodiments provided in this application, step S1022 includes: In the process of evolutionary optimization of feature subsets using the Binary Grasshopper Optimization Algorithm (BGOA), the initialization and evolution rules of BGOA are adjusted according to the priority of measurable physical variables, so that measurable physical variables with high priority are retained first. This invention develops a novel hybrid computational structure (i.e., enhanced evolutionary computation) for performing efficient combinatorial optimization on large-scale candidate feature variables. The Binary Grasshopper Optimization Algorithm (BGOA) is selected as the optimization framework because of its effective mechanism to resist local optima, and this mechanism is used to alleviate the local stagnation problem. At the same time, the aforementioned feature priority scores are used to improve the initialization and evolution rules of the original BGOA algorithm, improving the optimization efficiency of the algorithm in a semi-directed manner, thus forming an enhanced evolutionary computational structure, providing a solution to the optimization inefficiency problem.

[0075] More specifically, step S1022 includes: The feature combination optimization task can be modeled as a binary discrete optimization problem: from a set of candidate variables from a power system Choose a subset of variables to optimize a given fitness function. Among them, the use of Representative The search agent in the 1st The solution vector corresponding to the next iteration. , This is the total number of search agents. , It is the maximum number of iterations. , Indicates the first Several feature variables were selected. This indicates that the option was not selected.

[0076] In evolutionary computation, the solution vector is often represented by the position vectors of the search agents. The search agent swarm continuously updates its position vectors to find better solutions. In the BGOA algorithm, the position vector update is based on the step vector, similar to the velocity vector in the BPSO (Binary Particle Swarm Optimization) algorithm. For the... The search agent, in the 1st... The step vector in the next iteration can be represented as: The calculation method is as follows:

[0077] In the formula, It is the first With the Euclidean distance between search agents Indicates the first The search agent in the 1st The position vector in the next iteration. It is a function that defines the social strength of the search population:

[0078] In the formula, Indicates the intensity of attraction. Indicates the range of attractive force. When When the search agent i exerts a repulsive force on the search agent k, it exerts an attractive force on the search agent k.

[0079] Social force intensity function Detailed features are available Figure 6 To explain further, see the diagram. It was set to 1.5. It was set to 0.5. For example... Figure 6 As shown, when the distance between agents is within the interval [0, 2.079], the search agent... For search agents This manifests as a repulsive force; as the distance between agents continues to increase, the search agent... For search agents It manifests as gravity, when approaching It reaches its peak value at one time, and then gradually decreases and approaches 0.

[0080] Social force intensity function The main ideas are: (1) When the search agents are too close, other search agents should be excluded to avoid the population being too dense and reduce the risk of local stagnation; (2) When the distance between search agents increases, other agents should be appropriately attracted to the nearby area to avoid the group search being too dispersed and difficult to converge. Figure 6The social force strength function shown basically satisfies the two requirements mentioned above, but when the distance between search agents is too large, the attraction exerted by the function is almost zero. Therefore, to solve this problem, the distance between agents needs to be considered when calculating the social force strength. Mapped to the interval [1, 4], i.e. Furthermore, in equation (17), This is a decay coefficient used to gradually reduce the step size:

[0081] In the formula, This represents the current iteration number. It is the maximum number of iterations. and Set them to 50 and 0.00005 respectively. Use the sigmoid function and the step vector... The first one can be calculated The probability of each feature variable being selected in the next iteration:

[0082] In the formula, Given the sigmoid function, based on equation (20), the following position vector update formula of the original BGOA algorithm can be obtained:

[0083] In the formula, This represents a random number that is uniformly distributed in the interval [0, 1].

[0084] Based on the above framework, this invention further proposes modifying the initialization and iterative evolution rules of the BGOA method using the aforementioned candidate variable priority score metric, thereby forming an enhanced evolutionary computation method with higher optimization efficiency. Simultaneously, the social force mechanism, which has proven effective in the BGOA algorithm, is adopted to improve the local stagnation resistance of the designed method. The specific modifications are as follows: (1) Improved initialization rules: Although random initialization can effectively diversify the initial solutions, when faced with high-dimensional candidate features, a completely random approach may lead to a very slow convergence speed of the algorithm, or even local stagnation in the early stages of iteration. Therefore, in order to improve the quality of the initial solutions while maintaining the diversity of the search population, this invention proposes to use a semi-directional approach combining feature variable priority scores with random rules to initialize the position vector of the search agent. The improved initialization rules are as follows:

[0085] In the formula, For the first Standardized priority scores corresponding to each feature variable; These are the weighting coefficients. ; For the floor function, This represents a random number uniformly distributed within the interval [0, 1]. The above rule means that, after semi-directional guidance through the priority scores of feature variables, feature variables with higher priorities will have a greater chance of being considered in the initial stage; at the same time, the random component is still retained.

[0086] (2) Improved update evolution rules: The original BGOA algorithm is based on equations (17) to (21) for evolutionary updates. Although the random method described by equation (21) can increase the diversity of solutions and give the search agent a chance to escape local conditions, it may also produce inappropriate evolutionary directions, which will lead to difficulty in algorithm convergence and low efficiency. In addition, due to the purely random method, some feature variables with obvious importance may never be considered in the entire optimization process. Therefore, this invention proposes to use feature priority fraction semi-directional guidance for the evolutionary update process of the algorithm and retain the random rule, so that feature variables with higher priority have a greater chance of being considered in the iterative evolution process. The improved update evolution rules are as follows:

[0087] In the formula, Representative vector The first in One element, The best position vector found so far. and This represents two combined calculated values, and the calculation method is as follows:

[0088]

[0089] In the formula, The weighting coefficient is set to decrease linearly from 0.5 to 0.2 as the number of iterations increases, thereby gradually reducing the influence of the mMIFS-U feature priority score. For variables The corresponding priority score. Furthermore, in equation (23), The mutation rate, representing a linear decrease from 0.6 to 0 with increasing iteration number, is:

[0090] To clearly illustrate the above evolution rules, the rounding function can be used. Internally, it is understood as the overall chance of a feature variable being selected in the next iteration, where It can represent random chance. This represents algorithmic opportunities arising from evolutionary computation itself. The guiding opportunity is assigned by the feature priority score. The overall evolutionary rule is described by equation (23), which is divided into three cases: (a) This situation indicates that, for the first If a search agent has a total social attraction to it that is less than its total repulsion, that agent is encouraged to continue exploring with other search agents. In relatively distant search areas, the next value is determined based on random chance and guiding chance, as shown in equation (24), thereby bringing some feature variables with high priority scores back into consideration; (b) ,and This situation indicates that, for the first For each search agent, the total social attraction to them is greater than the total repulsion. However, to increase the diversity of the group and give it the opportunity to break out of its local constraints, in... In this case, the agent is still encouraged to make certain changes and determine its next value based on algorithmic opportunities and bootstrapping opportunities, as shown in equation (25); (c) ,and This situation indicates that for search agents Other search agents have a greater total social attraction than total repulsion towards it, and it does not need to undergo mutation; therefore, the agent is encouraged to move to the current optimal position vector. The surrounding area continues to evolve and iterate.

[0091] Figure 4 and Figure 5 This invention demonstrates two other main components of the proposed Wrapper-style feature combination optimization framework: a multivariate time series learning model and a fitness function. .

[0092] (1) MTS learning model in the feature combinatorial optimization framework: In each iteration of combinatorial optimization based on augmented evolutionary computation, the position vector returned by each search agent The corresponding subset of feature variables can be obtained. For the entire search agent population, a subset of feature variables can be extracted, namely... ,in , .based on and multidimensional time series supervised datasets , Each MTS learning model will be trained in parallel.

[0093] Subsequently, each trained MTS model will be invoked to perform overloaded event warning tests, thereby outputting overloaded event warning test results corresponding to each feature subset, serving as partial support for evaluating the quality of feature variable subsets (see...). Figure 4 The second part, represented by the second dashed box, is the high-dimensional feature variable combinatorial optimization based on enhanced evolutionary computation.

[0094] Specifically, the k-NN classifier based on Multidimensional Dynamic Time Warping (MDTW) is set as the MTS learning model in the feature combination optimization framework, which has good parallelism and the ability to consider MTS time information. Although the k-NN model is simple in principle and easy to implement, it has achieved highly competitive test performance in the field of time series classification and is often considered as the benchmark algorithm for multidimensional time series classification. MDTW is a method that can effectively measure the distance between multidimensional time series samples, and can eliminate the influence of unequal time series sample lengths and misaligned start times, improve the robustness of the distance measurement between samples, consider the dynamic change patterns of time series data, and lay the foundation for accurate multidimensional time series reload event early warning.

[0095] (2) Fitness Function in Feature Combination Optimization Framework: When constructing the corresponding subset evaluation criteria, this invention mainly considers the following two aspects: (a) the dimension of the feature variable subset; (b) the prediction performance of the multi-dimensional time series overloaded event early warning model when using the corresponding feature variable combination. Therefore, the following comprehensive fitness function can be formed:

[0096] In the formula, Representation and subset of characteristic variables Performance metrics for overloaded event alerting tests of relevant MTS learning models. Representing feature combinations Variable dimensions, This represents the total number of original candidate feature variables. These are the weighting coefficients.

[0097] Power grid overload events can be divided into three stages, therefore, overload event early warning can be considered a three-classification task. Therefore, this invention uses a weighted multi-class F-beta score. To construct performance metric functions :

[0098] In the formula, Representative and characteristic variable combination The F-beta score of the overloaded event warning of the relevant MTS learning model, , A larger value indicates better performance in heavy-load event warning. A set of event phase labels. , Indexed by event category, Indicates category as The proportion of the data sample in the total test data sample. and They respectively represent the use When used as an input variable, the overloaded event warning model is applied to the first... Precision and recall for event-type stage tags. It is an adjustable parameter used to control the weight given to precision and recall in this invention. Set to 1, for and Apply the same level of attention.

[0099] S103. Build a second learning model. Use a subset of target features and a supervised multidimensional time series dataset to train the second learning model and obtain the overloaded event prediction model.

[0100] After optimization through high-dimensional feature combination, it is possible to start from the optimal position. Extract optimized feature subsets .based on and power grid overload event supervision dataset The final heavy-load event warning model based on multi-dimensional time series learning can be trained offline. This enables rapid online early warning (see...) Figure 3 Part Three).

[0101] In some examples, the second learning model is a more complex model than the first learning model, such as LSTM or transformer, to ensure that the accuracy of the second learning model is higher.

[0102] This invention uses a synthetic dataset simulated on the IEEE 39-bus test system to analyze the effectiveness of the proposed high-dimensional feature combination optimization algorithm. The IEEE 39-bus system consists of 10 conventional generators, 12 two-winding transformers, and 34 AC lines. Based on the standard example, this invention modifies the original system to accommodate a 20% renewable energy penetration rate. Two photovoltaic power stations are added at buses 9 and 14, and two wind turbines are added at buses 5 and 26, each with an installed capacity of 500MW. The system topology and partitioning information are as follows: Figure 7 As shown.

[0103] Preparation of Multidimensional Time Series Supervised Dataset: To simulate continuous system operating conditions and use the S101 method to determine the overload state of the operating scenario at each time point, thereby obtaining the overload event state sequence, this example first uses a one-year (sampling interval) dataset from an actual power system. Load and renewable energy fluctuation data (15 min) were injected into the modified IEEE 39-bus test system. The AC optimal power flow method was used to determine the output of conventional generator units and the voltage of generator nodes at each moment, thus obtaining a year-long system operation scenario data. The photovoltaic power generation time-series data at buses 9 and 14, and the wind power time-series data at buses 5 and 26, were from different data sources to form a rich renewable energy output scenario. Simultaneously, the power injection time-series data for all load nodes also came from different actual load data sources. All injected time-series data were scaled according to the test system scale using a Min-Max method. The AC optimal power flow program was implemented in the PandaPower computing library using Python. The experiment obtained 34,080 continuous time-section system operating conditions (including control variables such as generator output and load power, and state variables such as PQ node voltage amplitude and line transmission power). Based on this system operating condition sequence, the original multidimensional time-series data can be formed. 426 feature variables were collected for each time-section, forming the variable candidate set. , =426. Subsequently, the overload state determination algorithm is used to determine the system overload state at each moment, forming an overload event state sequence (each with a length of 34080). This example considers the transmission cross-section between region 1 and region 2, such as... Figure 7As shown in the diagram. Region 2 is the power receiving area, receiving power through tie lines 1-39, 2-3, 18-3, and 16-15. In the overload determination procedure, this invention comprehensively considers AC power flow safety, N-1 static safety, and transient power angle stability rules. In the N-1 static safety verification procedure, this example considers N-1 faults as open circuit faults in tie lines 1-39, 2-3, 18-3, and 16-15. The N-1 static safety verification is also implemented using the PandaPower calculation library. In the transient power angle stability verification procedure, this example considers a three-phase metallic short circuit occurring 50% of the distance from the beginning of tie lines 1-39, 2-3, 18-3, and 16-15. The fault occurrence time is 1 second, the fault clearing time is 1.15 seconds, the transient time-domain simulation duration is set to 5 seconds, and the transient simulation step size is 20 ms. Transient time-domain simulations were implemented using Python in the Andes computing library. The generator used the GENROU model, the excitation system used the IEEE X11 type excitation model, and all loads adopted constant impedance models. The photovoltaic power station and wind turbine used the standard models provided by the Andes computing library. The power angle stability criterion is shown in the equation. Furthermore, the fault set can be adjusted and improved according to specific needs, but the proposed method structure will not be affected. Through the above steps, the overload determination result (overloaded or non-overloaded) for each moment of the system operation scenario can be obtained, forming an overload event state sequence as the basis for forward-looking sample labeling. A total of 1927 time segments were determined to be under overload conditions.

[0104] Based on the multidimensional time series data (variable dimension 426, time length 34080) and overloaded event state sequences (including overloaded or non-overloaded state markers, time length 34080) obtained from the above experiments, an MTS supervised dataset can be constructed using a look-ahead labeling strategy by setting corresponding observation time windows. The length of the observation time window is specified in the figure. and forward-looking parameters This will significantly impact the morphology of the MTS samples and will also affect the performance of the subsequent overload event warning model. Therefore, this invention considers the following methods for setting parameters. and .

[0105] (1) Configuration: Considering that the main function of the overload event early warning model studied in this invention is to provide early warning information about potential future risks to the system, thereby allowing dispatching and operation personnel sufficient time to take preventive measures, the forward-looking parameter... The settings should refer to the response time of commonly used adjustable resources so as to make full use of multiple resources for coordinated control during the prevention and control phase, thereby achieving better control results. Figure 11The response times of commonly used adjustable resources reported in the retrieved literature are listed. Here, response time refers to the time required for the resource to adjust its output power to the target value, which is different from the resource's reaction time. For example... Figure 11 As shown. This experiment will use the advance parameter. Setting it to 30 minutes allows for sufficient time for most resources to participate in coordinated prevention and control efforts, while also minimizing the need for [other measures]. Too large an amount will reduce the accuracy of the prediction.

[0106] (2) Setup: In this example, based on the same multidimensional time series data and overloaded event state sequences, different... A sliding observation time window (i.e., {480, 420, 360, 300, 240, 180, 120} minutes) was used to obtain MTS data matrix samples of different sizes. Data from months {1-3, 5-7, 9-11} were used as the training and validation sets, while data from months {4, 8, 12} were used as the test set. Subsequently, the former was divided into training and validation sets through non-repeating random sampling. During this process, The timeout period is set to 30 minutes and left unchanged. Based on this procedure, seven MTS supervised datasets containing samples of different sizes are obtained, N=426, L∈{32, 28, 24, 20, 16, 12, 8}. Then, using the training data from these seven datasets, the mMIFS-U criterion is used to calculate the feature priority score for each candidate variable. Based on the calculated feature priority scores, the top 30 variables in each dataset are selected, and overload event alert tests are performed on their respective validation datasets based on these variables to obtain results corresponding to different... The test results are used to select parameter settings that yield better results. The MTS classifier used in this process is a k-NN model based on MDTW distance. The grid search test results are as follows: Figure 8 As shown.

[0107] like Figure 8 As shown, when the time window length When the time window is 180 minutes (3 hours), the k-NN classifier based on MDTW distance achieves the second-highest accuracy of 92.521%; when the time window length is... The accuracy rate was highest, reaching 92.659%, when the time frame was 360 minutes. However, when... When the timeframe is 360 min, the total inference time on the validation dataset is 18.804 seconds, which is significantly higher than... We take 12.645 seconds at 180 min. Since a relatively small time window length can generate smaller MTS samples, it is beneficial for the efficiency of subsequent feature combination optimization, offline model training, and online model prediction. Furthermore, considering that... At 180 minutes, the model achieved near-first-place results in overload event warnings (including accuracy, recall, and precision) in grid search, while also having a shorter computation time. Therefore, this experiment sets the sliding time window length... The time limit is 180 min. All subsequent methods will be compared fairly under the same parameter settings.

[0108] Based on the above parameter settings ( =30 min、 =180 min), multidimensional time series data, and reloaded event state sequences can be used to obtain continuous MTS samples based on a sliding observation time window. During this process, a reloaded event development stage label can be attached to each MTS sample based on a look-ahead labeling strategy: normal stage (C0), early stage (C1), and reloaded stage (C2). Furthermore, this example uses an undersampling strategy to alleviate the class imbalance problem. Based on the above settings, this experiment ultimately generated 3548 supervised MTS samples, of which 1640, 710, and 1198 samples belong to the "normal stage," "early stage," and "reloaded stage," respectively. Subsequently, based on the month information, data from {1-3, 5-7, 9-11} were used as the training and validation sets (containing 2618 samples), and data from {4, 8, 12} were used as the test set (containing 930 samples). The former will be further divided into a training set (containing 1896 samples) and a validation set (containing 722 samples) through non-repeating random sampling. The training and validation data will be used for feature combination optimization and offline model training, while the test data will be used for online simulation of heavy-load event warning.

[0109] mMIFS-U Feature Priority Score Calculation: After acquiring the multidimensional time series supervised dataset, the mMIFS-U feature priority score calculation program is first executed to obtain and store the feature priority score for each candidate variable. This score serves as the basis for subsequent feature optimization calculations in the enhancement evolution process. This program only needs to be executed once offline. Specific information about the 426 candidate variables involved in this example is as follows: Figure 12 As shown. To prevent overfitting, the feature priority scores are calculated using only the training set data. As described in step S102 above, when calculating mutual information... as well as When compressing the time series data, mutual information calculation is performed based on the compressed variables. This experiment uses the following compressed variables: maximum, minimum, mean, standard deviation, cumulative sum, sum of variation, and median, and uses averaging for aggregation. Subsequently, the final feature priority score metric results are obtained, and their histogram is plotted as follows: Figure 8 As shown. Figure 9 As shown, the mMIFS-U feature priority scores effectively distinguish the importance rounds of candidate feature variables, with 135 feature variables having an importance score greater than 0.5, accounting for approximately 31.7% of the total. In subsequent experiments, this feature priority metric will be used to enhance iterative optimization of feature variable combinations. Figure 13 It displays information about the top 16 variables based on their feature priority scores.

[0110] Effectiveness Analysis of Feature Combination Optimization Algorithm: Based on the MTS supervised dataset mentioned above and the mMIFS-U feature priority scores calculated and stored in the previous section, this experiment executes a high-dimensional feature combination optimization program through the method described in step S102 above to mine variable combinations that are beneficial to the heavy-load event warning task. In order to test the effectiveness of the proposed enhanced evolutionary computation structure, the experiment first compares the proposed method with the Wrapper feature combination optimization benchmark method without the enhanced structure, including: (1) the original BGOA algorithm; (2) GA genetic algorithm (Genetic Algorithm); (3) BGWO algorithm (Binary GreyWolf Optimization). In the feature combination optimization process, the supervised dataset, MTS learning model (k-NN classifier based on MDTW distance) and fitness function (see Equation (27)) used by various methods are the same, and each method will be run 20 times to compare the average results. Except for the parameter settings mentioned in step S102, the other parameter settings are as follows Figure 14 As shown. The average convergence curves of the feature combination optimization of the four comparison methods are as follows. Figure 9 As shown.

[0111] like Figure 10As shown, with the help of the enhanced evolutionary computation structure based on feature priority scores designed in this invention, the improved feature combination optimization method outperforms the other three methods in terms of initial fitness, optimization efficiency, and final fitness score. Faced with 426 candidate feature variables in this example, the BGWO and GA methods exhibited slower evolution speeds (slow decrease in fitness value) in the mid-to-late search stages, while the BGOA method encountered local stagnation (almost no decrease in fitness value) in the early to mid-search stages. However, the method of this invention, under the semi-guided feature priority score approach, has a better fitness value in the initialization stage. The average initial fitness value of this invention is 0.1449, while that of BGOA, BGWO, and GA methods are 0.1645, 0.1640, and 0.1633, respectively. This is mainly because the semi-guided feature priority score approach can help eliminate some feature variables irrelevant to the prediction target at the beginning, thus replacing completely random initialization and generating a more suitable initial exploration direction. The experimental results demonstrate the effectiveness of the semi-guided feature priority score approach in the initialization stage. During the update and evolution phase, the method proposed in this invention, by combining algorithmic chance, random chance, and guiding chance (see equations (23)~(25)), enables the search agent to strengthen its exploration of more likely effective directions, thereby transitioning from coarse search to fine exploration as early as possible. Consequently, the MTS learning model (i.e., a k-NN classifier based on MDTW) embedded in the feature combination optimization process can perform overload event warning tests earlier with fewer variables, thereby improving optimization efficiency. Figure 10 As shown on the right, compared to the other three methods, the method proposed in this invention achieves the best final fitness value of 0.0566 with the same number of iterations, while BGOA, BGWO, and GA methods achieve 0.0959, 0.0802, and 0.0812, respectively. In particular, compared to the original BGOA method, the method proposed in this invention optimizes the final fitness value by nearly 40%, demonstrating the effectiveness of the enhanced evolutionary computation structure based on feature priority scores.

[0112] Figure 15 The statistical comparison results of the feature combination optimization performance of each method are listed, including: (1) the average CEP accuracy on the validation dataset based on the corresponding feature variable combination; (2) the average data dimension of the feature variable combination optimized by each method; (3) the average fitness function value corresponding to the feature variable combination optimized by each method; and (4) the average computation time of feature combination optimization. Figure 15As shown, the method proposed in this invention exhibits the fastest computational speed, with an average computation time of 1689.019 seconds, while the original BGOA method requires 2736.710 seconds (the computation time is reduced by nearly 38% through the enhanced evolutionary computation structure designed in this invention). Furthermore, the method of this invention combines the social force rules of the original BGOA method (see...) Figure 6 The proposed method, along with its designed enhanced evolutionary computation mechanism, demonstrates better resistance to local stagnation, resulting in improved performance in final fitness, CEP (validation set) accuracy, and dimensionality reduction. Specifically, the proposed method optimizes the feature variable combinations to an average dimensionality of 16.25, which, compared to the original data's dimensionality of 426, effectively reduces the parameter size of subsequent prediction models and the risk of overfitting.

[0113] Subsequently, based on the combination of feature variables obtained by each method, the performance of overload event warning in different MTS models was compared on the test dataset. In addition to BGOA, BGWO and GA methods, this experiment will also compare with the following commonly used methods: (1) No selection, that is, using all candidate feature variables for overload event warning test; (2) Empirical features, the power generation, load and line transmission power data of the entire system, which contain a total of 123 feature variables; (3) ReliefF method. The MTS model used for overload event warning test is a k-NN model (k=1) based on MDTW distance and an LSTM model. Among them, the LSTM model has a hidden layer of LSTM containing 50 neurons and a fully connected layer containing 50 neurons. Other settings are as follows: Dropout Rate: 20%, maximum number of iterations: 200, batch size: 20, loss function: cross-entropy loss, optimizer: Adam. For the ReliefF method, the number of selected features is set to 16, which is close to the average dimension of the feature variable combination optimized by the method of this invention (i.e. 16.25). Each method was run 20 times to report the average result, and the performance comparison results were as follows: Figure 16 As shown, the CEP average accuracy (AvgAcc) on the test dataset, the average online prediction time (AvgTp) for a single sample when using the k-NN model, and the average F-beta score (Avg) are included. The table also includes the average offline training time of the LSTM network. For intuitive comparison, the LSTM training time for each method has been converted into a relative multiple (AvgM) based on the method of this invention.

[0114] like Figure 16As shown, based on the feature variable combination optimized by the method of this invention, the MTS model achieved better heavy-load event warning performance on the test dataset, with higher average CEP accuracy and F-beta score. Compared to directly using all variables to construct the heavy-load event warning model, the MTS model optimized by feature combination achieves better test performance. Specifically, based on the feature variable combination optimized by the method of this invention, the k-NN and LSTM models achieved heavy-load event warning accuracies of 93.108% and 95.086%, respectively, and heavy-load event warning F-beta scores of 0.931 and 0.951 on the test dataset. In contrast, the model without feature combination optimization only achieved warning accuracies of 86.989% and 89.887%, and F-beta scores of 0.870 and 0.899, demonstrating the effectiveness of feature combination optimization (combination optimization) in improving the accuracy of the prediction model. Compared to empirical feature combinations, the method of this invention still achieved better heavy-load event warning test performance. While the empirical feature variable set is usually aligned with domain knowledge and has clear physical meaning, it may still contain some redundant variables, or even variables with low correlation to the prediction target. The presence of these variables may interfere with the CEP model learning the correct mapping relationship, thereby impairing the model's predictive performance. Furthermore, the data dimensionality of the empirical feature variable set is often still high. For the IEEE 39-bus system in this example, the empirical feature set still contains 123 data variables (power generation, load, and line transmission power data for the entire system). Too many input feature variables (relative to a limited number of samples) can easily increase the risk of model overfitting. Moreover, further reducing the data dimensionality of the empirical feature set usually requires a significant investment of expert manpower. This selection method is no longer sufficient to meet the current situation of rapidly increasing data volume and may result in information gaps due to unreasonable manual selection. The method of this invention uses an end-to-end processing structure, which can automatically optimize the optimal variable set based on the input data and candidate variable set to achieve better performance in the heavy load event early warning test. This not only improves the accuracy of the prediction model but also saves a significant amount of manpower. In addition, a large number of input feature variables also increases the subsequent model training time and online inference time.

[0115] like Figure 16As shown, for the k-NN model, the inference time of the model built based on empirical features is 1.925 times that of the model built based on variables selected using the method of this invention; while the training time of the former's LSTM model is 2.616 times that of the latter. This demonstrates that the method proposed in this invention not only improves model accuracy but also enhances model efficiency to some extent. For the ReliefF method commonly used in existing research, its performance in heavy-load event warning tests is also unsatisfactory, achieving only 89.785% and 90.194% event warning accuracy. This is mainly because the ReliefF method only considers the correlation between feature variables and the prediction target, while ignoring the redundancy between candidate features.

[0116] Meanwhile, the ReliefF method does not interact with the MTS learning model during its feature combination optimization process, and therefore cannot consider the actual predictive performance of the selected feature variables. Compared with BGOA, BGWO, and GA methods, the method of this invention still achieves significantly better performance in heavy-load event warning tests, demonstrating the effectiveness of the enhanced evolutionary computation structure based on feature priority scores. One of the feature variable combinations optimized using the method of this invention will be shown below. Figure 17 As shown. Compare it with... Figure 13 A comparison reveals that not all variables selected by the method of this invention come from [the present invention]. Figure 13 Some features with low mMIFS-U priority scores were also selected. This is mainly because in the method of this invention, feature priority scores only play a semi-directional guiding role; the selection of candidate variables does not entirely depend on them, but rather on a combination of algorithmic chance, random chance, and guiding chance. Therefore, features with high priority scores are not necessarily adopted in the method of this invention, and may still be eliminated during the iteration process based on the actual evaluation results.

[0117] In summary, the enhanced evolutionary computation structure based on feature priority scores can effectively improve the efficiency and results of feature combination optimization. The improved feature combination optimization method can help find more favorable variable combinations for overload event early warning, thereby effectively improving the accuracy and efficiency of the MTS overload event early warning model.

[0118] Figure 18 This is a structural block diagram of an online early warning device for overloaded power grid sections considering the ultimate transmission capacity, provided as an embodiment of this application. See also... Figure 18 ,include: The acquisition module 11 is used to acquire multidimensional measurable physical variables of the power grid under heavy load in the historical process, forming a supervised multidimensional time series dataset; Optimization module 12 is used to build the first learning model. It generates a feature subset through an evolutionary optimization method and trains the first learning model based on the feature subset and the multidimensional time series dataset. After the first learning model is trained, the evolutionary optimization method is used to calculate the fitness value of the first learning model based on the variable dimension of the feature subset and the prediction performance of the first learning model. The fitness value is used as the optimization objective to optimize the feature subset, and the feature subset with the largest fitness value is taken as the target feature subset. The early warning module 13 is used to build a second learning model. It uses a subset of target features and a supervised multidimensional time series dataset to train the second learning model and obtain a heavy-load event prediction model.

[0119] Figure 19 This is a structural block diagram of an electronic device provided according to an embodiment of this application. See also... Figure 19 Electronic devices may include Figure 18 The aforementioned online early warning device for overloaded power grid sections typically includes a processor 21 and a memory 22. The processor 21 may include one or more processing cores, such as a 4-core processor or an 8-core processor. The processor 21 can be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), or PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor is used to process data in the wake-up state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor used to process data in the standby state. The memory 22 may include one or more computer-readable storage media, which may be non-transitory. The memory 22 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage medium in memory 22 is used to store at least one instruction, which is executed by processor 21 to implement the online early warning method for overloaded power grid sections provided by an electronic device in the method embodiments of this application.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for online early warning of overload in power grid sections considering the ultimate transmission capacity, characterized in that, include: Acquire multidimensional measurable physical variables of the power grid under heavy load conditions in the historical process to form a supervised multidimensional time series dataset; A first learning model is built, and a feature subset is generated through an evolutionary optimization method. The first learning model is then trained based on the feature subset and a multidimensional time series dataset. After the first learning model is trained, the evolutionary optimization method is used to calculate the fitness value of the first learning model based on the variable dimension of the feature subset and the prediction performance of the first learning model. The fitness value is used as the optimization target to optimize the feature subset, and the feature subset with the largest fitness value is taken as the target feature subset. A second learning model was built, and trained using a subset of target features and a supervised multidimensional time series dataset to obtain a heavy-load event prediction model.

2. The online early warning method for overloaded power grid sections considering the ultimate transmission capacity according to claim 1, characterized in that, The method for determining the overload state includes: Based on the repetitive power flow method, the cross-sectional limit transmission capacity is calculated considering AC power flow security, N-1 static security, and transient stability rules. Whether it is a heavy load state is determined based on the cross-sectional limit transmission capacity and the detected cross-sectional transmission power.

3. The online early warning method for overloaded power grid sections considering the ultimate transmission capacity according to claim 1, characterized in that, The process of constructing a supervised multidimensional time series dataset includes: According to the data sampling interval By sampling the multidimensional measurable physical variables of the power grid, a historical sequence of multidimensional measurable physical variables that changes over time is obtained; Using observation time windows, with step sizes Multiple spatiotemporal data matrices are obtained by sampling on a historical sequence of multidimensional measurable physical variables; Each spatiotemporal data matrix is ​​assigned a category label to indicate the event state and form a supervised data format; and a prospective labeling strategy is used to assign time stage labels to the spatiotemporal data matrices, thereby forming a supervised multidimensional time series dataset.

4. The online early warning method for overloaded power grid sections considering the ultimate transmission capacity according to any one of claims 1 to 3, characterized in that, The steps of using an evolutionary optimization method to calculate the fitness value of the first learning model based on the variable dimension of the feature subset and the prediction performance of the first learning model, and then optimizing the feature subset with the fitness value as the optimization objective, and selecting the feature subset with the largest fitness value as the target feature subset, include: Based on the correlation between measurable physical variables and the early warning target of overload state, the priority of each measurable physical variable in the multidimensional time series dataset is determined. The binary grasshopper optimization algorithm is used, combined with the priority of measurable physical variables, to perform evolutionary optimization of the feature subset.

5. The online early warning method for overloaded power grid sections considering the ultimate transmission capacity according to claim 4, characterized in that, The priority of measurable physical variables is determined by adopting the mMIFS-U criterion and combining it with a sequential forward search strategy.

6. The online early warning method for overloaded power grid sections considering the ultimate transmission capacity according to claim 4, characterized in that, In the process of using the binary grasshopper optimization algorithm to evolve and optimize the feature subset, the initialization and evolution rules of the binary grasshopper optimization algorithm are adjusted according to the priority of the measurable physical variables, so that the measurable physical variables with high priority are retained first.

7. The online early warning method for overloaded power grid sections considering the ultimate transmission capacity according to any one of claims 1 to 3, characterized in that, The first learning model includes a k-NN classifier based on a multidimensional version of dynamic time warped distance.

8. An online early warning device for overloaded power grid sections considering the ultimate transmission capacity, characterized in that, include: The acquisition module is used to acquire multidimensional measurable physical variables of the power grid under heavy load conditions in the historical process, forming a supervised multidimensional time series dataset. The optimization module is used to build the first learning model. It generates a feature subset through an evolutionary optimization method and trains the first learning model based on the feature subset and the multidimensional time series dataset. After the first learning model is trained, the evolutionary optimization method is used to calculate the fitness value of the first learning model based on the variable dimension of the feature subset and the prediction performance of the first learning model. The fitness value is used as the optimization target to optimize the feature subset, and the feature subset with the largest fitness value is taken as the target feature subset. The early warning module is used to build a second learning model. It uses a subset of target features and a supervised multidimensional time series dataset to train the second learning model and obtain a heavy-load event prediction model.

9. An electronic device, characterized in that, This includes the online early warning device for overloaded power grid sections that takes into account the maximum transmission capacity, as described in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one piece of program code, which is executed by a processor to implement the online early warning method for overloaded power grid sections considering the ultimate transmission capacity as described in any one of claims 1 to 7.

Images