Deep learning-based time sequence production simulation multi-dimensional risk analysis method and system

Abstract

The invention discloses a time sequence production simulation multi-dimensional risk analysis method and system based on deep learning, and the method comprises the steps: constructing a time sequence production simulation model, and carrying out the calculation to obtain system operation state data; constructing a multi-dimensional risk index analysis system, and calculating an initial weight of each risk index; introducing a real-time state quantity corresponding to each risk index as a driving factor to correct the initial weight of each risk index to obtain a corrected dynamic weight; carrying out weighted fusion on the risk index values based on the dynamic weight to obtain a comprehensive risk value; forming a feature vector by the product of the drive factor normalized value corresponding to each risk index and the dynamic weight, taking the feature vector as input, taking the corresponding comprehensive risk value as output, training the GRU model, and taking the trained GRU model as a risk analysis model; and processing the operation data of the target system according to the steps to obtain a corresponding feature vector, and inputting the feature vector into the analysis model to obtain a comprehensive risk value of the target system. According to the invention, the sensitivity and accuracy of risk identification are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power system risk assessment based on artificial intelligence algorithm, and particularly relates to a time-series production simulation multi-dimensional risk analysis method and system based on deep learning. BACKGROUND

[0002] At present, new energy development and utilization has made remarkable progress, and establishing a new power system suitable for high proportion of new energy has become a key task. The output of renewable energy such as wind energy and photovoltaic energy has significant volatility and uncertainty, which leads to the dual challenges of "guaranteeing supply" and "utilization" in power balance of the power system. Therefore, it is urgent to build a multi-dimensional risk assessment method suitable for high proportion of new energy power system.

[0003] The large-scale access of new energy such as wind power and photovoltaic power has profoundly changed the inherent characteristics and risk pattern of the system, posing a severe challenge to the traditional risk assessment method. The traditional risk assessment system is mainly based on the stable and controllable characteristics of fossil fuel units, and it is difficult to effectively cope with the inherent randomness, volatility and intermittency of new energy generation. At the same time, the existing operation risk assessment index has no uniform standard. In the current literature, only one or several indicators such as load shedding risk, voltage out-of-limit risk, line active power out-of-limit risk and wind curtailment risk are often considered, and there is a lack of calculation and analysis of adaptive comprehensive indicators that can comprehensively consider various risks.

[0004] Therefore, a multi-dimensional risk analysis method that comprehensively considers various risks is urgently needed. SUMMARY

[0005] In order to fully consider the high proportion of new energy in the system and realize the rapid prediction of the comprehensive risk score of the system, the application establishes a multi-dimensional risk index analysis system based on time-series production simulation related data, and constructs a time-series neural network model for risk analysis. Specifically, it includes a time-series production simulation multi-dimensional risk analysis method and system based on deep learning.

[0006] In order to achieve the purpose of the application, the following technical solutions are specifically adopted.

[0007] In the first aspect, the application discloses a time-series production simulation multi-dimensional risk analysis method based on deep learning, which includes the following steps: S1, constructing a time-series production simulation model of the power system to calculate system operation state data; S2, based on the system operation state data, constructing a multi-dimensional risk index analysis system including multiple risk indicators, and calculating the initial weight of each risk indicator in the multi-dimensional risk index analysis system; S3, introduce real-time state quantities corresponding to each risk indicator as driving factors of each risk indicator, and correct the initial weights of each risk indicator based on each driving factor to obtain modified dynamic weights; based on the dynamic weights, each risk indicator value is weighted and fused to obtain a comprehensive risk value; S4, multiply the normalized value of each risk indicator corresponding to the driving factor and the dynamic weight corresponding to each risk indicator, and group all the multiplication results into a feature vector as input; the comprehensive risk value corresponding to the same moment is taken as output, the GRU network model is trained, and the trained GRU network model is taken as a system multi-dimensional risk analysis model; wherein, when the GRU network model is trained, a penalty weight is introduced in the loss function, and the penalty weight increases with the increase of the actual comprehensive risk value; S5, calculate the product of the normalized value of each risk indicator corresponding to the driving factor and the dynamic weight corresponding to each risk indicator in the real-time operation data of the target system, group all the multiplication results into a feature vector, input the feature vector into the system multi-dimensional risk analysis model, and obtain the comprehensive risk value of the target system.

[0008] Further preferably, The system operation state data includes wind curtailment rate, light curtailment rate, thermal power unit output, energy storage system output power, energy storage system state of charge, demand response load adjustment amount, and transmission line power flow; the plurality of risk indicators include wind curtailment rate, light curtailment rate, thermal power unit output, energy storage system output power, energy storage system state of charge, demand response load adjustment amount, and transmission line power flow; further preferably, In S2, the initial weights of each risk indicator in the multi-dimensional risk indicator analysis system are determined in the following manner: the first initial weight of each risk indicator is calculated based on the analytic hierarchy process; the second initial weight of each risk indicator is calculated based on the entropy method; the first initial weight and the second initial weight are weighted and fused to obtain the initial weight of each risk indicator.

[0009] Further preferably, In S3, the real-time state quantity corresponding to each risk indicator is introduced as the driving factor of each risk indicator, specifically including: The wind power output variation coefficient is taken as the driving factor of the wind curtailment rate; the light power output variation coefficient is taken as the driving factor of the light curtailment rate; the key line power flow utilization rate is taken as the driving factor of the key line power flow overrun degree; the load proximity is taken as the driving factor of the load loss amount proportion; the reactive power margin deficiency is taken as the driving factor of the voltage stability index.

[0010] Further preferably, In S3, the initial weight of each risk indicator is corrected based on each driving factor to obtain the modified dynamic weight, specifically including: Select typical operating conditions of the power system, apply random disturbances of 1% to 3% to each driving factor corresponding to each risk indicator in the typical operating conditions, and calculate the rate of change of the comprehensive risk value corresponding to each driving factor. The normalized value of the rate of change of the comprehensive risk value corresponding to each driving factor is used as the sensitivity coefficient corresponding to each driving factor. Multiply each driving factor by its corresponding sensitivity coefficient, and then add 1 to get the dynamic adjustment factor for each risk indicator corresponding to each driving factor. The dynamic adjustment factor of each risk indicator is multiplied by the initial weight of each risk indicator, and after normalization and smoothing, the adjusted dynamic weight of each risk indicator is obtained. More preferably, The rate of change of the overall risk value corresponding to each driving factor is determined as follows:

[0011] in, For the first i The rate of change of the combined risk value of each driving factor For the first typical working condition i One driving factor The applied random perturbation, The comprehensive risk value under typical working conditions corresponds to The change in quantity.

[0012] More preferably, In S4, the loss function is specifically:

[0013] in, N The total number of training samples, The penalty weight corresponding to the actual comprehensive risk value. To predict the overall risk value, This represents the actual overall risk value.

[0014] Secondly, this invention discloses a deep learning-based time-series production simulation multi-dimensional risk analysis system based on the aforementioned method, including a time-series production simulation and operation data calculation module, a multi-dimensional risk indicator analysis system construction module, a comprehensive risk value calculation module, a system multi-dimensional risk analysis model construction module, and a target system comprehensive risk value assessment module; The time-series production simulation and operation data calculation module constructs a time-series production simulation model of the power system and calculates the system operation status data. A multi-dimensional risk index analysis system construction module constructs a multi-dimensional risk index analysis system including multiple risk indexes based on the system operation state data, and calculates initial weights of each risk index in the multi-dimensional risk index analysis system; A comprehensive risk value calculation module introduces real-time state quantities corresponding to each risk index as driving factors of each risk index, and modifies the initial weights of each risk index based on each driving factor to obtain modified dynamic weights; based on the dynamic weights, each risk index value is weighted and fused to obtain a comprehensive risk value; A system multi-dimensional risk analysis model construction module multiplies the driving factors corresponding to each risk index and the dynamic weights to form a feature vector as input, and takes the corresponding comprehensive risk value at the same time as output, trains a GRU network model, and takes the trained GRU network model as a system multi-dimensional risk analysis model; wherein, when training the GRU network model, a penalty weight is introduced in the loss function, and the penalty weight increases with the increase of the actual comprehensive risk value; A target system comprehensive risk value evaluation module calculates the product of the normalized values of the driving factors corresponding to each risk index in the real-time operation data of the target system and the dynamic weights corresponding to each risk index, and forms a feature vector by multiplying all the results, and inputs the feature vector into the system multi-dimensional risk analysis model to obtain the target system comprehensive risk value.

[0015] In a third aspect, the present application provides a terminal, comprising a processor and a storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to perform the steps of the method of any one of the first aspect of the present application.

[0016] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method of any one of the first aspect of the present application.

[0017] Compared with the prior art, the present application has the following beneficial technical effects: 1) The present application proposes an adaptive dynamic weight updating method based on subjective and objective fusion. On the basis of the static basic weight obtained by the traditional analytic hierarchy process (AHP) and entropy method, the present application firstly introduces the operation state quantities such as new energy output fluctuation, load proximity, key line flow utilization rate and reactive power margin into the weight calculation process, constructs a dynamic correction factor, and obtains the instant weight through normalization and exponential smoothing. The dynamic weight mechanism can adjust the relative importance of each risk index in real time according to the system operation state, so that the risk assessment result is more in line with the true characteristics of the power system under the condition of high proportion of new energy, and the sensitivity and accuracy of risk identification are significantly improved.

[0018] 2) The application uses the GRU model to learn the time sequence variation characteristics, long and short term correlation and meteorological driving law of new energy output and load, so that the system can dynamically capture the risk evolution trend, and no longer rely on traditional static or simplified simulation method. The model can give a comprehensive risk score in advance when the risk trend such as continuous attenuation of wind and light output and rapid rise of load has not yet been significant, realize more timely risk warning, and at the same time can maintain high prediction accuracy in high volatility and high uncertainty operation environment, thereby overcoming the strong lag problem of traditional method, making the risk assessment truly dynamic, forward-looking and real-time.

[0019] The specific embodiment of the advantages of the application can be referred to the embodiment analysis part of the application. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a schematic diagram of the multi-dimensional risk analysis method process of the application based on deep learning time sequence production simulation; Figure 2 is a flow chart of the analytic hierarchy process of embodiment one of the application; Figure 3 is a basic neural network diagram of embodiment one of the application; Figure 4 is a neural network diagram after applying Dropout of embodiment one of the application; Figure 5 is a load power curve of embodiment two of the application; Figure 6 is a thermal power unit output of embodiment two of the application; Figure 7 is a photovoltaic actual consumption rate of embodiment two of the application; Figure 8 is a wind power actual consumption rate of embodiment two of the application; Figure 9 is a risk value score distribution diagram of embodiment two of the application; Figure 10 is a risk score time sequence of embodiment two of the application; Figure 11 is a risk area time proportion of embodiment two of the application; Figure 12 is a multi-dimensional time sequence neural network model structure diagram of embodiment two of the application; Figure 13 is a training and test loss curve of embodiment two of the application; Figure 14 is a model prediction and actual value scatter diagram of embodiment two of the application; Figure 15 is a prediction error distribution diagram of embodiment two of the application; Figure 16 is a comparison chart of the first 100 samples in Embodiment Two of the present application. DETAILED DESCRIPTION

[0021] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. The embodiments described in the present application are only a part of the embodiments of the present application, but not all the embodiments. Based on the spirit of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0022] As shown in Figure 1 The present application discloses a time-series production simulation multi-dimensional risk analysis method based on deep learning, comprising the following steps: S1, a time-series production simulation model of a power system is constructed, and system operation state data is calculated; the system operation state data includes wind curtailment rate, light curtailment rate, thermal power unit output, energy storage system output power, energy storage system state of charge, demand response load adjustment amount, and transmission line power flow; S2, based on the system operation state data, a multi-dimensional risk index analysis system including multiple risk indexes is constructed, and initial weights of each risk index in the multi-dimensional risk index analysis system are calculated; The multiple risk indexes include wind curtailment rate, light curtailment rate, critical line power flow overrun degree, load loss amount proportion, and voltage stability index.

[0023] The initial weights of each risk index in the multi-dimensional risk index analysis system are determined in the following manner: first initial weights of each risk index are calculated based on an analytic hierarchy process; second initial weights of each risk index are calculated based on an entropy method; and the first initial weights and the second initial weights are weighted and fused to obtain the initial weights of each risk index.

[0024] S3, real-time state quantities corresponding to each risk index are introduced as driving factors of each risk index, and the initial weights of each risk index are corrected based on each driving factor to obtain corrected dynamic weights; and each risk index value is weighted and fused based on the dynamic weights to obtain a comprehensive risk value; The introduction of the real-time state quantities corresponding to each risk index as the driving factors of each risk index specifically includes: The wind power output variation coefficient is taken as the driving factor of the wind curtailment rate; the photovoltaic output variation coefficient is taken as the driving factor of the light curtailment rate; the critical line power flow utilization rate is taken as the driving factor of the critical line power flow overrun degree; the load proximity is taken as the driving factor of the load loss amount proportion; and the reactive power margin deficiency is taken as the driving factor of the voltage stability index.

[0025] The process of correcting the initial weights of each risk indicator based on each driving factor to obtain the corrected dynamic weights specifically includes: Select typical operating conditions of the power system, apply random disturbances of 1% to 3% to each driving factor corresponding to each risk indicator in the typical operating conditions, and calculate the rate of change of the comprehensive risk value corresponding to each driving factor. The normalized value of the rate of change of the comprehensive risk value corresponding to each driving factor is used as the sensitivity coefficient corresponding to each driving factor. Multiply each driving factor by its corresponding sensitivity coefficient, and then add 1 to get the dynamic adjustment factor for each risk indicator corresponding to each driving factor. The dynamic adjustment factor of each risk indicator is multiplied by the initial weight of each risk indicator, and after normalization and smoothing, the adjusted dynamic weight of each risk indicator is obtained.

[0026] The rate of change of the overall risk value corresponding to each driving factor is determined as follows:

[0027] in, For the first i The rate of change of the combined risk value of each driving factor For the first typical working condition i One driving factor The applied random perturbation, The comprehensive risk value under typical working conditions corresponds to The change in quantity.

[0028] S4. Multiply the normalized value of the driving factor corresponding to each risk indicator with the dynamic weight corresponding to each risk indicator, and use the result of all multiplications to form a feature vector as input; use the comprehensive risk value corresponding to the same moment as output to train the GRU network model, and use the trained GRU network model as the system's multi-dimensional risk analysis model; wherein, when training the GRU network model, a penalty weight is introduced into the loss function, and the penalty weight increases as the actual comprehensive risk value increases; The loss function is specifically as follows:

[0029] in, N The total number of training samples, The penalty weight corresponding to the actual comprehensive risk value. To predict the overall risk value, This represents the actual overall risk value.

[0030] S5. Calculate the product of the normalized value of the driving factor corresponding to each risk indicator and the dynamic weight corresponding to each risk indicator in the real-time operation data of the target system. Combine all the product results into a feature vector and input the feature vector into the multi-dimensional risk analysis model of the system to obtain the comprehensive risk value of the target system.

[0031] The present invention also discloses a deep learning-based multi-dimensional risk analysis system for time-series production simulation based on the aforementioned method, including a time-series production simulation and operation data calculation module, a multi-dimensional risk indicator analysis system construction module, a comprehensive risk value calculation module, a system multi-dimensional risk analysis model construction module, and a target system comprehensive risk value assessment module; The time-series production simulation and operation data calculation module constructs a time-series production simulation model of the power system and calculates the system operation status data. The multi-dimensional risk indicator analysis system construction module constructs a multi-dimensional risk indicator analysis system including multiple risk indicators based on the system operation status data, and calculates the initial weight of each risk indicator in the multi-dimensional risk indicator analysis system. The comprehensive risk value calculation module introduces real-time state quantities corresponding to each risk indicator as driving factors for each risk indicator, and corrects the initial weights of each risk indicator based on each driving factor to obtain the corrected dynamic weights; based on the dynamic weights, the values ​​of each risk indicator are weighted and fused to obtain the comprehensive risk value. The system's multi-dimensional risk analysis model construction module takes the product of the driving factors corresponding to each risk indicator and the dynamic weights to form a feature vector as input, and takes the comprehensive risk value corresponding to the same moment as output to train the GRU network model. The trained GRU network model is then used as the system's multi-dimensional risk analysis model. When training the GRU network model, a penalty weight is introduced into the loss function, and the penalty weight increases as the actual comprehensive risk value increases. The target system comprehensive risk value assessment module calculates the product of the normalized value of the driving factor corresponding to each risk indicator and the dynamic weight corresponding to each risk indicator in the real-time operation data of the target system. All multiplication results are combined into a feature vector, and the feature vector is input into the system multi-dimensional risk analysis model to obtain the comprehensive risk value of the target system.

[0032] Example 1: like Figure 1 As shown, this invention discloses a multi-dimensional risk analysis method based on deep learning for time-series production simulation, comprising the following steps: S1. Construct a time-series production simulation model of the power system and calculate the system operating status data; The system operating state data includes a wind curtailment rate, a light curtailment rate, a thermal power unit output, an energy storage system output power, an energy storage system state of charge, a demand response load adjustment amount, and a transmission line power flow; The time sequence production simulation model of the power system specifically comprises: (1) A thermal power unit operation constraint model

[0033]

[0034]

[0035] In the formula, is the active power output of the thermal power unit in the time period . , is the minimum active power output and the maximum active power output of the thermal power unit . , is the up and down ramp rate limit of the thermal power unit . is the unit start-stop state (1 running, 0 shutdown); , is the minimum running and shutdown time of the thermal power unit .

[0036] (2) A new energy power station operation constraint model

[0037]

[0038] In the formula, , are the actual outputs of the wind power station and the photovoltaic power station in the time period . , are the maximum predicted outputs of the wind power station and the photovoltaic power station in the time period .

[0039] (3) An energy storage power station operation constraint model

[0040]

[0041]

[0042] In the formula, is the output power of the energy storage power station in the time period . , respectively represent the charging and discharging power of the energy storage power station at the time period ; , respectively represent the charging and discharging efficiency of the energy storage power station ; , respectively represent the charging and discharging state of the energy storage power station

[0043]

[0044]

[0045]

[0046] respectively represent the upper and lower limits of the state of charge of the energy storage power station ; respectively represent the state of charge of the energy storage power station at the beginning and end of the period ; respectively represent the rated capacity of the energy storage power station , respectively represent the upper and lower limits of the state of charge of the energy storage power station ; , respectively represent the state of charge of the energy storage power station at the beginning and end of the period ;

[0047] (4) Price-type demand response resource operation constraint model

[0048]

[0049]

[0050]

[0051]

[0052] respectively represent the minimum and maximum price change at the time ; , respectively represent the minimum and maximum load change at the time ; respectively represent the minimum and maximum load change at the time ; respectively represent the minimum and maximum load change at the time ; , respectively represent the minimum and maximum load change at the time .

[0053] (5) Incentive demand response resource operation constraint model

[0054]

[0055]

[0056] wherein, is actual adjustment of the electric quantity of the incentive demand response at the moment; , respectively represent positive and negative reserve capacity at the moment; , respectively represent state quantity of positive and negative reserve capacity at the moment; , respectively represent maximum capacity of positive and negative reserve.

[0057] System-level balance and safety constraints are as follows: (1) System power balance constraint

[0058] wherein, is predicted value of the power system load at the moment.

[0059] (2) System reserve capacity constraint

[0060]

[0061]

[0062]

[0063] wherein, , respectively represent positive and negative reserve demand; , respectively represent load positive and negative reserve coefficients; , respectively represent new energy positive and negative reserve coefficients; (3) Tie-line transmission power constraint

[0064] wherein, is tie-line output power; , The upper and lower limits of transmission power of the tie line, respectively.

[0065] In step A, the objective function is constructed as follows:

[0066] In the formula, , respectively represent the weight of wind power and photovoltaic power consumption.

[0067] Specifically, the system operation state data generation process is as follows: Run the above-mentioned refined time sequence production simulation model, input including annual source load prediction curve, unit parameter, network parameter, etc., obtain the system operation state data of 8760 hours in a year by solving the mixed integer linear programming (MILP) problem, and constitute the training sample set.

[0068]

[0069] S2, based on the system operation state data, a multi-dimensional risk index analysis system including multiple risk indexes is constructed, and the initial weight of each risk index in the multi-dimensional risk index analysis system is calculated; The multiple risk indexes include wind curtailment rate, light curtailment rate, critical line flow out-of-limit degree, load loss proportion and voltage stability index; the construction of the multi-dimensional adaptive risk assessment index system can realize accurate identification of complex risks from the perspective of system integrity.

[0070] Each risk index is calculated in the following manner: (1) Wind curtailment rate , light curtailment rate

[0071] The wind curtailment rate and the light curtailment rate refer to the percentage of the power that is forced to be abandoned by the wind farm or photovoltaic power station due to insufficient grid regulation capacity, power transmission congestion or safety constraints, etc. to the theoretical maximum power generation under the condition of sufficient wind and light resources.

[0072]

[0073]

[0074] In the formula, , respectively represent the wind curtailment power and the light curtailment power in a dispatching period, , respectively represent the theoretical power generation of the wind farm and the theoretical power generation of the photovoltaic power station in a dispatching period.

[0075] (2) Critical line flow out-of-limit degree

[0076] The critical line flow over-limit degree is used to evaluate whether the actual power flow of the transmission line exceeds its thermal stability limit or stable transmission limit by calculating the risk value of the critical line flow exceeding the safe range. The critical line refers to the line that has a significant impact on the system flow distribution or is monitored by the dispatch department, including long-term high-load operation lines, power grid channel hub lines, and historical congestion lines. For the critical line set, the critical line flow over-limit degree is:

[0077] wherein, is the active power of the line l in a dispatch cycle, is the maximum active power allowed to pass through the line l , and is the total number of critical lines participating in the over-limit calculation.

[0078] (3) Loss load proportion

[0079] The loss load proportion refers to the percentage of the total amount of system load that is interrupted due to insufficient power generation resources, network congestion, or extreme faults.

[0080]

[0081] wherein, is the system load shedding amount in a dispatch cycle, is the total system load in a dispatch cycle.

[0082] (4) Voltage stability index The voltage stability index is the ratio of the synchronous unit on-line power to the total system load, which essentially measures the "reactive power support capacity margin" of the system to maintain voltage level.

[0083]

[0084] wherein, is the synchronous unit on-line power, is the total system load in a dispatch cycle.

[0085] The power system risk evaluation index system fully considers the influence of subjective and objective factors, and calculates the first initial weight of each risk index based on the analytic hierarchy process. The second initial weight of each risk index is calculated based on the entropy method. The first initial weight and the second initial weight are weighted and fused to obtain the initial weight of each risk index. The calculation method is as follows: (1) Analytic hierarchy process The Analytic Hierarchy Process (AHP) is a system analysis method that decomposes complex decision problems into hierarchical structures. Its core idea is to systematize the decision problem by establishing a ladder-like hierarchical model. This method can quantify qualitative problems, effectively transforming the decision-maker's subjective judgment into objective weights through rigorous mathematical calculations. The specific structure of the AHP is as follows: Figure 2 As shown.

[0086] The weights of relevant indicators are calculated using the Analytic Hierarchy Process (AHP). First, a three-scale AHP is used to determine the relative importance of each risk indicator, constructing a consensus matrix. Then, the corresponding maximum eigenvalue and eigenvector are calculated. After normalization, the eigenvectors represent the initial weights of each risk indicator. .

[0087] (2) Entropy method The entropy method is an objective weighting method based on the theory of information entropy. Its basic idea is to determine the weights based on the degree of variation of each indicator value. The entropy method relies on the dispersion of the data itself to determine the weights, effectively avoiding bias caused by human factors.

[0088] The entropy method first constructs the target attribute matrix. In this invention, the operating status data for 8760 hours throughout the year is obtained by time-series production simulation, and each hour corresponds to a set of risk indicator values. Indicates the first i The first hour j One risk indicator. The target attribute matrix. H Column normalization:

[0089] in, m The sample size, in this embodiment, corresponds to the number of operating hours throughout the year. n The number of risk indicators, To indicate the first i The first hour j The normalized value corresponding to each risk indicator; Next, calculate the... j Information entropy value of each risk indicator :

[0090] Then calculate the first... j The second initial weight of each risk indicator :

[0091] in, For the first k The information entropy value of each risk indicator.

[0092] Finally, the first initial weight solved by the analytic hierarchy process and the second initial weight solved by the entropy value method are fused to obtain the initial weight of each risk index:

[0093] wherein, is an influence factor, and .

[0094] S3, introducing real-time state quantities corresponding to each risk index as driving factors of each risk index, and correcting the initial weight of each risk index based on each driving factor to obtain a corrected dynamic weight; based on the dynamic weight, the risk index values are weighted and fused to obtain a comprehensive risk value; The dynamic weight can dynamically adjust the comprehensive risk value with the change of the real-time operation state of the system, avoiding the problem that the traditional static weight cannot reflect the importance change of each risk index under different working conditions.

[0095] The application selects real-time state quantities including wind power output variation coefficient, photovoltaic output variation coefficient, load proximity, key line flow utilization rate, and reactive power margin as driving factors.

[0096] Specifically, the wind power output variation coefficient is taken as the driving factor of the wind curtailment rate; the photovoltaic output variation coefficient is taken as the driving factor of the light curtailment rate; the key line flow utilization rate is taken as the driving factor of the key line flow overrun degree; the load proximity is taken as the driving factor of the load loss amount proportion; and the reactive power margin deficiency is taken as the driving factor of the voltage stability index.

[0097] The calculation method of each driving factor is as follows: ① Wind power output variation coefficient

[0098] The rolling time window with a length of is used to calculate the wind power output sequence The wind power output mean value and the wind power output standard deviation are calculated.

[0099]

[0100] The wind power output variation coefficient reflects the new energy fluctuation intensity, and is consistent with the change trend of the wind curtailment rate, and is specifically:

[0101] ② Photovoltaic output variation coefficient

[0102] The photovoltaic output sequence is averaged with a rolling time window of length

[0103]

[0104] The coefficient of variation of photovoltaic output reflects the intensity of photovoltaic fluctuation, which is consistent with the change trend of curtailment rate, and is specifically:

[0105] ③ Key line power flow utilization rate

[0106] For the set of key lines , the key line power flow utilization rate is:

[0107]

[0108] wherein, denotes the active power of line at time t , and denotes the maximum active power allowed to pass through line .

[0109] ④ Load proximity

[0110] The load proximity is measured by the ratio of system load to available generation capacity:

[0111] wherein, denotes the total load of the system at time t , and is the available output capacity of online units at time t , which can be taken as the value after deducting the standby demand from the online installed capacity according to the dispatching model.

[0112] ⑤ Insufficiency of reactive power margin

[0113] The reactive power margin is defined as the difference between the available reactive power and the demand:

[0114]

[0115] wherein,​​​​ Indicates the system at time... t The total amount of reactive power required to maintain voltage levels This indicates the maximum reactive power support capacity that the system can provide. An increase indicates insufficient reactive power support and deterioration of voltage stability.

[0116] To eliminate the influence of dimensions, the driving factors are normalized by min–max according to the annual statistical range, specifically as follows:

[0117]

[0118] in, For the first i Each driving factor at time... t The original value; For the first i Each driving factor at time... t The normalized median value; For the first i Each driving factor at time... t The final normalized value after truncation; For the first i The minimum value of each driving factor in the annual time series data. For the first i The maximum value of each driving factor in the annual time series data. The value is fixed to avoid the denominator being 0; where It is determined by annual time-series simulation data to enhance the model's robustness to extreme operating conditions.

[0119] To quantify the marginal risk impact of each driving factor, this invention introduces a sensitivity coefficient. By analyzing the driving factors under typical operating conditions Apply a small random perturbation of 1% to 3% to calculate the rate of change of risk, specifically:

[0120] in, For the first i The rate of change of the combined risk value of each driving factor For the first typical working condition i One driving factor The applied random perturbation, The comprehensive risk value under typical working conditions corresponds to The change in quantity.

[0121] Furthermore, the sensitivity coefficient is obtained by normalizing the rate of change of risk, specifically:

[0122] wherein, is the sensitivity coefficient of the first i driving factor, satisfying and ; reflecting the sensitivity of different driving factors to the risk score.

[0123] Further, the present application introduces a dynamic correction factor to represent the immediate influence of the index under the current operating state, specifically:

[0124] wherein the first driving factor corresponds to the first index.

[0125] is the static initial weight of each risk index obtained by combining the analytic hierarchy process and the entropy method in S2, and the initial weight is denoted as , which represents the structural importance of each risk index in the overall evaluation system, which remains unchanged within a dispatching cycle (i.e. within the analysis period).

[0126] The present application obtains the unnormalized immediate weight by applying the dynamic correction factor to the static base weight:

[0127] To avoid the weight distortion caused by the tendency to zero under extreme working conditions, the present application sets a minimum cut-off threshold , and performs:

[0128] Subsequently, all weights are normalized to obtain the final adaptive dynamic weight :

[0129] wherein, m is the total number of immediate weights, is the first k immediate weight.

[0130] To suppress the short-term drastic changes in weight caused by the random fluctuations of new energy, the present application further adopts an exponential smoothing strategy as needed:

[0131] wherein, is the first i immediate weight obtained after normalization and smoothing.The dynamic weight of each risk index is corrected, is an exponential smoothing coefficient used for smoothing the time evolution of the weight, avoiding jumps caused by noise, and .

[0132] Through the above method, the weight system of the application is composed of static structural importance , marginal risk sensitivity of the index and real-time index value driven by the operating state , realizing the organic combination of "long-term structural weight" and "short-term real-time state weight", so that the weight can be dynamically self-adaptive adjusted with the change of system operating conditions, effectively improving the accuracy and practicability of multi-dimensional risk fusion.

[0133] S4, multiplying the normalized value of each driving factor corresponding to each risk index and the dynamic weight corresponding to each risk index, and combining all the multiplication results to form a feature vector as input; taking the comprehensive risk value corresponding to the same moment as output, training the GRU network model, and taking the trained GRU network model as a system multi-dimensional risk analysis model; wherein, when training the GRU network model, a penalty weight is introduced into the loss function, and the penalty weight increases with the increase of the actual comprehensive risk value; The GRU network of the application not only extracts the time correlation of sequence data, but also automatically learns the nonlinear coupling relationship between operating indexes, the amplification or inhibition effect of weight change on power system risk, and the dynamic law of power system state evolution from local change to overall risk, thereby realizing the mapping from "multi-index dynamic feature sequence" to "overall risk score output".

[0134] The GRU network model specifically includes: (1) Gated Recurrent Unit: Gated Recurrent Unit (GRU) is a kind of recurrent neural network specially designed to process sequence data. It is born to solve the gradient vanishing and explosion problem of traditional recurrent neural network (RNN).

[0135]

[0136]

[0137]

[0138]

[0139] Among them, h t- , ht These represent the current time step. t The candidate hidden states generated by the corresponding gating unit and the final output hidden states; z t To update the gate at the current time step t The output determines the amount of information that needs to be retained in the current candidate hidden state; r t To reset the door at the current time step t The output of determines the hidden state that the neural network should output from the previous time step at the current time step. h t-1 The amount of information discarded in the process; x t Indicates the current time step t The input features; tanh(·) is the hyperbolic tangent function. It is the sigmoid function; W z , W r , W h , U z , U r and U h All are parameter matrices.

[0140] (2) Rectified Linear Unit: The Rectified Linear Unit (ReLU) is an activation function widely used in deep learning. It enhances the expressive power of neural networks by introducing complex nonlinear mappings.

[0141] ReLU( x ) = max(0, x ) In the formula, when the input is greater than or equal to 0, the output is the input value; when the input is less than 0, the output is 0.

[0142] (3) Dropout: This is a regularization technique used in neural networks to prevent overfitting. Its core idea is to randomly and temporarily remove a portion of neurons from the network during training, including their input and output connections. A comparison of neural networks before and after applying Dropout is shown below. Figure 3 and Figure 4 As shown.

[0143] (4) Introduce a penalty weight into the loss function. The penalty weight increases with the increase of the actual comprehensive risk value, making the prediction model more sensitive to the prediction error of high risk. Specifically, when training the GRU model, the comprehensive risk value is divided into three risk levels, namely high risk, medium risk and low risk; the relevant risk level is divided as shown in the following table:

[0144] According to the risk level, the penalty weight is determined for subsequent training of the GRU neural network model, which is specifically divided as shown below:

[0145] Further, the loss function is defined as:

[0146] wherein, N is the total number of training samples, is the penalty weight corresponding to the true comprehensive risk value, is the predicted comprehensive risk value, is the actual comprehensive risk value.

[0147] Embodiment two: The application discloses a time-series production simulation multi-dimensional risk assessment method based on deep learning, comprising: In order to construct the training set of the artificial intelligence model, the application constructs a time-series production simulation model on the IEEE 39-node test system based on the actual source-load data of a certain central province in China. The model includes thermal power, wind power, photovoltaic, energy storage and demand response resources, and takes the highest new energy consumption rate as the optimization target to perform simulation calculation for 8760 hours in a year.

[0148] Obtain detailed operation data of the system in a year. The construction method of the training set is as follows: the input data of simulation, such as load, maximum predicted output of wind power and photovoltaic, are taken as input features, and the output results of simulation, such as total output of thermal power units, actual consumption rate of wind power, actual consumption rate of photovoltaic and line flow data, are taken as output labels, which together constitute a sample set for training the model, and learn the complex mapping relationship from the initial source-load condition to the optimal operation state of the system. The system load, thermal power unit output, actual consumption rate of photovoltaic, actual consumption rate of wind power are as shown in Figure 5 、 Figure 6 、 Figure 7 、 Figure 8 .

[0149] The application adopts five indexes of wind curtailment rate, light curtailment rate, voltage stability index, critical line flow overrun degree and load loss proportion. Firstly, the system risk score of each index is calculated according to the system source-load data, and then the weight of each index is calculated. The consistent judgment matrix of the weight of each index calculated by the analytic hierarchy process is as shown in the following table:

[0150] On the basis of constructing the consistent judgment matrix, the corresponding maximum eigenvalue and the corresponding eigenvectors are calculated, and after normalization processing, the eigenvectors are the subjective weights of the corresponding influence factors.

[0151] The entropy method first standardizes the original data to eliminate the influence of dimension; then calculates the proportion of different sample values under each index; then calculates the entropy value of each index, quantifies the information content of the index through the entropy value formula; then calculates the difference coefficient according to the entropy value and determines the subjective weight of each index. Finally, based on the calculation of subjective and objective weights, the adaptive weight is calculated to get the final result.

[0152] After obtaining the system comprehensive operation risk value by using the index weight and risk score, the risk value is further divided into risk levels. According to the As Low As Reasonably Practically (ALAPR) criterion, combined with the "State Grid Corporation of China Safety Accident Investigation Regulations", two risk dividing lines are obtained: the unacceptable risk level line and the negligible risk level line, which are used to divide the risk value into three levels. The related risk value score distribution diagram and risk level line are shown in Figure 9 、 Figure 10 、 Figure 11 .

[0153] According to the comprehensive analysis of the related risk value score distribution diagram, the system negligible risk level line is 0.1, and the unacceptable risk level line is 0.2. The related risk level division is shown in the following table:

[0154] According to the risk level, the risk penalty loss function penalty weight is determined, which is used for subsequent training of the GRU neural network model. The specific division is shown as follows:

[0155] The core architecture of the multi-dimensional risk assessment model is based on GRU. The model input is 5-dimensional system source and load data of 8760 samples. First, the feature space is transformed through the fully connected input layer with 64 neurons, and the nonlinearity is introduced through the activation function; then it enters the double-layer bidirectional GRU network layer, which contains 128 hidden units (bidirectional processing is realized by doubling the output dimension), and the dropout rate of 0.2 is used between each layer for regularization to prevent overfitting. After time step selection, the GRU layer output is further extracted through the fully connected hidden layer with 64 neurons to extract high-level features, and the activation function is used again with a dropout rate of 0.3 to enhance the generalization ability. Finally, the continuous risk score prediction value is generated through the single neuron regression output layer. Its basic structure diagram is as follows:Figure 12 The model is trained and tested on the dataset.

[0156] The dataset is divided into training and validation sets, and the historical true risk scores are used as the supervision labels. During training, the model uses a risk penalty loss function, and the Adam optimizer (initial learning rate is 1x10 -3 ) performs batch gradient descent. The predicted value is obtained by forward propagation in each training batch, and the gradient is calculated by back propagation to update the network parameters. To improve the stability of training, gradient clipping is introduced to limit the gradient norm, dropout, validation set monitoring, and learning rate decay are used to prevent overfitting, and the early stopping strategy is triggered to terminate training when the validation loss does not decrease for several consecutive rounds.

[0157] The input of the GRU model of the application is derived from the time series production simulation output of step A, mainly including the maximum predicted output sequence of wind power, the maximum predicted output sequence of photovoltaic, the system load prediction sequence, the state of charge of energy storage, and the key line flow. The corresponding input matrix dimension is 8760x5. The model output is the system comprehensive risk value defined in step B after dynamic weight fusion, which is a continuous time sequence regression result, reflecting the evolution of the overall risk of the system with the change of source and load.

[0158] Then the accuracy and convergence of the model are tested, aiming to evaluate the generalization ability of the model, and ensure that it can effectively predict on unseen data, not just memorize the training samples. At the same time, the convergence test can verify whether the training process is stable and complete, and help identify potential problems such as overfitting and underfitting.

[0159] Figure 13 The training loss and test loss change with the training rounds, respectively. The training loss represents the error of the model on the training set, reflecting the ability of the model to learn data patterns. The test loss represents the error of the model on the unseen test set, reflecting the generalization ability of the model. As shown in the figure, the loss curve tends to balance with the change of training rounds, indicating that the model is convergent. At the same time, there is no situation that the training loss continues to decrease and the test loss begins to rise, and the model does not exist overfitting. There is also no situation that both curves are high and slowly declining, and the model does not exist underfitting.

[0160] Figure 14 The relationship between the actual risk score and the model predicted risk score is represented, and a total of 8760 points are included, each representing a sample, wherein x The value on the axis represents the actual value of the sample, y The axis represents the predicted value of the sample. The red dotted line represents the perfect prediction line, and as shown in the figure, the data is basically near the diagonal line, and the model prediction is accurate. R The R-squared value measures the degree to which the model explains the variability of the data.R 2 Close to 1, the model can perfectly predict within the error allowed.

[0161] Figure 15 The figure shows the histogram of the model prediction error distribution, where x The axis is the difference between the predicted value and the actual value, that is, the prediction error, y The axis is the number of samples that occur with this error, and the red dotted line represents the number of samples that occur with this error. As can be seen from the figure, the error distribution is centered around 0, and the distribution shape is close to normal distribution, and the error is relatively concentrated, which indicates that the error is relatively stable.

[0162] Figure 16 The figure shows the comparison between the actual value and the predicted value of the first 100 samples of the model, where x The axis is the index of the first 100 samples of the system, y The axis represents the risk score value. Among them, the blue circle point line represents the actual value, and the orange square point line represents the predicted value. As can be seen from the figure, the model does not have systematic deviation.

[0163] The indicators used to measure the performance of the model mainly include six dimensions: the determination coefficient R 2 , the mean square error, the 10% accuracy, the mean absolute error (MAE), the root mean square error (RMSE), and the mean absolute percentage error (MAPE). The 10% accuracy refers to the proportion of samples with an error within 10%, and is used to measure the reliability of the model. The mean absolute error represents the average absolute error between the predicted value and the true value, which is not sensitive to outliers and the smaller the value is, the better. The root mean square error refers to the square root of the average value of the square of the prediction error, which punishes larger errors more severely, and the smaller the value is, the better. The mean absolute percentage error refers to the average percentage of the prediction error relative to the true value, which is convenient for comparing different data sets and is sensitive to true values close to 0, and the smaller the value is, the better. The performance indicators are shown in the following table:

[0164] Example three: The terminal provided by the embodiment of the application comprises a processor and a storage medium. The storage medium is used for storing instructions. The processor is used for operating according to the instructions to perform the steps of the method according to any one of the embodiments.

[0165] Example four: The embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the steps of the method in any one of the embodiments.

[0166] According to the scheme, the redundant backup protection for each line can be realized in the bus protection by the above processing, and the expected goal of improving the adaptability of the protection by cutting off the fault when the fault occurs at the end of the line is met.

[0167] The present disclosure can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium (media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present disclosure.

[0168] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or punched tape, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.

[0169] Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network can comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0170] Computer readable program instructions for carrying out operations of the present disclosure can be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like, and conventional procedural programming languages such as the "C" programming language or similar programming languages. The computer readable program instructions can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate array (FPGA), or programmable logic array (PLA) can execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present disclosure.

[0171] Finally, it should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, but not to limit it. Although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced, and any modification or replacement without departing from the spirit and scope of the present application should be covered in the protection scope of the claims of the present application.

Claims

1. A deep learning based multi-dimensional risk analysis method for time series production simulation, characterized in that, The method comprises the following steps: S1, constructing a time sequence production simulation model of a power system, and calculating system operation state data; S2, based on the system operation state data, constructing a multi-dimensional risk index analysis system comprising multiple risk indexes, and calculating initial weights of each risk index in the multi-dimensional risk index analysis system; S3, introducing real-time state quantities corresponding to each risk index as driving factors of each risk index, and modifying the initial weights of each risk index based on each driving factor to obtain modified dynamic weights; based on the dynamic weights, weighting and fusing each risk index value to obtain a comprehensive risk value; S4, multiplying the normalized values of the driving factors corresponding to each risk index by the dynamic weights corresponding to each risk index, and combining all the multiplication results into a feature vector as input; taking the comprehensive risk value corresponding to the same moment as output, training a GRU network model, and taking the trained GRU network model as a system multi-dimensional risk analysis model; wherein, when training the GRU network model, a penalty weight is introduced in the loss function, and the penalty weight increases with the increase of the actual comprehensive risk value; S5, calculating the product of the normalized values of the driving factors corresponding to each risk index in the real-time operation data of the target system and the dynamic weights corresponding to each risk index, combining all the multiplication results into a feature vector, inputting the feature vector into the system multi-dimensional risk analysis model, and obtaining a target system comprehensive risk value.

2. The deep learning-based time sequence production simulation multi-dimensional risk analysis method according to claim 1, wherein the system operation state data comprises wind curtailment rate, light curtailment rate, thermal power unit output, energy storage system output power, energy storage system state of charge, demand response load adjustment amount, and transmission line power flow; and the multiple risk indexes comprise wind curtailment rate, light curtailment rate, critical line power flow overrun degree, load loss amount proportion, and voltage stability index.

3. The deep learning-based time sequence production simulation multi-dimensional risk analysis method according to claim 2, wherein in S2, the initial weights of each risk index in the multi-dimensional risk index analysis system are determined in the following manner: first initial weights of each risk index are calculated based on an analytic hierarchy process; second initial weights of each risk index are calculated based on an entropy method; the first initial weights and the second initial weights are weighted and fused to obtain the initial weights of each risk index.

4. The deep learning-based time sequence production simulation multi-dimensional risk analysis method according to claim 2, wherein in S3, the real-time state quantities corresponding to each risk index are introduced as driving factors of each risk index, specifically comprising: a wind power output variation coefficient is taken as a driving factor of wind curtailment rate; a photovoltaic output variation coefficient is taken as a driving factor of light curtailment rate; a critical line power flow utilization rate is taken as a driving factor of critical line power flow overrun degree; a load proximity is taken as a driving factor of load loss amount proportion; and a reactive power margin deficiency degree is taken as a driving factor of voltage stability index. ​ ​ ​ 5. The deep learning-based multi-dimensional risk analysis method for time-series production simulation according to claim 4, wherein, in S3, the initial weight of each risk indicator is corrected based on each driving factor to obtain a corrected dynamic weight, and the correction specifically includes: selecting a typical operating condition of the power system, applying a random disturbance of 1% to 3% to each driving factor corresponding to each risk indicator in the typical operating condition, and calculating a comprehensive risk value change rate corresponding to each driving factor; taking the normalized value of the comprehensive risk value change rate corresponding to each driving factor as a sensitivity coefficient corresponding to each driving factor; multiplying each driving factor and the corresponding sensitivity coefficient, and adding 1 as a dynamic correction factor of each risk indicator corresponding to each driving factor; multiplying the dynamic correction factor of each risk indicator and the initial weight of each risk indicator, and performing normalization and smoothing processing to obtain the corrected dynamic weight of each risk indicator.

6. The deep learning-based multi-dimensional risk analysis method for time-series production simulation according to claim 5, wherein, the comprehensive risk value change rate corresponding to each driving factor is determined in the following manner: in, For the first i The rate of change of the combined risk value of each driving factor For the first typical working condition i One driving factor The applied random perturbation, The comprehensive risk value under typical working conditions corresponds to The change in quantity.

7. The deep learning-based multi-dimensional risk analysis method for time-series production simulation according to claim 5, wherein, in S4, the loss function is specifically: wherein, N is the total number of training samples, is the penalty weight corresponding to the actual comprehensive risk value, is the predicted comprehensive risk value, is the actual comprehensive risk value.

8. A deep learning-based multi-dimensional risk analysis system for time-series production simulation based on the method of any one of claims 1 to 7, comprising a time-series production simulation and operating data calculation module, a multi-dimensional risk indicator analysis system construction module, a comprehensive risk value calculation module, a system multi-dimensional risk analysis model construction module, and a target system comprehensive risk value evaluation module, wherein: the time-series production simulation and operating data calculation module constructs a time-series production simulation model of the power system and calculates system operating state data; the multi-dimensional risk indicator analysis system construction module constructs a multi-dimensional risk indicator analysis system including multiple risk indicators based on the system operating state data, and calculates the initial weight of each risk indicator in the multi-dimensional risk indicator analysis system; the comprehensive risk value calculation module introduces real-time state quantities corresponding to each risk indicator as driving factors of each risk indicator, and corrects the initial weight of each risk indicator based on each driving factor to obtain a corrected dynamic weight; based on the dynamic weight, the risk indicator values are weighted and fused to obtain a comprehensive risk value; the system multi-dimensional risk analysis model construction module takes the product of each risk indicator corresponding driving factor and the dynamic weight as a feature vector as input, takes the comprehensive risk value corresponding to the same moment as output, trains a GRU network model, and takes the trained GRU network model as a system multi-dimensional risk analysis model; wherein, when training the GRU network model, a penalty weight is introduced in the loss function, and the penalty weight increases with the increase of the actual comprehensive risk value. The target system comprehensive risk value evaluation module calculates the product of the normalized value of each risk index corresponding driving factor in the real-time operation data of the target system and the dynamic weight corresponding to each risk index, groups all the multiplication results into a feature vector, inputs the feature vector into a system multi-dimensional risk analysis model, and obtains a target system comprehensive risk value. 9.A computer device comprising a memory, a processor and a computer program stored on the memory; characterized in that, the processor executes the computer program to implement the steps of the method of any one of claims 1-7. 10.A computer readable storage medium having stored thereon a computer program, characterized in that, the computer program, when executed by a processor, implements the steps of the method of any one of claims 1-7.

Images