A self-reasoning multi-level integrated energy intelligent agent modeling method and system

Abstract

A self-inference multi-level integrated energy agent modeling method and system is used to optimize the coordinated operation of multiple agents in the urban integrated energy system. The integrated energy system of the urban area is divided into multiple agent models according to the level. Starting from the local operation constraints of each agent and the global operation cost target, a multi-agent coordinated optimization scheduling model describing the entire system is established. Then the rules and constraint information obtained by self-inference are encoded into a parameterized decision graph representation, and the optimal graph mapping model for multi-agent coordinated scheduling is obtained through graph deep learning. Due to the use of self-inference and graph learning technology, the model can effectively capture the inherent physical mechanism of complex systems, and realize fast and accurate graph mapping solutions for global optimization problems containing a large number of integer variables, thereby saving a lot of computing costs for daily coordinated operation plans, and providing a new and effective way to solve the multi-agent collaborative optimization of urban integrated energy systems.

Description

Technical Field

[0001] The present invention relates to the technology in the field of multi-level intelligent agent modeling and analysis of urban regional integrated energy systems, and specifically to an autonomous reasoning multi-level integrated energy intelligent agent modeling method and system. Background Art

[0002] Cities are concentrated places for production, consumption and exchange, showing the characteristics of high energy consumption intensity and high carbon emissions. Urban integrated energy systems can integrate multiple energy sources, improve energy utilization efficiency, and strongly support the reliable, efficient and environmentally friendly supply of urban energy.

[0003] The urban integrated energy system includes a variety of energy resources, energy equipment and participants. These factors together constitute a complex and dynamic system that requires efficient scheduling, control and optimization to ensure reliable and efficient energy supply. Traditional centralized scheduling and control methods involve the participation of a large number of energy entities and require highly coordinated operation. It is difficult to cope with the complexity and uncertainty of large-scale integrated energy systems. When faced with complex systems, the scheduling decision-making complexity is high and the solution calculation is slow.

[0004] An agent is a unit structure that can sense the environment, make decisions, and communicate with other agents. In the field of electric power, the "agent" in the system refers to power equipment, energy resources, or energy management systems that have a certain degree of intelligence and autonomous decision-making capabilities. These agents sense the environment, exchange information, and make autonomous decisions and collaborate based on the overall goals and constraints of the system to optimize the operating efficiency, stability, and reliability of the power system.

[0005] Prior art document 1 (CN117852710A) discloses a method and system for collaborative optimization and scheduling of a multi-park integrated energy system, which models a single park integrated energy system, and establishes an overall model of the multi-park integrated energy system in combination with the internal energy market clearing mechanism; takes the lowest overall operating cost of the multi-park integrated energy system as the optimization goal, and establishes the system economic operation objective function; converts the overall model of the multi-park integrated energy system into a Markov game model of a multi-agent system, and converts the system economic operation objective function into a reward function, and establishes a multi-agent reinforcement learning task; uses a decomposed multi-agent deep policy gradient algorithm to solve the multi-agent reinforcement learning task, and obtains the system's optimal scheduling strategy. Its disadvantage is that this method mainly relies on the design of the reward function and the convergence of the reinforcement learning algorithm, which may lead to low efficiency when facing large-scale mixed integer programming problems, and it is difficult to ensure the global optimization effect.

[0006] Prior art document 2 (CN114091879A) discloses a multi-park energy scheduling method and system based on deep reinforcement learning, S1: construct a distributed park integrated energy system model, propose an optimized scheduling architecture for distributed park integrated energy system with optimal economic operation as the goal, and design the energy interaction between the parks, which is divided into a multi-park sharing layer and a single park consumption layer; S2: on the basis of the proposed architecture, a multi-agent-based deep reinforcement learning algorithm is used to solve the dynamic scheduling problem of the integrated energy system, and a multi-agent deep reinforcement learning framework for the distributed park integrated energy system is built; S3: replace the objective function with a real-time reward function, and use the interaction between the agents in each park and the environment to find the optimal scheduling strategy for the distributed park integrated energy system, so as to avoid the problem that the traditional energy scheduling method cannot respond to the random changes of source and load in real time; S4: use the test set data for the trained agent to make scheduling decisions, and compare the target cost obtained with the target cost obtained by the solver of the integrated energy system model after linearization to prove the effectiveness of the algorithm. Its shortcomings are that although this method can handle random changes in source and load, the model's solution accuracy and speed are still limited when faced with complex system constraints and a large number of integer variables; this method is based on multi-agent reinforcement learning to solve scheduling problems, and multi-agent interaction may produce comprehensive energy scheduling results outside the safe and feasible domain. Summary of the invention

[0007] Aiming at the shortcomings of the existing modeling and solution methods of multi-agent integrated energy systems, which cannot make full use of multi-source heterogeneous data and lack autonomous reasoning capabilities, the present invention proposes a multi-source data fusion autonomous reasoning integrated energy system graph learning modeling method, which integrates heterogeneous data sources from multiple systems such as energy production, transmission, storage, and consumption, and constructs a mathematical optimization model that comprehensively describes the coordinated operation of multi-agents in the urban area integrated energy system, and then constructs its decision graph expression model, designs a graph deep learning method based on a graph convolutional neural network model to perform autonomous reasoning on the decision graph, and mines the graph mapping relationship between the physical constraints and decision results in the global optimization scheduling model. The rules and constraint information obtained by autonomous reasoning are encoded into a parameterized decision graph model as the input of the global optimization scheduling problem, thereby transforming the solution of the multi-agent optimization scheduling problem of the integrated energy system into a graph mapping function, and solving the optimal collaborative scheduling strategy of the multi-agent units.

[0008] The present invention adopts the following technical solutions. The first aspect of the present invention provides an autonomous reasoning multi-level integrated energy intelligent agent modeling method and system, comprising the following steps:

[0009] Step 1: Divide each multi-agent park according to the structure of the integrated energy system to obtain multiple agent units;

[0010] Step 2: Conduct operation optimization modeling for each intelligent unit, with the optimization goal of minimizing the overall operation cost;

[0011] Step 3: Transform the global optimization scheduling model with a mixed integer programming structure into a bipartite graph of undetermined parameters;

[0012] Step 4: Generate a loss function based on graph convolution operations and Bernoulli conditions, and find a bipartite graph with optimal parameters through graph deep learning;

[0013] Step 5: Map the original optimization scheduling model to an easy-to-solve graph model through a graph map to obtain the values ​​of integer variables, and use it to solve the remaining global optimization sub-problems containing only continuous variables.

[0014] Preferably, in step 1, different device agents are divided according to devices of different energy forms and defined as an intelligent agent unit.

[0015] Preferably, in step 2, each intelligent unit optimizes the dispatch of the controlled equipment in an economically optimal manner, and the intelligent unit includes an electricity operation intelligent agent (Electricity Agent, EA) and a gas energy intelligent agent (Gas Agent, GA).

[0016] Preferably, in step 3, the specific steps are as follows:

[0017] Step 3.1: Construct an autonomous reasoning graph for the established global optimization scheduling model of the regional integrated energy system and abstract it into the form of a simplest optimization model.

[0018] Step 3.2: Construct the simplest optimization model of the above integrated energy system optimization scheduling model into a graph mapping model.

[0019] Preferably, in step 3.2, the nodes are divided into two non-intersecting sets, one of which represents decision variables and the other represents constraints. A bipartite graph is constructed, wherein the node types include variable nodes, equality constraint nodes, and inequality constraint nodes. At the same time, variable boundary constraints, integer constraints, equality constraints, and inequality constraint flags are assigned to the attribute values ​​of variable nodes. Non-numerical attributes are incorporated into the bipartite graph structure through the one-hot encoding method to participate in the graph convolution operation;

[0020] Preferably, in step 4, after obtaining the bipartite graph data describing the mixed integer programming problem, a graph convolutional neural network is used to construct a conditional generation model to achieve learning from the bipartite graph data to generate high-quality feasible solutions to the intelligent agent power allocation problem, and assign a greater probability distribution to the feasible solution of the better objective function. The uncertainty represented in the model prediction is used to fix the values ​​of most integer variables of the multi-agent optimization model, and the remaining sub-mixed integer programming problems can be quickly solved based on the branch and bound method equipped with the existing solver, so as to obtain high-quality feasible solution distribution.

[0021] Preferably, bipartite graph data is used as model input data, and a graph convolutional neural network is used to infer node embedding vectors of the input bipartite graph structure, and then a multi-layer perceptron module is used to input a mixed integer programming problem as a conditionally independent 0-1 integer variable value probability distribution, and the operation includes a graph convolution operation and a multi-layer perceptron operation, and the output value range is a probability distribution of integer variable values ​​between 0 and 1:

[0022]

[0023] in:

[0024] A is the adjacency matrix of the input graph data,

[0025] λ is the coefficient,

[0026] I N is the identity matrix,

[0027] is the degree matrix,

[0028] H is the feature vector of the input graph data,

[0029] W is the coefficient matrix of the fully connected layer.

[0030] Preferably, except for the input data A, the remaining variables are the parameters to be learned in the graph mapping model, i.e., the internal parameters of the graph model. The purpose of graph learning is to use the internal parameters and mapping structure of the graph model to replace the solution of the global optimization problem. Therefore, graph learning needs to find the optimal internal parameters, train the graph mapping model based on the graph deep learning algorithm, solve the optimal graph parameters, and design the loss function based on the Bernoulli conditional distribution as follows:

[0031]

[0032] Where:

[0033] ω ij represents the weight parameter of the jth feasible solution in the i-th sample,

[0034] θ represents the internal parameter of the graph model, that is, N represents the total number of samples in the data set.i represents the number of feasible solutions in the i-th sample, M i represents the bipartite graph of the i-th sample,

[0035] p θ (x i，j |M i ) is a feasible solution x i，j The generation probability of reflects the contribution of the feasible solution to the objective function and is defined as:

[0036]

[0037] Where:

[0038] d represents the value of the dth dimension of the decision variable, and I represents the total dimension of the decision variable, that is, the number of equipment scheduling decisions in all parks at all times.

[0039] Preferably, in step 5, the specific steps include:

[0040] Step 5.1: Use the trained graph model to solve the regional integrated energy system optimization scheduling model. Input the parameters such as electricity, heat, gas load demand, photovoltaic power generation power prediction into the graph model respectively. The graph model outputs the integer variable value of the power scheduling plan for each intelligent park.

[0041] Step 5.2: Fix the integer variable values ​​of the graph mapping model, solve the remaining integrated energy system optimization scheduling model containing only continuous variables, and obtain the global scheduling plan.

[0042] The second aspect of the present invention provides an autonomous reasoning graph learning modeling system for multi-source data fusion, which is used to run the autonomous reasoning graph learning modeling method for multi-source data fusion, including:

[0043] Multi-agent division module: divide urban area multi-agent units by energy type;

[0044] Global optimization scheduling module: using the internal constraints and common constraints of intelligent agents to establish an optimization scheduling model for the multi-agent integrated energy system in urban areas;

[0045] Spectrum mapping model: construct a bipartite graph for the optimal dispatch model of multi-agent integrated energy system in urban areas;

[0046] Graph solution model: The original optimization scheduling model is mapped to an easy-to-solve graph model through a graph for rapid solution.

[0047] Compared with the prior art, the beneficial effects of the present invention include at least: by establishing a mapping relationship of the multi-level intelligent agent decision graph of the urban regional integrated energy system, fast and accurate optimization can be achieved, the solution efficiency is improved, and the dependence on the adjustment of algorithm parameters is reduced. It can achieve more accurate and stable global optimization in a larger-scale integrated energy system, and can reduce the dimension of the optimization problem with large-scale integer variables in the regional integrated energy system, reduce the complexity of decision-making, and effectively guide the solution of local scheduling problems. This method is more applicable, especially for intelligent agents in non-continuous decision domains, such as energy storage and capacitor groups; the expression of the graph also improves the flexibility and adaptability of the system, so that the system can be easily migrated to other multi-level intelligent agent systems with different equipment parameters but similar structures. By integrating technologies such as multi-source heterogeneous data sources, autonomous reasoning to discover regular constraints, and graph learning modeling, the present invention constructs a bipartite graph and uses a graph convolutional neural network to directly learn and optimize the graph data, which can effectively capture the inherent physical mechanisms of complex systems and can more effectively handle complex system constraints and large-scale integer variable problems. The present invention uses graph neural networks and autonomous reasoning to directly learn the physical laws and constraints of the system from data, improves the accuracy and robustness of multi-agent scheduling decisions, and can improve the convergence speed of multi-agent interaction iteration while giving priority to ensuring the solution accuracy, achieving high-quality and fast solutions to mixed integer programming problems, greatly improving the real-time and accuracy of scheduling decisions, and effectively supporting real-time dynamic scheduling of large-scale systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a schematic diagram of the structure of a multi-park integrated energy system;

[0049] Figure 2 This is a schematic diagram of the bipartite graph representation method for the power allocation and start-stop optimization problem of the multi-agent model;

[0050] Figure 3 Schematic diagram of the model generation for multi-agent optimization solution prediction conditions;

[0051] Figure 4 is the simulation environment parameter;

[0052] Figure 5 Graph model training convergence process;

[0053] Figure 6 It is a schematic diagram of the equipment dispatching scheme for the power grid of the integrated energy system;

[0054] Figure 7 It is a schematic diagram of the gas network equipment dispatching scheme for the integrated energy system;

[0055] Figure 8 This is a schematic diagram of the equipment dispatching scheme for the heat network of the integrated energy system;

[0056] Fig. 9 The decision learning accuracy results of the graph network for the local optimization problem in the park;

[0057] Fig.10 Flowchart of the autonomous reasoning graph learning modeling method for multi-source data fusion. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The described embodiments are only embodiments of a part of the present invention, not all embodiments. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0059] This embodiment relates to a method for establishing a multi-agent graph model in an integrated energy system, comprising the following steps:

[0060] Step 1: Divide each multi-agent park according to the structure of the integrated energy system;

[0061] like Figure 1 As shown, the multi-park integrated energy system includes at least: Park 1 and Park 2.

[0062] The present invention divides different equipment agents according to the equipment in different energy forms, and defines them as an intelligent unit. For example but not limited to: photovoltaics, energy storage, and electric heat pumps belong to one intelligent unit, while gas turbines and gas boilers belong to another intelligent unit. For parks, multiple energy equipment aggregation agents in an area are defined as a park. A park generally refers to a relatively closed or clear geographical area, including an industrial park, a science and technology park, a university campus or a commercial area. These parks usually have their own unique energy demand and supply characteristics, including electricity, heat, cold energy, natural gas, and may also include their own energy production facilities and energy storage equipment.

[0063] Energy production facilities include solar panels, wind power, and small gas turbines.

[0064] Step 2: Conduct operation optimization modeling, with the optimization goal of minimizing the overall operation cost;

[0065] The main body of the integrated energy system equipment operator is divided into energy forms and regarded as an intelligent unit. Each intelligent unit optimizes the dispatch of the controlled equipment at the economic optimum. In a park, the electricity, heat and gas equipment aggregators are respectively regarded as intelligent entities, and optimize the dispatch of the controlled equipment at the economic optimum:

[0066] Step 2.1 Set the scheduling goals and constraints of EA. The specific process is as follows:

[0067] The Electricity Agent (EA) includes photovoltaics, energy storage and electric heat pumps. The operating cost is composed of electricity purchase costs and equipment operating costs. The scheduling objectives are:

[0068]

[0069] in:

[0070] Φ is the set of EA equipment decision variables, that is, Φ={P grid,t ,P χ,t},t∈Τ,χ∈Ω ele

[0071] Ω ele is the EA equipment set, Τ is the scheduling period set

[0072] P grid,t is the electric energy purchased from the upper power grid during period t,

[0073] c ele,t is the price of electricity purchased from the upper power grid during period t,

[0074] P χ,t The electric energy generated or consumed by the EA equipment during period t,

[0075] c χ It is the operating cost of EA equipment per unit of power generation or energy consumption. EA constraints include: equipment operation constraints, power flow constraints, and power balance constraints.

[0076] a) Operation constraints of photovoltaic, energy storage and electric heat pump equipment:

[0077]

[0078] in:

[0079] P pv,i,t is the active power output of the photovoltaic device connected to node i at time t,

[0080] is the maximum power predicted by photovoltaic,

[0081] Q pv,i,t is the reactive power output of the photovoltaic inverter,

[0082] S pv,i is the photovoltaic inverter capacity,

[0083] SOC i,t is the energy storage charge state,

[0084] SOC i are the upper and lower limits of the energy storage charge state,

[0085] P ch,i,t , P dch,i,t is the energy storage charging / discharging power,

[0086] The upper limit of energy storage charging and discharging power,

[0087] H eb,i,t Producing heat for electric heat pumps,

[0088] η eb is the efficiency of the electric heat pump,

[0089] P eb,i,t is the operating power of the electric heat pump,

[0090] P eb,max is the power of the electric heat pump,

[0091] U dch,i,t , U ch,i,t is the charge and discharge state variable of the energy storage battery, U dch,i,t =1 for charging, U ch,i,t =1 is discharging. b) Power flow constraint:

[0092]

[0093] Where:

[0094] I ij,t ,P ij,t ,Q ij,t is the line ij current, active power, and reactive power,

[0095] r ij ,x ij is the resistance and reactance of the grid line ij,

[0096] V i is the upper / lower limit of the voltage amplitude at node i,

[0097] is the upper limit of the current of the grid line ij.

[0098] c) Power balance constraint, that is, the power network satisfies the active power balance of the main grid power, gas-fired power generation equipment, EA equipment and load participation respectively:

[0099]

[0100] Where:

[0101] Ω Eis the EA node set, j is the park number, Θ is the park set, P sell,j, Selling electricity to the park.

[0102] EA nodes represent a set of nodes in the power network topology.

[0103] Step 2.2 sets the scheduling objective function and constraints of GA. The specific process is as follows:

[0104] The objective function of the gas energy agent (GA) is to minimize the gas purchase cost and equipment operation cost. The objective function is:

[0105]

[0106] in:

[0107] Ξ is the set of GA decision variables, that is, Ξ={G grid,t ,P δ,t}t∈Τ,δ∈Ω gas ,

[0108] G grid,t is the gas purchase volume of the natural gas main grid during period t,

[0109] Ω gas is the GA equipment set, T is the scheduling period set,

[0110] G gas,t is the gas purchase volume during period t,

[0111] c gas,t is the natural gas price during period t,

[0112] P δ,t is the gas consumption of GA equipment during period t,

[0113] c δ It is the operating cost of GA equipment per unit gas consumption.

[0114] The constraints of GA include: gas turbine gas boiler operation constraints, gas network flow constraints, gas balance constraints, heat network flow constraints and thermal power balance.

[0115] a) Gas turbine and gas boiler operation constraints:

[0116]

[0117] Where:

[0118] P gt,i,t Represents the gas-to-electricity output of the gas turbine (GT),

[0119] Hgt,i,t Represents the heat output of GT gas heating.

[0120] G gt,i,t Enter the gas volume for GT,

[0121] η gt,e With η gt,e They are GT gas and heating efficiency,

[0122] P gb,i,t Represents the heat output of gas heating from gas boiler (GB),

[0123] P gt is the maximum and minimum output power of GT,

[0124] P gb is the maximum and minimum output power of GB;

[0125] b) Gas network flow constraints:

[0126]

[0127] Where:

[0128] i,j,k are network node indexes,

[0129] Ω C is the set of compressor nodes,

[0130] G ji,t ,G pack,ij,t is the flow and storage of gas network pipeline ij,

[0131] G C,ij,t is the flow rate through the compressor,

[0132] η C,zj is the compressor energy consumption coefficient,

[0133] π i,t is the air pressure at node i of the air network,

[0134] W ij is the Weymouth coefficient of the gas network pipeline ij,

[0135] K pack,ij is the storage coefficient of the gas network pipeline,

[0136] π i , is the upper / lower limit of the gas pressure at gas network node i,

[0137] κ, is the maximum / minimum compression ratio of the compressor,

[0138] It is the upper limit of gas network pipeline / compressor flow rate.

[0139] c) Gas balance constraint, that is, the gas flow of node i is the difference between the main network gas purchase volume and the gas consumption of equipment and load:

[0140]

[0141] Where:

[0142] Ω G It is a collection of gas network nodes;

[0143] The heat load power is provided by the electric-to-heat intelligent device and the gas-to-heat intelligent device, and needs to meet the heat network flow constraints and heat power balance:

[0144]

[0145] Where:

[0146] HC w is the specific heat capacity of the heating medium,

[0147] m ij,t is the flow rate of the heat network pipeline ij,

[0148] T s,i,t is the supply medium temperature,

[0149] T r,i,t is the reflux medium temperature,

[0150] L ij ,γ ij is the length / heat transfer coefficient of the heating network pipe ij,

[0151] T env,t is the external environment temperature,

[0152] T in,j is the temperature of the medium flowing into node j,

[0153] T out,j is the medium temperature outflowing from node j,

[0154] H load,i,t The heat load demand.

[0155] Step 2.3 Global scheduling objectives and operation constraints. The specific process is as follows:

[0156] The global scheduling goal is to minimize the total operating cost of all parks while satisfying the operating constraints of all parks, as shown below:

[0157]

[0158] Where j represents the park index, p j (Φ j )≤0 represents the equality constraint of park j, h j (Φ j )=0 represents the inequality constraint of park j, F(Φ 1 ,...Φ J )=0 represents the coupling constraint of multiple parks, that is, the power balance constraint of the power network. The global represents multiple parks, and the global scheduling plan represents the power scheduling plan of all equipment in multiple parks. The EA and GA in all parks are put into one model to solve the global scheduling target, and finally the optimal solution for the coordinated scheduling of multiple parks is obtained.

[0159] Step 3: Transform the global optimization scheduling model with a mixed integer programming structure into a bipartite graph of undetermined parameters. The specific process is as follows:

[0160] Step 3.1: Construct an autonomous reasoning graph for the global dispatch optimization problem established in 2.3. Abstract the global optimization dispatch model of the regional integrated energy system into the form of the simplest optimization model, that is, replace the global optimization model (10) with the following form, but do not change the content of the global optimization model:

[0161]

[0162] Where:

[0163] x is the decision variable of the global optimization scheduling problem, that is, x = {Φ j ,Ξ j}j=0,1,...J,

[0164] Ax≤b and l≤x≤u are the representations of global inequality constraints, that is, the matrix expression of all inequality constraints of the form Ax≤b and l≤x≤u in equations (2), (3), (6), and (7).

[0165] A represents the adjacency matrix of the input graph data,

[0166] b represents the inequality constraint bias vector,

[0167] l represents the lower bound vector of the decision variables,

[0168] u represents the upper bound vector of the decision variable,

[0169] Dx=e is the matrix expression of the equality constraints in equations (2), (3), (4), (6), (7), and (9).

[0170] D represents the coefficient matrix of the equality constraint,

[0171] e represents the right-hand side value vector of the equality constraint,

[0172] c T represents the coefficient vector of the objective function,

[0173] Represents the decision variable x={Φ j ,Ξ j The part of i∈I in}j=0,1,...J is a discontinuous decision, that is, the decision on the charging and discharging state of the energy storage battery in EA or GA.

[0174] Step 3.2: construct the above-mentioned integrated energy system optimization scheduling model into a graph mapping model. Figure 2 In the bipartite graph shown in , the nodes are divided into two disjoint sets, one of which represents decision variables and the other represents constraints. Figure 2 The bipartite graph is constructed in the manner shown in the figure. The node types include variable nodes, equality constraint nodes, and inequality constraint nodes.

[0175] a ij is the element in the inequality constraint matrix A of the global optimization problem of the integrated energy system, representing the variable x j The weight connected to the constraint node of the i-th inequality constraint, a ij =0 means no connection,

[0176] d ij is an element in the matrix D, representing the variable x j The weight connected to the i-th equality constraint node, d=0 means no connection.

[0177] At the same time, the variable boundary constraints (l≤x≤u), integer constraints (x i ∈Z, i∈I), equality constraints and inequality constraints mark the attribute values ​​of the variable nodes. The non-numerical attributes are incorporated into the bipartite graph structure through the one-hot encoding method to participate in the graph convolution operation.

[0178] The attribute value is added to the bipartite graph as a bipartite graph node vector element. The form of the bipartite graph data is as follows Figure 2 As shown in Figure 2. In graph theory, bipartite graph data consists of two main elements, the adjacency matrix and the eigenvector. The adjacency matrix contains Figure 2 a ij and d ij , the shape of the matrix is ​​(m+p)×n, m represents a ij The number of rows, p represents d ij The number of rows, n represents a ij and d ijThe adjacency matrix is ​​a vector of the right-hand boundary values ​​of the inequality constraints and the equality constraints and the non-numeric attributes. The non-numeric attribute s_n indicates the variable x. n Is it an integer variable?

[0179] Step 4: Generate a loss function based on graph convolution operations and Bernoulli conditions, and find a bipartite graph with optimal parameters through graph deep learning;

[0180] After obtaining the bipartite graph data describing the mixed integer programming problem, a conditional generation model is constructed using a graph convolutional neural network to learn and generate high-quality feasible solutions to the agent power allocation problem from the bipartite graph data, that is, assigning a greater probability distribution to the feasible solution of the better objective function. The uncertainty represented in the model prediction is used to fix the values ​​of most integer variables in the multi-agent optimization model, and the remaining sub-mixed integer programming problems can be quickly solved based on the branch and bound method equipped with the existing solver, thereby obtaining a high-quality feasible solution distribution.

[0181] The conditional generative model is constructed as Figure 3 As shown, the bipartite graph data is used as the model input data, and the graph convolutional neural network is used to infer the node embedding vector of the input bipartite graph structure [c 1 ,...c n ,b 1 ,...,b m ,e 1 ,...,e p ,s 1 ,...s n ], and then the multi-layer perceptron module is used to input the mixed integer programming problem as the conditional independent 0-1 integer variable value probability distribution. The operation includes graph convolution operation and multi-layer perceptron operation, and the output value range is the probability distribution of integer variable values ​​between 0 and 1:

[0182]

[0183] in:

[0184] A is the adjacency matrix of the input graph data,

[0185] λ is the coefficient,

[0186] I N is the identity matrix,

[0187] is the degree matrix,

[0188] H is the feature vector of the input graph data,

[0189] W is the coefficient matrix of the fully connected layer.

[0190] In addition to the input data A, the remaining variables are the parameters to be learned in the graph mapping model, that is, the internal parameters of the graph model. The purpose of graph learning is to use the internal parameters and mapping structure of the graph model to replace the solution of the global optimization problem, so graph learning needs to find the optimal internal parameters.

[0191] Based on the graph deep learning algorithm, the graph mapping model is trained to find the optimal graph parameters. Based on the Bernoulli conditional distribution, the loss function is designed as

[0192]

[0193] Where:

[0194] ω ij represents the weight parameter of the jth feasible solution in the i-th sample,

[0195] θ represents the internal parameter of the graph model, that is, N represents the total number of samples in the data set. i represents the number of feasible solutions in the i-th sample, M i Represents the bipartite graph of the i-th sample. ,

[0196] p θ (x i，j |M i ) is a feasible solution x i，j The generation probability of reflects the contribution of the feasible solution to the objective function and is defined as:

[0197]

[0198] Where:

[0199] d represents the value of the dth dimension of the decision variable, and I represents the total dimension of the decision variable, that is, the number of equipment scheduling decisions in all parks at all times.

[0200] Step 5: Map the original optimization scheduling model to an easy-to-solve graph model through a graph map to obtain the values ​​of integer variables, and use it to solve the remaining global optimization sub-problems containing only continuous variables.

[0201] Step 5.1: Use the trained graph model to solve the regional integrated energy system optimization scheduling model. Input the parameters such as electricity, heat, gas load demand, photovoltaic power generation power prediction into the graph model respectively. The graph model outputs the integer variable value of the power scheduling plan for each intelligent park.

[0202] Step 5.2: Fix the integer variable values ​​of the graph mapping model, solve the remaining integrated energy system optimization scheduling model containing only continuous variables, and obtain the global scheduling plan.

[0203] In order to verify the effectiveness of the present invention, a comprehensive energy system of a certain park is used as a simulation object, and a simulation verification analysis is performed based on the following parameters. The specific parameter settings are as follows: Figure 4 And as shown in Table 1.

[0204] The example test of the present invention includes the following specific steps:

[0205] 1. First, based on historical operation scenario data samples and known equipment parameters, a series of global optimization scheduling models are constructed as data set samples.

[0206] 2. The training set and test set are divided into 8:2 ratios. The input is the bipartite graph representing the global optimization scheduling model in the sample, and the output is the value of the integer variable in the global optimization scheduling model.

[0207] 3. Train the graph mapping model by minimizing the loss function (13). The training convergence process is as follows: Figure 5 shown.

[0208] 4. Test in a new scenario and input new scenario parameters, i.e. load forecast and renewable energy generation forecast results, e.g. Fig.10 As shown in the figure, the value of integer variables is predicted by graph mapping, and then the remaining sub-optimal scheduling problems containing continuous variables are solved. Finally, the optimal scheduling results of the integrated energy system are obtained. The energy scheduling results and energy balance results of the power grid, gas grid and heat network are shown in Figure 6 , Figure 7 , Figure 8 shown.

[0209] 5. Tested in the test set samples, the decision learning accuracy of the graph network for the local optimization problem of the park is as follows Fig. 9 As shown, it contains whether the predicted integer variable is feasible and the size of the gap from the optimal objective function value.

[0210] Table 1 Parameter settings of each device

[0211]

[0212] Embodiment 2 of the present invention provides an autonomous reasoning graph learning modeling system for multi-source data fusion, which is used to run the autonomous reasoning graph learning modeling method for multi-source data fusion described in Embodiment 1, including:

[0213] Multi-agent division module: divide urban area multi-agent units by energy type;

[0214] Global optimization scheduling module: using the internal constraints and common constraints of intelligent agents to establish an optimization scheduling model for the multi-agent integrated energy system in urban areas;

[0215] Spectrum mapping model: construct a bipartite graph for the optimal dispatch model of multi-agent integrated energy system in urban areas;

[0216] Graph solution model: The original optimization scheduling model is mapped to an easy-to-solve graph model through a graph for rapid solution.

[0217] Compared with the prior art, the beneficial effects of the present invention include at least: by establishing a mapping relationship of the multi-level intelligent agent decision graph of the urban regional integrated energy system, fast and accurate optimization can be achieved, the solution efficiency is improved, and the dependence on the adjustment of algorithm parameters is reduced. It can achieve more accurate and stable global optimization in a larger-scale integrated energy system, and can reduce the dimension of the optimization problem with large-scale integer variables in the regional integrated energy system, reduce the complexity of decision-making, and effectively guide the solution of local scheduling problems. This method is more applicable, especially for intelligent agents in non-continuous decision domains, such as energy storage and capacitor groups; the expression of the graph also improves the flexibility and adaptability of the system, so that the system can be easily migrated to other multi-level intelligent agent systems with different equipment parameters but similar structures. By integrating technologies such as multi-source heterogeneous data sources, autonomous reasoning to discover regular constraints, and graph learning modeling, the present invention constructs a bipartite graph and uses a graph convolutional neural network to directly learn and optimize the graph data, which can effectively capture the inherent physical mechanisms of complex systems and more effectively handle complex system constraints and large-scale integer variable problems. The present invention uses graph neural networks and autonomous reasoning to directly learn the physical laws and constraints of the system from data, thereby improving the accuracy and robustness of multi-agent scheduling decisions, and can improve the convergence speed of multi-agent interaction iterations while giving priority to ensuring the accuracy of the solution, thereby achieving high-quality and rapid solutions to mixed integer programming problems, greatly improving the real-time and accuracy of scheduling decisions, and effectively supporting real-time dynamic scheduling of large-scale systems.

[0218] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0219] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the present invention. Although the present invention is described in detail with reference to the above embodiments, those skilled in the art should understand that:

[0220] The specific embodiments of the present invention may still be modified or replaced by equivalents, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention shall be included in the scope of protection of the claims of the present invention.

Claims

1. An autonomous reasoning multi-level integrated energy agent modeling method, characterized in that: The steps include: Step 1: Divide each multi-agent park according to the structure of the integrated energy system to obtain multiple agent units; Step 2: Conduct operation optimization modeling for each intelligent unit, with the optimization goal of minimizing the overall operation cost; Step 3: Convert the global optimization scheduling model with mixed integer programming structure into a bipartite graph of undetermined parameters; Step 3 specifically includes: Step 3.1: Construct an autonomous reasoning graph for the established global optimization scheduling model of the regional integrated energy system, and abstract it into the form of a simplest optimization model; Step 3.2: Construct the simplest optimization model of the above-mentioned integrated energy system optimization scheduling model into a graph mapping model; In step 3.2, the nodes are divided into two non-intersecting sets, one of which represents decision variables and the other represents constraints; construct a bipartite graph, in which the node types include variable nodes, equality constraint nodes, and inequality constraint nodes; classify the variable boundary constraints, integer constraints, equality constraints, and inequality constraint signs into the attribute values ​​of the variable nodes; incorporate non-numerical attributes into the bipartite graph structure through the one-hot encoding method to participate in the graph convolution operation; Step 4: Based on the graph convolution operation and Bernoulli condition generation loss function, a bipartite graph with optimal parameters is found through graph deep learning; in step 4, after obtaining the bipartite graph data describing the mixed integer programming problem, a graph convolutional neural network is used to construct a conditional generation model; the uncertainty represented in the model prediction is used to fix the value of the integer variable of the multi-agent optimization model, and the remaining sub-mixed integer programming problems are solved based on the branch and bound method equipped with the existing solver to obtain a feasible solution allocation; Step 5: Map the original optimization scheduling model to an easy-to-solve graph model through a graph map to obtain the values ​​of integer variables, and use it to solve the remaining global optimization sub-problems containing only continuous variables.

2. The autonomous reasoning multi-level integrated energy intelligent agent modeling method according to claim 1, characterized in that: In step 1, different device agents are divided according to devices with different energy forms and defined as an intelligent unit.

3. The autonomous reasoning multi-level integrated energy intelligent agent modeling method according to claim 1 is characterized by: In step 2, each intelligent unit optimizes the dispatch of the controlled equipment in an economically optimal manner, and the intelligent units include the power operation intelligent agent and the gas energy intelligent agent.

4. The autonomous reasoning multi-level integrated energy intelligent agent modeling method according to claim 1 is characterized by: The bipartite graph data is used as the model input data, and the graph convolutional neural network is used to infer the node embedding vector of the input bipartite graph structure. Then, the multi-layer perceptron module is used to input the mixed integer programming problem as the conditional independent 0-1 integer variable value probability distribution. The operation includes graph convolution operation and multi-layer perceptron operation, and the output value range is the probability distribution of integer variable values ​​between 0 and 1: in: A is the adjacency matrix of the input graph data, λ is the coefficient, I N is the identity matrix, is the degree matrix, H is the feature vector of the input graph data, W is the coefficient matrix of the fully connected layer.

5. The autonomous reasoning multi-level integrated energy intelligent agent modeling method according to claim 4 is characterized by: In addition to the input data A, the remaining variables are the parameters to be learned in the graph mapping model; Based on the graph deep learning algorithm, the graph mapping model is trained to solve the optimal graph parameters. Based on the Bernoulli conditional distribution, the loss function is designed as Where: ω ij represents the weight parameter of the jth feasible solution in the i-th sample, θ represents the internal parameter of the graph model, that is, N represents the total number of samples in the data set. i represents the number of feasible solutions in the i-th sample, M i represents the bipartite graph of the i-th sample, p θ (x i，j |M i ) is a feasible solution x i，j The generation probability of reflects the contribution of the feasible solution to the objective function and is defined as: Where: d represents the value of the dth dimension of the decision variable, and I represents the total dimension of the decision variable, that is, the number of equipment scheduling decisions in all parks at all times.

6. The autonomous reasoning multi-level integrated energy intelligent agent modeling method according to claim 1 is characterized by: Step 5 includes: Step 5.1: Apply the trained graph model to solve the regional integrated energy system optimization scheduling model, input the electricity, heat, gas load demand, and photovoltaic power generation power prediction parameters into the graph model, and the graph model outputs the integer variable value of the power scheduling plan for each intelligent park; Step 5.2: Fix the integer variable values ​​of the graph mapping model, solve the remaining integrated energy system optimization scheduling model containing only continuous variables, and obtain the global scheduling plan.

7. An autonomous reasoning graph learning modeling system, used to run an autonomous reasoning multi-level integrated energy intelligent agent modeling method according to any one of claims 1-6, characterized in that: include: Multi-agent division module: divide urban area multi-agent units by energy type; Global optimization scheduling module: using the internal constraints and common constraints of intelligent agents to establish an optimization scheduling model for the multi-agent integrated energy system in urban areas; Spectrum mapping model: construct a bipartite graph for the optimal dispatch model of multi-agent integrated energy system in urban areas; Graph solution model: The original optimization scheduling model is mapped to an easy-to-solve graph model through a graph for rapid solution.

Images