Distributed optimization scheduling method and system for integrated energy system considering privacy protection

Abstract

The application discloses a kind of comprehensive energy system distributed optimization scheduling method and system considering privacy protection, it is related to comprehensive energy system technical field, including: with each equipment in system as node, with energy interaction between node as edge, construct the communication network topology diagram of comprehensive energy system;With total cost minimum as objective function, configure the constraint condition of system operation, build comprehensive energy system optimization scheduling model;Assume that communication network topology diagram is undirected connected graph, using the preset time distribution optimization algorithm of introducing time-varying gain, the optimization scheduling model is solved, so that comprehensive energy system converges to optimal solution from arbitrary initial state within preset time, obtain optimal optimization strategy;Wherein, in solving process, by encryption and decryption module based on deep learning adversarial network, the interaction variable between adjacent nodes is encrypted and decrypted, privacy protection is realized.The application can meet the fast convergence and privacy protection requirements of comprehensive energy system scheduling optimization solution.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of integrated energy systems, and particularly relates to a privacy-considered distributed optimization scheduling method and system for integrated energy systems. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] An integrated energy system is a typical multi-agent information-physical system, which internally couples multiple heterogeneous energies such as electricity, heat, gas, and cold to achieve multiple energy complementation and mutual aid and meet user energy demand. At the physical layer, the system contains multiple energy forms and supplies energy to the load through different energy conversion devices; at the information layer, each node in the system interacts with information to ensure the coordinated and efficient operation of the system. In the context of large energy consumption and high consumption of fossil energy, it is of great strategic significance to establish an integrated energy system with low carbon, environmental protection, and high energy utilization rate.

[0004] Currently, the optimization scheduling problem of an integrated energy system can be solved by centralized optimization and distributed optimization. The centralized optimization method requires global information and has large communication data volume and poor communication network robustness. Although the existing distributed optimization method can solve these problems to some extent, as the system scale expands, the optimization scheduling of the system requires higher convergence speed of the optimization algorithm and privacy protection, and the existing distributed optimization method cannot effectively cope with it. For the low-carbon economic optimization scheduling problem of a large-scale integrated energy system with higher convergence speed and privacy protection requirements, although the preset time algorithm for solving the high convergence speed problem has been proposed, the existing method does not consider the privacy protection problem in the solving process, which cannot complete the encryption task within the minimum time interval required by the algorithm, and the accuracy and efficiency of the existing encryption method are low, which cannot be directly applied to the algorithm solving process. SUMMARY

[0005] To solve the above problems of the prior art, the present application provides a privacy-considered distributed optimization scheduling method and system for integrated energy systems, which uses a preset time distributed optimization algorithm independent of the initial state of the system to solve the optimization scheduling strategy of the integrated energy system, and uses Lyapunov's second method to prove that the algorithm can guarantee the preset time convergence of the system from any initial state, and designs an encryption and decryption module to be applied to the solving process, thereby improving the privacy of the interactive information between nodes in the optimization solving process and meeting the privacy protection requirements of the scheduling optimization solution.

[0006] In a first aspect, the present application provides a privacy-considered distributed optimization scheduling method for integrated energy systems.

[0007] A privacy-protected integrated energy system distributed optimization scheduling method, comprising:

[0008] A communication network topology graph of the integrated energy system is constructed by taking each device in the system as a node and energy interaction between nodes as an edge.

[0009] An optimization scheduling model of the integrated energy system is constructed by taking the minimum total cost as an objective function and configuring constraint conditions of system operation.

[0010] The communication network topology graph is assumed to be an undirected connected graph, and a preset time distributed optimization algorithm with time-varying gain is adopted to solve the optimization scheduling model, so that the integrated energy system converges to an optimal solution from an arbitrary initial state within a preset time, and an optimal optimization strategy is obtained; wherein, in the solving process, the interaction variables between neighbor nodes are encrypted and decrypted through the encryption and decryption module based on the deep learning adversarial network, and privacy protection is realized.

[0011] In a further technical solution, the encryption and decryption module based on the deep learning adversarial network is implemented based on a multi-agent game framework, and includes an encrypter Alice, a decrypter Bob and an attacker Eve constructed based on a neural network.

[0012] The encrypter Alice and the decrypter Bob form a cooperative alliance, and the joint key is deployed in the neighbor nodes; wherein, the interaction variables between the neighbor nodes are taken as plaintext vectors, the encrypter Alice learns to generate ciphertext vectors by taking the plaintext vectors and the key as inputs, and the decrypter Bob learns to reconstruct the original plaintext vectors by taking the ciphertext vectors and the same key as inputs.

[0013] The attacker Eve forms an adversarial game with the cooperative alliance, and guides the encrypter Alice and the decrypter Bob to learn complex nonlinear encryption transformation through adversarial training.

[0014] In a further technical solution, the encrypter Alice, the decrypter Bob and the attacker Eve all adopt a multi-layer perceptron with deep residual as a basic framework, and the multi-layer perceptron with deep residual includes an input linear layer, a core network stacked by a plurality of residual blocks and an output linear layer.

[0015] Wherein, each residual block is defined as H(X) = X + F(X), F(X) is a residual mapping learned by the network layer, and X is the original input transmitted through the shortcut connection; the network layer includes linear layers, batch normalization layers and activation layers connected in turn.

[0016] In a further technical solution, the encryption and decryption module adopts a two-stage update strategy for training in each training batch, including:

[0017] In Phase 1, the network parameters of the encryptor Alice and the decryptor Bob are frozen. The ciphertext vector generated by the encryptor Alice is passed to the attacker Eve, and the weights of the encryptor Alice are not updated. Based on the minimization of the reconstruction loss of the attacker Eve, backpropagation is performed to update the network weights of the attacker Eve.

[0018] In Phase Two, the network parameters of attacker Eve are frozen, and encryptor Alice and decryptor Bob perform a complete encryption and decryption process. Backpropagation is then performed based on minimizing the composite loss function, while the network weights of encryptor Alice and decryptor Bob are updated.

[0019] A further technical solution involves end-to-end optimization of the encryption / decryption module using an asymmetric loss function. The composite loss function for the cooperative alliance formed by the encryptor Alice and the decryptor Bob is as follows:

[0020] ;

[0021] The L1 reconstruction loss function of the decryptor Bob for:

[0022] ;

[0023] The L1 reconstruction loss function of the attacker Eve for:

[0024] ;

[0025] Where U is the original plaintext vector, U' is the decryption result of decryptor Bob, and U'' is the cracking result of attacker Eve. This is a hyperparameter.

[0026] A further technical solution employs a pre-defined time-distributed optimization algorithm incorporating time-varying gain to solve the optimized scheduling model, as follows:

[0027] ;

[0028] in, , For gradient, Indicates the first i The cost function of each node Indicates the first i The power of each node; For projection operators, It is a closed convex set. For about Inequality constraint functions; definition For the Laplace operator of a graph, Represents the adjacency matrix. denotes the element in the adjacency matrix of the i row and the j column, the set of neighbor nodes of node i is denoted as N i}, 、 、 is an auxiliary variable; is a time-varying gain function introduced, is a constant gain.

[0029] Further technical solutions, the comprehensive energy system is a heat and electricity coupled energy system;

[0030] The target function is: minimizing the sum of the operation costs of all power supply devices, heat supply devices, heat and electricity coupled devices and energy storage devices;

[0031] The constraint conditions include: heat and electricity ratio constraint, energy supply and demand balance constraint, and safe operation constraint.

[0032] In a second aspect, the present application provides a comprehensive energy system distributed optimization scheduling system considering privacy protection.

[0033] A comprehensive energy system distributed optimization scheduling system considering privacy protection, comprising:

[0034] A system topology graph construction module, configured to construct a communication network topology graph of the comprehensive energy system by taking each device in the system as a node and energy interaction between nodes as an edge;

[0035] A scheduling model construction module, configured to construct a comprehensive energy system optimization scheduling model by taking minimizing total cost as a target function and configuring constraint conditions for system operation;

[0036] An optimal scheduling solution module, configured to assume that the communication network topology graph is a undirected connected graph, adopt a preset time distributed optimization algorithm introducing time-varying gain, solve the optimization scheduling model, make the comprehensive energy system converge to an optimal solution from an arbitrary initial state within a preset time, and obtain an optimal optimization strategy; wherein, in the solving process, the interaction variables between neighbor nodes are encrypted and decrypted by an encryption and decryption module based on a deep learning adversarial network, so as to realize privacy protection.

[0037] In a third aspect, the present application further provides an electronic device, comprising: a memory configured to store executable instructions;A processor configured to execute the executable instructions stored in the memory to implement the comprehensive energy system distributed optimization scheduling method considering privacy protection.

[0038] In a fourth aspect, the present application further provides a computer readable storage medium storing executable instructions for causing a processor to execute the executable instructions to implement the above-mentioned privacy-considered distributed optimization scheduling method for integrated energy systems.

[0039] The above one or more technical solutions have the following beneficial effects:

[0040] The present application provides a privacy-considered distributed optimization scheduling method and system for integrated energy systems, which introduces a time-varying gain design preset time distributed optimization algorithm for the typical constrained convex optimization problem of integrated energy system economic scheduling. The preset time distributed optimization algorithm independent of the initial state of the system is used to solve the optimization scheduling strategy of the integrated energy system, and the Lyapunov second method is used to prove that the algorithm can guarantee the preset time convergence of the system from any initial state. At the same time, a decryption module based on a deep learning adversarial network is designed and applied to the solving process to encrypt the interactive information in the process, so as to improve the privacy of the interactive information between nodes in the optimization solving process and meet the privacy protection requirements of the scheduling optimization solution.

[0041] The advantages of the additional aspects of the present application will be partially given in the following description, partially become obvious from the following description, or be known through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0042] The drawings accompanying the specification of the present application serve to provide a further understanding of the present application, and the illustrative embodiments of the present application and their descriptions serve to explain the present application, and do not constitute an improper limitation on the present application.

[0043] Figure 1 The logic block diagram of the encryption and decryption module proposed in the embodiment of the present application;

[0044] Figure 2 The network structure diagram of the deep residual multi-layer perception introduced in the embodiment of the present application;

[0045] Figure 3 The structure and communication network topology diagram of the thermoelectric coupling integrated energy system in the embodiment of the present application;

[0046] Figure 4 The distribution diagram of the decryption result of the decryptor Bob network in the embodiment of the present application; wherein, (a) is the distribution diagram of the decryption result of the Bob network to data 1, and (b) is the distribution diagram of the decryption result of the Bob network to data 2;

[0047] Figure 5is a distribution diagram of the decryption result of data 1 by Eve, and (b) is a distribution diagram of the decryption result of data 2 by Eve;

[0048] Figure 6 is a decryption error diagram of the decryptor Bob and the attacker Eve in the embodiment of the present application;

[0049] Figure 7 is a comparison diagram of decryption error results under different network structures in the embodiment of the present application;

[0050] Figure 8 is a power convergence curve diagram in the embodiment of the present application; (a) is a thermal power convergence curve, and (b) is an electric power convergence curve;

[0051] Figure 9 is a power balance curve diagram in the embodiment of the present application; (a) is a thermal power balance curve, and (b) is an electric power balance curve;

[0052] Figure 10 is a thermal energy 0-125s power convergence curve diagram in the embodiment of the present application;

[0053] Figure 11 is a power convergence curve comparison diagram under topology transformation or not in the embodiment of the present application; (a) is a power convergence curve comparison diagram under no topology transformation, and (b) is a power convergence curve comparison diagram under topology transformation;

[0054] Figure 12 is a comparison diagram of original information and encrypted information of interactive variables in the embodiment of the present application; (a) is a comparison diagram of original information and encrypted information of interactive variable y, and (b) is a comparison diagram of original information and encrypted information of interactive variable z;

[0055] Figure 13 is a comparison diagram of original information and decrypted information of the decryptor Bob in the embodiment of the present application. DETAILED DESCRIPTION

[0056] It should be noted that the following detailed description is exemplary only and is intended to provide further description of the present application, and is not intended to limit the exemplary embodiments according to the present application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. In addition, it should be understood that when the terms "comprise" and / or "include" are used in the present specification, there is a presence of the features, steps, operations, devices, components and / or combinations thereof.

[0057] Embodiment One

[0058] The embodiment provides a privacy-considered distributed optimization scheduling method for an integrated energy system, and specifically comprises the following steps.

[0059] Step S1, constructing a communication network topology graph of the integrated energy system by taking each device in the system as a node and energy interaction between the nodes as an edge; configuring constraint conditions for system operation as an objective function of minimum total cost, and constructing an optimization scheduling model for the integrated energy system.

[0060] Step S2, assuming that the communication network topology graph is a connected undirected graph, solving the optimization scheduling model by using a preset time distributed optimization algorithm with time-varying gain, so that the integrated energy system converges to an optimal solution from an arbitrary initial state within a preset time, and obtaining an optimal optimization strategy; wherein, in the solving process, the interaction variables between neighbor nodes are encrypted and decrypted by using a deep learning adversarial network-based encryption and decryption module, so as to realize privacy protection.

[0061] The privacy-considered distributed optimization scheduling method for an integrated energy system proposed in the embodiment is described in more detail through the following content.

[0062] Firstly, in step S1, a communication network topology graph of an integrated energy system (IES) is constructed, and an optimization scheduling model for the integrated energy system is established. Specifically, the integrated energy system is a typical multi-agent information-physical system, and the embodiment takes a typical heat and electricity coupled integrated energy system as an example for illustration. In the physical layer, the system contains multiple energy forms, and energy is supplied to the load through different energy conversion devices, so that the system contains a supply side and a demand side. The supply side can be divided into: a single power supply device, a single heat supply device, a heat and electricity coupled device and an energy storage device, and the demand side contains: a heat load, an electric load. In the information layer, each node in the system exchanges information to ensure the coordinated and efficient operation of the system, so that the communication network of the integrated energy system can be represented by a connected undirected graph , which takes each device in the system as a node and energy interaction between the nodes as an edge, uses to represent a node, uses to represent an edge, to represent an adjacency matrix, to represent an element in the i th row and the j th column of the adjacency matrix, if , , otherwise , and the neighbor node set of the node i is represented as N i}. n , m , l , qThe number of various types of devices on the supply side of the system, r , s The number of various types of devices on the demand side of the system.

[0063] Secondly, by solving the economic dispatch problem of IES, the optimal economic operation can be realized under the premise of safe and stable operation of the system. The optimization objective of IES is to minimize the sum of the operation costs of all power supply devices, heat supply devices, heat and power coupled devices, and energy storage devices, which can be expressed as:

[0064] (1)

[0065] (2)

[0066] (3)

[0067] (4)

[0068] (5)

[0069] (6)

[0070] (7)

[0071] wherein, C is the overall operation cost of the integrated energy system, represents the cost function of each subject, and each subject PH, P, H, S, DR represents a heat and power coupled device, a separate heat supply device, a separate power supply device, an energy storage device, and a demand response device, respectively; , represents the electric power and the heat power of the i-th subject, i represents the cost coefficient (i.e., the quadratic term coefficient) of the i-th subject. i In addition, the constraint conditions for system operation are configured. For the heat and power coupled device, the heat-to-power ratio constraint should be satisfied, which is:

[0072] and a penalty term is introduced to improve the model, so PH can be expressed as: C

[0073] (8) wherein,

[0074] represents the heat-to-power ratio, represents the penalty coefficient.

[0075] In the economic dispatch problem, the energy supply and demand balance constraint and the safe operation constraint should be satisfied, which can be expressed as:​​

[0076] (9)

[0077] (10)

[0078] wherein, represents the load demand of the j th subject.

[0079] (11)

[0080] (12)

[0081] wherein, represents the minimum and maximum of the output electric energy power, represents the minimum and maximum of the output thermal energy power, represents the ramp rate of the i th subject output electric energy power and output thermal energy power, represents the output electric energy power and thermal energy power at time t.

[0082] For the convenience of subsequent problem expression, the economic dispatch model of IES can be expressed as follows:

[0083] (13)

[0084] wherein, , , is a closed convex set, is an inequality constraint function about .

[0085] Secondly, in step S2, the embodiment proposes a preset time distributed optimization algorithm independent of the initial state of the system, which is used to solve the optimization scheduling model. By introducing a specific time-varying gain in the system, it can be ensured that the economic dispatching problem of the comprehensive energy system with constraints converges to the optimal solution from any initial state within a preset time, and the optimal optimization strategy is obtained. Considering the security of information interaction between neighbor nodes in the solving process, the embodiment also proposes an encryption and decryption module based on a deep learning adversarial network design. Based on the adversarial game theory, a customized three-body adversarial framework is designed, in which the core cryptography mapping is generated by a deep residual multi-layer perceptron (ResMLP) self-learning. The entire module is driven and optimized end-to-end through an asymmetric loss function designed to prioritize decryption fidelity, ensuring that neighbor nodes with key authorization can obtain nearly lossless reconstruction accuracy, while ensuring that encrypted data cannot be cracked by external nodes without keys in terms of calculation.

[0086] Specifically, in step S2, first, the communication network topology graph is a connected undirected graph, and a preset time distributed optimization algorithm with time-varying gain is used to solve the optimization scheduling model, so that the comprehensive energy system converges to the optimal solution from any initial state within a preset time, and the optimal optimization strategy is obtained.

[0087] For the convenience of subsequent description and proof, the following assumptions are given:

[0088] Assumption 1: For each agent i , the cost function C i is strongly convex, and there is a constant such that .

[0089] Assumption 2: The communication topology graph G is a connected undirected graph.

[0090] Assumption 3: Define the function which must be continuously differentiable on the interval [0, T c ) and satisfy , .

[0091] To handle inequality constraints in the model, the following lemma is given:

[0092] Lemma 1: Let be a closed convex set, and take the variable x, y in the domain, for the projection dynamic system whose projection operator satisfies the following variational inequality, for any point in the feasible region holds: .

[0093] The above problem (13) is a typical constrained distributed optimization problem. To solve the problem, the embodiment provides a preset time distributed optimization algorithm, which can be expressed as:

[0094] (14)

[0095] wherein, , is a gradient, denotes a cost function of the i-th node (i.e., principal part), i denotes power of the i-th node; is a projection operator, i is a closed convex set, is an inequality constraint function about ; it is defined that is a Laplacian operator of a graph, denotes an adjacency matrix, denotes an element in the i-th row and the j-th column of the adjacency matrix, and a neighbor node set of the node is denoted as , i , j . i i , N , , are auxiliary variables; is a time-varying gain function introduced, is a constant gain.

[0096] By introducing the projection operator to process the inequality constraint, a more accurate solution can be achieved, that is, compared with other algorithms that obtain an approximate solution, the above method can theoretically obtain an accurate solution, so as to guarantee a better scheduling effect.

[0097] Theorem 1: under the conditions of assumptions 1 to 3, the algorithm (14) can guarantee that the system converges to an optimal equilibrium point T c from any initial state within a preset time .

[0098] Proof: define an error variable , and consider the following Lyapunov candidate function:

[0099] (15)

[0100] Since​ Taking the derivative of (15) along the system trajectory, we have

[0101] (16)

[0102] Taking each term one by one, for the convenience of subsequent description, we use to represent According to Lemma 1, let , and the optimal point , thus we have

[0103] (17)

[0104] For the term containing , we have

[0105] (18)

[0106] Substituting the error variable and the KKT condition , we have

[0107] (19)

[0108] The KKT condition (Karush-Kuhn-Tucker condition) is a set of necessary conditions in optimization theory for solving nonlinear programming problems with equality and inequality constraints.

[0109] For the term containing , substituting the KKT condition , we have

[0110] (20)

[0111] For the term containing , we have

[0112] (21)

[0113] Substituting (19)-(21) into (16) and extracting the common factor , we have

[0114] (22)

[0115] According to the commutative law of vector inner product and the symmetry of Laplacian matrix (Assumption 2), all cross terms completely cancel out, and equation (22) can be simplified as

[0116] (23)

[0117] ​Inequality (23) implies that This means that the system state is bounded. Let S be the largest invariant set of in S , then and According to graph theory, means that there exists a constant vector c such that Substitute into the dynamics of , it implies that must be equal to , i.e. , thus Subsequently, according to the dynamics of and (under the condition of ), it can be known that must also be zero. Therefore, S is the only point in . According to the LaSalle invariance principle, the system is asymptotically stable.

[0118] For the algorithm proposed in this embodiment, the Lyapunov function is quadratic, and its derivative is bounded by a negative semi-definite quadratic form; according to the Lyapunov second method, there exists a positive constant such that satisfies the following differential inequality:

[0119] (24)

[0120] where is a constant related to the system parameters .

[0121] By introducing the time-varying gain , it can be obtained that:

[0122] (25)

[0123] Let , solving inequality (25) can obtain:

[0124] (26)

[0125] According to assumption 3, when , the divergence of ensures that when , , thus . This indicates that the algorithm can guarantee that the system is stable within the preset time Converge to the optimal value, and the theorem 1 is proved.

[0126] Secondly, considering that in the calculation process of algorithm (14), the neighbor nodes need to interact the information of variables y and z, the information is easy to be intercepted in the interaction process, and the privacy is poor, therefore, by encrypting the information of interaction, the privacy of the algorithm can be greatly improved, for this purpose, based on the multi-agent game idea, a deep learning encryption method is designed, and a safe and key-dependent data confidentiality transformation is mastered through learning. That is, in the solving process, the encryption and decryption module based on the deep learning adversarial network is used to encrypt and decrypt the interaction variables between the neighbor nodes, so as to realize privacy protection.

[0127] Specifically, inspired by the generative adversarial network (GAN) and game theory, the encryption and decryption module based on the deep learning adversarial network is constructed, which is realized based on the multi-agent game framework, which is composed of three functionally clear neural networks: encrypter (Alice), decrypter (Bob) and attacker (Eve). In this game framework, Alice and Bob form a cooperative alliance and are jointly deployed in the neighbor nodes with the key, and jointly pursue the accuracy of information restoration, while the alliance and Eve form an adversarial game, and through competitive pressure (i.e. adversarial training), the network learns a safe encryption strategy, and the logical relationship is as shown in Figure 1 .

[0128] The encrypter Alice, the decrypter Bob and the attacker Eve are all based on neural networks, and the functions of each neural network are as follows:

[0129] Alice (encrypter): taking the interaction variables between neighbor nodes as plaintext vectors, taking the plaintext vectors U and the key K as inputs, learning to generate ciphertext vectors W . Its mapping definition is: W = Alice( U , K ).

[0130] Bob (decrypter): as an authorized receiver, taking the ciphertext C and the same key K as inputs, learning to accurately reconstruct the original plaintext. Its mapping definition is: U' = Bob( W , K ).

[0131] Eve (attacker): represents intercepting C but without the keyK adversary whose goal is to reconstruct the plaintext from C alone. Its mapping is defined as: U'' = Eve( W ).

[0132] Considering that the data to be encrypted is essentially a low-dimensional vector lacking spatial structure, the embodiment adopts a multi-layer perceptron (MLP) as the basic architecture. To overcome the performance degradation and gradient vanishing problems that occur after stacking multiple layers in traditional deep MLPs, a residual learning mechanism is also innovatively introduced. The resulting network architecture, namely the deep residual MLP (ResMLP), is composed of an input linear layer, a core network stacked with multiple residual blocks, and an output linear layer. The structure is shown in Figure 2 , where each residual block is mathematically defined as H ( X )= X + F ( X ), F ( X ) is the residual mapping learned by the network layer, and X is the original input transmitted through a skip-connection. Inside each residual block, the network layer includes linear layers, batch normalization layers, and activation layers connected in sequence. The input matrix X is linearly transformed in the linear layer (the linear transformation formula is ) to output the matrix , which is then outputted by the normalization layer and the activation layer. The application of the batch normalization (Batch Normalization) technique can stabilize the data distribution of each layer input, alleviate the internal covariate shift, and thus accelerate the model convergence. Through the above deep residual design, the network is given the powerful ability to learn highly complex and nonlinear cryptographic transformations. Among them, is the weight matrix, is the bias matrix; is the matrix form of ; represents the mean and standard deviation of the matrix ; is a small constant, usually 0.00005; represents the parameter to be optimized.

[0133] Further, to achieve the dual goals of high precision and high security, the embodiment adopts a systematic training optimization strategy, including dynamic data generation and cosine annealing learning rate scheduling, for:

[0134] (1) To ensure that the model learns a general function mapping rather than overfitting to a specific dataset, a dynamic data generation strategy is adopted. In each training cycle, the system generates a new batch of random data covering the entire numerical range in real time, providing the model with a theoretically infinite and diverse training set. This strategy not only fundamentally avoids overfitting due to limited data sets, but also ensures that the final learned encryption transformation has strong generalization ability and robustness, and its security does not depend on any specific statistical characteristics of the plaintext data.

[0135] (2) Cosine annealing scheduler is used to dynamically adjust the learning rate. Start with a higher learning rate to promote rapid initial convergence, and then the learning rate will be smoothly decayed along the cosine curve. This allows the model to explore the loss function minimum point with finer steps in the later training period, which is crucial for improving the final decryption accuracy.

[0136] In the multi-agent game framework, by introducing the attacker Eve, the system is forced to autonomously explore and learn complex nonlinear transformations to hide any exploitable patterns in the data, a process similar to an automated, security-oriented algorithm discovery process. This method can greatly improve encryption security and reduce the risk of data being cracked, while the continuous adversarial training between the Alice / Bob combination and Eve ensures that the final generated encryption method is robust to its specific adversary, thus creating a security boundary that evolves and strengthens during training.

[0137] Considering that the encryption system not only needs to ensure that the decrypter can accurately restore the ciphertext with the key, but also needs to ensure that the attacker cannot accurately restore the ciphertext without the key. Therefore, the Alice / Bob alliance and Eve in this embodiment have different optimization objectives, and by introducing the hyperparameter λ, the asymmetric loss function is decoupled, and λ is used to precisely trade-off the security and security of the system.

[0138] Specifically, the goal of Alice and Bob alliance is to minimize the following composite loss function That is:

[0139] (27)

[0140] Where, and are the L1 reconstruction loss of Bob and Eve, respectively, U is the original plaintext, U' is the decryption result of Bob, U'' is the cracking result of Eve, and the hyperparameter λ is used to precisely trade-off the decryption accuracy and encryption security.

[0141] Meanwhile, an attacker Eve is independently trained to minimize her own reconstruction loss i.e.

[0142] (28)

[0143] The above asymmetric design ensures the implementation of the "precision first" principle and effectively guides the evolution direction of the three-body game.

[0144] In each training batch, a precise two-stage update strategy is followed to ensure the effective implementation of cooperation and confrontation, including:

[0145] Stage 1 is to update the attacker (Eve-Step), which aims to maximize Eve's cracking ability. First, the network parameters of Alice and Bob are frozen, Alice generates ciphertext and passes it to Eve, but the gradient flow of the computation graph is cut off to ensure that Alice's weights are not updated. Subsequently, the entire module is based on Eve's reconstruction loss backpropagation is performed, and only the network weights of Eve are updated.

[0146] Stage 2 is to update the communicators (Alice & Bob-Step), which aims to improve the precision and security of communication. At this time, the network parameters of Eve are frozen, Alice and Bob perform the complete encryption and decryption process, and the entire module is based on the composite loss function backpropagation is performed, and the network weights of Alice and Bob are simultaneously updated.

[0147] By alternately executing the above two steps in each batch, the encryption and decryption module is forced to learn to generate more difficult-to-crack ciphertexts under the constant pressure of Eve's growing strength, while maintaining the decryption ability on the Bob side, ultimately achieving a dynamic balance with high fidelity and high security.

[0148] The effectiveness and advancement of the method of the embodiment are verified through the following simulation examples. Among them, the simulation system is an improved thermal-electric coupling comprehensive energy system of IEEE30 power bus-14 thermal bus, and its structure and communication topology are as Figure 3As shown in the figure, G1, G2, G3, G4 are respectively generator sets, L1, L2, L3, L4, L5, L6, L7, L8, L9, L10, L11, L12, L13, L14 are respectively electric load devices, S1 and S2 are respectively electric energy storage devices, CHP1 and CHP2 are respectively CHP (Combined heat and power, combined heat and power) sets, P1 and P2 are respectively heat pump devices, B1 and B2 are respectively boiler devices, H1, H2, H3, H4, H5 and H6 are respectively thermal load devices, PV represents photovoltaic device, WT represents wind power device, EnergyHub (abbreviated as EH) represents energy hub, and the numerical reference 1-30 represents 30 system nodes. In addition, the simulation platform is a PC host: Intel(R) Xeon(R) Platinum 8358P@ 2.60GHz with 32 cores, 64GB RAM, NVIDIA A40 GPU with 48GB RAM; the simulation software is MATLAB R2023b and Python 3.8.

[0149] (1) Parameter setting

[0150] Figure 3 The simulation system shown contains two forms of thermal and electrical energy, and the scheduling time scales of the two forms of energy are different, so it is necessary to decouple the problem and preset different convergence times. The CHP (Combined heat and power, combined heat and power) is decoupled by the heat ratio, and the EH (Energy Hub, energy hub) contains multiple energy conversion devices, so its output state can be directly decoupled. The simulation system parameters are shown in Table 1 and Table 2.

[0151] Table 1: Parameters of thermal and electrical coupling units

[0152]

[0153] Table 2: Parameters of electric and thermal energy devices

[0154]

[0155] In addition, the distributed algorithm related parameters are as follows, and the initial values of the auxiliary variables are: , , , thermal load , electric load , adjustment coefficient , the time-varying gain function is:

[0156] (29)

[0157] Note 1: Although formula (29) does not strictly meet assumption 3, in order to avoid numerical singularity near t = T c , the gain function k(t) is processed, and by selecting a small enough σ (in this embodiment, σ = 10-6 ), the residual error of the system at t = T c can be ignored, and the numerical overflow phenomenon is avoided.

[0158] The parameter settings of the encryption system are as follows: the number of deep residual layers is 8, the hyperparameter is 0.001, the number of key bits is 16, and the initial value of the learning rate scheduler is 0.001.

[0159] (2) Effect verification of encryption and decryption modules

[0160] The three deep learning networks of Alice, Bob and Eve are trained by using the method and parameters of the embodiment, and the three networks are tested.

[0161] First, 1000 groups of two-dimensional data are randomly generated, decrypted by the Alice network, and then decrypted by the Bob and Eve networks, respectively, and the test effect is as shown in Figure 4 Figure 5 From the figure, it can be seen that the neural encryption system proposed in the embodiment shows extremely high fidelity on the authorized decryption channel. Figure 4 The scatter plot distribution of the Bob decryption result is closely and linearly distributed around the standard line (true value) y = x , showing a strong positive correlation, and the two are almost indistinguishable visually. This shows that the Bob network can restore the encrypted information almost losslessly in the case of having the correct key, and successfully achieves the primary goal of system design. Correspondingly, Figure 5 The scatter plot distribution of the Eve decryption result shows a disordered and diffuse cloud-like distribution, and there is no identifiable linear or nonlinear structure. This intuitively shows that the interaction between the prediction value of Eve and the true plaintext is independent, and the cracking performance is equivalent to random guessing. This strongly proves the security of the encryption system and effectively prevents information leakage.

[0162] Secondly, in order to more intuitively analyze the decryption accuracy of the encryption method proposed in the case of having the key, the embodiment uses the MAE index to compare and analyze the decryption errors of Bob and Eve, and the comparison result is as shown in Figure 6As shown, the bar height of Bob's decryption error is extremely low and always tends to the zero plane, while the decryption error of Eve shows a significant height and sharp fluctuations, forming a clear order of magnitude difference. This result again verifies the accuracy of the encryption module for authorized users and the robustness of the attacker from the error quantization perspective.

[0163] In summary, Figures 4-6 From the perspective of correlation and error quantization, the effectiveness of the proposed anti-encryption framework is intuitively verified, which clearly shows the bivalence of the encryption mapping: for the legitimate user holding the key, the mapping is an approximate identity transformation, ensuring the lossless restoration of information; for the attacker, the mapping degenerates into a random permutation, which cannot restore any valid information.

[0164] To verify that the proposed deep residual network structure (RESMLP) can improve accuracy, ordinary MLP networks / 3 layers / 5 layers / 8 layers / 12 layers RESMLP are used for model training, and 10000 groups of randomly generated data are used to test the decryption effect of Bob network, and the results are shown in Figure 7 As shown, RMSE, MAE, and MAPE are prediction evaluation indexes, which are root mean square error, mean absolute error, and mean absolute percentage error, respectively.

[0165] As can be seen from the results, the introduction of residual network can improve the accuracy of the model, but the network depth must match the intrinsic complexity of the problem and the amount of available data. In the above experiment, the model trained by an 8-layer network has the highest decryption accuracy. This shows that the residual network is not the more layers the better, and beyond a certain optimal point, there may be a risk of overfitting and increase the training cost of the model.

[0166] Next, set , other parameters remain unchanged, train multiple deep learning networks, and test the effect by using 10000 groups of randomly generated data, and the results are shown in Table 3.

[0167] Table 3: Comparison of decryption error under different hyperparameters

[0168]

[0169] As can be seen from Table 3, as the increases, the decryption error will increase, the data fidelity will decrease, but the anti-cracking ability will be stronger. Therefore, by adjusting the hyperparameter , the fidelity and anti-cracking weight of the encryption and decryption module can be adjusted.

[0170] To verify the encryption efficiency of the proposed encryption system, its performance was compared with that of homomorphic encryption algorithms commonly used in distributed optimization. In the comparative experiments, the key length was set to 16 bits, and the test environment was a virtual machine environment using an Intel(R) Xeon(R) Platinum8358P @ 2.60GHz with 1 core and 256MB RAM. To avoid the randomness of the experimental results, experiments were conducted on 200, 500, 1000, and 2000 sets of data using two different methods. The experimental results are shown in Table 4. The table shows that the proposed method is significantly superior to homomorphic encryption in terms of encryption efficiency, meeting the requirements for fast encryption.

[0171] Table 4: Comparison of Encryption Speeds

[0172]

[0173] (3) Performance of the preset time-distributed optimization algorithm

[0174] First, we discuss the convergence performance of the algorithm. Based on the above analysis, thermoelectric and thermal energy forms have different scheduling time scales. Here, we give a preset convergence time for the thermoelectric energy-related entities. T c1 =40s, the preset convergence time for the electrical energy-related entities is T c2 =10s. Other parameters are as described above. The system power curve is as follows. Figure 8 As shown, the system's supply and demand balance curve is as follows: Figure 9 As shown.

[0175] from Figure 8 , Figure 9 As can be seen, the algorithm proposed in this embodiment can guarantee that the system converges to the optimal value within a preset time. Furthermore, due to the introduction of the projection operator, it ensures that all entities within the system achieve supply and demand balance while satisfying inequality constraints. The above examples demonstrate that the proposed algorithm has good convergence.

[0176] Secondly, the plug-and-play nature of the algorithm is verified. This embodiment conducts the following experiment: 1) From 0 to 25 seconds... T c1 =10s, ;2) In t =25s, disconnect the 200kW heat load; 3) at t=50s, reconnect the 200kW heat load and set T c1 =15s; 4) At t=75s, disconnect device P2; 5) At t=100s, reconnect device P2. Simulation results are as follows: Figure 10 As shown. From Figure 10It can be seen that, in the case of load variation and equipment switching in the system, it can be automatically adjusted to the new optimal operating state, and the convergence time can be flexibly set according to the demand. This shows that the method has good plug-and-play characteristics.

[0177] Finally, it is verified that the proposed distributed algorithm can still meet the preset time convergence under the condition of dynamic topology. Taking the thermal energy related main body as an example, the preset time is set to T c1 =20s, and the initial network topology is as shown in Figure 3 At t=10s, the network connection between B1 and B2 is cut off, but the communication topology is still guaranteed to be an undirected connected graph. The power convergence curve is as shown in t At t=10s, the network connection between B1 and B2 is cut off, but the communication topology is still guaranteed to be an undirected connected graph. The power convergence curve is as shown in Figure 11 It can be seen from Figure 11 that at t=10s, the network topology in the system changes, causing the power convergence curve to fluctuate, but the system can still achieve preset time convergence under the new topology structure. The above results show that the algorithm proposed in this embodiment has good robustness under the condition of dynamic topology.

[0178] (4) Privacy and accuracy analysis

[0179] In order to verify that the designed encryption system can significantly improve the privacy of the interactive information in the distributed optimization process, the privacy protection performance of the encryption system is analyzed by taking the interaction between node B1 and node B2 as an example.

[0180] The auxiliary variables y and z that need to be interacted in the algorithm are encrypted by the encrypter respectively to obtain the variables used for interaction. The values before and after encryption are as shown in Figure 12 It can be seen from Figure 12 that there is a significant difference between the encrypted ciphertext and the unencrypted plaintext in both order of magnitude and trend, and this method can significantly improve the privacy of the algorithm.

[0181] In order to verify the accuracy of the decrypter and thus ensure that the algorithm can converge accurately, 15 points are taken continuously at t=1.25s, and the data decrypted by the decrypter is compared with the original plaintext, and the results are as shown in Figure 13 It can be seen from Figure 13 that the decrypted data and the original plaintext have very small errors, which can ensure that the algorithm converges to the optimal value and has good accuracy.

[0182] Embodiment Two

[0183] The embodiment provides a comprehensive energy system distributed optimization scheduling system considering privacy protection, which comprises:

[0184] A system topology graph construction module is configured to construct a communication network topology graph of the integrated energy system by taking each device in the system as a node and taking energy interaction between the nodes as an edge.

[0185] A scheduling model construction module is configured to construct an optimal scheduling model of the integrated energy system by taking the minimum total cost as an objective function and configuring constraint conditions for system operation.

[0186] An optimal scheduling solution module is configured to assume that the communication network topology graph is a directed connected graph, to solve the optimal scheduling model by using a preset time-distributed optimization algorithm with time-varying gain, to make the integrated energy system converge to an optimal solution from an arbitrary initial state within a preset time, and to obtain an optimal optimization strategy. In the solving process, the interaction variables between neighbor nodes are encrypted and decrypted by an encryption and decryption module based on a deep learning adversarial network to realize privacy protection.

[0187] Embodiment three

[0188] The embodiment provides an electronic device, comprising: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method provided in the embodiment.

[0189] Embodiment four

[0190] The embodiment also provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, cause the processor to perform the method provided in the embodiment.

[0191] The steps and method of embodiments two to four correspond to the method of embodiment one, and the specific implementation can be referred to the related description of embodiment one. The term "computer-readable storage medium" should be understood as including a single medium or multiple media of one or more instruction sets; and should also be understood as including any medium capable of storing, encoding or carrying instruction sets for execution by a processor and causing the processor to perform any method in the present application.

[0192] Those skilled in the art should understand that each module or step of the present application described above can be realized by a general computer device, and alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be respectively manufactured into individual integrated circuit modules, or a plurality of modules or steps among them can be manufactured into a single integrated circuit module. The present application is not limited to any specific combination of hardware and software.

[0193] The above merely describes the preferred embodiments of the present application, and the specific embodiments of the present application are described in conjunction with the drawings, but are not intended to limit the protection scope of the present application, and those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A method for privacy-protected distributed optimization scheduling of integrated energy systems, characterized in that, The application relates to a method for optimizing scheduling of a comprehensive energy system. The method comprises the following steps: constructing a communication network topology graph of the comprehensive energy system by taking each device in the system as a node and taking energy interaction between the nodes as an edge; constructing an optimization scheduling model of the comprehensive energy system by taking the minimum total cost as an objective function and configuring constraint conditions for system operation; assuming that the communication network topology graph is a non-directional connected graph, and solving the optimization scheduling model by adopting a preset time distribution optimization algorithm with time-varying gain, so that the comprehensive energy system converges to an optimal solution from an arbitrary initial state within a preset time, and an optimal optimization strategy is obtained; and in the solving process, the interaction variables between neighbor nodes are encrypted and decrypted by a decryption and encryption module based on a deep learning adversarial network, and privacy protection is realized. The decryption and encryption module based on the deep learning adversarial network is realized based on a multi-agent game framework and comprises an encrypter Alice, a decrypter Bob and an attacker Eve which are all constructed based on a neural network. The encrypter Alice and the decrypter Bob form a cooperative alliance, and a joint key is deployed in the neighbor nodes; wherein the interaction variables between the neighbor nodes are taken as plaintext vectors, the encrypter Alice learns to generate ciphertext vectors by taking the plaintext vectors and the key as inputs, and the decrypter Bob learns to reconstruct the original plaintext vectors by taking the ciphertext vectors and the same key as inputs. The attacker Eve forms an adversarial game with the cooperative alliance, and guides the encrypter Alice and the decrypter Bob to learn complex nonlinear encryption transformation through adversarial training. The optimal scheduling model is solved using a pre-defined time-distributed optimization algorithm that incorporates time-varying gain, as follows: ;in, , For gradient, Indicates the first i The cost function of each node , They represent the first i The electrical power and thermal power of each main body; For projection operators, It is a closed convex set. For about The safe operation inequality constraint function; definition For the Laplace operator of a graph, Represents the adjacency matrix. Represents the adjacency matrix of the nth element. i Line number j Element, node i The set of neighbor nodes is represented as { N i }, , , As an auxiliary variable; For the introduced time-varying gain function, The gain is constant. 2.The privacy-considered distributed optimization scheduling method for integrated energy system according to claim 1, wherein, Each residual block is defined as H(X) = X + F(X), F(X) is a residual mapping learned by a network layer, and X is an original input transmitted through a shortcut connection; the network layer comprises linear layers, batch normalization layers and activation layers which are connected in sequence. In each training batch, the encryption and decryption module is trained by adopting a two-stage updating strategy, which comprises the following steps: 3.The privacy preserving integrated energy system distributed optimization scheduling method of claim 1, wherein, In stage one, the network parameters of the encrypter Alice and the decrypter Bob are frozen, the ciphertext vectors generated by the encrypter Alice are transmitted to the attacker Eve, the weight of the encrypter Alice is not updated, the network weight of the attacker Eve is updated based on the reconstruction loss minimization of the attacker Eve, and the network weight of the attacker Eve is updated through back propagation; 4.The method of claim 2, wherein, In stage two, the network parameters of the attacker Eve are frozen, the encrypter Alice and the decrypter Bob perform a complete encryption and decryption process, and the network weights of the encrypter Alice and the decrypter Bob are updated through back propagation based on the minimization of a composite loss function. ​ ​ 5.The method of claim 4, wherein, The encryption and decryption module is optimized end-to-end through an asymmetric loss function, and a composite loss function of a cooperative alliance composed of the encryptor Alice and the decryptor Bob is as follows: ； ； ； wherein, , are the LI reconstruction loss of the decryptor Bob and the attacker Eve, respectively, U is the original plaintext vector, U' is the decryption result of the decryptor Bob, and U" is the cracking result of the attacker Eve, is a hyperparameter.

6. A privacy-protected integrated energy system distributed optimization scheduling system, characterized in that, The method comprises the steps of: The system topology graph construction module is configured to construct a communication network topology graph of the comprehensive energy system by taking each device in the system as a node and taking energy interaction between nodes as an edge. The comprehensive energy system is a heat and electricity coupled energy system, and an objective function of the system is to minimize the sum of operation costs of all power supply devices, heat supply devices, heat and electricity coupled devices and energy storage devices; constraint conditions include heat and electricity ratio constraints, energy supply and demand balance constraints and safe operation constraints. The scheduling model construction module is configured to construct an optimal scheduling model of the comprehensive energy system by taking minimization of total cost as an objective function and configuring constraint conditions for system operation. The optimal scheduling solving module is configured to assume that the communication network topology graph is a directed connected graph, to solve the optimal scheduling model by using a preset time-distributed optimization algorithm with time-varying gain, to make the comprehensive energy system converge to an optimal solution from an arbitrary initial state within a preset time, and to obtain an optimal optimization strategy; in the solving process, the encryption and decryption module based on deep learning adversarial network is used to encrypt and decrypt interaction variables between neighbor nodes, thereby achieving privacy protection. The encryption and decryption module based on the deep learning adversarial network is implemented based on a multi-agent game framework and comprises an encryptor Alice, a decryptor Bob and an attacker Eve constructed based on a neural network; the encryptor Alice, the decryptor Bob and the attacker Eve are all built by using a multi-layer perceptron with deep residual as a basic framework, and the multi-layer perceptron with deep residual comprises an input linear layer, a core network stacked by a plurality of residual blocks and an output linear layer. The preset time distribution optimization algorithm with time-varying gain is introduced to solve the optimization scheduling model, and the optimization scheduling model is solved as follows: ; wherein, , is a gradient, a cost function of the i th node, , respectively represent the electric power and the thermal power of the i th agent; is a projection operator, is a closed convex set, is a safe operation inequality constraint function about ; define as a Laplace operator of the graph, indicates an adjacency matrix, indicates the i th element in the j th row of the adjacency matrix, and the neighbor node set of the node i is represented as N i}, 、 、 is an auxiliary variable; is a time-varying gain function introduced, is a constant gain.

7. An electronic device, comprising: The method comprises the steps of: The memory is configured to store executable instructions. The processor is configured to execute the executable instructions stored in the memory to implement the method for distributed optimal scheduling of a comprehensive energy system considering privacy protection according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, Executable instructions are stored in the memory to cause the processor to execute the executable instructions to implement the method for distributed optimal scheduling of a comprehensive energy system considering privacy protection according to any one of claims 1-5.

Images