CN-122026517-A - Power system scheduling method for time sequence diffusion and analytic type functional synapse

Abstract

The invention provides a power system scheduling method for time sequence diffusion and analytic type functional synapse, which can solve the problems of scarce load data, low solving precision of non-convex nonlinear scheduling problems, unstable training of a pulse neural network, discrete output and difficult compromise of low energy consumption and environmental protection of a scheduling scheme in extreme weather. According to the method, an extreme weather load sample is generated through a time sequence diffusion model, scarce data is supplemented, a scheduling model with energy consumption and carbon emission as targets is constructed, the output of a preliminary unit is solved through a prediction-correction original-dual interior point method, the output is optimized through an analytic type functional synaptic enhanced pulse neural network, and the system constraint is ensured to be met through corrective measures. The method provided by the invention can improve the dispatching reliability under extreme working conditions, improve the non-convex problem solving precision, solve the defect of the impulse neural network, and is suitable for dispatching of the electric power system containing clean energy, and low energy consumption and environmental protection are considered.

Inventors

• YIN LINFEI
• GONG YANG
• LU QUAN

Assignees

• 广西大学

Dates

Publication Date

20260512

Application Date

20260203

Claims (1)

1. 1. A power system scheduling method of time sequence diffusion and analytic type functional synapse is characterized in that the method comprises the following steps in the using process: Collecting basic data, wherein the basic data comprises the requirements of each load node for 24 hours, unit parameters and extreme weather historical load data, and the unit parameters comprise the energy consumption coefficient and the carbon emission coefficient of a thermal power unit, the unit power generation energy consumption coefficient of a wind power unit, a photovoltaic power unit and a hydroelectric unit, the upper and lower output limits of various units and the climbing limit value of the thermal power unit; Time period of Load demand of the system The method comprises the following steps: (1) In the formula, Represent the first A plurality of nodes; the total number of the system nodes; Representing a summation symbol; Represent the first The individual node is at the first Load demand for the time period; representing load demand of economic dispatch model of system in time of day by matrix The method comprises the following steps: (2) In the formula, Representing the total number of time periods in a day; representing the load demand amount of the 1 st period; Representing the load demand amount of the 2 nd period; Represent the first Load demand for each time period; Represent the first Load demand for each time period; generating load samples in extreme weather through a time sequence diffusion model, and obtaining Gaussian distribution variables through gradually adding Gaussian noise into original load demands through a forward diffusion process, wherein the Gaussian distribution variables are expressed as a Markov chain: (3) In the formula, Representing a conditional probability distribution during forward diffusion; representing a gaussian distribution; Representing a current time step; representing the total number of time steps; representing the original load demand; is shown in the first Load requirements after adding noise variables in time steps; is shown in the first Load requirements after adding noise variables in time steps; Representing noise scheduling parameters in the diffusion process; representing the identity matrix; representing an open square root symbol; representing a product symbol; noise scheduling parameters Expressed in cosine scheduling as: (4) In the formula, Representing an offset parameter; representing the circumference ratio; Representing a cosine function; calculation of arbitrary time steps using a re-parameterization technique A kind of electronic device : (5) In the formula, Representing from step 1 to step 1 Is a noise scheduling parameter accumulated value; Is a gaussian noise matrix; the reverse process returns to the original load demand by gradually removing noise, expressed as: (6) In the formula, Representing a gaussian distribution; A conditional probability distribution representing a reverse generation process; Representing the denoising model in the first place Time step pair noisy data A predicted noise mean; representing the variance of the inverse process; parameters representing the denoising model; the denoising model comprises an input embedding module and a time decomposition reconstruction U-Net module, wherein the input embedding module uses a 1-D convolution layer to reverse intermediate results in the condition inversion process Embedding to generate embedded features The method comprises the following steps: (7) In the formula, Representing a 1D convolution operation; is shown in the first Load requirements after adding noise variables in time steps; discrete time step pair using sinusoidal position coding method Encoding and generating position embedding The method comprises the following steps: (8) In the formula, Representing a sinusoidal position coding method; embedding the generated location Further processing by two fully connected layers and an activation function to generate higher-dimensional continuous features The method comprises the following steps: (9) In the formula, Representation of Activating a function; Representing a fully connected layer; The time decomposition reconstruction U-Net module consists of 4 encoders and 4 decoders, wherein each encoder and each decoder consist of two convolution blocks and a self-attention mechanism, a 1D convolution layer is used in each convolution block for carrying out convolution operation on input data, time step coding information is processed by summation pairs through layer normalization transformation and is activated through Swish functions with noise coding signals passing through full connection layers, residual connection is introduced between the input layers and the output layers of the decoder, the input signals are directly added into the output of the encoder and are carried out convolution operation in the encoder, the convolution layers in the encoder and the decoder comprise jump connection, the jump connection and the summation layer fuse the characteristics of the corresponding layers in the encoder and the characteristics in the decoder together and input the characteristics in the decoder into the self-attention layer, and the prediction noise is obtained through the output of the two full connection layers; The self-attention mechanism captures long-distance dependency relationship on time sequence, input noise is input into each encoder and decoder through the full connection layer, so that the input noise is fused at different positions and denoising is performed at different time steps to different extents, and the calculation formula of the self-attention mechanism is as follows: (10) In the formula, As a self-attention mechanism function; is an input matrix; Is that Activating a function; is a query matrix; is a key matrix; is a value matrix; Is a transpose of the key matrix; Is the dimension of the key vector; Representing a key vector corresponding to the dimension; the noise estimation loss function is: (11) In the formula, Is shown in the first Time-step generation of raw data Is a desired operation of (1); Represents the L2 norm square; representing the noise predicted by the denoising model, and generating by reverse sampling after training Load sample matrix in extreme weather Wherein, the method comprises the steps of, Representation generation of the first A group load sample matrix; Representing the total number of load samples generated by the back sampling; Step (iii) of constructing a power system dispatch mathematical model Electric power system The bench set is at the first Total output power of time period The method comprises the following steps: (12) In the formula, Represent the first The bench set is at the first Output power of the time period; Representing the total number of units of the power system; Represent the first The power generating units are arranged; In the first place Total output power of time period power system unit Meeting the power balance constraint is: (13) In the formula, Is shown in the first Total output power of the time period; Represent the first Load demand of time period; First, the The bench set is at the first Output power of time period Meeting climbing constraint is: (14) In the formula, Represent the first A lower climbing limit value of the bench unit; Represent the first The climbing limit value of the bench unit; Represent the first The bench set is at the first Output power of the time period; Represent the first The bench set is at the first Output power of the time period; First, the The bench set is at the first Output power of time period Meeting the upper and lower limit constraints is: (15) In the formula, Represent the first The lower limit of the output power of the station unit; Represent the first The upper limit of the output power of the station unit; Scheduling fitness function value with minimum carbon emission and primary energy loss of power generation in time period The method comprises the following steps: (16) In the formula, Represent the first A time period; Represent the first A station set; is the first Scheduling quadratic term energy consumption coefficient of the station unit; is the first Scheduling quadratic term carbon emission coefficient of the station unit; is the first A scheduling primary energy consumption coefficient of the station unit; is the first A primary carbon emission coefficient of the station unit is scheduled; is the first The energy consumption coefficient of a scheduling constant item of the station unit; is the first A scheduling constant term carbon emission coefficient of the station unit; Represent the first The bench set is at the first Output power of the time period; Sum symbols; Step (iv), obtaining a search direction through linearizing Karush-Kuhn-Tucker system, and synchronously updating an original variable and a dual variable; the quadratic programming problem is in the standard form: (17) In the formula, Is an objective function in which A positive definite matrix formed by the quadratic term coefficients of all the units; Vector matrixes formed by primary term coefficients of all units; Is original as The row vectors are corresponding to the units in the first place Output of time period The constraint type is a linear inequality constraint Constraint of linear equation Upper and lower limit constraints of sum components Wherein, the method comprises the steps of, A coefficient matrix representing inequality constraints; A constant term matrix representing inequality constraints; A coefficient matrix representing the constraint of the equation; a constant term matrix representing an equality constraint; representing the minimum output of each unit; Representing the maximum output of each unit; The original-dual Karush-Kuhn-Tucker condition introduces a standard form of relaxation variable: (18) In the formula, A transpose of the coefficient matrix representing the inequality constraint; A transpose of the coefficient matrix representing the equality constraint; To relax the variables so that ; As a dual variable of the inequality constraint, ; Constraining the dual variables for the equation; representing constructing a diagonal matrix function; representing a diagonal matrix of relaxation variables; representing an inequality constraint dual variable diagonal matrix; representing an all 1 vector; definition of dual residuals Equation constrained original residual Inequality constraint original residual And complementary residual The method comprises the following steps: (19) In the formula, Is a dual residual error; And Original residuals corresponding to the equality constraint and the inequality constraint, respectively; Is the complementary residual; Definition of dual gap The method comprises the following steps: (20) In the formula, Is the number of inequalities; representing a transpose of the relaxation variable matrix; For a pair of Performing first-order Taylor expansion to obtain an incremental equation set as follows: (21) In the formula, 、 、 And Respectively obtained by solving The initial incremental solution of each unit output of the dimension vector, the initial incremental solution of the inequality constraint dual variable, the initial incremental solution of the equality constraint dual variable and the initial incremental solution of the relaxation variable are obtained by solving, and the initial incremental solutions jointly represent the affine direction of the next step; calculating affine original maximum feasible step length according to an initial increment solution obtained by solving an increment equation set Affine dual maximum feasible step size And dual gap in affine case : (22) In the formula, To take the minimum function; Using complementary residuals in the dual gap correction equation (20) in affine case Corrected complementary residual The method comprises the following steps: (23) (24) In the formula, Is the corrected complementary residual error; Representing an obstacle parameter; solving the corrected increment equation set to obtain the increment solution of the final direction Initial incremental solution of output of each unit of dimension vector Initial delta solution for inequality constraint dual variables Initial delta solution for equality constraint dual variables And initial delta solution of relaxation variables Calculating the actual step length in the final direction The method comprises the following steps: (25) In the formula, The maximum feasible step length is the final original; the maximum feasible step length of the final dual; is a scaling factor; Calculating the final unit in the first unit according to the actual step length Output at moment of time Dual variable constrained by inequality Dual variable of equality constraint Relaxation variable The method comprises the following steps: (26) In the formula, 、 、 And Each unit solved after updating iteration is respectively in the first Corresponding to the moment of force Dimension row vectors, inequality constrained dual variables, equality constrained dual variables, and relaxation variables; capturing time sequence characteristics by adopting a pulse neural network, replacing synaptic weights by using spline functions in a Kolmogorov-Arnold network to solve the problem of non-leadership of the pulse neural network, improving the expression force and training stability of a model, and outputting the output force of an optimal unit; The optimized solution vector and the load sample are normalized and then input into an enhanced pulse neural network, the network model comprises a pulse coding layer, a synaptic layer and a leakage integration release neuron layer, the pulse decoding layer is introduced into a Kolmogorov-Arnold network to replace synaptic weight The method comprises the following steps: (27) In the formula, Is a spline function; is a learnable spline function coefficient; The number of spline functions; is an input signal; at a time step When the load value is normalized, the discrete pulse input is obtained through the pulse coding layer Neurons Receiving neurons from a previous layer Pulse input of (2) After that, neurons The output burst current is expressed as: (28) In the formula, Is a neuron At the moment of time Is used to control the current of the current source, indicating the total inrush current Is the first of (2) A component; to be replaced synaptic weights, neurons are represented To neurons Is used for the connection strength of the steel wire; is a neuron Is a pulse output of (a); membrane potential under continuous time in enhanced pulsed neural networks incorporating Kolmogorov-Arnold network synapses The mathematical model of the leakage integral firing neuron is described as: (29) In the formula, At the moment of time for neurons Is a membrane potential of (a); as a function of the time constant of the film, , In the form of a film resistor, Is a film capacitor; Is at rest potential; Is the output current of the total synapse; rule of pulse delivery when membrane potential Reaching a threshold value When the pulse is generated, the membrane potential is reset to the resting potential ; Training the enhanced impulse neural network through a time back propagation method, and outputting an optimal unit output matrix The method comprises the following steps: (30) In the formula, The optimal output of the 1 st machine set in the 1 st period is represented; the optimal output force of the 1 st unit in the 2 nd period is represented; Indicating that the 1 st machine set is at the 1 st machine set Optimal output of the time period; The optimal output of the 1 st machine set in the 24 th period is represented; The optimal output of the 2 nd machine set in the 1 st period is represented; The optimal output force of the 2 nd machine set in the 2 nd period is represented; Indicating that the 2 nd machine set is at the 2 nd machine set Optimal output of the time period; the optimal output of the 2 nd machine set in the 24 th time period is represented; Represent the first The optimal output of the bench unit in the 1 st period; Represent the first The optimal output force of the bench unit in the 2 nd period; Represent the first The bench set is at the first Optimal output of the time period; Represent the first The optimal output of the bench set in 24 th period; Represent the first The optimal output of the bench unit in the 1 st period; Represent the first The optimal output force of the bench unit in the 2 nd period; Represent the first The bench set is at the first Optimal output of the time period; Represent the first The optimal output of the bench set in 24 th period; step (vi) period of time Deviation of total optimized output from load demand The method comprises the following steps: (31) In the formula, Denoted as the first Power balance deviation of the time period; Represent the first Total optimized output for the time period; Represent the first Load demand for the time period; First, the The proportion of the output of the machine set to the total output of the system The method comprises the following steps: (32) In the formula, Represent the first The output of the bench unit; representing the total output of all units of the system; According to the deviation direction, the unit output duty ratio and the upper and lower limit constraints of unit output, the actual output of each unit is carried out And (3) correcting: (33) In the formula, Is a maximum function; To take the minimum function; Represent the first Minimum output of the bench unit; Represent the first Maximum output of the bench unit; Representing an output matrix of an optimal unit output by the enhanced pulse neural network; Represent the first The output of the machine set accounts for the proportion of the total output of the system; the representation is shown as the first Power balance deviation of the time period; finally, each unit output matrix The method comprises the following steps: (34) In the formula, The optimal output of the 1 st machine set in the 1 st period after correction is shown; the optimal output force of the 1 st machine set in the 2 nd period after correction is shown; Indicating that the 1 st machine set is at the 1 st machine set after correction Optimal output of the time period; the optimal output of the 1 st machine set in the 24 th time period after correction is shown; the optimal output of the 2 nd machine set in the 1 st period after correction is shown; The optimal output force of the 2 nd machine set in the 2 nd period after correction is shown; Indicating that the 2 nd machine set is at the 2 nd machine set after correction Optimal output of the time period; the optimal output of the 2 nd machine set after correction in the 24 th time period is shown; indicating post-correction item The optimal output of the bench unit in the 1 st period; indicating post-correction item The optimal output force of the bench unit in the 2 nd period; indicating post-correction item The bench set is at the first Optimal output of the time period; indicating post-correction item The optimal output of the bench set in 24 th period; indicating post-correction item The optimal output of the bench unit in the 1 st period; indicating post-correction item The optimal output force of the bench unit in the 2 nd period; indicating post-correction item The bench set is at the first Optimal output of the time period; indicating post-correction item The optimal output of the bench set in 24 th period.

Description

Power system scheduling method for time sequence diffusion and analytic type functional synapse Technical Field The invention belongs to the fields of power system dispatching, new energy utilization and intelligent optimization algorithms, and relates to an economic dispatching method which is suitable for dispatching clean energy power systems containing thermal power, wind power, photovoltaic power and hydropower. Background The existing power system dispatching is based on a traditional heuristic method or a single analytic method, when extreme weather occurs, the problem that the dispatching robustness is poor due to the fact that load data are scarce is solved, the load distribution deviation and uncertainty are difficult to deal with, and the situation that optimization convergence is difficult or a dispatching scheme is not feasible is caused. In addition, the existing data enhancement method is difficult to capture nonlinear characteristics of extreme loads, the generated samples deviate from actual samples and cannot effectively supplement load data in extreme weather, gradient disappearance problems occur in the existing generation model training process, accuracy and diversity of generated samples are insufficient, reliable scheduling decisions are difficult to support, the existing single-target optimization method is low in precision and slow in convergence speed when solving the nonlinear problems of non-convexity, the problem of carbon emission is not considered, the defects of gradient disappearance and discrete output exist in the existing impulse neural network training, the method is difficult to directly apply to scheduling decisions, and an effective optimization mechanism is lacked to promote stability and precision of the impulse neural network training. Therefore, the power system scheduling method for time sequence diffusion and analytic type functional synapse solves the problems that in extreme weather, load data is insufficient, data enhancement and generation model effect is limited, non-convex problem solving precision is low, scheduling scheme is difficult to consider low energy consumption and environmental protection, pulse neural network training is unstable and output is discrete. Disclosure of Invention The invention provides a power system scheduling method of time sequence diffusion and analytic type functional synapse, which comprises the following steps in the using process: Collecting basic data, wherein the basic data comprises the requirements of each load node for 24 hours, unit parameters and extreme weather historical load data, and the unit parameters comprise the energy consumption coefficient and the carbon emission coefficient of a thermal power unit, the unit power generation energy consumption coefficient of a wind power unit, a photovoltaic power unit and a hydroelectric unit, the upper and lower output limits of various units and the climbing limit value of the thermal power unit; Time period of Load demand of the systemThe method comprises the following steps: (1) In the formula, Represent the firstA plurality of nodes; the total number of the system nodes; Representing a summation symbol; Represent the first The individual node is at the firstLoad demand for the time period; representing load demand of economic dispatch model of system in time of day by matrix The method comprises the following steps: (2) In the formula, Representing the total number of time periods in a day; representing the load demand amount of the 1 st period; Representing the load demand amount of the 2 nd period; Represent the first Load demand for each time period; Represent the first Load demand for each time period; generating load samples in extreme weather through a time sequence diffusion model, and obtaining Gaussian distribution variables through gradually adding Gaussian noise into original load demands through a forward diffusion process, wherein the Gaussian distribution variables are expressed as a Markov chain: (3) In the formula, Representing a conditional probability distribution during forward diffusion; representing a gaussian distribution; Representing a current time step; representing the total number of time steps; representing the original load demand; is shown in the first Load requirements after adding noise variables in time steps; is shown in the first Load requirements after adding noise variables in time steps; Representing noise scheduling parameters in the diffusion process; representing the identity matrix; representing an open square root symbol; representing a product symbol; noise scheduling parameters Expressed in cosine scheduling as: (4) In the formula, Representing an offset parameter; representing the circumference ratio; Representing a cosine function; calculation of arbitrary time steps using a re-parameterization technique A kind of electronic device: (5) In the formula,Representing from step 1 to step 1Is a noise scheduling pa