CN-122026520-A - Flexible straight central power distribution method and system based on model predictive control

CN122026520ACN 122026520 ACN122026520 ACN 122026520ACN-122026520-A

Abstract

The invention belongs to the technical field of power systems, and discloses a soft straight central power distribution method and a system based on model prediction control, wherein an LSTM and attention mechanism double-branch prediction model is built through a multisource fusion dynamic association model, the LSTM captures time sequence correlation, an attention mechanism carries out layered weighted optimization on key time periods such as fluctuation, faults and the like, and a topology parameter correction item is matched for correcting a result in real time; the prediction model parameters can be optimized by the reinforcement learning agent in an iterative mode, the prediction error can be effectively reduced, the prediction model can be dynamically adapted to the characteristics of strong power correlation and variable working conditions of flexible straight central multiport, a multi-objective collaborative optimization system is designed, a fuzzy and reinforcement learning fusion algorithm is adopted, each objective weight coefficient is dynamically adjusted according to the real-time working condition of the power grid, and the multi-objective conflict is resolved by combining a Pareto optimal solution screening mechanism.

Inventors

ZHAO YUANYOU
WANG FEI
YAN YU
MIN QIANG
TAN JIN
ZHANG JIANSHENG
LI MENG
LIU YONGJIANG
YANG SHUANG

Assignees

大唐贵州发耳发电有限公司

Dates

Publication Date: 20260512
Application Date: 20260225

Claims (9)

1. The soft straight central power distribution method based on model predictive control is characterized by comprising the following steps of: The coupling modeling stage comprises the steps of collecting various operation information related to new energy power generation, an energy storage system, a flexible direct current converter, regional load, line and power grid faults, realizing self-identification of flexible direct central topological parameters by adopting a deep reinforcement learning intelligent agent based on kirchhoff's law and a converter mathematical model, synchronously constructing a multi-source fusion dynamic association model incorporating new energy output fluctuation prediction error compensation items, and combining a multi-source data cross checking mechanism; The prediction modeling stage comprises the steps of performing short-term prediction on new energy output and load demands and capturing dynamic change rules by adopting a double-branch prediction model combining LSTM (least squares) and an attention mechanism based on a multisource fusion dynamic association model, optimizing prediction accuracy by using an attention mechanism and a topology parameter correction term, adaptively adjusting prediction time domain and step length by combining multidimensional index, and optimizing prediction model parameters by using deep reinforcement learning agent iteration synchronously; The target optimization stage comprises the steps of designing a multi-target collaborative optimization target function containing targets related to new energy consumption, energy conservation, environmental protection, power grid stability, energy storage balance and conversion efficiency based on prediction data, dynamically adjusting each target weight coefficient according to real-time operation conditions of the power grid by adopting a fuzzy and reinforcement learning fusion dynamic weight distribution algorithm, and solving multi-target conflict by combining a Pareto optimal solution screening mechanism; The constraint construction stage comprises the steps of constructing a layered robust constraint system comprising hard constraint and flexible constraint, enhancing constraint robustness through a mixed regularization item of a mixed type of L1 and L2, and dynamically adjusting a regularization coefficient; integrating the multi-objective collaborative optimization function and constraint conditions, adopting a multi-algorithm fused quick solution algorithm, reducing the solution space, accelerating the convergence speed, synchronously checking the solution result and optimizing the solution parameters; the feedback correction stage is based on a model predictive control rolling optimization principle, a double closed loop feedback correction system is constructed by taking a control sequence which is rapidly solved and output as a core, relevant parameters are corrected in real time through various deviation indexes, and a fault grading response mechanism is set; And in the verification and optimization stage, a full-scene simulation platform and a reduced-scale experiment platform are built, a typical operation condition is simulated to verify a power distribution method, and parameters of each stage are optimized through a deep reinforcement learning intelligent body by combining a verification result.
2. The flexible-direct-center power distribution method based on model predictive control according to claim 1, wherein the operation data collected in the coupling modeling stage comprises real-time output power, output fluctuation rate and prediction error of a new energy power generation unit, SOC state of an energy storage system, charge-discharge power limit and SOC change rate, output power, direct-current voltage, loss coefficient and conversion efficiency of a flexible-direct converter, real-time demand and load mutation early warning of loads of all areas, line impedance parameter, line loss and power grid fault type early warning information, the deep reinforcement learning agent is used for iteratively optimizing topology parameter identification precision through real-time interaction with a flexible-direct-center operation environment, and the multi-source data cross checking mechanism is used for guaranteeing that an associated model is in agreement with actual operation conditions.
3. The flexible straight central power distribution method based on model prediction control according to claim 2 is characterized in that in the prediction modeling stage, a first branch of a double-branch prediction model captures dynamic change rules and long-term time sequence relativity of new energy output and load through an LSTM network, a second branch focuses on a key period through an attention mechanism and optimizes a prediction result, a topology parameter correction term corrects the prediction result in real time, the multidimensional index comprises a prediction error threshold, a new energy output fluctuation rate and a power grid operation condition, a prediction time domain and a step length are adaptively adjusted according to three indexes, and the prediction model parameters are self-updated on line through a deep reinforcement learning intelligent body and adapt to dynamic change of a working condition.
4. The soft straight central power distribution method based on model predictive control according to claim 3 is characterized in that a multi-objective collaborative optimization system in the objective optimization stage is used for minimizing new energy light and wind abandoning rate, minimizing total system loss, minimizing direct current bus voltage fluctuation amplitude, balancing energy storage system SOC to be four core targets and maximizing conversion efficiency to be auxiliary targets, the dynamic weight distribution algorithm is used for constructing a multi-rule self-adaptive adjustment mechanism through a fuzzy control algorithm, deep reinforcement learning agent iterative optimization fuzzy control rules and weight thresholds, and the optimal solution screening mechanism is used for realizing self-adaptive switching of optimal power distribution strategies under different working conditions.
5. The flexible straight center power distribution method based on model predictive control according to claim 4 is characterized in that in the layered robust constraint system, the hard constraint comprises the output power, current and voltage limits of a flexible straight converter, the charge and discharge power and SOC limits of an energy storage system, the current capacity limit of a power transmission line and a robust compensation term of uncertainty of line parameters, the flexible constraint sets a dynamic self-adaptive adjustment mechanism of a direct current bus voltage fluctuation range, small range deviation is allowed to exist in new energy output distribution, the mixed regularization term is of a mixed type of L1 and L2 and is used for coping with multiple interference, and the coefficient is combined with topology parameters after reinforcement learning optimization to dynamically adjust.
6. The flexible straight center power distribution method based on model predictive control according to claim 5, wherein the rapid solving algorithm is an algorithm of mixed integer second order cone planning, a cutting plane and a self-adaptive dichotomy, the nonlinear optimization problem can be converted into a convex optimization problem, a solving space is reduced through a cutting plane mechanism, the self-adaptive dichotomy adjusts iteration step length according to solving precision and computing efficiency, the iteration step length is converged to a Pareto optimal solution, and an optimal power distribution control sequence of each port is output, and the solving parameters are self-optimized on line according to long-term solving data through a deep reinforcement learning agent.
7. The soft straight central power distribution method based on model predictive control according to claim 6, wherein in the feedback correction stage, only the current moment power distribution instruction of an optimal control sequence is executed, the rest sequence is used as an initial reference, prediction deviation, model error and constraint deviation are calculated through comparison of real operation data and prediction data collected in real time, three types of deviation are corrected in real time by adopting a proportional integral and reinforcement learning fusion correction algorithm, proportional integral correction parameters and next time domain related parameters are optimized in a deep reinforcement learning agent iteration mode, the fault grading response mechanism switches emergency power distribution strategies according to fault types and severity, strategy smooth switching is achieved after fault recovery, and corrected parameters and strategies are fed back to the stage.
8. The soft straight central power distribution method based on model predictive control according to claim 7 is characterized in that in the verification optimization stage, a full-scene simulation platform simulates basic working conditions, extreme working conditions and fault working conditions, mainly verifies five core performance indexes including new energy consumption rate, system loss, voltage fluctuation amplitude, response time and fault recovery speed, a deep reinforcement learning intelligent agent can perform full-flow self-adaptive iterative optimization on core parameters of each stage including prediction model parameters, dynamic weight coefficients, solving algorithm parameters and constraint condition thresholds, a scale experiment platform is connected with an actual simulation unit and energy storage equipment to simulate engineering actual environment, verify algorithm feasibility and stability, optimize engineering suitability and verify optimized parameters to be fed back to the stage.
9. The flexible direct neutral power distribution system based on model predictive control, based on the method of claim 8, comprising: The data acquisition unit is used for acquiring flexible-direct central multi-port full-dimensional operation data and comprises new energy power generation, an energy storage system, a flexible-direct converter, regional loads, line and power grid fault related information; The coupling modeling unit is used for realizing topology parameter self-identification by adopting a deep reinforcement learning agent based on kirchhoff's law and a mathematical model of the converter, constructing a multisource fusion dynamic association model and combining a data cross checking mechanism; The prediction modeling unit relies on the association model to perform short-term prediction through an LSTM and attention mechanism double-branch structure, and the prediction time domain and the step length are adaptively adjusted by combining the multi-dimension index, so that the on-line self-updating of the prediction model parameters is realized; The target optimization unit is used for designing a multi-target collaborative optimization function, dynamically adjusting target weight by adopting a fuzzy and reinforcement learning fusion algorithm, and solving multi-target conflict through an optimal solution screening mechanism; The constraint construction unit is used for constructing a layered robust constraint system based on equipment operation limit and power grid safety requirements, enhancing robustness through mixed regularization items and providing constraint boundaries; The rapid solving unit integrates the optimization target and the constraint condition, adopts a rapid solving strategy of MISOCP, a cutting plane and a self-adaptive dichotomy, outputs a power distribution control sequence of each port and optimizes solving parameters; The feedback correction unit is used for constructing a double closed loop correction system based on a rolling optimization principle, correcting parameters in real time and executing a fault grading response mechanism to form a control closed loop; And the verification optimizing unit is used for verifying the feasibility of the system through full scene simulation and physical experiments, and optimizing parameters of each unit by combining verification results, so that the control performance and engineering suitability are improved.

Description

Flexible straight central power distribution method and system based on model predictive control Technical Field The invention belongs to the technical field of power systems, and particularly relates to a flexible straight central power distribution method and system based on model predictive control. Background The global energy structure is forward clean and low-carbon, the large-scale grid connection of new energy is in a development situation, and the flexible direct current transmission (flexible direct current) technology is a core technical support for the remote transmission of new energy and the interconnection of regional power grids by virtue of the advantages of being capable of independently controlling active and reactive power, having no commutation failure risk, adapting to the grid connection of weak power grids and the like. The flexible straight center is used as an energy scheduling core of the multi-terminal flexible straight system, and the power distribution rationality of the flexible straight center directly influences the new energy consumption efficiency, the power grid operation stability and the power supply economy. The conventional common methods such as droop control and constant power control have obvious short plates, are low in response speed, are difficult to quickly track the intermittence and fluctuation of new energy output, are easy to cause overlarge power distribution deviation, cause voltage fluctuation of a direct current bus and further threaten stable operation of a system, are insufficient in robustness, are easy to be instable in control strategy under complex working conditions such as power grid faults and load abrupt change, cannot adaptively adjust a power distribution scheme, and are high in new energy light and wind rejection rate and limited in power grid operation efficiency. The model predictive control is used as an advanced nonlinear control method, can directly process the problems of multi-constraint and multi-objective optimization, has fast dynamic response and high control precision, is gradually applied to the field of power system control, but when the model predictive control is applied to flexible-straight central power distribution, the traditional scheme mainly adopts a fixed predictive time domain and a weight coefficient, and still has a plurality of technical problems that the model predictive control cannot adapt to the characteristics of strong power correlation and dynamic change of working conditions of flexible-straight central multiport, and the predictive model does not fuse multisource operation information, has insufficient prediction precision and large calculation amount in the solving process, and is difficult to meet the real-time control requirement. Disclosure of Invention The invention aims to provide a soft straight central power distribution method and system based on model predictive control, so as to solve the problems in the background art. In order to achieve the above purpose, the invention provides a soft straight central power distribution method based on model predictive control, which comprises the following steps: preferably, the coupling modeling stage collects flexible-direct central multi-port full-dimensional operation data, including real-time output force, output fluctuation rate and prediction error data of a new energy power generation unit, and the SOC state, charge-discharge power limit and SOC change rate of an energy storage system, the output power, direct-current voltage, loss coefficient and conversion efficiency of a flexible-direct converter, real-time requirements of loads of all areas, load mutation early warning, line impedance parameters, line loss and grid fault type early warning information; Based on kirchhoff's law and a mathematical model of the converter, the self-identification of the topological parameters of the flexible and straight center is realized by adopting a deep reinforcement learning agent, the agent interacts with the running environment of the flexible and straight center in real time, the identification precision of the topological parameters is optimized in an iterative manner, a multisource fusion dynamic association model is constructed synchronously in the process, and a dynamic compensation item of the output fluctuation prediction error of the new energy source is brought into association relation; And combining a real-time cross checking mechanism of the multi-source data, so that the correlation model takes into account the influence of multiple factors and is in agreement with the actual operation condition. The multi-source fusion dynamic association model takes a flexible straight central topological structure as a basic framework, fuses six-dimensional data of new energy power generation, an energy storage system, a flexible straight converter, regional load, line operation and power grid faults, performs space-time fusion on the dimensional data accord