Power network operation and maintenance decision-making method and system based on artificial intelligence

Abstract

The invention discloses a power network operation and maintenance decision-making method and system based on an artificial intelligence, and relates to the field of power system operation and maintenance, and the method comprises the steps: obtaining a historical operation feature sequence in a past preset time range after the operation time of a power network reaches an operation and maintenance period, performing operation feature prediction in the maintenance time period based on an artificial intelligence body to obtain a predicted operation feature sequence; performing power operation and maintenance resource demand prediction according to the historical operation feature sequence based on an artificial intelligence body; randomly selecting a plurality of first operation and maintenance windows, indexing to obtain a plurality of first resource scheduling times corresponding to the predicted demand resources, processing the plurality of first power loss parameters, and calculating to obtain a plurality of first operation and maintenance scores; and according to the plurality of first operation and maintenance scores, continuing selection optimization of the operation and maintenance window to obtain an optimal operation and maintenance window as an operation and maintenance decision result. The problems that in the prior art, dynamic adjustment cannot be conducted to adapt to a complex operation environment, operation and maintenance are not timely, and power loss is large are solved.

Description

Technical Field

[0001] This invention relates to the field of power system operation and maintenance, and specifically to a power network operation and maintenance decision-making method and system based on artificial intelligence agents. Background Technology

[0002] Existing technologies lack dynamic prediction and optimization of operating status, rely on human experience, cannot dynamically adjust to complex operating environments, and are prone to maintenance losses due to misjudgment, resulting in untimely operation and maintenance, low resource scheduling efficiency, and significant power loss.

[0003] In summary, a method is needed that can comprehensively predict operational characteristics and resource requirements, and optimize the selection of operation and maintenance windows based on multiple objectives. Summary of the Invention

[0004] This application provides a power network operation and maintenance decision-making method and system based on artificial intelligence agents, aiming to solve the problems in existing technologies that cannot dynamically adjust to adapt to complex operating environments, are prone to errors in maintenance judgment, resulting in untimely operation and maintenance and significant power losses.

[0005] In view of the above problems, this application provides a power network operation and maintenance decision-making method and system based on artificial intelligence agents.

[0006] Firstly, this application provides a power network operation and maintenance decision-making method based on artificial intelligence agents, including: After the power network has reached the maintenance cycle, historical operation feature sequences within a preset time range are obtained, and operation feature predictions are made within the maintenance period based on artificial intelligence to obtain predicted operation feature sequences. Based on artificial intelligence, the power operation and maintenance resource demand is predicted according to the historical operation feature sequence to obtain the predicted resource demand. During the maintenance period, multiple first maintenance windows are randomly selected, multiple first resource scheduling times corresponding to the predicted demand resources are obtained by indexing, and multiple first power loss parameters are processed based on the predicted operation feature sequence index to calculate multiple first maintenance scores. Based on multiple first-level operation and maintenance scores, the selection and optimization of operation and maintenance windows are further carried out to obtain the optimal operation and maintenance window as the operation and maintenance decision result. In the optimization process, resource similarity weights and power similarity weights are configured according to the historical operation feature sequence to calculate and merge the similarity of operation and maintenance windows.

[0007] Secondly, this application provides a power network operation and maintenance decision-making system based on artificial intelligence agents, including: The operation feature sequence prediction module is used to obtain historical operation feature sequences within a preset time range after the power network operation time reaches the maintenance cycle, and to predict the operation features within the maintenance period based on the artificial intelligence agent to obtain the predicted operation feature sequence. The demand resource prediction module is used to predict the demand for power operation and maintenance resources based on the historical operating characteristic sequence using an artificial intelligence agent, and to obtain the predicted demand resources. The first operation and maintenance score calculation module is used to randomly select multiple first operation and maintenance windows within the maintenance period, index and obtain multiple first resource scheduling times corresponding to the predicted demand resources, and process multiple first power loss parameters based on the predicted operation feature sequence index to calculate multiple first operation and maintenance scores. The operation and maintenance window optimization module is used to further optimize the selection of operation and maintenance windows based on multiple first operation and maintenance scores, and obtain the optimal operation and maintenance window as the operation and maintenance decision result. In the optimization process, resource similarity weights and power similarity weights are configured according to the historical operation feature sequence to calculate and merge the similarity of operation and maintenance windows.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application first improves the accuracy and standardization of operational feature prediction by forecasting operational feature sequences, providing high-quality data support for subsequent operation and maintenance decisions. Second, based on these historical operational feature sequences, an artificial intelligence agent is used to obtain predicted resource demands, establishing a dynamic mapping relationship between operational status and resource demands, avoiding resource redundancy or shortage, and providing a scientific basis for resource scheduling plans. Simultaneously, the first resource scheduling time corresponding to the predicted resource demands is obtained through indexing, and the first operation and maintenance score is calculated to achieve preliminary screening of operation and maintenance windows. The random window selection method ensures a global exploration of operation and maintenance windows, avoiding the blindness of manual window selection. Finally, resource similarity weights and power similarity weights are configured, and window iterative optimization is achieved through similarity calculation, dynamically adapting to different load fluctuations, narrowing the optimal window search range, and ultimately obtaining the optimal operation and maintenance window. This realizes full-process automation of power network operation and maintenance from data collection, forecasting, evaluation to decision-making, effectively solving problems such as traditional operation and maintenance relying on manual labor, inability to dynamically adapt to complex environments, untimely operation and maintenance, and large power losses, significantly improving the reliability, economy, and stability of power grid operation and maintenance. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A flowchart illustrating a power network operation and maintenance decision-making method based on artificial intelligence agents; Figure 2 This is a schematic diagram of the structure of a power network operation and maintenance decision-making system based on artificial intelligence agents.

[0011] The labels in the attached diagram are explained as follows: The module includes: 11 for predicting the running feature sequence; 12 for predicting the demand resources; 13 for calculating the first operation and maintenance score; and 14 for optimizing the operation and maintenance window. Detailed Implementation

[0012] This application provides a power network operation and maintenance decision-making method and system based on artificial intelligence agents, which is used to address the problems in existing technologies that cannot dynamically adjust to adapt to complex operating environments and are prone to untimely operation and maintenance and significant power loss due to errors in maintenance judgment.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, this application provides a power network operation and maintenance decision-making method based on artificial intelligence agents, the method comprising: S10: After the power network operation time reaches the maintenance cycle, obtain the historical operation feature sequence within the past preset time range, and predict the operation feature within the maintenance period based on the artificial intelligence agent to obtain the predicted operation feature sequence. In this embodiment, the operating characteristics of the power network within a preset time range are retrieved from the power network monitoring system or database, and the historical operating characteristic sequence is obtained after data processing. The load sequence within the maintenance window is then input into a pre-trained artificial intelligence model to predict the operating characteristic sequence.

[0016] The preset time range can be the past month, acquiring historical characteristic operation sequences for the past month, including hourly recorded power load values. The maintenance period is the predicted time range of the AI ​​agent.

[0017] Step S10 in the method provided in this application embodiment includes: After the power network has reached its maintenance cycle, a sequence of historical operating characteristics within a preset time range is obtained, wherein each operating characteristic includes power load. Invoke an AI-based predictive agent that predicts operational features; The historical running feature sequence is input into the running feature prediction agent, and the predicted running feature sequence is output.

[0018] In this embodiment of the application, after the power network has reached the maintenance cycle, a historical sequence of operating characteristics within a preset time range is obtained, wherein each operating characteristic includes power load.

[0019] Electricity load refers to the total amount of electricity required or consumed by a power system or power equipment at a certain moment or over a period of time, usually expressed as electrical power. Based on the time of load occurrence, electricity load can also be divided into peak load, minimum load, and average load.

[0020] Specifically, when the power network reaches a preset maintenance cycle, operational data from the past week, including power load values, is retrieved from the historical database. This retrieved operational data is then standardized to form a regular historical operational characteristic sequence. This provides a high-quality data foundation for subsequent intelligent forecasting.

[0021] Secondly, based on the preset configuration information, the latest version of the running feature prediction agent is called from the model management platform.

[0022] Artificial intelligence construction refers to the process of building an intelligent system with autonomous perception, learning, decision-making, and execution capabilities through specific technologies, tools, and methods. Invocation refers to the process of loading a pre-trained model into a computer system, preparing the runtime environment, and executing the prediction task, including model loading, parameter initialization, and runtime environment configuration.

[0023] Specifically, the predictive agent for operational characteristics can be constructed using advanced artificial intelligence technology to meet the needs of power load forecasting and provide intelligent predictions for the future. Through artificial intelligence, the network structure of the agent is specially designed to adapt to the specific requirements of power load forecasting, address long-term dependency issues, and capture multi-period patterns.

[0024] Finally, the historical running feature sequence is input into the running feature prediction agent, and the predicted running feature sequence is output.

[0025] Among them, the predicted running feature sequence is the structured prediction result output by the agent, which contains the running feature value at each prediction time point in the future, and constitutes a complete prediction sequence within the maintenance period.

[0026] Specifically, the historical operational feature sequence is preprocessed to transform it into the input format required by the predictive agent. Based on the power load at multiple times over the past 7 days, the power load at multiple times during future maintenance periods is predicted. The output is then denormalized, transformed, and timestamped to form a complete predictive operational feature sequence. The input to the predictive agent can include batch size, time step, and feature dimension, i.e., power load value.

[0027] For example, the input data is [1,168,1], where 1 represents the batch size; 168 represents the time step, 7 days × 24 hours; and 1 represents the feature dimension, which is the power load value. The predicted running feature sequence is obtained by predicting through the running feature prediction agent.

[0028] In step S10 of the method provided in this application embodiment, the step of building the feature prediction agent includes: Based on the past operation monitoring data of the power network, a set of historical operation feature sequences of samples is collected, and the operation features of each sample during the maintenance period are collected to obtain a set of predicted operation feature sequences of samples. Based on artificial intelligence, an input layer, hidden layer, fully connected layer and output layer are constructed within the intelligent body that predicts operational features. The historical operational feature sequence set and the predicted operational feature sequence set of the samples are used to conduct supervised training on the operational feature prediction agent until the test converges, thus completing the setup and configuring it in the power network operation and maintenance center.

[0029] In this embodiment of the application, the operation monitoring data of the power network in the past month are first collected based on a preset time window to obtain a set of sample historical operation feature sequences. Then, the operation features of each sample historical operation feature sequence are collected during the maintenance period to obtain a set of sample predicted operation feature sequences.

[0030] For example, 30 days of sample sequences were collected, with a sampling interval of 15 minutes.

[0031] Secondly, based on artificial intelligence, an input layer, hidden layer, fully connected layer, and output layer are constructed within the intelligent body that predicts operational features.

[0032] For example, an agent for predicting features is built based on LSTM.

[0033] First, data preparation involves preprocessing the historical running feature sequence set and the predicted running feature sequence set of the samples. By standardizing, the parameters are normalized to the range of [-1,1] or [0,1] to eliminate the difference in units. The sliding window technique is then used to divide the historical running feature sequence set and the predicted running feature sequence set of the samples into corresponding input-output pairs, which serve as training samples for the LSTM model.

[0034] Secondly, the model is constructed, consisting of an input layer, hidden layers, fully connected layers, and an output layer within the feature prediction agent. The network structure comprises an input layer, three hidden layers, two fully connected layers, and an output layer. The input layer receives the set of historical running feature sequences and the set of predicted running feature sequences as input. The hidden layers, located between the input and output layers, are responsible for extracting abstract features from the batch size, time steps, and feature dimensions of the input. The fully connected layers use the ReLU activation function to integrate and transform the features extracted by the hidden layers. The output layer outputs the predicted future parameters or the device state classification result.

[0035] Next, model training is performed, with the sample set divided into training, validation, and test sets in a 7:2:1 ratio. The mean squared error (MSE) is used as the loss function during training, and the Adam optimizer is used for parameter updates. The initial learning rate is set to 0.001, and a learning rate decay strategy is employed. Model performance is evaluated on the validation set after each training epoch. Training is terminated when the validation set loss no longer decreases for 10 consecutive epochs. Finally, a final evaluation is performed on the test set. A mean absolute percentage error (MASE) of less than 2.5% in load prediction is considered a valid operational resource prediction agent.

[0036] Finally, the model is deployed in the power network operation and maintenance center to collect equipment data in real time and output prediction results. When the predicted parameter value exceeds the normal operation constraints, an early warning is triggered, realizing intelligent monitoring and fault prevention of the operating status of electrical equipment.

[0037] In this embodiment, a high-quality data foundation is provided for prediction through systematic data acquisition and processing. The use of a specially trained predictive agent significantly improves the accuracy and reliability of feature prediction, automates and standardizes the prediction process, reduces manual intervention, and increases work efficiency.

[0038] S20: Based on the artificial intelligence entity, predict the power operation and maintenance resource demand according to the historical operation characteristic sequence, and obtain the predicted resource demand. Predicted resource demand refers to the types and quantities of resources required during operation and maintenance, as predicted by an artificial intelligence agent.

[0039] In this embodiment, an operation and maintenance resource prediction agent is used to input the acquired historical operational characteristic sequence into the agent. Based on patterns in the historical data, the operation and maintenance resource prediction agent predicts the resources that may be needed for maintenance. If the power equipment had excessive power load in the past, the more resources would be required, and the more resources would be allocated subsequently.

[0040] Step S20 in the method provided in this application embodiment includes: The operation and maintenance resource prediction intelligent agent built on artificial intelligence is invoked. The operation and maintenance resource prediction intelligent agent is trained using a set of sample historical operation feature sequences and a set of sample resource requirements. Each sample resource requirement includes the type of resource requirement. The historical operational feature sequence is input into the operation and maintenance resource prediction agent, which outputs the predicted resource demand.

[0041] In this embodiment, an AI-based operation and maintenance resource prediction agent is first invoked. This agent is trained using a set of historical operational feature sequences and a set of resource requirements from samples; each sample's resource requirement includes the type of resource required.

[0042] Among them, the operation and maintenance resource prediction intelligent agent refers to an artificial intelligence model specifically designed to predict resource demand, and its training is based on historical operational characteristics and resource usage data. The sample demand resource set refers to the resource usage records in the training data, including resource type and quantity. Demand resource type refers to the classification of resources, such as manpower, equipment, materials, etc.

[0043] Specifically, an operational resource prediction agent is trained using a set of historical operational feature sequences and a set of resource requirements from the samples. The trained agent learns from the data on resource requirements from the samples and outputs a prediction result that also includes the type of resource required. Subsequently, by calling the specially trained operational resource prediction agent, accurate prediction of resource requirements can be achieved.

[0044] Secondly, the historical operational characteristic sequence is input into the operation and maintenance resource prediction agent, which outputs the predicted resource demand.

[0045] Specifically, the historical operational feature sequence is input into the operation and maintenance resource prediction agent, the transformed feature tensor is input into the resource prediction agent, the model generates preliminary prediction results through forward computation, and then the prediction results are post-processed to finally output a structured predicted resource demand, including resource type and estimated quantity.

[0046] For example, inputting the power load operation characteristics of the past 30 days, the model outputs the predicted demand: 3 maintenance engineers, 1 insulation tester, and 1 crane.

[0047] In this embodiment, by invoking an intelligent agent for predicting operational resources, a mapping from operational characteristics to resource requirements is achieved, providing data support for operational resource planning and improving the accuracy of resource forecasting. This overcomes the limitations of traditional resource estimation relying on manual experience, enabling dynamic matching of resource demand with the operating status of the power network.

[0048] After successfully invoking the resource prediction agent, historical operational data needs to be input into the model to perform predictive inference. To address the above issues, this application randomly selects multiple first maintenance windows within the maintenance period, indexes and obtains multiple first resource scheduling times corresponding to the predicted resource demand, and processes multiple first power loss parameters based on the predicted operational feature sequence index to calculate multiple first maintenance scores.

[0049] S30: Randomly select multiple first operation and maintenance windows during the maintenance period, index and obtain multiple first resource scheduling times corresponding to the predicted demand resources, and process multiple first power loss parameters based on the predicted operation feature sequence index to calculate multiple first operation and maintenance scores. In this embodiment of the application, a maintenance time window is randomly selected, the time for resource scheduling within each time window is obtained, and the power loss caused by power outage maintenance is obtained. Based on the resource scheduling time and power loss, the first power loss parameter of the first maintenance window is calculated, and a large number of possible maintenance windows within the maintenance period are initially screened through the first maintenance score.

[0050] The first maintenance window is a randomly selected time interval within the maintenance period. The further the maintenance time window is from peak hours, such as working hours, the longer the resource scheduling time and the smaller the power loss; the closer the maintenance time window is to peak hours, the shorter the resource scheduling time and the greater the power loss within the maintenance time window.

[0051] Step S30 in the method provided in this application embodiment includes: Multiple first maintenance windows are randomly selected during the maintenance period, and these multiple first maintenance windows are not the same. Multiple first operation and maintenance windows are input into the resource scheduling table, and the predicted demand resources are indexed to obtain multiple first resource scheduling times. The resource scheduling table includes a mapping relationship between a sample operation and maintenance window set and multiple sample resource scheduling time sets of various resources. The sum of the resource scheduling times of the indexed predicted demand resources is calculated as the first resource scheduling time. Based on multiple first operation and maintenance windows, multiple first operation features are obtained by indexing within the predicted operation feature sequence, which serve as multiple first power loss parameters; Based on the multiple first resource scheduling times and multiple first power loss parameters, multiple first operation and maintenance scores are calculated.

[0052] In this embodiment of the application, multiple first maintenance windows are randomly selected during the maintenance period, and the multiple first maintenance windows are different.

[0053] Specifically, multiple first maintenance windows are obtained by randomly selecting and uniformly sampling across all possible maintenance windows, while ensuring that these multiple first maintenance windows are distinct. This guarantees the effectiveness of the random selection and avoids invalid data. The distinct first maintenance windows ensure diversity and prevent getting trapped in local optima.

[0054] For example, 50 non-overlapping 1-hour time periods are selected without replacement as the first maintenance window using a random number generator.

[0055] Secondly, multiple first operation and maintenance windows are input into the resource scheduling table, and the index processing predicts the resource demand to obtain multiple first resource scheduling times. The resource scheduling table includes the mapping relationship between the sample operation and maintenance window set and multiple sample resource scheduling time sets for various resources. The sum of the resource scheduling times for the indexed predicted demand resources is calculated as the first resource scheduling time.

[0056] The resource scheduling table is a mapping table that stores the correspondence between operation and maintenance windows and resource scheduling times. The sample operation and maintenance window set refers to historically used operation and maintenance window samples. The sample resource scheduling time set is the set of time used by historical operation and maintenance windows during resource scheduling; it represents the resource scheduling time records corresponding to the sample operation and maintenance window set. The first resource scheduling time is the total resource scheduling time retrieved from the resource scheduling table for a specific operation and maintenance window. The shorter the time, the more thorough the prior resource preparation for the operation and maintenance window, the better the operation and maintenance situation, and thus, the more optimal the corresponding window.

[0057] Specifically, the resource scheduling table stores the actual time spent scheduling various resources under different past operation and maintenance windows. After inputting the first operation and maintenance window, the corresponding resource scheduling time record for that window can be retrieved based on the mapping relationship between the sample operation and maintenance window set and multiple sample resource scheduling time sets. Then, all resource scheduling times retrieved from the resource scheduling table are summed to obtain the first resource scheduling time.

[0058] For example, the first maintenance window is from 08:00 to 12:00 on October 18th. Operational characteristics: Weekday morning rush hour, scheduling time for maintenance engineers: 1.5 hours; Operational characteristics: Weekday, scheduling time for insulation testers: 1.0 hour; Operational characteristics: Urban weekday daytime, scheduling time for cranes: 2.5 hours. First resource scheduling time = 1.5h + 1.0h + 2.5h = 5.0h.

[0059] Furthermore, based on multiple first operation and maintenance windows, multiple first operation features are obtained by indexing within the predicted operation feature sequence, serving as multiple first power loss parameters.

[0060] Specifically, based on the time of the first maintenance window, the load data corresponding to that time period is extracted from the predicted operation feature sequence, and the power loss due to the suspension of operation during the first maintenance window is calculated and used as the first power loss parameter.

[0061] The first power loss parameter is calculated as: average load during the operation and maintenance period × operation and maintenance duration. It represents the power loss caused by power outages during this period. A larger first power loss parameter indicates that the first operation and maintenance window is closer to the peak load period. A smaller first power loss parameter indicates a more favorable window.

[0062] The power loss varies depending on the operation and maintenance period: during off-peak hours, the power loss is small and the interference with the power grid is minimal; during peak hours, the power loss is large and the interference with the power grid is also significant. Therefore, by using the first power loss parameter, excessive power loss can be avoided and power outage losses can be reduced.

[0063] For example, the first maintenance window is from 08:00 to 12:00 on October 18th. The load values ​​for every 15 minutes within the 1-hour window are extracted from the predicted operating characteristic sequence [8:00, 55kW, 8:15, 110kW, ..., 11:45, 65kW]. The average load data is 75kW, and the first power loss parameter = 75kW × 4h = 300kW.

[0064] Furthermore, multiple first operation and maintenance scores are obtained based on multiple first resource scheduling times and multiple first power loss parameters.

[0065] The primary maintenance score is calculated by fusing two indicators with different dimensions: resource scheduling time and power loss parameters. This score is used to rank all candidate windows. The higher the primary maintenance score, the better the overall performance of the maintenance window, the better the maintenance effect of the candidate window, and the greater the probability of it being selected as the optimal window.

[0066] Specifically, by using a weighted fusion method to calculate the first resource scheduling time and the first power loss parameter, the corresponding first operation and maintenance score can be obtained.

[0067] Step S30 of the method provided in this application embodiment calculates multiple first operation and maintenance scores based on the multiple first resource scheduling times and multiple first power loss parameters, including: Obtain baseline resource scheduling time and baseline power loss parameters; Calculate the ratio of the baseline resource scheduling time to the multiple first resource scheduling times to obtain multiple first resource scheduling coefficients; The ratios of the baseline power loss parameter and the multiple first power loss parameters are calculated respectively to obtain multiple first power coefficients; Multiple first operation and maintenance scores are calculated based on multiple first resource scheduling coefficients and multiple first power coefficients.

[0068] In this embodiment of the application, the baseline resource scheduling time and baseline power loss parameters are first obtained.

[0069] The baseline resource scheduling time is set based on the historical average scheduling time and used as a standard value for comparison. The baseline power loss parameter is set based on the historical average loss value.

[0070] Specifically, the first resource scheduling time and the first power loss parameter can be compared with the benchmark resource scheduling time and the benchmark power loss parameter respectively. By comparing with the benchmark value, the resource scheduling efficiency and the degree of power loss can be obtained.

[0071] For example, the baseline resource scheduling time is 6 hours and the baseline power loss parameter is 460kW. Secondly, the first resource scheduling coefficient is calculated, and the ratio of the baseline resource scheduling time to the first resource scheduling time is used as the first resource scheduling coefficient.

[0072] Specifically, the first resource scheduling coefficient = baseline resource scheduling time / first resource scheduling time, which reflects scheduling efficiency. The larger the first resource scheduling coefficient, the faster the scheduling efficiency.

[0073] For example, the first resource scheduling time is 5 hours, and the first resource scheduling coefficient is 6 / 5 = 1.2.

[0074] Next, calculate the first power coefficient, which is the ratio of the baseline power loss parameter to the first power loss parameter.

[0075] Specifically, the first power coefficient = baseline power loss parameter / first power loss parameter, which reflects the degree of power loss. The larger the first power coefficient, the smaller the power loss.

[0076] For example, the first power loss parameter is 300kW, and the first power coefficient is 460 / 300≈1.53.

[0077] Finally, based on multiple first resource scheduling coefficients and multiple first power coefficients, multiple first operation and maintenance scores are calculated.

[0078] Specifically, the first resource scheduling time and the first power loss parameter are calculated using a weighted scoring method or a normalized weighted method to obtain the corresponding first operation and maintenance score. Normalization is achieved by dividing the resource scheduling time and power loss parameter of each window by the maximum value among all candidate windows, eliminating dimensions and obtaining a normalization coefficient with a value between 0 and 1.

[0079] The first maintenance score is calculated as w1 × first resource scheduling coefficient + w2 × first power coefficient. A higher first maintenance score indicates a better window selection, guiding the choice of an efficient window. w1 and w2 are weights. Weights are assigned based on resource scheduling efficiency and the impact of reducing power losses on grid maintenance; these weights can be configured by the user according to their specific scenario.

[0080] If the first maintenance score is low, it indicates that the overall performance of the maintenance window is worse. If the first maintenance score is high, the overall performance of the maintenance window is better, which can guide the selection of an efficient window.

[0081] For example, w1 and w2 are 0.6 and 0.4 respectively, and the first maintenance score = 0.6 × 1.2 + 0.4 × 1.53 ≈ 1.33.

[0082] By weighted calculation, the first maintenance score of multiple first maintenance windows can be obtained. By comparing the first maintenance scores, the first maintenance window with the highest first maintenance score can be selected to perform optimal maintenance window search. By integrating resource scheduling time index and power loss index, the final result is fast resource scheduling and low power loss, thus improving maintenance efficiency.

[0083] In this embodiment, random selection ensures global exploration, and weighted calculation, combining resource efficiency and power loss, evaluates the window's operational effectiveness, facilitating comprehensive operational decision-making. Simultaneously, a dual-dimensional evaluation using resource scheduling time and power loss parameters comprehensively considers operational economy and grid stability, making the initial screening results more objective. A standardized scoring calculation system, taking into account both scheduling efficiency and power loss, lays a unified evaluation foundation for subsequent optimization steps.

[0084] After obtaining multiple operation and maintenance scores, further optimization of window selection is needed. To address the above issues, this application continues to optimize the selection of operation and maintenance windows based on multiple first operation and maintenance scores to obtain the optimal operation and maintenance window as the operation and maintenance decision result. In the optimization process, resource similarity weights and power similarity weights are configured according to historical operating characteristic sequences to calculate and merge the similarity of operation and maintenance windows.

[0085] S40: Based on multiple first maintenance scores, continue to optimize the selection of maintenance windows to obtain the optimal maintenance window as the maintenance decision result. In the optimization process, resource similarity weights and power similarity weights are configured according to the historical operation feature sequence to calculate and merge the similarity of maintenance windows.

[0086] In this embodiment, based on multiple first maintenance scores, a second maintenance window is generated by selecting a maintenance window. During the optimization process, the similarity of the maintenance windows is analyzed, and the two windows with the highest similarity are merged. Finally, the optimal maintenance window with the highest score is selected as the maintenance decision result. Through dynamic optimization and similarity processing, the accuracy and efficiency of decision-making are improved. The weight configuration is based on historical data, making the optimization more closely aligned with actual operational characteristics.

[0087] Step S40 in the method provided in this application embodiment includes: During the maintenance period, multiple first maintenance windows are adjusted by window step size to obtain multiple second maintenance windows, and multiple second maintenance scores are obtained through processing. Based on the historical operation feature sequence, resource similarity weights and power similarity weights are configured, and the window similarity of multiple first operation and maintenance windows and multiple second operation and maintenance windows is calculated. Filter the first and second maintenance windows with the highest similarity, delete the maintenance windows with low maintenance scores, and obtain the remaining maintenance windows. Continue iterative optimization until convergence, obtaining the optimal operation and maintenance window with the highest operation and maintenance score, which serves as the operation and maintenance decision result.

[0088] In this embodiment of the application, firstly, during the maintenance period, multiple first maintenance windows are adjusted by window step size to obtain multiple second maintenance windows, and multiple second maintenance scores are obtained through processing.

[0089] Specifically, during the maintenance period, after obtaining multiple first maintenance windows and their scores, multiple adjacent candidate windows are generated for each first maintenance window using a preset window step size as the adjustment unit, which serve as second maintenance windows.

[0090] The window step size is the smallest time unit for moving the maintenance window on the timeline, typically set to a fixed interval of 15 minutes or 30 minutes. The preset window step size can be set before selecting the second maintenance window.

[0091] Smaller step sizes can improve search accuracy, but excessively small step sizes can increase computational load, so they cannot be set too small. Larger step sizes can speed up convergence, but excessively large step sizes can reduce accuracy, so they cannot be set too large. A suitable step size should be chosen to adjust the window movement. In this screening process, the preset window step size is 15 minutes.

[0092] Then, for each newly generated second maintenance window, the resource scheduling time calculation, power loss parameter evaluation, and maintenance score calculation are repeated to obtain the second maintenance score for the second maintenance window.

[0093] For example, by sliding the window in 15-minute increments, multiple second maintenance windows can be generated, which can be adjusted forward by 15 minutes [07:45-11:45] and backward by 15 minutes [08:15-12:15].

[0094] Secondly, based on the historical operation characteristic sequence, resource similarity weights and power similarity weights are configured, and then the similarity of multiple windows is calculated based on multiple first operation and maintenance windows and multiple second operation and maintenance windows.

[0095] In the historical operating characteristic sequence, the higher the historical load, the more important the maintenance quality, the greater the resource similarity weight, and the smaller the corresponding power similarity weight. The weights can be adjusted accordingly using the historical operating characteristic sequence to ensure that the configured weights are suitable for the operating environment.

[0096] Furthermore, the first and second maintenance windows with the highest similarity are selected, and the maintenance windows with low maintenance scores are deleted to obtain the remaining maintenance windows.

[0097] By comparing the similarity of all first and second maintenance windows, the maintenance window with the highest similarity is obtained. After completing the comparison of all similarities, multiple maintenance windows with the highest similarity can be obtained. Among these multiple maintenance windows with the highest similarity, maintenance scores are used for further filtering.

[0098] Specifically, by filtering for similarity, K first and second maintenance windows with the highest similarity are obtained. Based on the maintenance score, the K first and second maintenance windows with the highest similarity are filtered, and maintenance windows with maintenance scores less than 1 are deleted, while maintenance windows with high maintenance scores are retained, resulting in multiple maintenance windows with high maintenance scores.

[0099] Furthermore, iterative optimization continues until convergence, obtaining the optimal operation and maintenance window with the highest operation and maintenance score, which serves as the operation and maintenance decision result.

[0100] The operation and maintenance (O&M) score is used as a complete screening process. If multiple O&M windows are obtained after one screening, the screening process is repeated. This iteration is repeated multiple times, and the change in the O&M score during each iteration is calculated as the criterion for whether to continue the iteration. The maximum number of iterations is reached when the O&M score does not change significantly after 10 consecutive iterations. The remaining O&M windows from the K selected O&M windows are the optimal O&M windows. The optimal O&M window with the highest score is selected as the O&M decision result, and the final O&M operation can be performed based on this decision.

[0101] Step S40 of the method provided in this application embodiment, based on the historical operation feature sequence, configures resource similarity weights and power similarity weights, and calculates the window similarity of multiple first operation and maintenance windows and multiple second operation and maintenance windows, including: Calculate the mean of the historical operating feature sequence to obtain the historical average operating feature, and calculate the ratio of this ratio to the maximum operating feature as the resource similarity weight; Based on the resource similarity weights, the power similarity weights are calculated. Calculate the similarity of resource scheduling time and power loss parameters for each first maintenance window and each second maintenance window to obtain multiple resource similarities and multiple power similarities. Perform weighted calculation based on the resource similarity weight and power similarity weight to obtain multiple window similarities.

[0102] In this embodiment of the application, the mean of the historical operating feature sequence is calculated to obtain the historical average operating feature, and the ratio of the average operating feature to the maximum operating feature is calculated as the resource similarity weight. Specifically, the historical operating characteristic sequence is first cleaned, preprocessed, and statistically calculated to determine its mean, which is then used as the historical average operating characteristic. The ratio of the historical average operating characteristic to the maximum operating characteristic is then used as the resource similarity weight. This ratio reflects the load fluctuation characteristics of the power grid operation, providing a quantitative basis for weight allocation.

[0103] For example, the maximum rated load of the power grid is used as the maximum operating characteristic. The historical operating characteristic sequence is 850kW, while the maximum operating characteristic is 1350kW. The ratio is 850 / 1350≈0.63, that is, the resource similarity weight is 0.63.

[0104] Secondly, based on the resource similarity weights, the power similarity weights are calculated.

[0105] Specifically, the power similarity weight is calculated as 1 minus the resource similarity weight, and the sum of the two weights equals 1, ensuring the proportionality and comparability of the similarity calculation. The weights are set based on the trade-off between resource scheduling efficiency and power economy, allowing the calculation to adapt to different operating environments.

[0106] For example, the power similarity weight = (1-0.63) = 0.37.

[0107] Finally, the similarity of resource scheduling time and power loss parameters for each first maintenance window and each second maintenance window is calculated to obtain multiple resource similarities and multiple power similarities. Based on the resource similarity weight and power similarity weight, a weighted calculation is performed to obtain multiple window similarities.

[0108] Specifically, for each pair of first and second maintenance windows, the similarity of resource scheduling times is calculated. For resource scheduling time, the resource similarity between the first and second maintenance windows is calculated as: 1 - |Difference in resource scheduling time between the first and second maintenance windows| / Resource scheduling time of the second maintenance window. A similar formula is used for power loss parameters: Power similarity between the first and second maintenance windows is calculated as: 1 - |Difference in power loss parameter between the first and second maintenance windows| / Power loss parameter of the second maintenance window.

[0109] Based on the calculated resource similarity weights and power similarity weights, the resource similarity and power similarity are weighted and calculated to obtain the window similarity between the first and second maintenance windows. Window similarity = resource similarity weight × resource similarity + power similarity weight × power similarity.

[0110] By calculating window similarity using weighted averages, the importance of different characteristic dimensions of resource scheduling time and power loss in the overall evaluation is demonstrated. Higher window similarity allows for faster filtering; when high-scoring maintenance windows are obtained, similarly scored maintenance windows can be quickly identified, facilitating window optimization and reducing decision-making time.

[0111] For example, the first resource scheduling time is 5 hours, the second resource scheduling time is 5.8 hours, and the resource similarity is 1 - |5 - 5.8| / 5.8 ≈ 0.86; the first power loss parameter is 300kW, the second power loss parameter is 350kW, and the power similarity is 1 - |300 - 350| / 350 ≈ 0.857. Next, the system uses the previously determined weights for weighted synthesis: window similarity = 0.63 × 0.86 + 0.37 × 0.857 ≈ 0.86.

[0112] In this embodiment, the optimal window search range is gradually narrowed through window step size adjustment and iterative optimization, avoiding the limitations of traditional fixed window selection and achieving dynamic optimization and accuracy improvement of the operation and maintenance window. Secondly, by configuring resource similarity weights and power similarity weights, the adaptability and rationality of window optimization are improved. Similarity calculation can adapt to different load fluctuation scenarios, enhancing the adaptability and rationality of window optimization. Finally, by filtering the window with the highest similarity and deleting low-score windows, the number of redundant candidate windows is reduced, the amount of iterative computation is decreased, ensuring the final output of the optimal operation and maintenance window and improving the execution effect of operation and maintenance decisions.

[0113] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects: First, systematic data acquisition and processing provide a high-quality data foundation for prediction. Employing specially trained predictive agents improves the accuracy and reliability of feature predictions, reduces human intervention, and increases work efficiency.

[0114] Secondly, by invoking an intelligent agent for predicting operational resources, a mapping from operational characteristics to resource requirements is achieved, providing data support for operational resource planning and improving the accuracy of resource forecasting. This overcomes the limitations of traditional resource estimation relying on manual experience, enabling dynamic matching of resource demand with the operational status of the power network.

[0115] Furthermore, random selection ensures global exploration, and weighted calculations comprehensively consider resource efficiency and power loss to evaluate the operational effectiveness of the window. Simultaneously, a dual-dimensional evaluation using resource scheduling time and power loss parameters comprehensively considers operational economy and grid stability, making the initial screening results more objective. A standardized scoring calculation system is constructed, balancing generality and individual needs, laying a unified evaluation foundation for subsequent optimization steps. Finally, by adjusting the window step size and iteratively optimizing, the search range of the optimal window is gradually narrowed, avoiding the limitations of traditional fixed window selection and achieving dynamic optimization and accuracy improvement of the operation and maintenance window. Secondly, by configuring resource similarity weights and power similarity weights, the adaptability and rationality of window optimization are improved. Similarity calculation can adapt to different load fluctuation scenarios, enhancing the adaptability and rationality of window optimization. Finally, by selecting the window with the highest similarity and deleting low-score windows, the number of redundant candidate windows is reduced, the amount of iterative computation is decreased, ensuring the final output of the optimal operation and maintenance window and improving the execution effect of operation and maintenance decisions.

[0116] Example 2, as Figure 2 As shown, this application provides a power network operation and maintenance decision-making system based on artificial intelligence agents, the system comprising: The operation feature sequence prediction module 11 is used to obtain historical operation feature sequences within a preset time range after the power network operation time reaches the operation and maintenance cycle, and to predict the operation features within the maintenance period based on the artificial intelligence entity to obtain the predicted operation feature sequence. Demand resource prediction module 12 is used to predict the demand for power operation and maintenance resources based on artificial intelligence and the historical operation feature sequence, and obtain the predicted demand resources. The first operation and maintenance score calculation module 13 is used to randomly select multiple first operation and maintenance windows within the maintenance period, index and obtain multiple first resource scheduling times corresponding to the predicted demand resources, and process multiple first power loss parameters based on the predicted operation feature sequence index to calculate multiple first operation and maintenance scores. The operation and maintenance window optimization module 14 is used to continue to optimize the selection of operation and maintenance windows based on multiple first operation and maintenance scores, and obtain the optimal operation and maintenance window as the operation and maintenance decision result. In the optimization process, resource similarity weights and power similarity weights are configured according to the historical operation feature sequence to calculate and merge the similarity of operation and maintenance windows.

[0117] In one embodiment, the feature sequence prediction module 11 is run to: After the power network has reached its maintenance cycle, a sequence of historical operating characteristics within a preset time range is obtained, wherein each operating characteristic includes power load. Invoke an AI-based predictive agent that predicts operational features; The historical running feature sequence is input into the running feature prediction agent, and the predicted running feature sequence is output.

[0118] The steps for building the feature prediction agent include: Based on the past operation monitoring data of the power network, a set of historical operation feature sequences of samples is collected, and the operation features of each sample during the maintenance period are collected to obtain a set of predicted operation feature sequences of samples. Based on artificial intelligence, an input layer, hidden layer, fully connected layer and output layer are constructed within the intelligent body that predicts operational features. The historical operational feature sequence set and the predicted operational feature sequence set of the samples are used to conduct supervised training on the operational feature prediction agent until the test converges, thus completing the setup and configuring it in the power network operation and maintenance center.

[0119] In one embodiment, the demand resource prediction module 12 is used for: The operation and maintenance resource prediction intelligent agent built on artificial intelligence is invoked. The operation and maintenance resource prediction intelligent agent is trained using a set of sample historical operation feature sequences and a set of sample resource requirements. Each sample resource requirement includes the type of resource requirement. The historical operational feature sequence is input into the operation and maintenance resource prediction agent, which outputs the predicted resource demand.

[0120] In one embodiment, the first operation and maintenance score calculation module 13 is used for: Multiple first maintenance windows are randomly selected during the maintenance period, and these multiple first maintenance windows are not the same. Multiple first operation and maintenance windows are input into the resource scheduling table, and the predicted demand resources are indexed to obtain multiple first resource scheduling times. The resource scheduling table includes a mapping relationship between a sample operation and maintenance window set and multiple sample resource scheduling time sets of various resources. The sum of the resource scheduling times of the indexed predicted demand resources is calculated as the first resource scheduling time. Based on multiple first operation and maintenance windows, multiple first operation features are obtained by indexing within the predicted operation feature sequence, which serve as multiple first power loss parameters; Based on the multiple first resource scheduling times and multiple first power loss parameters, multiple first operation and maintenance scores are calculated.

[0121] Among them, based on the plurality of first resource scheduling times and the plurality of first power loss parameters, a plurality of first operation and maintenance scores are calculated, including: Obtain baseline resource scheduling time and baseline power loss parameters; Calculate the ratio of the baseline resource scheduling time to the multiple first resource scheduling times to obtain multiple first resource scheduling coefficients; The ratios of the baseline power loss parameter and the multiple first power loss parameters are calculated respectively to obtain multiple first power coefficients; Multiple first operation and maintenance scores are calculated based on multiple first resource scheduling coefficients and multiple first power coefficients.

[0122] In one embodiment, the operation and maintenance window optimization module 14 is used for: During the maintenance period, multiple first maintenance windows are adjusted by window step size to obtain multiple second maintenance windows, and multiple second maintenance scores are obtained through processing. Based on the historical operation feature sequence, resource similarity weights and power similarity weights are configured, and the window similarity of multiple first operation and maintenance windows and multiple second operation and maintenance windows is calculated. Filter the first and second maintenance windows with the highest similarity, delete the maintenance windows with low maintenance scores, and obtain the remaining maintenance windows. Continue iterative optimization until convergence, obtaining the optimal operation and maintenance window with the highest operation and maintenance score, which serves as the operation and maintenance decision result.

[0123] Specifically, based on the historical operational feature sequence, resource similarity weights and power similarity weights are configured to calculate the window similarity of multiple first and second operation and maintenance windows, including: Calculate the mean of the historical operating feature sequence to obtain the historical average operating feature, and calculate the ratio of this ratio to the maximum operating feature as the resource similarity weight; Based on the resource similarity weights, the power similarity weights are calculated. Calculate the similarity of resource scheduling time and power loss parameters for each first maintenance window and each second maintenance window to obtain multiple resource similarities and multiple power similarities. Perform weighted calculation based on the resource similarity weight and power similarity weight to obtain multiple window similarities.

[0124] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects: First, the feature sequence prediction module 11 provides a high-quality data foundation for prediction through systematic data acquisition and processing. A specially trained prediction agent is employed to improve the accuracy and reliability of feature prediction, reduce manual intervention, and increase work efficiency.

[0125] Secondly, by utilizing the demand resource prediction module 12 and invoking the operation and maintenance resource prediction intelligent agent, a mapping from operational characteristics to resource demand is achieved, providing data support for operation and maintenance resource planning and improving the accuracy of resource forecasting. This overcomes the limitations of traditional resource estimation relying on manual experience, enabling dynamic matching of resource demand with the operating status of the power network.

[0126] Furthermore, utilizing the first operation and maintenance score calculation module 13, global exploration was ensured through random selection. Weighted calculations were then used to comprehensively assess the operation and maintenance effectiveness of the window, considering both resource efficiency and power loss. Simultaneously, a dual-dimensional evaluation of resource scheduling time and power loss parameters was conducted to comprehensively consider both operation and maintenance economics and grid stability, making the initial screening results more objective. A standardized score calculation system was constructed, balancing generality and individual needs, laying a unified evaluation foundation for subsequent optimization steps.

[0127] Finally, using the operation and maintenance window optimization module 14, the search range for the optimal window is gradually narrowed through window step size adjustment and iterative optimization, avoiding the limitations of traditional fixed window selection and achieving dynamic optimization and accuracy improvement of the operation and maintenance window. Secondly, by configuring resource similarity weights and power similarity weights, the adaptability and rationality of window optimization are improved. Similarity calculation can adapt to different load fluctuation scenarios, enhancing the adaptability and rationality of window optimization. Finally, by filtering the window with the highest similarity and deleting low-score windows, the number of redundant candidate windows is reduced, the amount of iterative computation is decreased, and the final output optimal operation and maintenance window ensures high resource scheduling efficiency and improves the execution effect of operation and maintenance decisions.

[0128] In summary, this application addresses the issues of low accuracy and poor adaptability of traditional manual forecasting by replacing human experience-based judgment with an intelligent agent for predicting operational resources. Secondly, it establishes a multi-dimensional quantitative evaluation and iterative optimization mechanism to ensure that the optimal operational window achieves global optimization in terms of resource efficiency and power loss, avoiding the operational delays and excessive losses caused by traditional fixed windows or manual window selection. Simultaneously, through accurate resource demand forecasting and window optimization, it reduces total operational costs and improves the reliability and economy of power grid operation. Ultimately, it achieves dynamic adaptation and intelligent risk warning, enhancing the stability of power grid operation and maintenance.

[0129] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0130] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An artificial intelligence agent-based power network operation and maintenance decision method, characterized in that, The method comprises: After the operation time of the power network reaches the operation and maintenance period, a historical operation characteristic sequence in a preset time range in the past is acquired, an operation characteristic prediction in a maintenance time period is performed based on an artificial intelligent agent, and a predicted operation characteristic sequence is obtained; Based on the artificial intelligent agent, a power operation and maintenance resource demand prediction is performed according to the historical operation characteristic sequence, and a predicted demand resource is obtained; A plurality of first operation and maintenance windows are randomly selected in the maintenance time period, a plurality of first resource scheduling times corresponding to the predicted demand resource are acquired by indexing, and a plurality of first power loss parameters are processed by indexing based on the predicted operation characteristic sequence, and a plurality of first operation and maintenance scores are calculated and obtained; According to the plurality of first operation and maintenance scores, selection optimization of the operation and maintenance window is continuously performed, and an optimal operation and maintenance window is obtained as an operation and maintenance decision result, wherein, in the optimization process, resource similarity weights and power similarity weights are configured according to the historical operation characteristic sequence, and similarity calculation and merging of the operation and maintenance window are performed.

2. The method of claim 1, wherein the method further comprises: After the operation time of the power network reaches the operation and maintenance period, a historical operation characteristic sequence in a preset time range in the past is acquired, an operation characteristic prediction in a maintenance time period is performed based on an artificial intelligent agent, and a predicted operation characteristic sequence is obtained, comprising: After the operation time of the power network reaches the operation and maintenance period, a historical operation characteristic sequence in a preset time range in the past is acquired, wherein each operation characteristic comprises power load; An operation characteristic prediction intelligent agent based on artificial intelligence is called; The historical operation characteristic sequence is input into the operation characteristic prediction intelligent agent, and a predicted operation characteristic sequence is output.

3. The method of claim 2, wherein the method further comprises: The building step of the operation characteristic prediction intelligent agent comprises: Based on the operation monitoring data in the past time of the power network, a sample historical operation characteristic sequence set is collected, and operation characteristics in a maintenance time period after each sample historical operation characteristic sequence is collected, a sample predicted operation characteristic sequence set is obtained; Based on artificial intelligence, an input layer, a hidden layer, a full connection layer and an output layer in the operation characteristic prediction intelligent agent are constructed; The sample historical operation characteristic sequence set and the sample predicted operation characteristic sequence set are used to supervise the training of the operation characteristic prediction intelligent agent until test convergence, the building is completed, and the operation characteristic prediction intelligent agent is configured in the power network operation and maintenance center.

4. The method of claim 1, wherein the method further comprises: Based on the artificial intelligent agent, a power operation and maintenance resource demand prediction is performed according to the historical operation characteristic sequence, and a predicted demand resource is obtained, comprising: An operation and maintenance resource prediction intelligent agent based on artificial intelligence is called, wherein the operation and maintenance resource prediction intelligent agent is trained by using a sample historical operation characteristic sequence set and a sample demand resource set, and each sample demand resource comprises a demand resource type; The historical operation characteristic sequence is input into the operation and maintenance resource prediction intelligent agent, and a predicted demand resource is output.

5. The method of claim 1, wherein the method further comprises: A plurality of first operation and maintenance windows are randomly selected in the maintenance time period, a plurality of first resource scheduling times corresponding to the predicted demand resource are acquired by indexing, and a plurality of first power loss parameters are processed by indexing based on the predicted operation characteristic sequence, and a plurality of first operation and maintenance scores are calculated and obtained, comprising: A plurality of first operation and maintenance windows are randomly selected in the maintenance time period, wherein the plurality of first operation and maintenance windows are different; The first plurality of operation and maintenance window inputs are indexed into a resource scheduling table to obtain a first plurality of resource scheduling times by processing the predicted demand resources, wherein the resource scheduling table includes a mapping relationship between a sample operation and maintenance window set and a plurality of sample resource scheduling time sets of various resources, a sum of resource scheduling times of the indexed predicted demand resources is calculated as the first plurality of resource scheduling times; According to the plurality of first operation and maintenance windows, a plurality of first operation characteristics are obtained by indexing in the predicted operation characteristic sequence as a plurality of first power loss parameters; According to the plurality of first resource scheduling times and the plurality of first power loss parameters, a plurality of first operation and maintenance scores are calculated and obtained.

6. The method of claim 5, wherein the method further comprises: According to the plurality of first resource scheduling times and the plurality of first power loss parameters, a plurality of first operation and maintenance scores are calculated and obtained, including: obtaining a reference resource scheduling time and a reference power loss parameter; respectively calculating the ratio of the reference resource scheduling time and the plurality of first resource scheduling times to obtain a plurality of first resource scheduling coefficients; respectively calculating the ratio of the reference power loss parameter and the plurality of first power loss parameters to obtain a plurality of first power coefficients; According to the plurality of first resource scheduling coefficients and the plurality of first power coefficients, a plurality of first operation and maintenance scores are calculated and obtained.

7. The method of claim 1, wherein the method further comprises: According to the plurality of first operation and maintenance scores, the selection and optimization of the operation and maintenance window is continued to obtain an optimal operation and maintenance window as an operation and maintenance decision result, including: In the maintenance time period, a plurality of second operation and maintenance windows are obtained by adjusting the plurality of first operation and maintenance windows with a window step to obtain a plurality of second operation and maintenance scores; According to the historical operation characteristic sequence, resource similarity weights and power similarity weights are configured to calculate a plurality of window similarities of the plurality of first operation and maintenance windows and the plurality of second operation and maintenance windows; The first operation and maintenance window and the second operation and maintenance window with the largest similarity are screened, and the operation and maintenance window with a small operation and maintenance score is deleted to obtain a plurality of retained operation and maintenance windows; Iterative optimization is continued until convergence to obtain an optimal operation and maintenance window with the largest operation and maintenance score as an operation and maintenance decision result.

8. The method of claim 7, wherein the method further comprises: According to the historical operation characteristic sequence, resource similarity weights and power similarity weights are configured to calculate a plurality of window similarities of the plurality of first operation and maintenance windows and the plurality of second operation and maintenance windows, including: The mean of the historical operation characteristic sequence is calculated to obtain a historical average operation characteristic, and the ratio to the maximum operation characteristic is calculated as a resource similarity weight; According to the resource similarity weight, a power similarity weight is calculated; The similarity of the resource scheduling time and the power loss parameter of each first operation and maintenance window and each second operation and maintenance window is calculated to obtain a plurality of resource similarities and a plurality of power similarities, and the plurality of window similarities are obtained by weighted calculation based on the resource similarity weight and the power similarity weight.

9. An artificial agent-based power network operation and maintenance decision system, characterized in that, The system for implementing the artificial intelligence-based power network operation and maintenance decision method of any one of claims 1-8, the system comprising: An operation characteristic sequence prediction module is configured to obtain a historical operation characteristic sequence within a past preset time range after the power network operation time reaches an operation and maintenance cycle, and to predict an operation characteristic within a maintenance time period based on an artificial intelligence agent to obtain a predicted operation characteristic sequence; An operation and maintenance window selection and optimization module is configured to select a plurality of operation and maintenance windows within the maintenance time period based on the predicted operation characteristic sequence, and to calculate a plurality of operation and maintenance scores of the plurality of operation and maintenance windows based on the historical operation characteristic sequence and the predicted operation characteristic sequence. The demand resource prediction module is configured to predict power operation and maintenance resource demand based on an artificial agent according to the historical operation characteristic sequence, and obtain predicted demand resources. The first operation and maintenance score calculation module is configured to randomly select a plurality of first operation and maintenance windows in a maintenance time period, index a plurality of first resource scheduling times corresponding to the predicted demand resources, and index process a plurality of first power loss parameters based on the predicted operation characteristic sequence, and calculate a plurality of first operation and maintenance scores. The operation and maintenance window optimization module is configured to continue to optimize the selection of the operation and maintenance window according to the plurality of first operation and maintenance scores, and obtain an optimal operation and maintenance window as an operation and maintenance decision result. In the optimization process, resource similarity weights and power similarity weights are configured according to the historical operation characteristic sequence, and similarity calculation and merging of the operation and maintenance window are performed.

Images