CN-121787031-B - Distributed energy system closed-loop scene generation method and system

CN121787031BCN 121787031 BCN121787031 BCN 121787031BCN-121787031-B

Abstract

The invention belongs to the technical field of energy systems, and provides a method and a system for generating a closed-loop scene of a distributed energy system, wherein the method and the system take the comprehensive operation cost of the minimized distributed energy system as an objective function, introduce operation constraint and energy balance constraint of equipment, and construct an operation optimization model; the method comprises the steps of constructing a generating model, utilizing an operation optimizing model, taking a historical real data sequence as deterministic input, carrying out optimal power flow calculation to obtain optimal system scheduling cost under a historical real condition as a datum reference value, utilizing the generating model to generate an initial simulation scene, utilizing the initial simulation scene as input, utilizing the operation optimizing model to solve, calculating the optimal system scheduling cost corresponding to each simulation scene, carrying out iteration optimization until training conditions are met, and utilizing the finally optimized generating model to generate a typical scene set. The invention realizes the aim of improving the economic benefit and the operation reliability of the random optimal scheduling of the distributed energy system.

Inventors

LI KE
CHEN SHUAIBING
Mou Yuchen
WANG HAIYANG

Assignees

山东大学

Dates

Publication Date: 20260512
Application Date: 20260305

Claims (9)

1. A method for generating a closed-loop scene of a distributed energy system is characterized by comprising the following steps: Acquiring a historical real data sequence of a distributed energy system, taking the comprehensive operation cost of the minimized distributed energy system as an objective function, introducing operation constraint and energy balance constraint of equipment, and constructing an operation optimization model of the distributed energy system; constructing a generating model, randomly initializing network parameters of the generating model, and carrying out optimal power flow calculation by taking the historical real data sequence as deterministic input by utilizing the operation optimizing model to obtain a datum reference value; Generating initial simulation scenes by using the generation model, solving by using the operation optimization model, and calculating the optimal system scheduling cost corresponding to each simulation scene; Constructing a composite loss function, wherein the composite loss function comprises a cost-based loss term for measuring the difference between the optimal system scheduling cost and the base reference value and a regularization term for restricting the generation of scene statistical characteristics; Generating model parameters based on the feedback optimization scene, calculating the gradient of the composite loss function relative to the network parameters of the generated model by adopting an optimization algorithm, and updating the network parameters of the generated model based on the back propagation of the gradient information; Iterative updating and optimizing the network parameters of the generated model until the training requirement is met, and obtaining the finally optimized generated model; generating a typical scene set for random optimization scheduling of the distributed energy system by utilizing the final optimized generation model; The process for constructing the generation model comprises the steps of constructing the generation model of the double-layer framework, generating scenes meeting requirements by an upper layer through generating a minimum-maximum game of a generator and a discriminator of an countermeasure network, and carrying out daily and real-time two-stage scheduling optimization on each generation scene by a lower layer, calculating the deviation between the comprehensive cost and the ideal cost of the scenes and feeding the deviation back to the upper layer, and guiding the generator to optimize the generation direction of the scenes.
2. The method for generating the closed-loop scene of the distributed energy system according to claim 1, wherein the process of constructing the operation optimization model of the distributed energy system by taking the minimized integrated operation cost of the distributed energy system as an objective function and introducing the operation constraint and the energy balance constraint of the equipment comprises the steps of taking the minimized integrated operation cost of the distributed energy system as the objective function, wherein the integrated operation cost of the distributed energy system comprises the operation cost, the abandoned wind and abandoned light punishment cost, the energy storage cost, the carbon transaction cost, the demand response cost and the abandoned load punishment cost; The operation constraint of the equipment comprises the efficiency constraint of the energy conversion equipment, the efficiency constraint of the energy storage equipment and the maximum transmission power constraint of interaction with a power grid and an air network; the energy balance constraints include electrical load, thermal load, and cold load power balance constraints.
3. The method for generating the closed-loop scene of the distributed energy system according to claim 1, wherein the process of generating the scene meeting the requirements by the upper layer through generating the minimum and maximum games of a generator and a discriminator of an countermeasure network comprises the steps of taking historical data as a training set x, taking random noise z as input by a generator G, and simulating to generate data G (z); On the basis of generating the countermeasure network, introducing a condition label, wherein the condition label comprises at least one of a source load type, a time period characteristic and a history deviation, and a discriminator D judges the similarity of generated data and real data distribution and judges whether the condition label is satisfied.
4. A method of generating a closed loop scene of a distributed energy system as claimed in claim 3, wherein generating a loss function against the network is: ; The loss function of the generator G is generated, As a loss function of the arbiter D, And The functions of the arbiter and the generator respectively, Generating a data distribution; for a true data distribution, For historical data, y is a conditional label, 、 Respectively representing expectations; the similarity between the real data and the historical data sample distribution is represented by introducing the Wasserstein distance, and the calculation method of the Wasserstein distance meets the following conditions: ; Representation of The method satisfies a 1-Lipschitz function, wherein the upper limit of the derivative expected value difference is 1, so that the network can normally perform gradient optimization, and the sup is a function for solving the minimum upper limit, and gradient penalty is introduced to replace original weight clipping, and is as follows: ; The objective function of the minimum and maximum game is: ; Wherein, the Representing the gradient penalty coefficients.
5. A distributed energy system closed-loop scene generation method according to claim 3, wherein the lower layer scheduling problem objective function is expressed as a standard strong convex quadratic format; Constraint conditions are kept unchanged, a dual variable is introduced to embed the constraint into an objective function, and a Lagrange function of a day-ahead optimization problem is as follows: ; wherein the matrix Vector(s) Coefficient vectors that are original linear cost terms; The energy balance equation constrains the dual variables, Corresponding to the electric, thermal and cold energy balance constraints respectively, Constraint of the dual variables for inequality, and Reflecting the degree of constraint tension, Representing the scene to be generated, And Are all A is an inequality-constrained coefficient matrix, G is an equality-constrained coefficient matrix, Is historical data.
6. The method for generating the closed-loop scene of the distributed energy system according to claim 1, wherein the global problem of the objective function at the lower layer is written as a set of a plurality of sub-problems according to the number of generated scenes, a progressive hedging algorithm is adopted to solve the global parameter consensus problem, an augmented Lagrange function is constructed, global consensus parameters, dual multipliers and punishment coefficients are introduced, convergence consistency of the parameters of each sub-problem is achieved through iterative processes of sub-problem solving, global parameter updating and punishment coefficient adjustment, and meanwhile, the total average cost of two phases is taken as a convergence criterion to ensure achievement of a global optimization target.
7. The method of claim 6, wherein a bias penalty term and a dual multiplier are introduced for each sub-problem, and a consensus constraint is embedded into the sub-problem objective function to obtain an augmented lagrangian function: ; Wherein, the For the generator parameters in the d-th sub-problem, The original loss for the d-th sub-problem contains a squared term value guide loss; A dual multiplier for the kth iteration; For the L2 penalty term, The larger the penalty coefficient for the kth iteration, the stronger the penalty for the bias; The global consensus parameter for the kth iteration is initially Is fixed in the kth iteration And Each sub-problem independently solves local optimal parameters The method comprises the following steps: ; loss gradient per sub-problem The method consists of a distribution matching loss gradient and a value guiding loss gradient: ; the original loss for the d-th sub-problem contains the squared term value-oriented loss with respect to the generator parameters Solving a local optimal solution: ; Wherein, the For learning rate, the magnitude of each parameter update is controlled; The sub-problem solving only depends on the scheduling cost gradient of the self-generated scene, and other sub-problem information is not needed; after all the sub-problems are solved, the weighted average updates the global consensus parameters: ; wherein D is the total number of sub-questions; updating the dual multiplier according to deviation feedback, wherein the dual multiplier is used for strengthening common-knowledge constraint, and the larger the deviation is, the larger the adjustment amplitude of the dual multiplier is: ; If it is Deviation from , Will increase, forcing by penalty term in the next iteration To the direction of The dual multipliers are in unsigned limitation, positive deviation corresponds to positive punishment, and negative deviation corresponds to negative punishment; Penalty coefficient And a self-adaptive adjustment strategy is adopted to balance convergence speed and stability: ; in order to amplify the coefficient by punishment coefficient, the constant with the value larger than 1 is adopted, the punishment is slightly increased for each iteration, the constraint is strengthened, And when the maximum parameter deviation exceeds the threshold value, amplifying a penalty coefficient to accelerate convergence.
8. The method for generating a closed-loop scene of a distributed energy system according to claim 6, wherein in the process of solving the global parameter consensus problem by adopting a progressive hedging algorithm, when the maximum deviation between local parameters and global parameters of all sub-problems is smaller than a threshold value and the fluctuation of global average scheduling cost is smaller than the threshold value, iteration of the progressive hedging algorithm is stopped, and parameters of a generator are updated according to a solving result.
9. A distributed energy system closed loop scene generation system, comprising: the operation optimization model construction module is configured to acquire a historical real data sequence of the distributed energy system, take the minimum comprehensive operation cost of the distributed energy system as an objective function, introduce operation constraint and energy balance constraint of equipment and construct an operation optimization model of the distributed energy system; the benchmark reference value calculation module is configured to construct a generation model, randomly initialize network parameters of the generation model, and utilize the operation optimization model to perform optimal power flow calculation by taking the historical real data sequence as deterministic input so as to obtain benchmark reference values; the initial simulation scene generation module is configured to generate initial simulation scenes by using the generation model, solve the initial simulation scenes by using the operation optimization model, and calculate the optimal system scheduling cost corresponding to each simulation scene; A composite loss function construction module configured to construct a composite loss function, the composite loss function including a cost-based loss term for measuring a difference between the optimal system scheduling cost and the baseline reference value, and a regularization term for constraining generation of scene statistics; The system comprises a generating model training module, a generating model parameter, an iterative updating optimization module, a generating model parameter generating module, a generating model generating module, a generating model generating module and a generating model generating module, wherein the generating model training module is configured to generate model parameters based on a feedback optimization scene, calculate the gradient of a composite loss function relative to the generating model network parameters by adopting an optimization algorithm, and update the generating model network parameters based on the gradient information in a back propagation mode; The scene set generation module is configured to generate a typical scene set for random optimization scheduling of the distributed energy system by utilizing the finally optimized generation model; The process for constructing the generation model comprises the steps of constructing the generation model of the double-layer framework, generating scenes meeting requirements by an upper layer through generating a minimum-maximum game of a generator and a discriminator of an countermeasure network, and carrying out daily and real-time two-stage scheduling optimization on each generation scene by a lower layer, calculating the deviation between the comprehensive cost and the ideal cost of the scenes and feeding the deviation back to the upper layer, and guiding the generator to optimize the generation direction of the scenes.

Description

Distributed energy system closed-loop scene generation method and system Technical Field The invention belongs to the technical field of multi-energy complementary distributed energy systems, and particularly relates to a method and a system for generating a closed-loop scene of a distributed energy system. Background The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art. With the increasing permeability of renewable energy sources, the optimal scheduling of distributed energy systems faces unprecedented challenges. The wind-solar energy output has extremely strong intermittence, volatility and randomness, and the power output of the wind-solar energy output presents complex space-time distribution characteristics. At the same time, the customer side load demand also shows a high degree of uncertainty driven by electricity price signals, behavioral patterns and environmental factors. The multi-dimensional uncertainty of the source-load sides makes the traditional scheduling method based on deterministic prediction difficult to adapt to actual running requirements. If the scheduling policy cannot accurately capture the fluctuations, the problems of unbalanced supply and demand, overload of equipment, wind and light abandoning or load shedding and the like are very easy to be caused. In order to realize the optimal operation of the system in a complex random environment, a random optimal scheduling method based on scene analysis is widely applied. The method has the core idea that a continuous probability distribution space for approximately describing the source load uncertainty variable is constructed by constructing a limited number of typical discrete scene sets, so that a complex stochastic programming problem is converted into a deterministic programming problem to be solved. It can be seen that the quality of the constructed scene set directly determines the reliability and economy of the random optimal scheduling result. A high quality set of scenes should be able to accurately capture the key features of the uncertainty factors, leading to an optimization model that yields the best scheduling strategy that can accommodate the various possible scenarios in the future. Currently, for the scene generation technology of the distributed energy system, the existing research mainly tends to adopt a data-driven generation model method, such as generating a deep learning model of a challenge Network (GENERATIVE ADVERSARIAL Network, GAN), a variation self-encoder and the like. The basic development thought of the prior art is to collect a large amount of historical source load operation data as training samples, construct a complex generator model, and continuously adjust model parameters through a specific training algorithm, so as to enable the simulated scene data output by the generator to approximate the characteristics of historical real data as much as possible on statistical indexes such as probability density distribution, time sequence autocorrelation, cross correlation among multiple variables and the like. The technical logic of the bottom layer is considered that the requirements of the follow-up optimal scheduling can be met as long as the generated scene is enough to restore the history rule on the statistical level. However, through in-depth analysis and practical verification, the main stream of the prior art has a fundamental methodology defect in practical application, namely, the generation process of the scene and the subsequent optimal scheduling application process are in an open-loop state of mutual splitting. In particular, the core focus of existing generative models when constructing a loss function during the training phase is too limited to minimizing the difference between the generated data distribution and the historical real data distribution, for example, in an effort to reduce the value of both on some statistical distance indicators. The method for simply pursuing the similarity of the statistical level seriously ignores the physical characteristics and economic targets of the downstream distributed energy system optimization scheduling model. In the actual complex distributed energy system optimization problem, due to the existence of a large number of nonlinear equipment operation constraints, complex energy coupling relations and specific economic objective functions, the influence of scene inputs of different characteristics on the final scheduling result is nonlinear and has great difference. Some scene details with insignificant deviation in statistical indexes can be greatly amplified after complex calculation conduction of an optimization model, and finally the obtained scheduling scheme is significantly higher in economic cost and cannot be executed at all physically. The scene set generated by the prior art, although fitting the statistical rule of the historical data in a