Intelligent edge control method for information energy system based on self-adaptive dynamic programming

Abstract

The invention provides an information energy system intelligent edge control method based on adaptive dynamic programming, and relates to the technical field of information energy system optimization control. Compared with an existing control technology, the control strategy can be automatically adjusted according to the system state and the real-time change of the environment, the control strategy is continuously iterated and optimized through the online learning and dynamic planning principle, the system can pursue the optimal performance in the changing environment all the time, and compared with a static preset strategy, the optimal performance of the system is improved. Complex dynamic scenes can be dealt with better, and the stability and the control effect of the system can be ensured even under the condition that parameters change. And the adaptive dynamic planning method shows good performance when being applied to the solution of the optimization problem of the nonlinear system, can perfectly overcome the adverse effect of curse of dimensionality, and can be more close to the actual industrial requirements undoubtedly.

Description

Technical Field

[0001] The present invention belongs to the technical field of information energy system optimization control, and in particular to an information energy system intelligent edge control method based on adaptive dynamic programming. Background Art

[0002] Energy Internet can realize the large-scale development and utilization of clean energy, reduce carbon emissions, and help cope with climate change by integrating various types of distributed renewable energy. Its proposal and development are both the development trend of the energy system itself and the external requirements for the energy system. Under the joint action of internal and external factors, it has distinct contemporary significance and huge development space. The optimization and control of the integrated energy system is actually the development and extension of smart grid optimization technology, reflecting the intersection and integration of the two major disciplines of automatic control and energy application. How to optimize the control of the energy Internet has received widespread attention.

[0003] For the optimization and regulation of energy Internet systems, the steady-state model of the system is generally analyzed and the approximate optimal solution is solved. For complex static models, the approximate optimal solution is often found through iteration. In each round of iteration, based on the current solution, the objective function and the gradient or Jacobian matrix under the constraints are calculated. This information is used to update the decision variables (i.e., the control parameters) and gradually approach the global or local optimal solution through iteration. Check whether the new solution meets all the constraints. If not, it may be necessary to introduce projection operations or other correction mechanisms to ensure the validity of the solution. When the preset convergence standard is reached, the iteration is stopped, and the last solution obtained is used as the approximate optimal solution.

[0004] In the modeling and analysis of integrated energy systems, the use of algebraic equations to characterize is a common means of solving the problem of optimal scheduling of multi-energy flow networks. Integrated energy systems involve the conversion, transmission and storage of multiple energy forms such as electricity, heat, cold and gas. These processes are described and solved through a series of algebraic equations, which can obtain the optimal scheduling strategy of the integrated energy system to ensure that the system operates efficiently, economically and environmentally friendly. For each node (such as power plants, substations, user terminals, etc.), based on the principle of energy conservation, the balance equation between the supply and consumption of various types of energy (electricity, heat, etc.) is established; for various energy networks (such as power grids, heat networks, and natural gas pipelines), according to the principle of energy transmission, the continuous energy flow equation can be discretized into algebraic form; the working characteristics of equipment (such as generators, boilers, energy storage devices, etc.) can be reflected by algebraic equations, including but not limited to the relationship between the output power of the equipment and the input fuel, the equipment performance curve, the startup or shutdown logic, the minimum stable operating level, the maximum power limit, etc.

[0005] Define one or more objective functions, usually expressed as minimizing total cost, minimizing carbon emissions, maximizing energy utilization, etc. These objective functions are also in the form of algebraic expressions. When faced with uncertain factors such as energy prices, demand forecasts, and weather conditions, introduce probability distributions or interval variables to transform uncertainty problems into deterministic optimization problems.

[0006] Linearization methods are often used to solve problems. The application of linearization methods is mainly used to simplify complex nonlinear models and facilitate the rapid and effective solution of optimization problems. For example, economic scheduling problems can usually be transformed into a linear programming problem, the goal is to minimize the total cost or maximize the benefits while satisfying various constraints (such as supply and demand balance, upper and lower limits of equipment operation, environmental emission limits, etc.).

[0007] Most of the current research results are based on steady-state or quasi-steady-state models of energy internet, and there is little analysis of dynamic models of integrated energy systems. For the optimization and regulation of energy internet systems, the steady-state model of the system is generally analyzed and approximate optimal solutions are solved. The solutions obtained are of limited utility for the ever-changing actual energy network. Because the dynamics of energy systems in real scenarios will change due to the influence of production and life processes, the optimization solutions established by the steady-state model are difficult to adapt to the regulation process in new scenarios and new environments.

[0008] Moreover, most of the existing literature uses algebraic equations to characterize the integrated energy system, but there are few studies on the construction of the system's state space model, especially the interconnected large-scale system model of the integrated energy system has not yet been constructed. The research mainly focuses on the dynamic model of independent generators, wind turbines, and energy storage batteries, etc., without fully considering the interconnection effects between multiple energy sources. As a result, it is impossible to construct an interconnected network dynamic model that can reflect the multi-energy coupling relationship within the integrated energy system, which restricts the development of related control technologies.

[0009] There have been many studies that consider the Energy Internet as a multi-agent system, but they are still based on iterative solutions to the optimal solution, which makes it difficult to achieve online operation. Using iterative methods to solve the approximate optimal solution of static models is difficult to adapt to the dynamic changes and strong randomness of the Energy Internet. As the structure of the Energy Internet system becomes increasingly complex and the number of nodes continues to increase, the computational burden it faces will increase exponentially, bringing great difficulties to the solution, and the "dimensionality curse" problem is becoming increasingly prominent.

[0010] Although the use of linearization methods reduces the difficulty and amount of calculation to a certain extent, some nonlinear characteristics of the original problem are often ignored during the linearization process, which may lead to inaccurate solutions, especially when nonlinear characteristics significantly affect system performance. Many dynamic characteristics in the energy Internet (such as fluctuations in new energy output, uncertainty in load demand, equipment efficiency curves, etc.) cannot be fully captured by linear models. Linear models are difficult to effectively deal with the impact of environmental changes and uncertainties, especially the random characteristics and emergencies faced in the energy Internet. The simplified model is difficult to truly reflect the characteristics of the system, and the resulting calculation errors are also a problem that cannot be ignored.

[0011] The optimization problem of the energy Internet is essentially a constrained optimization problem. Traditional control schemes are not adaptable to changes in the energy system structure and new energy devices, and it is difficult to meet the current development needs of the energy Internet. Summary of the invention

[0012] In view of the deficiencies in the prior art, the present invention proposes an intelligent edge control method for an information energy system based on adaptive dynamic programming to solve the optimization scheduling problem of the information energy system.

[0013] The technical solution of the present invention is:

[0014] The intelligent edge control method of the information energy system based on adaptive dynamic programming includes the following steps:

[0015] Step 1: Model the gas grid system and power grid system in the electric-gas coupled energy system respectively, and then obtain the dynamic model of the electric-gas coupled energy system, and use the neural network to approximate the function terms in the dynamic model of the electric-gas coupled energy system to obtain the approximate dynamic mathematical model of the electric-gas coupled energy system.

[0016] The electric-gas coupled energy system includes an electric power network system and a gas network system. The gas network system includes a plurality of gas network subsystems, and the electric power network system includes a plurality of electric power network subsystems. Each gas network subsystem corresponds to a power network subsystem. Such a group of corresponding gas network subsystems and power network subsystems is called an energy subsystem.

[0017] Step 1.1: Model the gas grid system and obtain the dynamic model of each gas grid subsystem.

[0018] For containing N The electric-gas coupled energy system with energy subsystems, each energy subsystem is regarded as an energy node. Energy Nodes , allowing gas to flow to the energy node All energy nodes are called energy nodes Upstream nodes, energy nodes The upstream node set of , Energy Node The upstream node of is the number of the upstream node in the upstream node set, is the number of upstream nodes, let The mid-upstream nodes Energy Nodes The input gas pressure increment is , Represents the upstream node set The The upstream node Energy Nodes The input gas pressure increment is set from the upstream node The mid-upstream nodes Energy Nodes The output gas flow increment is , Represents the upstream node set The The upstream node Energy Nodes Output gas flow increment.

[0019] Let the gas flow from the energy node All energy nodes that flow to are called nodes Downstream nodes, energy nodes The set of downstream nodes is , Energy Node The downstream node of is the number of the downstream node in the downstream node set, is the number of downstream nodes, let Energy Nodes Downstream node collection The gas pressure increment input at all downstream nodes is , Energy Node To the downstream node set The gas pressure increment input from the downstream node is Energy Nodes Downstream node collection The gas flow increment output by all downstream nodes is , Energy Node To the downstream node set The gas flow increment output by the downstream node.

[0020] Order The gas grid state vector of energy nodes is , according to the gas network state vector, construct the The dynamic model of the gas grid subsystem is: (6);

[0021] in, The first The status of each gas grid subsystem, for The time derivative, for The time derivative, for The time derivative, , , , , , For the upstream node to The gas flow increment input by each energy node, For the The gas pressure increment output by an energy node to the downstream node, , , For the upstream node and The length of the natural gas pipeline between energy nodes, For the The length of the natural gas pipeline between the energy node and the downstream node, is the flow value at the steady-state operation point, is the pressure value at the steady-state operating point, is the speed of sound in natural gas, is the cross-sectional area of ​​the natural gas pipeline, Indicates the friction factor when gas flows in a natural gas pipeline. is the diameter of the natural gas pipeline, where For interconnected items, .

[0022] Step 1.2: Model the power network system and obtain the dynamic model of each power network subsystem.

[0023] No. The dynamic model of the power network subsystem in the energy subsystem is: (8);

[0024] in, For the The frequency increment of the generator set in the power network subsystem in the energy node, for The time derivative, For the The output power increment of the generator set in the power network subsystem in the energy node, for The time derivative, For the The increment of the speed regulator valve opening of the generator set in the power network subsystem in the energy node, for Derivative with respect to time; , , , To control the input variables, , For the The state vector of the power network subsystem in each energy node, , , , For interconnected items, is the global state of the electric-gas coupled energy system, For the The state vector of the energy subsystem, For the The time constant of the generator set in the power network subsystem in the energy node, For feedback adjustment gain, For the The time constant of the speed regulator in the power network subsystem in each energy node, For the The gain constant of the generator set in the power network subsystem in the energy node, For the The time constant of the steam turbine in the power network subsystem in the energy node, For the The gain constant of the steam turbine in the power network subsystem in the energy node, For the The increase in electric power provided by the power network subsystem in each energy node.

[0025] Step 1.3: Integrate the dynamic model of the gas grid subsystem and the dynamic model of the power grid subsystem to construct a dynamic model of each energy subsystem in the electricity-gas coupled energy system.

[0026] The dynamic model of each energy subsystem in the electric-gas coupled energy system is: (9);

[0027] in, for The time derivative, , For the The state vector of the energy subsystem elements, , Indicates The state vector of the energy subsystem The influence of each element on the dynamic process inside the gas grid subsystem. For interconnected items No. elements, , is the input control vector, is the matrix of interconnected items.

[0028] Step 1.4: Use a neural network to approximate the function terms of the dynamic model of each energy subsystem in the electric-gas coupled energy system to obtain an approximate dynamic mathematical model of each energy subsystem.

[0029] The approximate dynamic mathematical model of each energy subsystem is: (10);

[0030] in, is a stable matrix, Designed for size The matrix of The size is The zero matrix of , is the ideal value of the neural network weights and , is the activation function, ,in is the approximation error of the neural network.

[0031] Step 2: Design an identifier of the electric-gas coupled energy system based on the approximate dynamic mathematical model of the electric-gas coupled energy system.

[0032] The identifier of the electric-gas coupled energy system is: (11);

[0033] in, For the The estimated value of the state vector of each energy subsystem, is the estimated value of the state vector elements, for Derivative with respect to time; is the activation function estimate, is the estimated value of the neural network weight, is the dimension of the vector of neural network weight estimates, for The estimated value of is the Hurwitz matrix.

[0034] The weight adaptation law of the identifier of the electric-gas coupled energy system is: (13);

[0035] in, for The time derivative, is the designed symmetric positive definite matrix, is the learning rate, , , is a positive definite matrix, is the state error of the identifier: (12);

[0036] in, For the The time derivative of the state error of an energy subsystem.

[0037] Step 3: Treat each energy subsystem in the electric-gas coupled energy system as an isolated subsystem, and use the judgment network to update and iterate to solve the event-triggered approximate optimal control strategy of the electric-gas coupled energy system. At the same time, based on the identifier of the electric-gas coupled energy system obtained in step 2, the weight adaptation law of the judgment network is designed.

[0038] Step 3.1: Define the performance index of the isolated subsystem and obtain the time-driven Hamiltonian function based on the performance index of the isolated subsystem.

[0039] The performance indicators of the isolated subsystem are: (14);

[0040] in, For the The performance index of an isolated subsystem is is the utility function, and are all symmetric positive definite matrices, For time.

[0041] The time-driven Hamiltonian function is: (16);

[0042] in, is the time-driven Hamiltonian function; , is a smooth function, yes about The partial derivative function of For the Performance indicators of an isolated subsystem.

[0043] Step 3.2: Define the optimal value function of the performance index of the isolated subsystem, and based on the optimal value function of the performance index of the isolated subsystem and the time-driven Hamiltonian function obtained in step 3.1, obtain the optimal control strategy and the corresponding time-triggered HJB equation.

[0044] The optimal value function of the performance index of the isolated subsystem is: (17);

[0045] in, For the The optimal value function of the performance index of an isolated subsystem.

[0046] Setting the optimal value function of the performance index of the isolated subsystem Existence and in The above is differentiable, and the optimal control strategy is: (18);

[0047] in, is the optimal value function of the performance index of the isolated subsystem about The partial derivative of Input variables for optimal control.

[0048] The corresponding time-triggered HJB equation is: (19);

[0049] in, is the time-triggered HJB equation, is the utility function of the optimal control strategy.

[0050] Step 3.3: Based on the optimal control strategy obtained in step 3.2, establish the event-triggered optimal control strategy and the corresponding event-triggered HJB equation.

[0051] The event triggers the optimal control strategy as follows: (20);

[0052] in, is the event-triggered optimal control vector, , is the sampling state, For the A trigger moment, For the A trigger moment.

[0053] The event-triggered HJB equation is: (twenty one);

[0054] in, For event-triggered HJB equation, is the utility function of the event-triggered optimal control strategy.

[0055] Step 3.4: Use the neural network to approximate the optimal value function of the performance index of the isolated subsystem, obtain the optimal value function approximated by the neural network, and then obtain a new expression of the event-triggered optimal control strategy.

[0056] The optimal value function of the neural network approximation is: (twenty two);

[0057] in, is the optimal value function approximated using a neural network, is the optimal weight of the neural network, is the number of neurons in the hidden layer, is the activation function of the neural network, is the approximation error of the neural network.

[0058] The new expression of the event-triggered optimal control strategy is: (twenty four);

[0059] in, and They are and about The partial derivative of .

[0060] Step 3.5: Use the judgment network to approximately solve the new expression of the event-triggered optimal control strategy to obtain the event-triggered approximate optimal control strategy.

[0061] Using a judgement network The output of is used to approximate the event-triggered optimal control strategy, where is the optimal weight of the neural network The estimated value of To judge the output value of the network, is the activation function for judging the network.

[0062] The event-triggered approximate optimal control strategy is: (25);

[0063] in, is the approximate optimal control vector, is the estimated value of the optimal weight of the neural network.

[0064] Step 3.6: Based on the identifier of the electric-gas coupled energy system obtained in step 2, the error function of the event-triggered approximate optimal control strategy is minimized by estimating the optimal weights of the neural network in the approximate event-triggered approximate optimal control strategy of the judgment network, and the weight adaptation law of the judgment network is designed.

[0065] According to the dynamic , the approximate Hamiltonian equation of the event-triggered approximate optimal control strategy is obtained as: (26);

[0066] in, is the approximate Hamiltonian equation for the event-triggered approximate optimal control strategy, is the utility function of the event-triggered approximate optimal control strategy, yes about The partial derivative of The parameters to be defined.

[0067] definition and weight error , is a function of the state estimate, the approximate optimal control vector, and the optimal weight of the neural network, and is defined as .

[0068] The error function of the event-triggered approximate optimal control strategy is: ,in , and then design the adaptive law of the evaluation network as: (27);

[0069] In the formula, for The time derivative, is the learning rate, , , is the weight error.

[0070] Step 4: Utilize the event-triggered approximate optimal control strategy of the electric-gas coupled energy system and the weight adaptive law of the evaluation network to realize the control of the electric-gas coupled energy system.

[0071] Compared with the prior art, the advantages of the present invention are:

[0072] Compared with existing control technologies, the intelligent edge control method of information energy system based on adaptive dynamic programming can automatically adjust the control strategy according to the real-time changes of system status and environment. Through online learning and dynamic programming principles, the control strategy is continuously optimized iteratively, so that the system can always pursue the best performance in a changing environment. Compared with static preset strategies, it can better cope with complex dynamic scenarios and ensure the stability and control effect of the system even when parameters change. In addition, the adaptive dynamic programming method has shown good performance in solving nonlinear system optimization problems, which can perfectly overcome the adverse effects of the "dimensionality disaster" and is undoubtedly closer to actual industrial needs.

[0073] Fine adaptive control helps avoid overload of energy equipment and reduce failure rate, thereby increasing equipment life and the economy of overall system operation. It also helps better coordinate and optimize the conversion and distribution of different energy sources in view of the conversion and interaction of various energy forms in information energy systems, improves overall energy utilization efficiency, reduces losses, and promotes energy conservation and emission reduction.

[0074] The intelligent edge control method of information energy system based on adaptive dynamic programming breaks through the limitations of traditional control methods, realizes a higher degree of automation and intelligent control, effectively improves the overall efficiency and stability of the system, and provides a more advanced solution for the optimized operation of information energy system. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 This is a flow chart of an intelligent edge control method for an information energy system based on adaptive dynamic programming in an embodiment of the present invention. DETAILED DESCRIPTION

[0076] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0077] The present invention combines key technologies such as intelligent and adaptive control algorithms with the actual needs of information energy systems, and is aimed at a series of problems such as management, scheduling, protection and economic operation of energy Internet. It is an energy Internet control technology based on adaptive dynamic programming method.

[0078] Different from the traditional centralized control mode, decentralized control breaks down the optimization problem of a large system into local problems of each subsystem, greatly reducing the computational pressure of a single controller. It is also in line with the trend of distributed and interconnected modern energy systems.

[0079] Different from the previous static and quasi-static control technologies, the present invention constructs a dynamic model of the electric-gas coupled energy system for analysis. The intelligent edge control technology of the information energy system based on adaptive dynamic programming makes full use of the theory of adaptive dynamic programming, and can adjust the control strategy in real time according to the actual operating conditions, adapting to the changeable and complex characteristics of the information energy system, and embodying a strong self-learning and evolutionary ability. Whether the internal state of the system changes or the external environment fluctuates, it can respond quickly and find the optimal or suboptimal control strategy.

[0080] Intelligent edge control method of information energy system based on adaptive dynamic programming, such as Figure 1 As shown, the following steps are included:

[0081] Step 1: Consider the secondary optimal control problem of the electric-gas coupled energy system, the purpose of which is to make the state deviation of the electric-gas coupled energy system tend to 0, model the gas grid system and the power network system in the electric-gas coupled energy system respectively, and obtain the dynamic model of the electric-gas coupled energy system by comprehensive simplification, and use neural networks to approximate the function terms in the dynamic model of the electric-gas coupled energy system to obtain an approximate dynamic mathematical model of the electric-gas coupled energy system; the electric-gas coupled energy system is a comprehensive energy system in which the power network system and the gas grid system are interconnected and interacted with each other through specific technical means, the gas grid system includes several gas grid subsystems, the power network system includes several power network subsystems, and each gas grid subsystem corresponds to one power network subsystem, and such a group of corresponding gas grid subsystems and power network subsystems is called an energy subsystem.

[0082] Step 1.1: Model the gas grid system and obtain the dynamic model of each gas grid subsystem.

[0083] The main consideration is the slow transient process caused by the injection or output of natural gas and the fluctuation of natural gas load in the natural gas pipeline. The transmission of natural gas along the pipeline is assumed to be a constant temperature process. At this time, the energy conservation equation often considered in the natural gas pipeline network can be ignored. It is assumed that when the load demand fluctuates, the flow direction of the gas in the natural gas pipeline will not change. The flow direction is determined by the scheduling strategy and is not within the scope of this consideration. In order to respond to the uncertainty of renewable energy, the output of the gas unit needs to be adjusted in time at the three-level control level, and the output of the gas unit is affected by the flow and pressure of each node in the natural gas pipeline network. When the output changes, the flow and pressure demand of each node in the natural gas pipeline network is bound to change. This requires adjusting the compression ratio of the compressor in the natural gas pipeline network to achieve the stability of the pressure and flow in the pipeline, thereby ensuring the stable operation of the entire pipeline network.

[0084] Specifically: N The electric-gas coupled energy system with energy subsystems is constructed by treating each energy subsystem as an energy node. To another energy node Natural gas pipeline status model: (1);

[0085] in, and is the number of the energy node, For Energy nodes to the The gas pressure increment input by the energy node, For Energy nodes to the The gas flow increment output by each energy node, For Energy nodes to the The time derivative of the gas pressure increment input by the energy node, For Energy nodes to the The time derivative of the gas flow increment output by an energy node, is the speed of sound in natural gas, is the cross-sectional area of ​​the natural gas pipeline, For the Energy nodes and The length of the natural gas pipeline between energy nodes, Indicates the friction factor when gas flows in a natural gas pipeline. is the flow value at the steady-state operation point, is the pressure value at the steady-state operating point, is the diameter of the natural gas pipeline, For the Energy nodes to the The gas flow increment input by each energy node, For the Energy nodes to the The gas pressure increment output by an energy node.

[0086] make (2);

[0087] make (3);

[0088] in, For the The energy nodes connected to the first The column vector of the gas pressure increment input by the energy nodes, For the The energy nodes connected to the first The column vector of gas flow increments output by energy nodes, Indicates that in the electric-gas coupled energy system The number of the energy node connected to the energy nodes, Indicates that The number of the energy node in the set of energy nodes connected to the energy nodes, and , For the The number of energy nodes connected to each energy node; Indicates Energy nodes to the The gas pressure increment input by the energy node, Indicates Energy nodes to the The gas flow increment output by each energy node.

[0089] From equations (1), (2) and (3), we can get: (4); (5);

[0090] in, is a column vector The time derivative, is a column vector The time derivative, , , , , , .

[0091] For Energy Nodes , allowing gas to flow to the energy node All energy nodes are called energy nodes Upstream nodes, energy nodes The upstream node set of , Energy Node The upstream node of is the number of the upstream node in the upstream node set, is the number of upstream nodes, correspondingly, let The mid-upstream nodes Energy Nodes The input gas pressure increment is , Represents the upstream node set The The upstream node Energy Nodes The input gas pressure increment is set from the upstream node The mid-upstream nodes Energy Nodes The output gas flow increment is , Represents the upstream node set The The upstream node Energy Nodes Output gas flow increment.

[0092] Let the gas flow from the energy node All energy nodes that flow to are called nodes Downstream nodes, energy nodes The set of downstream nodes is , Energy Node The downstream node of is the number of the downstream node in the downstream node set, is the number of downstream nodes, correspondingly, let Energy Nodes Downstream node collection The gas pressure increment input at all downstream nodes is , Energy Node To the downstream node set The gas pressure increment input from the downstream node is Energy Nodes Downstream node collection The gas flow increment output by all downstream nodes is , Energy Node To the downstream node set The gas flow increment output by the downstream node.

[0093] Order The gas grid state vector of energy nodes is According to the gas network state vector, construct the first The dynamic model of the gas grid subsystem is: (6);

[0094] in, The first The status of each gas grid subsystem, for The time derivative, for The time derivative, for The time derivative, , , , , , For the upstream node to The gas flow increment input by each energy node, For the The gas pressure increment output by an energy node to the downstream node, , , For the upstream node and The length of the natural gas pipeline between energy nodes, For the The length of the natural gas pipeline between the energy node and the downstream node, is the flow value at the steady-state operation point, is the pressure value at the steady-state operating point, where For interconnected items, interconnected items For outflow energy node The gas flow increment and the gas flow into the energy node The gas pressure increment is not the energy node. The flow and pressure conditions, where .

[0095] Step 1.2: Model the power network system and obtain the dynamic model of each power network subsystem.

[0096] Specifically: For the For the power network subsystem in each energy node, the control process is formulated as follows: (7);

[0097] in, For the i The state vector of the power network subsystem in each energy node, for The time derivative, For the The frequency increment of the generator set in the power network subsystem in the energy node, for The time derivative, For the The output power increment of the generator set in the power network subsystem in the energy node, for The time derivative, For the The increment of the speed regulator valve opening of the generator set in the power network subsystem in the energy node, for Derivative with respect to time; For the The time constant of the generator set in the power network subsystem in the energy node, For feedback adjustment gain, For the The time constant of the speed regulator in the power network subsystem in each energy node, For the The gain constant of the generator set in the power network subsystem in the energy node, For the The time constant of the steam turbine in the power network subsystem in the energy node, For the The gain constant of the steam turbine in the power network subsystem in the energy node, To control the input variables, For the The increase in electric power provided by the power network subsystem in each energy node.

[0098] make , , , for The state vector is composed of for and The column vector formed by for , and The column vector is composed of , , , , , , and other interconnected items , which is the global state of the electric-gas coupled energy system A complex nonlinear function, where For the The state vector of the energy subsystem is Accurate mathematical modeling is extremely difficult.

[0099] In summary, The dynamic model of the power network subsystem in the energy subsystem can be simplified as follows: (8);

[0100] Step 1.3: Integrate the dynamic model of the gas grid subsystem and the dynamic model of the power grid subsystem to construct a dynamic model of each energy subsystem in the electricity-gas coupled energy system.

[0101] (9);

[0102] in, for The time derivative, , Indicates The state vector of the energy subsystem The influence of various elements on the dynamic process inside the gas grid subsystem, such as Indicates The state vector of the energy subsystem The influence of each element on the dynamic process inside the gas grid subsystem. For the The state vector of the energy subsystem elements, ,like For the i The state vector of the energy subsystem elements, , , , , ..., , , ..., ; For the The unknown internal dynamics of the gas grid subsystem, For interconnected items No. elements, , is the input control vector, is the matrix of interconnected items.

[0103] Step 1.4: Use a neural network to approximate the function terms of the dynamic model of each energy subsystem in the electric-gas coupled energy system to obtain an approximate dynamic mathematical model of each energy subsystem.

[0104] Introducing the Stability Matrix , Designed for size The matrix of The size is The zero matrix of .

[0105] Approximation with Neural Networks The approximate dynamic mathematical model of the energy subsystem is: (10);

[0106] in, is an integer, and , , is the ideal value of the neural network weights, is the activation function, ,in is the approximation error of the neural network.

[0107] Step 2: Design an identifier of the electric-gas coupled energy system based on the approximate dynamic mathematical model of the electric-gas coupled energy system.

[0108] make is a Hurwitz matrix, then for any given positive definite matrix , there must be a unique positive definite matrix satisfy .

[0109] The identifier of the electric-gas coupled energy system is: (11);

[0110] in, For the The estimated value of the state vector of each energy subsystem, is the estimated value of the state vector elements, for Derivative with respect to time; is the activation function estimate, is the estimated value of the neural network weight, is the dimension of the vector of neural network weight estimates, for The estimated value of .

[0111] remember , , and the state error of the identifier of the electric-gas coupled energy system is The elements are in For the The first elements, the simplified expression combining the two formulas is: (12);

[0112] For the The time derivative of the state error of an energy subsystem.

[0113] The weight adaptation law of the identifier of the electric-gas coupled energy system is: (13);

[0114] in, for The time derivative, is the designed symmetric positive definite matrix, is the learning rate, is the state error of the identifier.

[0115] Set any and , in a tight set The ideal value of the neural network weights , activation function and neural network approximation error are all norm-bounded, that is, , , , , and They are the bounds of the ideal value norm of the neural network weights, the bounds of the activation function norm, and the bounds of the approximation error norm, all of which are positive constants.

[0116] In the electric-gas coupled energy system under study, the above assumptions are met and the designed identifier is used. When the corresponding identifier weight adaptive law is adopted, the state estimation error and weight estimation error are all uniformly ultimately bounded.

[0117] The identifier can reconstruct the unknown dynamics of the electric-gas coupled energy system online. The reconstructed energy subsystem is isolated, that is, there is no interconnection term. The dynamics of the isolated subsystem corresponding to the energy subsystem of the unknown electric-gas coupled energy system is ,along with Approaching , the energy subsystem dynamics identified by the identifier can be regarded as an approximate form of the isolated subsystem dynamics. It can be reduced by increasing the number of hidden layer nodes in the neural network.

[0118] Step 3: Treat each energy subsystem in the electric-gas coupled energy system as an isolated subsystem, and use the judgment network to update and iterate to solve the event-triggered approximate optimal control strategy of the electric-gas coupled energy system. At the same time, the weight adaptive law of the judgment network is designed based on the identifier of the electric-gas coupled energy system obtained in step 2; the isolated subsystem has no interconnection relationship with other subsystems, that is, there is no interconnection term in the dynamic model of the isolated subsystem, which is opposite to the interconnected subsystem.

[0119] Step 3.1: Define the performance index of the isolated subsystem and obtain the time-driven Hamiltonian function based on the performance index of the isolated subsystem.

[0120] Under certain conditions, the optimal control strategy of an isolated system is also the decentralized control strategy of the original interconnected system. Using the adaptive dynamic programming method to solve the optimal control strategy of the corresponding isolated subsystem can achieve the purpose of stabilizing the system.

[0121] The performance indicators of the isolated subsystem are: (14);

[0122] in, For the The performance index of an isolated subsystem is is the utility function, and are all symmetric positive definite matrices, For time.

[0123] set up exist If the above is continuously differentiable, then the infinitesimal form of the above formula can be obtained: (15);

[0124] in, yes about The partial derivative function of For the Performance indicators of an isolated subsystem.

[0125] For the convenience of expression, , is a smooth function, so the time-driven Hamiltonian function can be expressed as: (16);

[0126] in, is the time-driven Hamiltonian function.

[0127] Step 3.2: Define the optimal value function of the performance index of the isolated subsystem, and based on the optimal value function of the performance index of the isolated subsystem and the time-driven Hamiltonian function obtained in step 3.1, obtain the optimal control strategy and the corresponding time-triggered HJB equation.

[0128] First, define The optimal value function of the performance index of an isolated subsystem is: (17);

[0129] in, For the The optimal value function of the performance index of an isolated subsystem.

[0130] Setting the optimal value function of the performance index of the isolated subsystem Existence and in The above is differentiable, and the optimal control strategy is: (18);

[0131] in, is the optimal value function of the performance index of the isolated subsystem about The partial derivative of Input variables for optimal control.

[0132] The corresponding time-triggered HJB equation is: (19);

[0133] in, is the time-triggered HJB equation, is the utility function of the optimal control strategy, .

[0134] Step 3.3: In order to reduce the communication and computing burden of the entire electric-gas coupled energy system, consider using an event-triggered control mechanism. Based on the optimal control strategy obtained in step 3.2, an event-triggered optimal control strategy and the corresponding event-triggered HJB equation are established.

[0135] The event triggers the optimal control strategy as follows: (20);

[0136] in, is the event-triggered optimal control vector, , is the sampling state, For the A trigger moment, For the At the same time, the time-triggered HJB equation changes to the event-triggered form, that is, the event-triggered HJB equation is: (twenty one);

[0137] in, For event-triggered HJB equation, is the utility function of the event-triggered optimal control strategy.

[0138] It should be pointed out that the event triggering error is introduced , the event-triggered HJB equation (21) is not equal to zero.

[0139] Step 3.4: Use the neural network to approximate the optimal value function of the performance index of the isolated subsystem, obtain the optimal value function approximated by the neural network, and then obtain a new expression of the event-triggered optimal control strategy.

[0140] Continue to derive and set: optimal control strategy About event triggering error is Lipschitz continuous, that is, for any , there exists a positive constant Make , is the optimal control strategy derived from the actual state of the isolated subsystem, is the optimal control strategy derived from the sampled state, It is the sampling state.

[0141] Since the time-triggered HJB equation is a nonlinear partial differential equation, generally speaking, its optimal value function is It is difficult to obtain analytically, so the adaptive dynamic programming method is used here to approximate the optimal value function.

[0142] With the help of neural network, the optimal value function can be expressed as: (twenty two);

[0143] in, is the optimal value function approximated using a neural network, is the optimal weight of the neural network, is the number of neurons in the hidden layer, and is the activation function of the neural network, is the approximation error of the neural network.

[0144] Based on the above derivation, the new expression of the optimal control strategy is: (twenty three);

[0145] and They are and about The partial derivative of .

[0146] Similarly, the event-triggered optimal control strategy can be expressed as: (twenty four);

[0147] in, and They are and about The partial derivative of .

[0148] Step 3.5: Use the judgment network to approximately solve the new expression of the event-triggered optimal control strategy to obtain the event-triggered approximate optimal control strategy.

[0149] because is unknown, using the judgment network The output of is used to approximate the event-triggered optimal control strategy, where is the optimal weight of the neural network The estimated value of To judge the output value of the network, , the event-triggered approximate optimal control strategy finally designed is: (25);

[0150] in, is the approximate optimal control vector, is the estimated value of the optimal weight of the neural network.

[0151] It has been proved that the lower bound of the event triggering time interval exists. When the designed event triggering control mechanism is used, the Zeno phenomenon will not occur. At the same time, the cutoff triggering operation is introduced to avoid triggering unnecessary events. That is, when the system state has converged to the triggering threshold size range, no more triggering events will be generated.

[0152] Step 3.6: Based on the identifier of the electric-gas coupled energy system obtained in step 2, the error function of the event-triggered approximate optimal control strategy is minimized by estimating the optimal weights of the neural network in the approximate event-triggered approximate optimal control strategy of the evaluation network, and at the same time, a weight adaptation law of the evaluation network is designed to adjust the approximation effect of the neural network.

[0153] The weight adaptive law predicts the action value or directly outputs the optimal action based on the input state information. By continuously updating the weight, the performance indicators of the system are gradually optimized and eventually tend to the optimal solution. When solving the optimal control problem of the corresponding isolated subsystem, the optimal control strategy solved by the adaptive dynamic programming method can stabilize the system.

[0154] According to the dynamics of the identifier , the approximate Hamiltonian equation of the event-triggered approximate optimal control strategy is obtained as: (26);

[0155] in, is the approximate Hamiltonian equation for the event-triggered approximate optimal control strategy, which represents a function of the estimated values ​​of the state, the approximate optimal control vector, and the optimal weights of the neural network. is the utility function of the event-triggered approximate optimal control strategy, yes about The partial derivative of The parameters to be defined.

[0156] In addition, the definition and weight error , is a function of the state estimate, the approximate optimal control vector, and the optimal weights of the neural network.

[0157] Subtracting the two equations gives .

[0158] The goal of adaptive criterion design is to find an estimate of the optimal weights To minimize the error function ,in , using the normalized gradient descent algorithm, design an estimate of the optimal weights The adaptive law is: (27);

[0159] In the formula, for The time derivative, is the learning rate, , , is the weight error.

[0160] Due to the weight is a fixed value, The dynamics can be expressed as ,in for The derivative with respect to time.

[0161] For the energy subsystem under study, the continuous excitation condition is maintained and the designed adaptive law of the judgment network weight is used to ,exist , so that When the designed event-based decentralized control strategy is adopted, the energy subsystem state , Sampling status And judge the network weight estimation error They are all uniformly bounded. By choosing an appropriate Lyapunov function, it is proved that the designed closed-loop control system is stable, and the adaptive dynamic programming method is used to solve the optimal control strategy of the corresponding isolated subsystem to achieve the purpose of stabilizing the system.

[0162] Step 4: Utilize the event-triggered approximate optimal control strategy of the electric-gas coupled energy system and the weight adaptive law of the evaluation network to realize the control of the electric-gas coupled energy system.

[0163] Considering the high coupling and complex nonlinear characteristics of the electric-gas coupled energy system, a neural network model is used to reconstruct the unknown internal dynamics of the electric-gas coupled energy system. Then, an adaptive dynamic programming method is used to learn the event-triggered approximate optimal control strategy. At the same time, an event-triggered mechanism is adopted to reduce the communication burden between regions.

[0164] The above steps derive the approximate optimal control strategy, which is a set of action selection or decision rules that can achieve the optimal performance indicators (cost, efficiency, etc.) of the event-triggered approximate optimal control strategy system under constraints to achieve the predetermined goal. In the actual event-triggered approximate optimal control strategy system, it is necessary to obtain the current state of the system through sensors or other measuring devices, collect key data (such as gas pressure, temperature, flow, etc.) of the event-triggered approximate optimal control strategy system through sensors to obtain the sampling state of the system, conduct a preliminary analysis of the collected data, and determine whether the system deviates from the ideal operating state. In order to achieve the optimal performance indicators, the optimal control action to be taken in the current state is calculated according to the obtained approximate optimal control strategy. Specific control instructions are output through digital-to-analog conversion. These instructions usually include operation commands for various actuators (such as valves, pumps, motors, etc.) in the steam turbine, adjust the valve opening to change the gas pressure and flow, adjust the generator output, and change the operating state of the system.

[0165] An important feature of the adaptive dynamic programming method is that it can learn online, that is, it continuously improves its decisions as new data arrives. For the electric-gas coupled energy system, this means that it can flexibly adjust the control parameters according to real-time load demand, weather forecast, market price and other factors to ensure that the electric-gas coupled energy system always operates in the optimal state.

Claims

1. An intelligent edge control method for an information energy system based on adaptive dynamic programming, characterized in that: The following steps are involved: Step 1: Model the gas grid system and the power grid system in the electric-gas coupled energy system respectively, and then obtain the dynamic model of the electric-gas coupled energy system, and use the neural network to approximate the function terms in the dynamic model of the electric-gas coupled energy system to obtain the approximate dynamic mathematical model of the electric-gas coupled energy system; The electric-gas coupled energy system includes an electric power network system and a gas network system, the gas network system includes a plurality of gas network subsystems, the electric power network system includes a plurality of electric power network subsystems, and each gas network subsystem corresponds to a power network subsystem, and such a group of corresponding gas network subsystems and electric power network subsystems is called an energy subsystem; Step 2: Design an identifier of the electric-gas coupled energy system according to the approximate dynamic mathematical model of the electric-gas coupled energy system; Step 3: Treat each energy subsystem in the electric-gas coupled energy system as an isolated subsystem, and use the judgment network to update and iterate to solve the event-triggered approximate optimal control strategy of the electric-gas coupled energy system. At the same time, the weight adaptive law of the judgment network is designed based on the identifier of the electric-gas coupled energy system obtained in step 2; Step 4: Utilize the event-triggered approximate optimal control strategy of the electric-gas coupled energy system and the weight adaptive law of the evaluation network to realize the control of the electric-gas coupled energy system.

2. The intelligent edge control method of information energy system based on adaptive dynamic programming according to claim 1 is characterized in that: The step 1 specifically includes: Step 1.1: Model the gas grid system and obtain the dynamic model of each gas grid subsystem; For containing N The electric-gas coupled energy system with energy subsystems, each energy subsystem is regarded as an energy node. Energy Nodes , allowing gas to flow to the energy node All energy nodes are called energy nodes Upstream nodes, energy nodes The upstream node set of , Energy Node The upstream node of is the number of the upstream node in the upstream node set, is the number of upstream nodes, let The mid-upstream nodes Energy Nodes The input gas pressure increment is , Represents the upstream node set The The upstream node Energy Nodes The input gas pressure increment is set from the upstream node The mid-upstream nodes Energy Nodes The output gas flow increment is , Represents the upstream node set The The upstream node Energy Nodes Output gas flow increment; Let the gas flow from the energy node All energy nodes that flow to are called nodes Downstream nodes, energy nodes The set of downstream nodes is , Energy Node The downstream node of is the number of the downstream node in the downstream node set, is the number of downstream nodes, let Energy Nodes Downstream node collection The gas pressure increment input at all downstream nodes is , Energy Node To the downstream node set The gas pressure increment input from the downstream node is Energy Nodes Downstream node collection The gas flow increment output by all downstream nodes is , Energy Node To the downstream node set The gas flow increment output by the downstream node; Order The gas grid state vector of energy nodes is , according to the gas network state vector, construct the The dynamic model of the gas grid subsystem is: （6）； in, The first The status of each gas grid subsystem, for The time derivative, for The time derivative, for The time derivative, , , , , , For the upstream node to The gas flow increment input by each energy node, For the The gas pressure increment output by an energy node to the downstream node, , , For the upstream node and The length of the natural gas pipeline between energy nodes, For the The length of the natural gas pipeline between the energy node and the downstream node, is the flow value at the steady-state operation point, is the pressure value at the steady-state operating point, is the speed of sound in natural gas, is the cross-sectional area of ​​the natural gas pipeline, Indicates the friction factor when gas flows in a natural gas pipeline. is the diameter of the natural gas pipeline, where For interconnected items, ; Step 1.2: Model the power network system and obtain a dynamic model of each power network subsystem; No. The dynamic model of the power network subsystem in the energy subsystem is: （8）； in, For the The frequency increment of the generator set in the power network subsystem in the energy node, for The time derivative, For the The output power increment of the generator set in the power network subsystem in the energy node, for The time derivative, For the The increment of the speed regulator valve opening of the generator set in the power network subsystem in the energy node, for Derivative with respect to time; , , , To control the input variables, , For the The state vector of the power network subsystem in each energy node, , , , For interconnected items, is the global state of the electric-gas coupled energy system, For the The state vector of the energy subsystem, For the The time constant of the generator set in the power network subsystem in the energy node, For feedback adjustment gain, For the The time constant of the speed regulator in the power network subsystem in each energy node, For the The gain constant of the generator set in the power network subsystem in the energy node, For the The time constant of the steam turbine in the power network subsystem in the energy node, For the The gain constant of the steam turbine in the power network subsystem in the energy node, For the The increase in electric power provided by the power network subsystem in each energy node; Step 1.3: Integrate the dynamic model of the gas grid subsystem and the dynamic model of the power grid subsystem to construct a dynamic model of each energy subsystem in the electricity-gas coupled energy system; The dynamic model of each energy subsystem in the electric-gas coupled energy system is: （9）； in, for The time derivative, , For the The state vector of the energy subsystem elements, , Indicates The state vector of the energy subsystem The influence of each element on the dynamic process inside the gas grid subsystem. For interconnected items No. elements, , is the input control vector, is the matrix of interconnected items; Step 1.4: Use a neural network to approximate the function terms of the dynamic model of each energy subsystem in the electric-gas coupled energy system to obtain an approximate dynamic mathematical model of each energy subsystem; The approximate dynamic mathematical model of each energy subsystem is: （10）； in, is a stable matrix, Designed for size The matrix of The size is The zero matrix of , is the ideal value of the neural network weights and , is the activation function, ,in is the approximation error of the neural network.

3. The intelligent edge control method of information energy system based on adaptive dynamic programming according to claim 2 is characterized in that: The identifier of the electric-gas coupled energy system in step 2 is: （11）； in, For the The estimated value of the state vector of each energy subsystem, is the estimated value of the state vector elements, for Derivative with respect to time; is the activation function estimate, is the estimated value of the neural network weight, is the dimension of the vector of neural network weight estimates, for The estimated value of is the Hurwitz matrix; The weight adaptation law of the identifier of the electric-gas coupled energy system is: （13）； in, for The time derivative, is the designed symmetric positive definite matrix, is the learning rate, , , is a positive definite matrix, is the state error of the identifier: （12）； in, For the The time derivative of the state error of an energy subsystem.

4. The intelligent edge control method of information energy system based on adaptive dynamic programming according to claim 3 is characterized in that: Step 3 specifically includes: Step 3.1: Define the performance index of the isolated subsystem, and obtain the time-driven Hamiltonian function according to the performance index of the isolated subsystem; Step 3.2: Define the optimal value function of the performance index of the isolated subsystem, and based on the optimal value function of the performance index of the isolated subsystem and the time-driven Hamiltonian function obtained in step 3.1, obtain the optimal control strategy and the corresponding time-triggered HJB equation; Step 3.3: Establish an event-triggered optimal control strategy and the corresponding event-triggered HJB equation based on the optimal control strategy obtained in step 3.2; Step 3.4: Use the neural network to approximate the optimal value function of the performance index of the isolated subsystem, obtain the optimal value function approximated by the neural network, and then obtain a new expression of the event-triggered optimal control strategy; Step 3.5: Use the judgment network to approximately solve the new expression of the event-triggered optimal control strategy to obtain the event-triggered approximate optimal control strategy; Step 3.6: Based on the identifier of the electric-gas coupled energy system obtained in step 2, the error function of the event-triggered approximate optimal control strategy is minimized by estimating the optimal weights of the neural network in the approximate event-triggered approximate optimal control strategy of the judgment network, and the weight adaptation law of the judgment network is designed.

5. The intelligent edge control method of information energy system based on adaptive dynamic programming according to claim 4 is characterized in that: The performance indicators of the isolated subsystem described in step 3.1 are: （14）； in, For the The performance index of an isolated subsystem is is the utility function, and are all symmetric positive definite matrices, For time; The time-driven Hamiltonian function is: （16）； in, is the time-driven Hamiltonian function; , is a smooth function, yes about The partial derivative function of For the Performance indicators of an isolated subsystem.

6. The intelligent edge control method of information energy system based on adaptive dynamic programming according to claim 5 is characterized in that: The optimal value function of the performance index of the isolated subsystem described in step 3.2 is: （17）； in, For the The optimal value function of the performance index of an isolated subsystem; Setting the optimal value function of the performance index of the isolated subsystem Existence and in The above is differentiable, and the optimal control strategy is: （18）； in, is the optimal value function of the performance index of the isolated subsystem about The partial derivative of Input variables for optimal control; The corresponding time-triggered HJB equation is: （19）； in, is the time-triggered HJB equation, is the utility function of the optimal control strategy.

7. The intelligent edge control method of information energy system based on adaptive dynamic programming according to claim 6 is characterized in that: The event-triggered optimal control strategy described in step 3.3 is: （20）； in, is the event-triggered optimal control vector, , is the sampling state, For the A trigger moment, For the A trigger moment; The event-triggered HJB equation is: （21）； in, For event-triggered HJB equation, is the utility function of the event-triggered optimal control strategy.

8. The intelligent edge control method of information energy system based on adaptive dynamic programming according to claim 7 is characterized in that: The optimal value function of the neural network approximation in step 3.4 is: （22）； in, is the optimal value function approximated using a neural network, is the optimal weight of the neural network, is the number of neurons in the hidden layer, is the activation function of the neural network, is the approximation error of the neural network; The new expression of the event-triggered optimal control strategy is: （24）； in, and They are and about The partial derivative of .

9. The intelligent edge control method of information energy system based on adaptive dynamic programming according to claim 8 is characterized in that: Step 3.5: Using the judgement network The output of is used to approximate the event-triggered optimal control strategy, where is the optimal weight of the neural network The estimated value of To judge the output value of the network, To judge the network activation function; The event-triggered approximate optimal control strategy is: （25）； in, is the approximate optimal control vector, is the estimated value of the optimal weight of the neural network.

10. The intelligent edge control method of information energy system based on adaptive dynamic programming according to claim 9 is characterized in that: Step 3.6 Based on the dynamics of the identifier , the approximate Hamiltonian equation of the event-triggered approximate optimal control strategy is obtained as: （26）； in, is the approximate Hamiltonian equation for the event-triggered approximate optimal control strategy, is the utility function of the event-triggered approximate optimal control strategy, yes about The partial derivative of For the defined parameters; definition and weight error , is a function of the state estimate, the approximate optimal control vector, and the optimal weight of the neural network, and is defined as ; The error function of the event-triggered approximate optimal control strategy is: ,in , and then design the adaptive law of the evaluation network as: （27）； In the formula, for The time derivative, is the learning rate, , , is the weight error.

Images