A reliability assessment method for islanded microgrid considering source and load uncertainty

Abstract

The present invention belongs to the technical field of microgrid reliability assessment, and discloses an islanded microgrid reliability assessment method considering source-load uncertainty, firstly establishing an uncertainty model of an islanded microgrid system; using a CNN‑BILSTM‑MATT hybrid model optimized by a RIME algorithm to predict wind and solar loads; combining an adaptive bandwidth kernel density estimation method to perform interval prediction of wind and solar power; using Monte Carlo sampling to sample wind and solar output within a wind and solar prediction interval, combining wind and solar output status to obtain the actual power of simulated wind and solar; combining power load prediction data, conventional generators, and energy storage status to calculate microgrid reliability indicators. The present invention takes into account the uncertainty of wind and solar prediction, uses a hybrid model to improve the accuracy of wind and solar output prediction, realizes accurate prediction of microgrid reliability, and provides microgrid operators with more accurate and reliable operation status monitoring, thereby optimizing the stable operation and resource scheduling of microgrids.

Description

Technical Field

[0001] The present invention belongs to the technical field of microgrid reliability assessment, and in particular relates to an islanded microgrid reliability assessment method taking source-load uncertainty into consideration. Background Art

[0002] With the widespread use of renewable energy and the intelligent development of power systems, the stability and reliability of islanded microgrids, as an independently operated power supply system, are of great significance for ensuring power supply in remote areas and improving energy efficiency. However, islanded microgrids face many challenges in actual operation, especially under the influence of source and load uncertainty. How to accurately evaluate the reliability of microgrids has become an urgent problem to be solved.

[0003] With the widespread application of renewable energy in microgrids, its inherent volatility and uncertainty have brought new challenges to the reliability and stability of the system. Existing microgrid reliability assessment methods are mainly based on deterministic models and uncertainty models. However, even though existing technologies have tried to take these uncertainties into account, problems such as insufficient prediction accuracy, high model complexity, and high computational costs still exist. In addition, existing methods are still insufficient in terms of real-time and adaptability, and it is difficult to meet the urgent needs of isolated microgrids for high reliability assessment. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides an islanded microgrid reliability assessment method considering source-load uncertainty, which not only considers multiple aspects such as source-load uncertainty, unit status, and energy storage life, but also adopts the RIME-CNN -BiLSTM -MATT -ABKDE model to improve the accuracy of wind and solar power output prediction, thereby realizing accurate prediction of microgrid reliability, providing microgrid operators with more accurate and reliable operation status monitoring and load demand forecasting, thereby optimizing the stable operation and resource scheduling of the microgrid.

[0005] The present invention provides an island microgrid reliability assessment method considering source-load uncertainty, comprising the following steps:

[0006] S1. Establish an uncertainty model for the islanded microgrid system, including: using a "fault-repair" time series model to simulate the life process of the micro gas turbine unit, using a "normal-deficit-outage" time series model combined with wind power and photovoltaic prediction errors to simulate wind and solar uncertainty, establishing a load uncertainty model that takes into account load prediction errors, and establishing an energy storage charging and discharging model that takes into account the energy storage life;

[0007] S2, using the frost optimization algorithm RIME to optimize the CNN-BILSTM-MATT hybrid model composed of a convolutional neural network (CNN), a bidirectional long short-term memory network (BILSTM) and a multi-head attention mechanism (MATT) to predict wind, solar and load;

[0008] S3. Adaptive Bandwidth Kernel Density Estimation (ABKDE) is used to calculate the probability density function of the prediction error, and the interval prediction of wind and solar power is performed by combining the CNN-BILSTM-MATT hybrid model and the probability statistics of the error;

[0009] S4. Monte Carlo sampling is used to sample wind and solar power output within the wind and solar power prediction interval, and combined with the wind and solar power normal-deficit-outage timing model, the actual power of the simulated wind and solar power at this time is obtained;

[0010] S5. According to the uncertainty model of the microgrid established in S1, the operating status of the micro gas turbine unit and the energy storage is obtained. Combined with the power load forecast data, the reliability index load loss probability and the expected value of power shortage are calculated, and the reliability of the microgrid is evaluated through the calculated reliability index.

[0011] Furthermore, in S1, a “fault-repair” time series model is used to simulate the life process of the micro gas turbine unit, specifically:

[0012] The power generation model of the micro gas turbine unit is represented by a two-state Markov model, and the failure rate and repair rate of the generator are expressed as and Micro gas turbine Trouble-free working time and troubleshooting time They are:

[0013] (1),

[0014] (2),

[0015] In the formula, and is a random variable uniformly distributed on the interval [0, 1]; is the failure rate of micro gas turbine i, is the repair rate of micro gas turbine i.

[0016] Furthermore, in S1, the “normal-deficit-outage” time series model is used to simulate the uncertainty of wind and solar power combined with the wind power and photovoltaic prediction errors, specifically:

[0017] The Markov method is defined by the transition probability, the random variable from the current moment to the next moment, the state Transfer to state Probability for:

[0018] (3),

[0019] In the formula, represents a discrete random variable at time t, is the possible state value of the random variable, i = 1, 2, ...; i is a certain moment, j is the next moment of i;

[0020] random variable The value is The probability is:

[0021] (4),

[0022] In the formula, is the state value of the random variable at time k, Represents a discrete random variable The value is probability;

[0023] Assuming the number of states is n, then:

[0024] (5),

[0025] State transfer matrix of wind turbines and photovoltaics It is expressed as follows:

[0026] (6),

[0027] (7),

[0028] The statistical value of P is affected by the number of samples. When the number of samples is sufficient, a state transfer matrix that meets certain error requirements is obtained, and its statistical value will be infinitely close to the true value.

[0029] The duration of each output state is obtained according to the transfer rate between different outputs ,Right now:

[0030] (8),

[0031] Among them, 8760 is the number of hours in 365 days in a year.

[0032] The new energy power generation system, i.e. wind power and photovoltaic power, is divided into three states: operation, shortage and shutdown;

[0033] State 1 is normal operation r: in this state, there are no faulty components inside the new energy unit, and the new energy power generation system operates in a planned manner;

[0034] State 2, i.e., shortage d: some constraints in the new energy system cannot meet the conditions. After adjustment, it operates at a certain proportion of the original plan;

[0035] State 3 is shutdown: In this state, there will be a temporary shutdown due to failure of internal system components.

[0036] Assume that the initial state of the new energy unit is According to the definition of Markov matrix, the state of the new energy unit after unit time is expressed as:

[0037] (9),

[0038] After m unit time, the state transfer of the new energy unit is expressed as:

[0039] (10),

[0040] The initial state of the new energy unit and its state transfer matrix are known to obtain the steady-state value of its state distribution at each moment; the probability formulas of the three states of operation, shortage and shutdown are expressed as:

[0041] (11),

[0042] In the formula, is the transition rate between normal operation, shortage and outage states, and its subscript represents the direction of state transition; is the probability of being in operation, shortage and outage status;

[0043] The average duration of stays in normal operation, shortage and outage is:

[0044] (12),

[0045] In the formula, The duration of normal operation, shortage and outage.

[0046] As new energy is affected by factors such as natural wind and light intensity, the output of new energy is characterized by volatility and uncertainty. The prediction error of new energy output is described by a normal distribution with a zero mean. The relationship between the prediction error of new energy output and the predicted value is as follows:

[0047] (13),

[0048] (14),

[0049] In the formula, and They are The actual and predicted values ​​of renewable energy output during the period; It is the prediction error of new energy output; represents the standard deviation of the PV output prediction error;

[0050] Finally, the power output taking into account the operating status of the new energy power generation system is obtained The expression is:

[0051] (15),

[0052] in:

[0053] (16),

[0054] In the formula, is the state function of the new energy unit; They are the three status sequences of normal operation, shortage and shutdown of new energy units.

[0055] Furthermore, in S1, a load uncertainty model considering load forecasting error is established, specifically:

[0056] The load forecast error follows a normal distribution with a zero mean, and the relationship between the load forecast error and the forecast value is established as follows:

[0057] (17),

[0058] (18),

[0059] In the formula, and They are The actual and predicted values ​​of system load during the time period; is the load forecast error; represents the standard deviation of load forecast error; is the proportionality factor that characterizes the accuracy of the prediction tool.

[0060] Furthermore, in S1, an energy storage charging and discharging model considering the energy storage life is established. In the energy storage system ESS charging and discharging model, the energy storage charging process is:

[0061] (19),

[0062] The discharge process is:

[0063] (20),

[0064] In the formula, and They are The power and stored energy of ESS at any moment; and They are The maximum charging and discharging power of ESS at the moment; and are the upper and lower limits of ESS storage energy respectively.

[0065] Charging or discharging the energy storage system will cause losses. Taking the depth of discharge as the main consideration, the number of battery cycles is inversely proportional to the depth of discharge and will decrease as the depth of discharge increases. The battery life model is:

[0066] (twenty one),

[0067] In the formula, is the actual discharge depth of the battery, is the rated discharge depth; , are the number of cycles corresponding to the rated discharge depth and the actual discharge depth respectively; according to the equivalent cycle life method, the actual number of cycles is converted to the rated number of cycles as follows:

[0068] (twenty two),

[0069] In the formula, is the equivalent number of cycles after conversion; the estimated result of the battery service life is:

[0070] (twenty three),

[0071] In the formula, is the estimated result of the battery service life, and N is the number of research periods of the battery in one year.

[0072] Furthermore, in S2, the CNN-BILSTM-MATT hybrid model is constructed as follows:

[0073] The CNN model consists of multiple convolutional layers, activation functions, pooling layers, and fully connected layers, and is trained using the back-propagation algorithm. During the training process, CNN minimizes the loss function by adjusting weight parameters.

[0074] Convolution calculation method:

[0075] (twenty four),

[0076] In the formula, is the convolution operation; The sequence calculated for the convolution; is the weight of the convolution kernel; Indicates the offset; is the activation function;

[0077] BILSTM processes sequence data by running two LSTM models simultaneously in the time dimension, one from front to back, that is, forward, and the other from back to front, that is, reverse. The LSTM model includes a forget gate, an input gate, and an output gate. The relationship between the current state of the LSTM model and the state at the previous moment is:

[0078] (25),

[0079] In the formula, is the sigmoid activation function; tanh is the hyperbolic tangent activation function; for Input of time; for The hidden state of the moment; , , , is the weight matrix of LSTM; , , , is the bias of LSTM; , , , For The forgetting threshold, input threshold, output threshold, and state matrix at the moment;

[0080] For the BILSTM model:

[0081] (26),

[0082] In the formula, for The forward hidden layer state at that moment; for The hidden layer state at the moment backward; , are the weight matrix and bias term respectively; for The output at each moment; during the BILSTM model training process, the multi-head attention mechanism is introduced as a supplement to BILSTM; the vector output by BILSTM is converted into three dimensions through three different mapping operations. The input matrix of is: query Q, key K and value V, and its attention output matrix is:

[0083] (27),

[0084] In the formula, It is the feature dimension of each key, used for weight scaling, and is normalized to the interval [0, 1] by softmax.

[0085] The multi-head attention mechanism divides the time series into subspaces, each head performs self-attention calculation on the subspace; then, The results of the individual heads are spliced ​​and integrated to obtain multiple heads. Each head is spliced ​​and the final result is obtained through linear transformation:

[0086] (28),

[0087] In the formula, , , Respectively , , The weight matrix of

[0088] (29),

[0089] In the formula, represents the weight of the linear transformation; represents the first Size; Represents a splicing operation; For the final output result, more feature information can be learned from different spaces.

[0090] Furthermore, the RIME algorithm is used to optimize the CNN-BILSTM-MATT hybrid model, specifically:

[0091] S2-1, initialize the frost population and parameters, including the number of layers, number of neurons and learning rate of the neural network;

[0092] S2-2, implement the soft frost ice search strategy to calculate the location of each soft frost ice particle; the soft frost ice search strategy updates the location of the particle, which is the local search part of the algorithm and helps the algorithm to perform detailed search near the currently found better solution; the calculation formula for soft frost ice particles is:

[0093] (30),

[0094] The parameters are expressed as:

[0095] (31),

[0096] (32),

[0097] (33),

[0098] In the formula, is the new position of the particle after the update; Indicates The first Particles; For the frost population The best frost agent in Particles; It is a random number in the range of (-1,1); is the angle between the frost movement direction and the horizontal direction; and Together, the direction of particle motion will change with the number of iterations; is another random coefficient, and Similarly, providing randomness to the search process; is the adhesion coefficient, which affects the condensation probability of the agent and increases with the number of iterations; and Respectively represent The upper and lower bounds of the dimension parameters are the constraints of the optimization problem; It is an environmental factor, which follows the number of iterations to simulate the influence of the external environment and is used to ensure the convergence of the algorithm; is the adhesion, a random number in the range of (0,1), used to control the distance between the centers of two frost particles; Indicates rounding; To control the number of segments of the step function, the default value is 5; is the current iteration number; is the maximum number of iterations of the algorithm;

[0099] S2-3. Implement the hard frost ice puncture mechanism to exchange the frost ice particles of the algorithm and absorb the boundaries of the new search positions to ensure that the search positions do not exceed the search space described by the problem; the hard frost ice puncture mechanism is the global search strategy of the algorithm. It increases the diversity of the population by introducing new and different particle positions to prevent the algorithm from falling into the local optimal solution; the exchange formula between particles in the hard frost ice puncture mechanism is:

[0100] (34),

[0101] In the formula, It is a random number in the range of (-1,1); Represents the normalized value of the current agent’s fitness value, indicating the The chance of a frost agent being selected; Representative A set or vector of solutions or states;

[0102] S2-4. Implement a positive greedy selection mechanism to compare the fitness value of the updated frost particle position with that before the update, where the fitness value is an indicator to measure the quality of the solution; if the updated fitness value is better than the value before the update, a replacement occurs, thereby obtaining the optimal position of the frost particle, that is, the global optimal solution; if the number of population iterations reaches the maximum or the fitness value is less than the error limit, stop the iteration and output the optimal solution, otherwise continue the iteration.

[0103] Furthermore, S3 is specifically:

[0104] For a set of n independently distributed power prediction error data , and its corresponding probability density function is as follows:

[0105] (35),

[0106] In the formula, represents the probability density function; is the number of power prediction error samples; is the window bandwidth; Represents the kernel function used for kernel density estimation; is the independent variable of the probability density function;

[0107] A smooth, continuous, and obviously unimodal Gaussian kernel function is selected as the kernel function of kernel density estimation. The calculation formula of the power prediction probability density function based on kernel density estimation is:

[0108] (36),

[0109] Among them, the bandwidth h has a greater impact on the accuracy of kernel density estimation than the kernel function. In theory, the bandwidth should be infinitely close to zero, but if the bandwidth is too small, there will be too few points participating in the fitting in the neighborhood, affecting the distribution of the estimated model. If the bandwidth is too large, the distribution curve will be too smooth and the actual distribution of the prediction error cannot be accurately estimated. In order to quantitatively realize the optimal bandwidth calculation, the mean integrated squared error (MISE) indicator is used. The MISE expression is:

[0110] (37),

[0111] In the formula, is the forecast error estimate; is the actual value of the prediction error; To find the mean function;

[0112] The adaptive bandwidth kernel density estimation method is obtained by correcting the relevant parameters of the fixed bandwidth kernel density estimation through a suitable bandwidth correction function. The calculation formula is:

[0113] (38),

[0114] (39),

[0115] In the formula, is the error probability density function under adaptive bandwidth; is the adaptive revision coefficient of bandwidth; is the sensitivity parameter, ranging from 0 to 1; Indicates bandwidth is The kernel density estimation function is Represents the total number of prediction error samples.

[0116] Furthermore, in S4, Monte Carlo sampling is used to sample the wind and solar power output within the wind and solar power prediction interval, and the expression is:

[0117] (40),

[0118] In the formula, is the analog sampling power at that moment; and For the moment The lower and upper limits of the power range covered; Generate a random number (0-1) for uniform distribution; combine the sampled power with the wind and solar output status to obtain the actual power of the simulated wind and solar at this time, which is:

[0119] (41),

[0120] In the formula, To simulate the actual power of the wind and solar power generator sets.

[0121] Furthermore, in S5, the power load forecast data, conventional generators and energy storage status are combined to calculate the microgrid reliability index, which is:

[0122] S5-1, data initialization, forming a load level of 8760h and determining the full time series simulation duration;

[0123] S5-2. Monte Carlo is used to sample the power within the prediction interval, and the simulated power considering the state of the wind and solar power generating units is obtained in combination with its operating status; if the simulated power is greater than the load demand, the energy storage is charged to store excess energy; if it is less than the load demand, the normal working duration TTF and fault repair time TTR of the gas turbine unit are obtained according to the Markov method, and they are arranged in sequence to form a sequence of the operating state duration of each gas turbine unit in the total simulation time, and the gas turbine outputs power to supply the remaining load; when the combined output of wind, solar and gas turbines is still less than the load demand, the energy storage is discharged; if the output of wind, solar, gas turbine and energy storage still cannot meet the load demand, the load is powered off, that is, the load is lost;

[0124] S5-3, determine whether the total simulation time is greater than the number of simulation years multiplied by 8760 hours; if so, execute S5-4, otherwise jump to S5-2;

[0125] S5-4, calculating reliability indicators, wherein the reliability indicators include load loss probability LOLP and expected power shortage value EENS of the system;

[0126] The LOLP calculation formula is:

[0127] (42),

[0128] (43),

[0129] In the formula, is the total simulation time; for The power demand of the load at any given moment; for The sum of wind power, photovoltaic, gas turbine and energy storage output power at the moment;

[0130] The EENS calculation formula is:

[0131] (44),

[0132] In the formula, is the set of states where load loss occurs; System Status probability; For the system status The amount of load loss under.

[0133] The beneficial effects described in the present invention are as follows: the present invention not only improves the prediction accuracy of wind and solar power output by adopting the CNN-BILSTM-MATT model optimized by the RIME algorithm, but also comprehensively considers the random fluctuations and uncertainty factors in the microgrid by establishing a comprehensive uncertainty model of the microgrid; in addition, combined with the adaptive bandwidth kernel density estimation (ABKDE) method, the present invention can provide a more accurate prediction of wind and solar power intervals and enhance the ability of risk management; using the Monte Carlo sampling technology, the present invention can sample within the prediction interval to obtain a simulated wind and solar power that is closer to the actual operating conditions, thereby realizing an accurate assessment of the reliability of the microgrid. The present invention also provides microgrid operators with a new tool for optimizing operating status monitoring and load demand forecasting, which helps to improve the stability of the system and reduce operating costs. The present invention, with its flexibility in technical implementation, can adapt to different microgrid configurations and operating conditions, solves the deficiencies of the prior art in prediction accuracy, uncertainty processing, risk assessment and real-time optimization, and provides a more effective and reliable method for the reliability assessment of isolated microgrids. BRIEF DESCRIPTION OF THE DRAWINGS

[0134] Figure 1 is a flow chart of the present invention;

[0135] Figure 2 This is the CNN-BILSTM-MATT prediction flow chart;

[0136] Figure 3 Comparison of wind power prediction using different models; (a) is the wind power prediction using the CNN model, (b) is the wind power prediction using the BILSTM model, (c) is the wind power prediction using the CNN-BILSTM-MATT model, and (d) is the wind power prediction using the RIME-CNN-BILSTM -MATT model.

[0137] Figure 4 This is a comparison chart of prediction error distribution between adaptive and fixed bandwidth kernel density estimation methods;

[0138] Figure 5 Flowchart for microgrid reliability assessment. DETAILED DESCRIPTION

[0139] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0140] like Figure 1 The present invention provides an island microgrid reliability assessment method considering source-load uncertainty, comprising the following steps:

[0141] S1. Establish an uncertainty model for an islanded microgrid system: Use a "fault-repair" time series model to simulate the life cycle of a micro gas turbine unit; use a "normal-deficit-outage" time series model combined with wind power and photovoltaic prediction errors to simulate wind and solar uncertainty; establish a load uncertainty model that takes into account load prediction errors; establish an energy storage charging and discharging model that takes into account the energy storage life;

[0142] S2, using the frost optimization algorithm RIME to optimize the CNN-BILSTM-MATT hybrid model composed of a convolutional neural network (CNN), a bidirectional long short-term memory network (BILSTM) and a multi-head attention mechanism (MATT) to predict wind, solar and load;

[0143] S3. Adaptive Bandwidth Kernel Density Estimation (ABKDE) is used to calculate the probability density function of the prediction error, and the interval prediction of wind and solar power is performed by combining the CNN-BILSTM-MATT hybrid model and the probability statistics of the error;

[0144] S4. Monte Carlo sampling is used to sample wind and solar power output within the wind and solar power prediction interval, and combined with the wind and solar power normal-deficit-outage timing model, the actual power of the simulated wind and solar power at this time is obtained;

[0145] S5. According to the uncertainty model of the microgrid established in S1, the operating status of the micro gas turbine unit and the energy storage is obtained. Combined with the power load forecast data, the probability of load loss and the expected value of power shortage are calculated, and the reliability of the microgrid is evaluated through the calculated reliability indicators.

[0146] In S1, an uncertainty model of the islanded microgrid system is established: the “failure-repair” time series model is used to simulate the life process of the micro gas turbine unit. The power generation model of the micro gas turbine unit can be represented by a two-state Markov model. The failure rate and repair rate of the generator are respectively expressed as and Micro gas turbine Trouble-free working time and troubleshooting time They are:

[0147] (1),

[0148] (2),

[0149] In the formula, and is a random variable with uniform distribution on the interval [0, 1].

[0150] The “normal-deficit-outage” time series model is used to simulate the uncertainty of wind and solar power combined with wind power and photovoltaic prediction errors, specifically:

[0151] The Markov chain can obtain the probability and duration of each state through its transition rate according to the known state. The present invention uses this method to simulate multi-state wind turbines and photovoltaics. The Markov method is defined by the transition probability. The random variable changes from the current moment to the next moment. Transfer to state Probability for:

[0152] (3),

[0153] In the formula, represents a discrete random variable at time t, is the possible state value of the random variable, i = 1, 2, ...;

[0154] random variable The value is The probability is:

[0155] (4),

[0156] In the formula, Represents a discrete random variable The value is probability.

[0157] Assuming the number of states is n, then:

[0158] (5),

[0159] State transfer matrix of wind turbines and photovoltaics It is expressed as follows:

[0160] (6),

[0161] (7),

[0162] The statistical value of P is affected by the number of samples. When the number of samples is sufficient, a state transfer matrix that meets certain error requirements can be obtained, and its statistical value will be infinitely close to the true value.

[0163] The duration of each output state is obtained according to the transfer rate between different outputs ,Right now:

[0164] (8);

[0165] The new energy power generation system is divided into three states: operation, shortage and shutdown. The transfer rate of operation, shortage and shutdown states is (Unit: times / year) It can be obtained from the historical fault data of wind power and photovoltaic stations:

[0166] State 1 (normal operation): In this state, there are no faulty components inside the new energy unit, and the new energy power generation system operates in a planned manner;

[0167] State 2 (deficient d): some constraints in the new energy system cannot be met. After adjustment, it operates at a certain proportion of the original plan;

[0168] State 3 (outage): In this state, there will be a temporary outage due to a failure of internal system components.

[0169] Assume that the initial state of the new energy unit is , according to the definition of Markov matrix, the state of the new energy unit after unit time can be expressed as:

[0170] (9),

[0171] After m unit time, the state transfer of the new energy unit is expressed as:

[0172] (10).

[0173] Given the initial state of the wind turbine and its state transfer matrix, the steady-state value of its state distribution at each moment can be obtained. The probability formulas for the three states of operation, shortage and shutdown can be expressed as:

[0174] (11),

[0175] In the formula, is the transition rate between normal operation, shortage and outage states, and its subscript represents the direction of state transition; is the probability of being in operation, shortage and outage status;

[0176] The average duration of stays in normal operation, shortage and outage is:

[0177] (12),

[0178] In the formula, The duration of normal operation, shortage and outage.

[0179] As new energy is affected by factors such as natural wind and light intensity, the output of new energy is characterized by volatility and uncertainty. The prediction error of new energy output can be described by a normal distribution with a zero mean. The relationship between the prediction error of new energy output and the predicted value is as follows:

[0180] (13),

[0181] (14),

[0182] In the formula, and They are The actual and predicted values ​​of renewable energy output during the period; It is the prediction error of new energy output; represents the standard deviation of the PV output prediction error;

[0183] Finally, the power output taking into account the operating status of the new energy power generation system is obtained The expression is:

[0184] (15),

[0185] in:

[0186] (16),

[0187] In the formula, is the state function of the new energy unit; They are the three status sequences of normal operation, shortage and shutdown of new energy units.

[0188] In S1, a load uncertainty model considering load forecasting error is established, specifically:

[0189] The load forecast error is usually considered to obey a normal distribution with a mean of zero. The relationship between the load forecast error and the forecast value is established as follows:

[0190] (17),

[0191] (18),

[0192] In the formula, and They are The actual and predicted values ​​of system load during the time period; is the load forecast error; represents the standard deviation of load forecast error; is the proportionality factor that characterizes the accuracy of the prediction tool.

[0193] An energy storage charging and discharging model considering the energy storage life is established. In the energy storage system (ESS) charging and discharging model, the energy storage charging process is:

[0194] (19),

[0195] The discharge process is:

[0196] (20),

[0197] In the formula, and They are The power and stored energy of ESS at any moment; and They are The maximum charging and discharging power of ESS at the moment; and are the upper and lower limits of ESS storage energy respectively.

[0198] Charging or discharging the energy storage system will cause losses. Taking the depth of discharge as the main consideration, the number of battery cycles is inversely proportional to the depth of discharge and will decrease as the depth of discharge increases. The battery life model is:

[0199] (twenty one),

[0200] In the formula, is the actual discharge depth of the battery, is the rated discharge depth; , are the number of cycles corresponding to the rated discharge depth and the actual discharge depth respectively; according to the equivalent cycle life method, the actual number of cycles is converted to the rated number of cycles as follows:

[0201] (twenty two),

[0202] In the formula, is the equivalent number of cycles after conversion; the estimated result of the battery service life is:

[0203] (twenty three),

[0204] In the formula, is the estimated result of the battery service life, and N is the number of research periods of the battery in one year.

[0205] S2 uses the CNN-BILSTM-MATT hybrid model optimized by the RIME algorithm to predict wind, solar and load. The CNN model usually consists of multiple convolutional layers, activation functions, pooling layers and fully connected layers, and is trained using the back propagation algorithm. During the training process, CNN minimizes the loss function by adjusting the weight parameters so that the model can better predict the target.

[0206] Convolution calculation method:

[0207] (twenty four),

[0208] In the formula, is the convolution operation; The sequence calculated for the convolution; is the weight of the convolution kernel; Indicates the offset; is the activation function.

[0209] BILSTM processes sequence data by running two LSTM models simultaneously in the time dimension, one from front to back (forward) and one from back to front (reverse). By splicing or merging information from both directions, BILSTM can understand and represent sequence data more comprehensively.

[0210] The LSTM model mainly introduces the "gate" mechanism, which includes three gates: forget gate, input gate, and output gate. Through the combination of these gates, LSTM can effectively store and retrieve information in the sequence, thereby better capturing long-term dependencies; the relationship between its current state and the state at the previous moment is:

[0211] (25),

[0212] In the formula, is the sigmoid activation function; tanh is the hyperbolic tangent activation function; for Input of time; for The hidden state of the moment; , , , is the weight matrix of LSTM; , , , is the bias of LSTM; , , , For The forgetting threshold, input threshold, output threshold, and state matrix at the current moment.

[0213] For the BILSTM model:

[0214] (26),

[0215] In the formula, for The forward hidden layer state at that moment; for The hidden layer state at the moment backward; , are the weight matrix and bias term respectively; for Output at the moment.

[0216] In order to solve the problem of information loss due to long sequences during BILSTM model training, this paper introduces a multi-head attention mechanism as a supplement to BILSTM to more effectively mine the long-distance data features of time series with correlations. The vector output by BILSTM is converted into three dimensions through three different mapping operations. The input matrix is: query Q(Query), key K(Key) and value V(Value), and its attention output matrix is ​​shown in the formula:

[0217] (27)

[0218] In the formula, It is the feature dimension of each key, used for weight scaling, and is normalized to the interval [0, 1] by softmax.

[0219] The multi-head attention mechanism divides the time series into subspaces, each head performs self-attention calculation on the subspace to enhance its expressive power; then, The results of the individual heads are spliced ​​and integrated to obtain multiple heads. Each head is spliced ​​and the final result is obtained through linear transformation:

[0220] (28),

[0221] In the formula, , , Respectively , , The weight matrix of

[0222] (29),

[0223] In the formula, represents the weight of the linear transformation; represents the first Size; Represents a splicing operation; For the final output result, more feature information can be learned from different spaces.

[0224] RIME is used to optimize the key parameters of the CNN-BILSTM-MATT model, including the number of layers, number of neurons and learning rate of the neural network. The specific search strategy of the RIME algorithm is to simulate the frost formation process, find the optimal parameter combination through global and local search strategies, and improve the learning ability and prediction accuracy of the model. The optimization method is as follows: first initialize the frost population and parameters, then implement the soft frost search strategy, calculate the position of each soft frost particle, and then implement the hard frost puncture mechanism to exchange the particles of the algorithm, and absorb the new search position by the boundary to ensure that the search position does not exceed the search space described by the problem, implement the forward greedy selection mechanism, compare the updated fitness value with the previous one, and if the updated fitness value is better than the previous one, then a replacement occurs, thereby obtaining the global optimal solution. If the number of population iterations reaches the maximum or the fitness value is less than the error limit, the iteration is stopped and the optimal solution is output, otherwise the iteration continues; the calculation formula for soft frost particles is:

[0225] (30),

[0226] The parameters are expressed as:

[0227] (31),

[0228] (32),

[0229] (33),

[0230] In the formula, is the new position of the particle after the update; Indicates The first Particles; For the frost population The best frost agent in Particles; It is a random number in the range of (-1,1); is the angle between the frost movement direction and the horizontal direction; and Together, the direction of particle motion will change with the number of iterations; is another random coefficient, and Similarly, providing randomness to the search process; is the adhesion coefficient, which affects the condensation probability of the agent and increases with the number of iterations; and Respectively represent The upper and lower bounds of the dimension parameters are the constraints of the optimization problem; It is an environmental factor, which follows the number of iterations to simulate the influence of the external environment and is used to ensure the convergence of the algorithm; is the adhesion, a random number in the range of (0,1), used to control the distance between the centers of two frost particles; Indicates rounding; To control the number of segments of the step function, the default value is 5; is the current iteration number; is the maximum number of iterations of the algorithm.

[0231] Under strong wind conditions, the growth of hard frost is simpler and more regular, while the growth of soft frost is more random. Hard frost grows like a snowball in the same direction and is prone to crossover. Therefore, the hard frost puncture mechanism can improve the convergence of the algorithm and the ability to jump out of the local optimum. The exchange formula between particles in the hard frost puncture mechanism is:

[0232] (34),

[0233] In the formula, It is a random number in the range of (-1,1); Represents the normalized value of the current agent’s fitness value, indicating the The chance of a frost agent being selected; Representative A set or vector of solutions or states.

[0234] Figure 2 The figure shows the prediction flow chart of CNN-BILSTM-MATT. When the CNN-BILSTM-MATT model is used to predict wind power, photovoltaic power and load respectively, the CNN layer extracts the spatial features of wind, photovoltaic and load data respectively; the BILSTM layer trains multiple LSTM units from the forward and backward directions respectively. The BILSTM model combines the forward LSTM and the backward LSTM, and reads the forward and backward information features of the historical data through the hidden layer at the same time, explores the intrinsic connection between the current data and the past and future data, and integrates the multi-head attention mechanism to give different weights to each feature to enhance the important features, thus achieving accurate prediction of wind, photovoltaic and load.

[0235] Using root mean square error, mean absolute error, and coefficient of determination R 2The prediction model is evaluated in detail as the error evaluation index of the model, and based on the evaluation results, the RIME algorithm is used to adjust and optimize the CNN-BILSTM-MATT model to further improve the prediction accuracy of the prediction model.

[0236] (35),

[0237] (36),

[0238] (37),

[0239] In the formula, is the number of samples; For the predicted values; For the Actual value; is the average value of the actual value; R 2 The closer it is to 1, the better the prediction effect of the model.

[0240] Figure 3 The following is a comparison chart of wind power prediction using different models, where: Figure 3 (a) is the wind power prediction diagram using the CNN model. Figure 3 (b) is the wind power prediction diagram using the BILSTM model. Figure 3 (c) in the figure is the wind power prediction diagram using the CNN-BILSTM-MATT model. Figure 3 (d) in the figure is the wind power diagram predicted by RIME-CNN-BILSTM-MATT model; the error evaluation index of the model is shown in Table 1.

[0241] Table 1 Comparison of wind power prediction error evaluation indicators of different models

[0242]

[0243] From Table 1 and Figure 3 It can be seen that the RMSE and MAE of the error evaluation indicators of the proposed RIME-CNN-BILSTM-MATT model are smaller than those of other models. 2 Compared with other models, it is closer to 1, which shows that the prediction effect of RIME-CNN-BILSTM-MATT is better than other models.

[0244] In S3, the probability density function of the prediction error is calculated using the adaptive bandwidth kernel density estimation method, and the interval prediction of wind and solar power is performed by combining the wind and solar point prediction model and the probability statistics of the error, as follows:

[0245] The probability density prediction obtains the probability density function of the prediction object, and then obtains the cumulative distribution function of the prediction object on this basis to obtain the future probability density of the prediction object.

[0246] For a set of n independently distributed power prediction error data , and its corresponding probability density function is as follows:

[0247] (38),

[0248] In the formula, represents the probability density function; is the number of power prediction error samples; is the window bandwidth; Represents the kernel function used for kernel density estimation; the present invention selects a smooth, continuous, and obviously unimodal Gaussian kernel function as the kernel function for kernel density estimation, and the calculation formula of the Gaussian kernel function is:

[0249] (39),

[0250] The calculation formula of power prediction probability density function based on kernel density estimation is:

[0251] (40),

[0252] Among them, bandwidth Compared with the kernel function, it has a greater impact on the accuracy of kernel density estimation. In theory, the bandwidth should be infinitely close to zero, but if the bandwidth is too small, there will be too few points in the neighborhood involved in the fitting, affecting the distribution of the estimated model; if the bandwidth is too large, the distribution curve will be too smooth and the actual distribution of the prediction error cannot be accurately estimated; in order to quantitatively realize the optimal bandwidth calculation, the mean integrated square error is proposed, and the MISE expression is:

[0253] (41),

[0254] In the formula, is the forecast error estimate; is the actual value of the prediction error; To find the mean function.

[0255] The probability density calculated by the fixed bandwidth kernel density estimation may be quite different from the actual distribution, so it is hoped that the bandwidth can change with the change of data, and be smaller in places with dense data and larger in places with sparse data; therefore, the adaptive bandwidth kernel density estimation method can be obtained by correcting the relevant parameters of the fixed bandwidth kernel density estimation through a suitable bandwidth correction function, and the calculation formula is:

[0256] (42),

[0257] (43),

[0258] In the formula, is the error probability density function under adaptive bandwidth; is the adaptive revision coefficient of bandwidth; is the sensitivity parameter, ranging from 0 to 1; Indicates bandwidth is The kernel density estimation function is Represents the total number of prediction error samples.

[0259] Figure 4 The figure shows the comparison of prediction error distribution between adaptive and fixed bandwidth kernel density estimation methods. Figure 4 It is known that the adaptive bandwidth kernel density estimation can dynamically adjust the bandwidth according to the characteristics of the data, thereby providing more appropriate smoothness in different regions. Therefore, the adaptive bandwidth kernel density estimation curve is smoother than the fixed bandwidth curve, indicating that the adaptive bandwidth method can better capture the overall distribution trend of the data without being affected by local fluctuations.

[0260] In S4, Monte Carlo sampling is used to sample wind and solar power output within the wind and solar power prediction interval. The expression is:

[0261] (44),

[0262] In the formula, is the analog sampling power at that moment; and For the moment The lower and upper limits of the power range covered; Generate a random number (0-1) for uniform distribution; combine the sampled power with the wind and solar output status to obtain the actual power of the simulated wind and solar at this time, which is:

[0263] (45),

[0264] In the formula, To simulate the actual power of the wind and solar power generator sets.

[0265] In S5, the reliability index of the microgrid is calculated by combining the power load forecast data, conventional generators and energy storage status. Figure 5 The figure shows a microgrid reliability assessment flow chart, the specific steps of which are:

[0266] S5-1, data initialization, forming a load level of 8760h and determining the full time series simulation duration;

[0267] S5-2. Monte Carlo is used to sample the power within the prediction interval, and the simulated power considering the state of the wind and solar power generating units is obtained in combination with its operating status; if it is greater than the load demand, the energy storage is charged to store excess energy; if it is less than the load demand, the normal working duration TTF and fault repair time TTR of the gas turbine unit are obtained according to the Markov method, and they are arranged in sequence to form a sequence of the operating state duration of each gas turbine unit in the total simulation time, and the gas turbine outputs power to supply the remaining load; when the combined output of wind, solar and gas turbines is still less than the load demand, the energy storage is discharged; if the output of the three still cannot meet the load demand, the load is powered off, that is, the load is lost;

[0268] S5-3, determine whether the total simulation time is greater than the number of simulation years multiplied by 8760 hours; if so, execute step S5-4, otherwise jump to step S5-2;

[0269] S5-4. Calculate reliability indicators, where the reliability indicators include load loss probability LOLP and expected power shortage value EENS of the system.

[0270] The present invention selects indicators for evaluating the reliability of the entire system including loss of load probability (LOLP) and expected electric not supplied (EENS); wherein the LOLP calculation formula is:

[0271] (46),

[0272] (47),

[0273] In the formula, is the total simulation time; for The power demand of the load at any given moment; for The sum of wind power, photovoltaic, gas turbine and energy storage output power at the moment;

[0274] The EENS calculation formula is:

[0275] (48),

[0276] In the formula, is the set of states where load loss occurs; System Status probability; For the system status The amount of load loss under.

[0277] This embodiment selects different models to predict wind power under the condition that the photovoltaic output remains unchanged, and calculates the microgrid reliability index as shown in Table 2.

[0278] Table 2 Microgrid reliability indicators under wind power prediction by different models

[0279]

[0280] The data in Table 2 show that when using different models, the microgrid equipped with energy storage system is better than the system without energy storage in terms of LOLP and EENS indicators, indicating that the energy storage system plays an important role in improving the reliability of microgrids. And with the improvement of model accuracy, the LOLP and EENS indicators are getting closer and closer to the true value, indicating that the use of RIME-CNN-BILSTM-MATT-ABKDE model effectively improves the accuracy of wind and solar output prediction and realizes accurate prediction of microgrid reliability.

[0281] The above description is only a preferred embodiment of the present invention and is not intended to be a further limitation of the present invention. All equivalent changes made using the contents of the present specification and drawings are within the protection scope of the present invention.

Claims

1. A reliability assessment method for an islanded microgrid considering source-load uncertainty, characterized in that: The steps include: S1. Establish an uncertainty model for the islanded microgrid system, including using a fault-repair timing model to simulate the life process of the micro gas turbine unit, using a normal-deficit-outage timing model combined with wind power and photovoltaic prediction errors to simulate wind and solar uncertainty, establishing a load uncertainty model that takes into account load prediction errors, and establishing an energy storage charging and discharging model that takes into account the energy storage life; Among them, the normal-deficit-outage time series model is used to combine wind power and photovoltaic prediction errors to simulate wind and solar uncertainty, specifically: The Markov method is defined by the transition probability, the random variable from the current moment to the next moment, the state Transfer to state Probability for: （3）， In the formula, represents a discrete random variable at time t, is the state value of the random variable, i = 1, 2, ...; i is a certain moment, j is the next moment of i; random variable The value is The probability is: （4）， In the formula, is the state value of the random variable at time k, Represents a discrete random variable The value is probability; Assuming the number of states is n, then: （5）， State transfer matrix of wind turbines and photovoltaics It is expressed as follows: （6）， （7）， The duration of each output state is obtained according to the transfer rate between different outputs ,Right now: （8）； The new energy power generation system, i.e. wind power and photovoltaic power, is divided into three states: operation, shortage and shutdown; State 1 is normal operation r: in this state, there are no faulty components inside the new energy unit, and the new energy power generation system operates in a planned manner; State 2, i.e., shortage d: some constraints in the new energy system cannot meet the conditions. After adjustment, it operates at a certain proportion of the original plan; State 3 is shutdown: In this state, there will be a temporary shutdown due to the failure of internal components of the system; Assume that the initial state of the new energy unit is , the state of the new energy unit after unit time is expressed as: （9）， After m unit time, the state transfer of the new energy unit is expressed as: （10）； The initial state of the new energy unit and its state transfer matrix are known to obtain the steady-state value of its state distribution at each moment; the probability formulas of the three states of operation, shortage and shutdown are expressed as: （11）， In the formula, is the transition rate between normal operation, shortage and outage states, and its subscript represents the direction of state transition; is the probability of being in operation, shortage and outage status; The average duration of stays in normal operation, shortage and outage is: （12）， In the formula, The duration of normal operation, shortage and outage; The forecast error of renewable energy output is described by a normal distribution with a mean of zero. The relationship between the forecast error of renewable energy output and the forecast value is as follows: （13）， （14）， In the formula, and They are The actual and predicted values ​​of renewable energy output during the period; It is the prediction error of new energy output; represents the standard deviation of the PV output prediction error; Finally, the power output taking into account the operating status of the new energy power generation system is obtained The expression is: （15）， in: （16）， In the formula, is the state function of the new energy unit; They are three status sequences: normal operation, shortage and shutdown of new energy units; S2, using the CNN-BILSTM-MATT hybrid model composed of a convolutional neural network (CNN), a bidirectional long short-term memory network (BILSTM), and a multi-head attention mechanism (MATT) optimized by the frost optimization algorithm RIME to predict wind, solar, and load; S3. Use adaptive bandwidth kernel density estimation ABKDE to calculate the probability density function of the prediction error, and combine the CNN-BILSTM-MATT hybrid model and the probability statistics of the error to perform interval prediction of wind and solar power; S4. Monte Carlo sampling is used to sample wind and solar power output within the wind and solar power prediction interval, and combined with the wind and solar power normal-deficit-outage timing model, the actual power of the simulated wind and solar power at this time is obtained; S5. According to the uncertainty model of the microgrid established in S1, the operating status of the micro gas turbine unit and the energy storage is obtained. Combined with the power load forecast data and the actual power of the simulated wind and solar power, the reliability index load loss probability and the expected value of power shortage are calculated. The reliability of the microgrid is evaluated through the calculated reliability index.

2. The reliability assessment method of an islanded microgrid considering source-load uncertainty according to claim 1 is characterized in that: In S1, the fault-repair time series model is used to simulate the life process of the micro gas turbine unit, specifically: The power generation model of the micro gas turbine unit is represented by a two-state Markov model, and the failure rate and repair rate of the generator are expressed as and Micro gas turbine Trouble-free working time and troubleshooting time They are: （1）， （2）， In the formula, and is a random variable uniformly distributed on the interval [0, 1]; is the failure rate of micro gas turbine i, is the repair rate of micro gas turbine i.

3. The reliability assessment method of an islanded microgrid considering source-load uncertainty according to claim 1 is characterized in that: In S1, a load uncertainty model considering load forecasting error is established, specifically: The load forecast error follows a normal distribution with a zero mean, and the relationship between the load forecast error and the forecast value is established as follows: （17）， （18）， In the formula, and They are The actual and predicted values ​​of system load during the time period; is the load forecast error; represents the standard deviation of load forecast error; is the proportionality factor that characterizes the accuracy of the prediction tool.

4. The reliability assessment method of an islanded microgrid considering source-load uncertainty according to claim 1 is characterized in that: In S1, an energy storage charging and discharging model considering the energy storage life is established. In the energy storage charging and discharging model, the energy storage charging process is: （19）， The discharge process is: （20）， In the formula, and They are The power and stored energy of the ESS at all times; and They are Maximum charging and discharging power of energy storage ESS at all times; and are the upper and lower limits of energy storage of the energy storage ESS respectively; The battery life model is: （21）， In the formula, is the actual discharge depth of the battery, is the rated depth of discharge; , are the number of cycles corresponding to the rated discharge depth and the actual discharge depth respectively; According to the equivalent cycle life method, the actual number of cycles is converted to the rated number of cycles as follows: （22）， In the formula, is the equivalent number of cycles after conversion; The estimated result of battery life is: （23）， In the formula, is the estimated result of the battery service life, and N is the number of research periods of the battery in one year.

5. The reliability assessment method of an islanded microgrid considering source-load uncertainty according to claim 1, characterized in that: In S2, the CNN-BILSTM-MATT hybrid model is constructed as follows: The CNN model consists of multiple convolutional layers, activation functions, pooling layers, and fully connected layers, and is trained using the back-propagation algorithm. During the training process, CNN minimizes the loss function by adjusting weight parameters. Convolution calculation method: （24）， In the formula, is the convolution operation; The sequence calculated for the convolution; is the weight of the convolution kernel; Indicates the offset; is the activation function; BILSTM processes sequence data by running two LSTM models simultaneously in the time dimension, one from front to back (forward) and the other from back to front (reverse); The LSTM model includes a forget gate, an input gate, and an output gate. The relationship between the current state of the LSTM model and the state at the previous moment is: （25）， In the formula, is the sigmoid activation function; tanh is the hyperbolic tangent activation function; for Input of time; for The hidden state of the moment; , , , is the weight matrix of LSTM; , , , is the bias of LSTM; , , , For The forgetting threshold, input threshold, output threshold, and state matrix at the moment; For the BILSTM model: （26）， In the formula, for The forward hidden layer state at that moment; for The hidden layer state at the moment backward; , are the weight matrix and bias term respectively; for Output at the moment; During the BILSTM model training process, the multi-head attention mechanism is introduced as a supplement to BILSTM; the vector output by BILSTM is converted into three dimensions through three different mapping operations. The input matrix of is: query Q, key K and value V, and its attention output matrix is: （27）， In the formula, is the feature dimension of each key, used for weight scaling, and normalized to the interval [0, 1] by softmax; The multi-head attention mechanism divides the time series into subspaces, each head performs self-attention calculation on the subspace; then, The results of the individual heads are spliced ​​and integrated to obtain multiple heads. Each head is spliced ​​and the final result is obtained through linear transformation: （28）， In the formula, , , Respectively , , The weight matrix of （29）， In the formula, represents the weight of the linear transformation; represents the first Size; Represents a splicing operation; For the final output result, more feature information can be learned from different spaces.

6. A reliability assessment method for an islanded microgrid considering source-load uncertainty according to claim 5, characterized in that: The RIME algorithm is used to optimize the CNN-BILSTM-MATT hybrid model, specifically: S2-1, initialize the frost population and parameters, including the number of layers, number of neurons and learning rate of the neural network; S2-2, implement the soft frost ice search strategy to calculate the location of each soft frost ice particle; The calculation formula of soft frost ice particles is: （30）， The parameters are expressed as: （31）， （32）， （33）， In the formula, is the new position of the particle after the update; Indicates The first Particles; For the frost population The best frost agent in Particles; It is a random number in the range of (-1,1); is the angle between the frost movement direction and the horizontal direction; and Together, the direction of particle motion will change with the number of iterations; is another random coefficient, providing randomness for the search process; is the adhesion coefficient; and Respectively represent Upper and lower bounds on dimensional parameters; For environmental factors; is the adhesion; Indicates rounding; To control the number of segments of the step function; is the current iteration number; is the maximum number of iterations of the algorithm; S2-3. Implement the hard frost puncture mechanism to exchange the frost particles of the algorithm and perform boundary absorption on the new search position to ensure that the search position does not exceed the search space described in the problem; The exchange formula between particles in the hard frost ice penetration mechanism is: （34）， In the formula, It is a random number in the range of (-1,1); Represents the normalized value of the current agent’s fitness value, indicating the The chance of a frost agent being selected; Representative A set or vector of solutions or states; S2-4. Implement a positive greedy selection mechanism to compare the fitness value of the updated frost particle position with that before the update, where the fitness value is an indicator to measure the quality of the solution; if the updated fitness value is better than the value before the update, a replacement occurs, thereby obtaining the optimal position of the frost particle, that is, the global optimal solution; if the number of population iterations reaches the maximum or the fitness value is less than the error limit, stop the iteration and output the optimal solution, otherwise continue the iteration.

7. The reliability assessment method of an islanded microgrid considering source-load uncertainty according to claim 1 is characterized in that: S3 is specifically: For a set of n independently distributed power prediction error data , and its corresponding probability density function is as follows: （35）， In the formula, represents the probability density function; is the number of power prediction error samples; is the window bandwidth; Represents the kernel function used for kernel density estimation; is the independent variable of the probability density function; The calculation formula of power prediction probability density function based on kernel density estimation is: （36）， In order to quantitatively realize the optimal bandwidth calculation, the mean integrated square error MISE indicator is used. The MISE expression is: （37）， In the formula, is the forecast error estimate; is the actual value of the prediction error; To find the mean function; The bandwidth correction function is used to correct the relevant parameters of the fixed bandwidth kernel density estimation to obtain the adaptive bandwidth kernel density estimation method. The calculation formula is: （38）， （39）， In the formula, is the error probability density function under adaptive bandwidth; is the adaptive revision coefficient of bandwidth; is the sensitivity parameter, ranging from 0 to 1; Indicates bandwidth is The kernel density estimation function is Represents the total number of prediction error samples.

8. The reliability assessment method of an islanded microgrid considering source-load uncertainty according to claim 1, characterized in that: In S4, Monte Carlo sampling is used to sample wind and solar power output within the wind and solar power prediction interval. The expression is: （40）， In the formula, is the analog sampling power at that moment; and For the moment The lower and upper limits of the power range covered; Generate random numbers (0-1) for uniform distribution; The sampled power is combined with the wind and solar power output status to obtain the actual power of the simulated wind and solar power at this time, and the formula is: （41）， In the formula, To simulate the actual power of the wind and solar power generator sets.

9. The reliability assessment method of an islanded microgrid considering source-load uncertainty according to claim 1, characterized in that: In S5, the reliability index of the microgrid is calculated by combining the power load forecast data, conventional generators and energy storage status, which is: S5-1, data initialization, forming a load level of 8760h and determining the full time series simulation duration; S5-2. Use Monte Carlo to sample the power within the prediction interval, and combine its operating status to obtain the simulated power considering the status of the wind and solar power generator set; If the simulated power is greater than the load demand, the energy storage is charged to store the excess energy; If it is less than the load demand, the normal working duration TTF and fault repair time TTR of the gas turbine unit are obtained according to the Markov method, and they are arranged in sequence to form the operating state duration sequence of each gas turbine unit in the total simulation time, and the gas turbine outputs power to supply the remaining load; When the combined output of wind, solar and gas turbines is still less than the load demand, the energy storage is discharged; If the output of wind, solar, gas turbines and energy storage still cannot meet the load demand, the load will be cut off, that is, the load will be lost; S5-3, determine whether the total simulation time is greater than the number of simulation years multiplied by 8760 hours; if so, execute S5-4, otherwise jump to S5-2; S5-4, calculating reliability indicators, wherein the reliability indicators include load loss probability LOLP and expected power shortage value EENS of the system; The LOLP calculation formula is: （42）， （43）， In the formula, is the total simulation time; for The power demand of the load at any moment; for The sum of wind power, photovoltaic, gas turbine and energy storage output power at the moment; The EENS calculation formula is: （44）， In the formula, is the set of states where load loss occurs; System Status probability; For the system status The amount of load loss under.

Images