Power grid data exception decision-making system based on deep learning

Abstract

The invention discloses a power grid data exception decision-making system based on deep learning, and aims to realize safe and stable control of a power system. The core process of the system comprises: data acquisition and evaluation: acquiring real-time data of a power grid, and constructing a state vector; the state quantitative evaluation unit calculates a steady state deviation entropy; according to the index, high-dimensional data is condensed into a single risk scalar, the degree that the system deviates from an ideal steady state is dynamically and globally evaluated, and risk grading is carried out according to an entropy threshold value; decision prediction: generating a potential decision action set by a decision effect prediction unit, and deducing and predicting a future prediction deviation entropy corresponding to each action by using a power grid state evolution model; and closed-loop control: a closed-loop decision control unit combines the two deviation entropies to construct a reward function, determines an optimal decision action by using a deep reinforcement learning model, finally restrains the intensity according to the current risk level, and outputs a hierarchical control instruction. According to the system, dynamic evaluation and intelligent optimal decision-making of the power grid operation risk are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power system safety and stability control and artificial intelligence technology application, in particular to a power grid data anomaly decision system based on deep learning. BACKGROUND

[0002] With the increasing complexity of modern power grid structure, the continuous expansion of new energy access scale, and the high-dimensional, variable and strong coupling characteristics of power grid operation state, the risk of safe and stable operation of power grid has significantly increased, and the limitations of traditional power grid decision and control methods are prominent.

[0003] Shortcomings in risk assessment: traditional monitoring systems usually rely on static, isolated single-parameter thresholds for alarm judgment; this method is difficult to condense the comprehensive deviation of numerous operating parameters in the power grid into a unified, dynamic risk indicator; when multiple parameters simultaneously deviate slightly but consistently, the traditional method often fails to identify the overall system risk, resulting in a lack of foresight and global perspective in risk assessment;

[0004] Lag and suboptimality in decision response: traditional control strategies rely on artificially set fixed rules or can only make short-sighted decisions based on the current state; this decision-making approach cannot predict the cascading effects or secondary risks that control actions may cause in the future, especially when dealing with complex abnormal states, inappropriate intervention may exacerbate system instability;

[0005] Rigid control intensity: the control instructions of decision output lack adaptive adjustment mechanisms matching the real-time risk level, making it difficult to switch between flexible fine-tuning and strong intervention;

[0006] Therefore, how to achieve dynamic and accurate quantitative assessment of power grid operation state, and overcome the short-sightedness of traditional decision-making through forward-looking prediction and adaptive control to maintain power grid dynamic balance quickly and accurately at minimal cost, has become a key technical problem that needs to be solved to ensure the safe and stable operation of modern power systems. SUMMARY

[0007] To solve the above technical problems, the present application provides a power grid data anomaly decision system based on deep learning, specifically, the technical scheme of the present application includes:

[0008] A data acquisition unit for acquiring real-time data of key nodes of the power grid to construct an original state vector;

[0009] A state quantitative evaluation unit for normalizing the original state vector to generate a normalized state vector, and calculating a steady-state deviation entropy based on the normalized state vector and a preset importance weight, and then classifying the system risk according to a preset entropy threshold to obtain the current risk level;

[0010] a decision effect prediction unit configured to generate a set of potential decision actions based on controllable resources of the power grid, and to deduce each decision action in the set of potential decision actions by using a preset power grid state evolution prediction model to obtain a corresponding future prediction deviation entropy of each decision action;

[0011] a closed-loop decision control unit configured to construct a decision reward function by combining the steady-state deviation entropy and the future prediction deviation entropy, to determine an optimal decision action by using a deep reinforcement learning model, and finally to constrain the intensity of the optimal decision action according to the current risk level to output a hierarchical control instruction.

[0012] Preferably, the process of normalization processing performed by the state quantitative evaluation unit comprises: for each parameter in the original state vector, calculating the absolute value of the difference between the real-time measurement value and the ideal steady-state center value, and standardizing the absolute value by using the safety operation boundary value set for the parameter to obtain a dimensionless relative value representing the degree of deviation.

[0013] Preferably, the state quantitative evaluation unit further combines the dimensionless relative values obtained for each parameter to construct a normalized state vector.

[0014] Preferably, the process of calculating the steady-state deviation entropy by the state quantitative evaluation unit comprises: multiplying each component of the normalized state vector by a corresponding importance weight, and further multiplying the component by a logarithmic value, finally summing and taking a negative value to obtain the steady-state deviation entropy.

[0015] Preferably, the process of classifying the system risk by the state quantitative evaluation unit comprises: comparing and analyzing the calculated steady-state deviation entropy with a preset safe steady-state threshold and a pre-warning unstable-state threshold; when the steady-state deviation entropy is not greater than the safe steady-state threshold, the current risk level is determined as a safe steady state; when the steady-state deviation entropy is greater than the safe steady-state threshold and not greater than the pre-warning unstable-state threshold, the current risk level is determined as a pre-warning unstable state; when the steady-state deviation entropy is greater than the pre-warning unstable-state threshold, the current risk level is determined as an emergency unstable state.

[0016] Preferably, the process of deducing each decision action by the decision effect prediction unit comprises: taking the original state vector and the decision action as inputs, and outputting a future original state vector by the power grid state evolution prediction model; performing normalization processing on the future original state vector to obtain a future normalized state vector; and finally calculating the future prediction deviation entropy based on the future normalized state vector and the importance weight.

[0017] Preferably, the process of constructing the decision reward function by the closed-loop decision control unit comprises: weighting and summing the steady-state deviation entropy and the future prediction deviation entropy by using a preset balance factor, and setting the negative value of the weighted sum result as the decision reward function.

[0018] Preferably, the balance factor is dynamically adjusted according to the current risk level; wherein when the current risk level is an emergency instability state, the value of the balance factor is increased; and when the current risk level is a pre-warning instability state, the value of the balance factor is decreased.

[0019] Preferably, the process of outputting the hierarchical control instruction by the closed-loop decision control unit comprises: when the current risk level is a pre-warning instability state, imposing a first-level constraint on the adjustment range of the optimal decision action to output a flexible fine-tuning instruction; and when the current risk level is an emergency instability state, imposing a second-level constraint on the adjustment range of the optimal decision action to output a strong intervention instruction.

[0020] Compared with the prior art, the present application has the following beneficial effects:

[0021] 1. The present system condenses the high-dimensional and multi-parameter state data in the power grid into a single and unified risk scalar by introducing a steady-state deviation entropy index; the index can objectively reflect the degree of deviation of the system from the ideal steady state, overcoming the shortcoming of traditional methods relying on static and isolated threshold values, and realizing dynamic and global evaluation of the operation risk of the power grid;

[0022] 2. The present system uses a prediction model based on a graph neural network, which can deeply predict the consequences that each potential control action may produce in the future period of time and quantify the consequences as future prediction deviation entropy; the decision model takes minimizing the current and future weighted risk as the goal, ensuring the forward-looking nature of the control strategy and effectively avoiding secondary risks caused by improper intervention;

[0023] 3. The present system can flexibly switch the control strategy focus between minimizing the current risk and minimizing the future risk by designing an adaptive decision reward function and dynamically adjusting the balance factor according to the current risk level; this enables the system to make situation-aware optimal decisions when facing different risk situations;

[0024] 4. The control instruction finally output by the present system will impose different intensity constraints according to the current risk level; a flexible fine-tuning instruction is output in the pre-warning state to avoid secondary impact on the power grid; a strong intervention instruction is output in the emergency state to ensure that the system has the ability to decisively suppress the instability trend; this mechanism guarantees the proportionality and effectiveness of the control behavior. BRIEF DESCRIPTION OF DRAWINGS

[0025] The present application will be further explained in conjunction with the drawings and embodiments:

[0026] Figure 1 is a structural diagram of the system of the present application. DETAILED DESCRIPTION

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0028] Example 1:

[0029] Please see Figure 1 A deep learning-based power grid data anomaly decision-making system includes:

[0030] The data acquisition unit is used to collect real-time data from key nodes of the power grid in order to construct the original state vector.

[0031] The state quantification assessment unit is used to normalize the original state vector to generate a normalized state vector, and calculate the steady-state deviation entropy based on the normalized state vector and the preset importance weight. Then, the system risk is classified according to the preset entropy threshold to obtain the current risk level.

[0032] The decision effect prediction unit is used to generate a set of potential decision actions based on the controllable resources of the power grid, and to use a preset power grid state evolution prediction model to deduce each decision action in the set of potential decision actions, so as to obtain the future prediction deviation entropy corresponding to each decision action.

[0033] The closed-loop decision control unit is used to construct a decision reward function by combining steady-state deviation entropy and future prediction deviation entropy, and to determine the optimal decision action using a deep reinforcement learning model. Finally, the optimal decision action is subjected to strength constraints based on the current risk level in order to output hierarchical control commands.

[0034] This invention provides a deep learning-based power grid data anomaly decision-making system. The system aims to achieve rapid, accurate, and forward-looking decision responses to power grid anomalies through precise quantitative assessment of power grid operating status, forward-looking prediction of decision consequences, and adaptive closed-loop control, thereby maintaining the dynamic balance of the power grid at minimal cost. The system includes a data acquisition unit, a state quantitative assessment unit, a decision effect prediction unit, and a closed-loop decision control unit.

[0035] The data acquisition unit aims to provide real-time and accurate power grid status observation data for the entire decision-making system. In this embodiment, the data acquisition unit is implemented through a Wide Area Measurement System (WAMS) deployed in the power grid and its built-in Phasor Measurement Unit (PMU). This unit continuously acquires multi-dimensional real-time data of preset key transmission nodes and generator nodes in the power grid, including but not limited to node voltage amplitudes. Phase angle System frequency And the active current of key branches These collected raw measurements were combined to construct a comprehensive representation of the current moment. raw state vector of power grid operating state ; raw state vector is a vector containing real-time measurement values of N key parameters, which serves as the most original data basis for subsequent state assessment and decision-making, and its source is the real-time collection of the data acquisition unit;

[0036] state quantification assessment unit, whose core purpose is to convert the high-dimensional, multi-dimensional original state vector into a quantitative index that can intuitively reflect the overall stability of the system; in this embodiment, the unit normalizes the original state vector output by the data acquisition unit to generate a dimensionless normalized state vector ; normalized state vector Each component in it represents the relative degree of deviation of the corresponding parameter from its ideal steady-state center, and its role is to eliminate the interference of different physical dimensions on subsequent model calculations; then, the unit calculates the steady-state deviation entropy at the current time based on the normalized state vector and a set of preset importance weights ; importance weight is a coefficient value set for each state parameter according to the power grid operating procedures and offline simulation analysis, which reflects its influence on the overall stability of the power grid, and its role is to distinguish the risk level of different parameter instability during comprehensive assessment, and its source is prior expert knowledge and data analysis; steady-state deviation entropy is a single scalar index based on information entropy theory, which is used to quantify the chaotic degree of the system deviating from the ideal steady state, and its role is to simplify the complex multivariate monitoring problem into the evaluation of a single comprehensive risk index; the state quantification assessment unit analyzes the calculated steady-state deviation entropy according to the preset entropy threshold, i.e. the safe steady state threshold and the early warning instability state threshold , classifies the system risk, and thus obtains the current risk level; the current risk level is a discrete division of the dangerous degree of the current state of the system, such as safe steady state, early warning instability state or emergency instability state, and its role is to provide a direct basis for the intensity selection of subsequent decision-making control;

[0037] decision effect prediction unit, whose purpose is to predict the long-term cascading effect that each potential control and regulation action may cause, and to avoid secondary risks caused by short-sighted decision-making; in this embodiment, the unit generates a potential decision action set containing multiple discrete adjustment instructions based on the real-time state and adjustable range of the available regulation resources of the power grid is a set of control operations that the system can currently perform; this unit adopts a preset power grid state evolution prediction model to the set of potential decision actions Each decision action is forward-decomposed; the power grid state evolution prediction model In this embodiment, it is a deep learning model based on a graph neural network (GNN) and trained through massive historical and simulation data, which can receive the current original state vector and the decision action as input, and output the predicted original state vector at a certain time in the future ; To enable those skilled in the art to implement, the GNN model here is specifically constructed as follows: the buses in the power grid are taken as the nodes of the graph, and the transmission lines are taken as the edges to construct the power grid topology graph; the GNN network adopts an architecture containing three layers of GraphSAGE convolutional layers and two layers of fully connected layers, and uses ReLU as the activation function; In the training stage, the mean square error is used as the loss function, and the Adam optimizer is used to iteratively optimize the model parameters; The prediction model inherently contains the power flow physical law and dynamic evolution characteristics of the power grid; It should be noted that the prediction model in this embodiment is a deterministic model designed to efficiently capture the main evolution trend triggered by the decision action; In other embodiments, to more accurately depict the influence of random disturbances in the power grid, the model can also be replaced by a stochastic prediction model, whose output is the probability distribution of the future state, and the subsequent decision effect evaluation is based on the expected value of the distribution;

[0038] For each of the future state vectors decomposed, the unit will call the same processing logic as the state quantization evaluation unit to convert it into the corresponding future predicted deviation entropy ; The future predicted deviation entropy is the steady-state deviation entropy value that the system is expected to reach at time after executing a specific decision action, and its role is to provide a quantitative, directly comparable risk scalar for the long-term effect of the decision action;

[0039] The closed-loop decision control unit aims to autonomously select and execute the optimal control strategy based on the comprehensive consideration of the current risk and future risk; In this embodiment, the unit adopts a deep reinforcement learning (DRL) framework; Specifically, this embodiment uses a deep Q-network algorithm; Its Q-network is a multilayer perceptron that receives the current state vector as input and outputs a vector, with each element of the vector corresponding to a potential decision action set Q value of each action; it combines the current steady-state deviation entropy calculated by the state quantification evaluation unit with the future prediction deviation entropy output by the decision effect prediction unit , to build a decision reward function ; the decision reward function is used to evaluate the pros and cons of executing any potential decision action in the current state, and is designed to guide the deep reinforcement learning model to learn a control strategy that can minimize both current and future risks; the pre-trained deep reinforcement learning model determines an optimal decision action from the potential decision action set according to the reward function, aiming to maximize the long-term cumulative reward ; the optimal decision action is the control instruction calculated by the model in the current state, which can optimally balance the current and future system stability; to ensure the accuracy and safety of control, this unit will constrain the execution intensity of the optimal decision action according to the current risk level evaluated by the state quantification evaluation unit, to output a hierarchical control instruction; the hierarchical control instruction is the final execution instruction generated by differentiating the intensity of the optimal decision action according to the risk level, for example, flexible fine-tuning in low risk and strong intervention in high risk;

[0040] The above-mentioned units work together to build a complete technical closed loop from data perception, state quantification, risk prediction to closed-loop decision making; compared with traditional power grid decision systems that rely on static thresholds and artificial rules, the system realizes dynamic and accurate quantification of the overall risk of the power grid; by introducing deep prediction of decision consequences, the short-sightedness of traditional decision methods is overcome, and secondary disasters caused by improper intervention are effectively avoided; finally, through deep reinforcement learning and hierarchical control, the system realizes fast, adaptive and risk level matched accurate response to abnormal states; this significantly improves the ability of the power grid to handle complex abnormal events, while ensuring the safe and stable operation of the power grid and minimizing control costs.

[0041] Embodiment 2:

[0042] The process of normalization processing by the state quantification evaluation unit includes: for each parameter in the original state vector, calculating the absolute value of the difference between the real-time measurement value and the ideal steady-state center value, and using the safety operation boundary value set for the parameter to standardize, to obtain a dimensionless relative value representing the degree of deviation.

[0043] The state quantification evaluation unit also combines the dimensionless relative values obtained for each parameter to construct a normalized state vector.

[0044] This embodiment is a detailed explanation of the normalization process performed by the state quantization evaluation unit in Embodiment 1; this process aims to normalize the original state vectors from the data acquisition unit, which have different physical units and numerical ranges. The various parameters are converted into a unified, dimensionless relative deviation index, providing a standardized input for the subsequent calculation of steady-state deviation entropy;

[0045] When the state quantization evaluation unit performs normalization processing, it considers the original state vector. The parameters in Obtain its real-time measurement value Compared with its preset ideal steady-state center value Ideal steady-state central value This is the theoretical optimal value of the parameter when the power grid is operating at its most stable and efficient level, derived from power grid operation standards or design specifications; the absolute value of the difference between these two values ​​is calculated. ;Utilize the safe operating boundary value set for this parameter Standardization processing; safe operation boundary values It is the operating limit value that this parameter can be set according to industry standards such as the Guidelines for the Safety and Stability of Power Systems;

[0046] The specific standardization process is achieved through formulas. Achieve; where, For parameters The normalized state value is a non-negative dimensionless relative value, calculated by this formula; Indicates parameters The absolute value of the difference between the real-time measured value and its ideal steady-state center value is collected by the data acquisition unit. and preset Calculated; This indicates the maximum permissible deviation from the ideal center to the safety boundary, determined by a preset value. and The calculation yielded the following result; the technical motivation behind this formula lies in the fact that it not only quantifies the absolute magnitude of the deviation but also considers it within the acceptable operating range of the parameter itself, making the deviations of parameters from different physical dimensions comparable. To ensure the robustness of the calculation, the system verifies the safe operating boundary values ​​of all parameters during initialization. All are greater than their ideal steady-state center value To avoid the denominator being zero; if the denominator is zero, a minimum value is added to the denominator. To ensure computational stability;

[0047] The state quantization evaluation unit is the original state vector. Each parameter in the calculation yields its corresponding dimensionless relative value. After that, it will obtain these. , They are combined to form a normalized state vector. ;

[0048] Through the above-described specific normalization method, this invention not only solves the technical problem that different physical dimensions cannot be directly compared, but more importantly, it gives the normalization result a clear physical meaning, namely the relative position of the current state point within the range from the ideal center to the safety boundary. This standardized processing method provides a more accurate and reliable data foundation for subsequent system-wide risk assessment based on information entropy theory, and its assessment results can more realistically reflect the actual safety margin of the power grid.

[0049] Example 3:

[0050] The process of calculating the steady-state deviation entropy by the state quantification evaluation unit includes: multiplying each component of the normalized state vector by its corresponding importance weight, and further multiplying it by the logarithm of the component, and finally summing the results and taking the negative value to obtain the steady-state deviation entropy.

[0051] The process by which the state quantification assessment unit classifies system risks includes: comparing and analyzing the calculated steady-state deviation entropy with preset safe and stable state thresholds and early warning instability thresholds; when the steady-state deviation entropy is not greater than the safe and stable state threshold, the current risk level is determined to be a safe and stable state; when the steady-state deviation entropy is greater than the safe and stable state threshold but not greater than the early warning instability threshold, the current risk level is determined to be an early warning instability state; when the steady-state deviation entropy is greater than the early warning instability threshold, the current risk level is determined to be an emergency instability state.

[0052] This embodiment describes the state quantization evaluation unit after obtaining the normalized state vector. The following section explains the specific implementation method for calculating steady-state deviation entropy and classifying system risks.

[0053] The process of calculating the steady-state deviation entropy by the state quantization evaluation unit is to normalize the multidimensional state vector. This can be consolidated into a single indicator that can characterize the overall degree of disorder or chaos in the system. This process is achieved by introducing information entropy theory. To apply entropy theory, the normalized state vector needs to be... The components are converted into probability form, and the calculation method is as follows: ,in Represents the first The proportion of each parameter's deviation in the total deviation; its calculation is based on the steady-state deviation entropy calculation model. Complete; in the formula, For the system at time The steady-state deviation entropy of the system, which is a dimensionless scalar, the greater its value, the higher the degree of deviation from the ideal steady state of the system; is the normalized state vector ; is the th component of the normalized state vector , and is the importance weight of the th parameter. To ensure its objectivity and reproducibility, the importance weight of each parameter is determined as follows: a large number of historical data samples containing normal and unstable states of the system are collected; a logistic regression classifier is trained using these samples to distinguish between the two states; the regression coefficients of each input parameter obtained after training the classifier are normalized, and the result is used as the importance weight of the corresponding parameter

[0054] ; The technical motivation of this formula is to solve the problem of multi-index comprehensive evaluation; to ensure mathematical completeness, it is agreed in this embodiment that when a certain deviation proportion , the corresponding term takes its limit value 0 in mathematics; the input of this formula is the normalized state vector calculated in the previous step and the preset importance weight , and the output is directly input into the subsequent risk classification module as the core state indicator;

[0055] The process of the state quantitative evaluation unit for classifying system risks is to map the continuously changing steady-state deviation entropy into discrete risk levels with clear operational implications; this process is achieved by comparing and analyzing the calculated with the preset safe stable state threshold and the early warning unstable state threshold ; the setting of the thresholds has sufficient technical basis, and its source is the statistical analysis of a large number of historical operation data of a specific power grid to determine the thresholds; is usually set as the 95th percentile of the entropy value distribution in the normal operating state, and is determined according to the entropy value data that has caused an alarm or a small range of faults in the history; when , the current risk level is determined as a safe stable state; when , the current risk level is determined as an early warning unstable state; when , the current risk level is determined as an emergency unstable state;

[0056] By introducing a steady-state deviation entropy calculation method based on information entropy, this invention can more profoundly reveal the overall operational risks of the system. Combined with a tiered threshold set based on historical data statistical analysis, the risk level classification is not only qualitative but also quantitative, objective, and statistically significant. This design enables the system to accurately distinguish between normal fluctuations, potential risks, and urgent threats, providing a precise and reliable basis for subsequent decision-making interventions of appropriate intensity, avoiding frequent false alarms or missed alarms caused by improper threshold settings in traditional systems. Furthermore, to further capture specific risk patterns, this system can also calculate multiple steady-state deviation entropies based on different parameter subsets in parallel, such as voltage stability entropy and frequency stability entropy, thereby providing more multi-dimensional risk insights for decision-making, supplementing a single comprehensive risk indicator.

[0057] Example 4:

[0058] The process of the decision effect prediction unit extrapolating each decision action includes: taking the original state vector and the decision action as input, outputting the future original state vector through the power grid state evolution prediction model; normalizing the future original state vector to obtain the future normalized state vector; and finally calculating the future prediction deviation entropy based on the future normalized state vector and importance weights.

[0059] This embodiment is a detailed explanation of the deduction process of each decision action by the decision effect prediction unit in Embodiment 1. The core of this process is to use the power grid state evolution prediction model to prospectively evaluate the possible consequences of any control action in the future and quantify the consequences as a future prediction deviation entropy comparable to the current risk index.

[0060] Decision effect prediction unit for potential decision action set Every decision action The deduction process involves a series of closely linked steps; the original state vector at the current moment... and decision actions to be evaluated As input, the power grid state evolution prediction model The forward computation can be represented as follows: Output the future The future original state vector after one time step The predicted future original state vector Further processing is required; the system will invoke the normalization processing method defined in Examples 2 and 3 to process the data. Each component in the vector is standardized to obtain the future normalized state vector. This step ensures that the evaluation of future states uses the same benchmark as the evaluation of current states; the future normalized state vector is then used. and a preset importance weight Substitute the defined steady-state deviation entropy calculation formula, and calculate the decision action corresponding future prediction deviation entropy ;

[0061] The core innovation of this embodiment is its forward-looking nature; the present application builds a deep learning prediction model that can simulate the dynamics of the power grid, and combines a unified entropy evaluation system, so that the system can estimate the consequences that may be caused by any operation before it is executed; this forward-looking prediction and quantitative evaluation capability of the cascading effect of decision-making can effectively help the system to filter out decisions that can solve problems without causing secondary disasters, thereby avoiding short-sighted decisions and significantly improving the long-term stability of the power grid.

[0062] Example 5:

[0063] The process of the closed-loop decision control unit for constructing the decision reward function includes: weighting and summing the steady-state deviation entropy and the future prediction deviation entropy by using a preset balance factor, and setting the negative value of the weighted sum result as the decision reward function.

[0064] The balance factor is dynamically adjusted according to the current risk level; when the current risk level is an emergency instability state, the value of the balance factor is increased; when the current risk level is a pre-warning instability state, the value of the balance factor is decreased.

[0065] This embodiment is a specific implementation of how the closed-loop decision control unit in Example 1 constructs the decision reward function and uses it for decision optimization;

[0066] The process of the closed-loop decision control unit for constructing the decision reward function is to provide a clear optimization target for the deep reinforcement learning agent, guiding it to learn an optimal control strategy that takes into account both the current urgency and the future stability; the immediate reward function is used to evaluate the reward or punishment obtained after performing action in state In this embodiment, the function is designed as an adaptive decision reward function that takes into account the current and future ; in the formula, is the immediate reward obtained after performing action in state The optimization goal is to maximize , which is equivalent to minimizing the weighted sum of entropies; is the steady-state deviation entropy before performing action, which is calculated by the state quantitative evaluation unit; is the future prediction deviation entropy predicted after performing action , which is calculated by the decision effect prediction unit; is a tunable parameter between 0 and 1, whose value is dynamically adjusted according to the current risk level;

[0067] To make the decision strategy more adaptive, the value of is dynamically adjusted according to the current risk level; when the current risk level is the emergency instability state, the system is in a highly dangerous state, at this time, the system will increase the balance factor , so that the reward function focuses more on minimizing the current steady-state deviation entropy , guiding the agent to prefer strong intervention measures that can most quickly and effectively alleviate the current crisis; when the current risk level is the early warning instability state, the system still has sufficient adjustment margin, at this time, the system will lower the balance factor , so that the reward function focuses more on minimizing the future predicted deviation entropy , guiding the agent to choose those long-term optimal flexible fine-tuning measures;

[0068] Through the design of this adaptive reward function, the invention provides a smart optimization guide for deep reinforcement learning agents; the dynamically adjusted balance factor makes the decision logic no longer rigid, but can flexibly switch the strategy focus between immediate effect and long-term effect according to the real-time risk status of the power grid; this enables the system to make decisive decisions when facing serious threats, and to think carefully when dealing with potential risks, thereby realizing a situation-aware adaptive optimal control.

[0069] Embodiment 6:

[0070] The process of the closed-loop decision control unit outputting the hierarchical control instruction includes: when the current risk level is the early warning instability state, imposing a first-level constraint on the adjustment amplitude of the optimal decision action to output a flexible fine-tuning instruction; when the current risk level is the emergency instability state, imposing a second-level constraint on the adjustment amplitude of the optimal decision action to output a strong intervention instruction.

[0071] This embodiment is a description of the specific implementation of how the closed-loop decision control unit in embodiment 1 or 5 outputs the final hierarchical control instruction according to the current risk level after determining the optimal decision action .

[0072] After the deep reinforcement learning model determines the theoretically optimal decision action , the closed-loop decision control unit will perform a strength constraint step before outputting the final instruction, and the constraint strength of this step is directly related to the current risk level evaluated by the state quantification evaluation unit;

[0073] When the current risk level is evaluated as the early warning instability state, corresponding to satisfies the optimal decision action is adjusted by the closed-loop decision control unit to a first level of adjustment range; the first level of adjustment range is a small allowed adjustment range or a change rate limit; the decision action after the adjustment will be output as a flexible fine-tuning instruction;

[0074] When the current risk level is assessed as an emergency unstable state, corresponding to the case of the optimal decision action is adjusted by the closed-loop decision control unit to a second level of adjustment range; the second level of adjustment range is much wider than the first level of adjustment range, allowing a larger adjustment of the power grid resources; the decision action after the adjustment will be output as a strong intervention instruction to suppress the spread of the unstable trend as quickly as possible; the hierarchical control instruction will be parsed by the control interface unit of the system, and according to the pre-set instruction mapping rule library, the abstract decision action will be converted into specific control parameters for specific devices, and finally through the power grid remote communication system, it is issued to the target device for execution;

[0075] The output mechanism of such hierarchical control instruction ensures the proportionality and appropriateness of the control behavior; when the risk is low, unnecessary drastic operation is avoided to cause secondary impact on the power grid, ensuring the delicacy of the control; when the risk is extremely high, it ensures that the system has enough authority and ability to execute strong intervention measures, ensuring the effectiveness of the decision; the design makes the whole closed-loop control system pursue the optimal decision while taking into account the safety regulations and stability requirements of the power grid operation, making it more practical for engineering.

[0076] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.​​

Claims

1. A deep learning-based power grid data anomaly decision system, characterized in that, include: The data acquisition unit is used to collect real-time data from key nodes of the power grid in order to construct the original state vector. The state quantification assessment unit is used to normalize the original state vector to generate a normalized state vector, and calculate the steady-state deviation entropy based on the normalized state vector and the preset importance weight. Then, the system risk is classified according to the preset entropy threshold to obtain the current risk level. The decision effect prediction unit is used to generate a set of potential decision actions based on the controllable resources of the power grid, and to use a preset power grid state evolution prediction model to deduce each decision action in the set of potential decision actions, so as to obtain the future prediction deviation entropy corresponding to each decision action. The closed-loop decision control unit is used to construct a decision reward function by combining steady-state deviation entropy and future prediction deviation entropy, and to determine the optimal decision action using a deep reinforcement learning model. Finally, the optimal decision action is subjected to strength constraints based on the current risk level in order to output hierarchical control commands.

2. The deep learning-based power grid data anomaly decision system of claim 1, wherein, The normalization process of the state quantification evaluation unit includes: for each parameter in the original state vector, calculating the absolute value of the difference between the real-time measured value and the ideal steady-state center value, and standardizing it using the safe operation boundary value set for the parameter to obtain a dimensionless relative value representing the degree of deviation.

3. The deep learning-based power grid data anomaly decision system of claim 2, wherein, The state quantification evaluation unit will also combine the dimensionless relative values ​​obtained from each parameter to construct a normalized state vector.

4. The deep learning-based power grid data anomaly decision system of claim 3, wherein, The process of calculating the steady-state deviation entropy by the state quantification evaluation unit includes: multiplying each component of the normalized state vector by its corresponding importance weight, and further multiplying it by the logarithm of the component, and finally summing the results and taking the negative value to obtain the steady-state deviation entropy.

5. The deep learning based power grid data anomaly decision system of claim 4, wherein, The process by which the state quantification assessment unit classifies system risks includes: comparing and analyzing the calculated steady-state deviation entropy with preset safe and stable state thresholds and early warning instability thresholds; when the steady-state deviation entropy is not greater than the safe and stable state threshold, the current risk level is determined as a safe and stable state; when the steady-state deviation entropy is greater than the safe and stable state threshold but not greater than the early warning instability threshold, the current risk level is determined as an early warning instability state; when the steady-state deviation entropy is greater than the early warning instability threshold, the current risk level is determined as an emergency instability state.

6. The deep learning based power grid data anomaly decision system of claim 1, wherein, The process by which the decision effect prediction unit extrapolates each decision action includes: taking the original state vector and the decision action as input, outputting the future original state vector through the power grid state evolution prediction model; normalizing the future original state vector to obtain the future normalized state vector; and finally calculating the future prediction deviation entropy based on the future normalized state vector and the importance weight.

7. The deep learning based power grid data anomaly decision system of claim 1, wherein, The process of constructing the decision reward function in the closed-loop decision control unit includes: using a preset balance factor to perform a weighted summation of the steady-state deviation entropy and the future prediction deviation entropy, and setting the negative value of the weighted summation result as the decision reward function.

8. The deep learning-based power grid data anomaly decision system of claim 7, wherein, The balance factor is dynamically adjusted based on the current risk level; specifically, when the current risk level is in an emergency instability state, the value of the balance factor is increased; when the current risk level is in a warning instability state, the value of the balance factor is decreased.

9. A deep learning-based power grid data anomaly decision-making system according to claim 1 or 5, characterized in that, The process of the closed-loop decision control unit outputting hierarchical control commands includes: when the current risk level is a warning instability state, applying a first-level constraint to the adjustment range of the optimal decision action to output a flexible fine-tuning command; when the current risk level is an emergency instability state, applying a second-level constraint to the adjustment range of the optimal decision action to output a strong intervention command.

Images