A Low-Carbon Power Generation Scheduling Method Based on Graph Attention Networks and Sequence Models

Abstract

The present application discloses a low-carbon power generation scheduling method based on a graph attention network and a sequence model, which relates to the technical field of power generation scheduling. The method comprises: constructing a power generation scheduling bipartite graph based on carbon emission decision variables and power generation constraints, and using the carbon emission evaluation calculation formula and power generation constraints as sequence data; extracting the structural features of the power generation scheduling bipartite graph based on the graph attention network, and extracting the dependency features between the sequence data based on the long short-term memory network; transferring the structural features to the first self-supervised loss function of the graph attention network training to perform graph attention network learning, and transferring the dependency features to the second self-supervised loss function of the long short-term memory network training to perform long short-term memory network learning to obtain a target model; inputting test set data into the target model to obtain predicted data on the power generation of each generator set. This improves the generalization ability of the model when dealing with complex power generation scheduling constraint problems.

Description

Technical Field

[0001] This application relates to the field of power generation dispatching technology, and in particular to a low-carbon power generation dispatching method based on graph attention networks and sequence models. Background Technology

[0002] The goal of low-carbon power generation dispatch is to formulate optimal generator power generation strategies to minimize carbon emissions during power generation while meeting electricity demand, thereby achieving the objective of low-carbon power generation. This requires considering factors such as the carbon emission coefficient of each generator unit, its actual output power for each time period, power limits, the total power range of the power plant, and the overall power demand of the system, among other characteristic parameters of each generator unit. However, the uncertainty and volatility of renewable energy processing, the coupling and coordinated dispatch of multiple energy systems, and the contradiction between carbon emission constraints and economic objectives make low-carbon power generation dispatch difficult to achieve.

[0003] Common low-carbon power generation dispatch methods mainly target thermal power generation. Since their goal is to achieve system stability and economy, their objective function focuses on fuel cost, without taking into account the carbon emissions generated by renewable energy sources such as wind, hydro, and solar power in the power grid system, and neglecting the integrated control of multiple energy sources.

[0004] To consider the integrated dispatch of multiple energy sources, heuristic algorithms and reinforcement learning methods can be used to solve complex optimization problems in low-carbon power generation dispatch by combining the automatic decision-making capabilities of heuristic search and reinforcement learning. For example, a comprehensive energy system model can be established with system economy, environmental protection, and output imbalance as optimization objectives, and a particle concentration evaluation operator can be introduced to improve the particle swarm optimization algorithm to solve the model; alternatively, a dynamic economic dispatch model can be developed with the objective of minimizing the economic cost of distribution network operation, and deep reinforcement learning can be used to solve it.

[0005] However, since heuristic algorithms rely primarily on heuristic rules or experience for searching, and reinforcement learning typically requires extensive interactive training, they are not only computationally expensive, but their convergence and efficiency in high-dimensional constrained spaces still need significant improvement, and finding the global optimum cannot be guaranteed.

[0006] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0007] The main purpose of this application is to provide a low-carbon power generation scheduling method based on graph attention networks and sequence models. It aims to solve the technical problems that commonly used low-carbon power generation scheduling methods are not only computationally expensive, but also have to be significantly improved in terms of convergence and efficiency in high-dimensional constrained spaces, and cannot guarantee finding the global optimal solution.

[0008] To achieve the above objectives, this application proposes a low-carbon power generation scheduling method based on graph attention networks and sequence models. The method includes:

[0009] The power generation of generator sets of each power generation type in the power generation system is obtained, as well as the power generation limit of each generator set in the power generation system, the power generation limit of each power plant, and the power generation limit of the power generation terminal.

[0010] The power generation of each generator set is used as the carbon emission decision variable in the carbon emission assessment calculation formula, and corresponding power constraints are constructed based on the power generation limit of the generator set, the power generation limit of the power plant and the power generation limit at the power generation end. The power constraints include a first power generation constraint, a second power generation constraint and a third power generation constraint.

[0011] A bipartite graph for power generation scheduling is constructed based on the carbon emission decision variables and the power generation constraints, and the carbon emission assessment formula and the power generation constraints are used as sequence data.

[0012] Based on graph attention networks, the structural features of the power generation scheduling bipartite graph are extracted, and based on long short-term memory networks, the dependency features between the sequence data are extracted.

[0013] The structural features are passed to the first self-supervised loss function for training the graph attention network to perform graph attention network learning, and the dependency features are passed to the second self-supervised loss function for training the long short-term memory network to perform long short-term memory network learning, so as to obtain the target model;

[0014] Input the test set data into the target model to obtain the predicted power generation data of each generator set.

[0015] In one embodiment, the steps of constructing a bipartite power generation scheduling graph based on the carbon emission decision variables and the power generation constraints, and using the carbon emission assessment formula and the power generation constraints as sequential data, include:

[0016] The low-carbon power generation scheduling problem is constructed by combining the carbon emission assessment formula and the power generation constraints.

[0017] The low-carbon power generation scheduling problem is encoded to obtain the power generation scheduling bipartite graph, which includes carbon emission decision variable nodes, power generation constraint nodes, and edges connecting the carbon emission decision variable nodes and the power generation constraint nodes.

[0018] A target vector is formed based on the carbon emission assessment formula, and a constraint vector is formed based on the first power generation constraint, the second power generation constraint, and the third power generation constraint. The sequence data includes the target vector and the constraint vector.

[0019] In one embodiment, the step of extracting the structural features of the generation scheduling bipartite graph based on a graph attention network includes:

[0020] Based on the first-level message transmission from the carbon emission decision variable node to the power generation constraint node, the impact of the carbon emission decision variable on the power generation constraint is captured;

[0021] Based on the second-level message transmission from the power generation constraint node to the carbon emission decision variable node, the reaction of the power generation constraint on the carbon emission decision variable is determined.

[0022] Based on the impact of the carbon emission decision variables on power generation constraints and the reaction of power generation constraints on carbon emission decision variables, the structural characteristics of the power generation dispatch bipartite graph are obtained.

[0023] In one embodiment, the step of extracting dependency features between the sequence data based on the Long Short-Term Memory network includes:

[0024] The target vector and the constraint vector are passed to the long short-term memory network to determine the candidate memory state and the gating value of the output gate;

[0025] The hidden state is obtained based on the candidate memory state and the gating value of the output gate;

[0026] The dependency features are determined based on the hidden state.

[0027] In one embodiment, before the steps of passing the structural features to the first self-supervised loss function for training the graph attention network and performing graph attention network learning, and passing the dependency features to the second self-supervised loss function for training the long short-term memory network and performing long short-term memory network learning, the method further includes:

[0028] The power generation constraint is incorporated as a penalty term into the carbon emission assessment formula to obtain the first self-supervised loss function; and...

[0029] Determine how to minimize the difference between the current dual estimate and the predicted dual variable, and construct the second self-supervised loss function based on the difference.

[0030] In one embodiment, the steps of passing the structural features to a first self-supervised loss function for training the graph attention network, performing graph attention network learning, and passing the dependency features to a second self-supervised loss function for training the long short-term memory network, performing long short-term memory network learning, to obtain the target model include:

[0031] The structural features are passed to the graph attention network to obtain the predicted carbon emission decision variables;

[0032] The predicted carbon emission decision variables are passed to the first self-supervised loss function, and the model parameters of the graph attention network are optimized using the gradient descent method to obtain an updated graph attention network; and,

[0033] The dependency features are passed to the long short-term memory network to obtain the predicted dual variable;

[0034] The predicted dual variable is passed to the second self-supervised loss function, and the model parameters of the long short-term memory network are updated by gradient to obtain the updated long short-term memory network.

[0035] The target model is obtained based on the updated graph attention network and the updated long short-term memory network.

[0036] In one embodiment, the step of inputting test set data into the target model to obtain predicted data on the power generation of each generator set includes:

[0037] Input the test set data into the target model for forward propagation calculation;

[0038] The data output by the target model is subjected to inverse normalization to obtain the predicted power generation data of each generator set.

[0039] In one embodiment, the power generation type includes fossil fuel power generation, nuclear power generation, and renewable energy power generation; the carbon emission assessment formula consists of carbon emission calculation formulas for fossil fuel power generation, nuclear power generation, and renewable energy power generation; the step of using the power generation of each of the generator sets as the carbon emission decision variable in the carbon emission assessment calculation formula includes:

[0040] The power generation of the fossil fuel power generation generator set is used as the carbon emission decision variable in the calculation formula of the carbon emission of fossil fuel power generation;

[0041] The power generation of the nuclear power generation generator set is used as the carbon emission decision variable in the calculation formula of the nuclear power generation carbon emissions;

[0042] The power generation of the renewable energy generator sets is used as the carbon emission decision variable in the calculation formula for the carbon emissions of renewable energy generation.

[0043] In addition, to achieve the above objectives, this application also proposes a low-carbon power generation scheduling device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of a low-carbon power generation scheduling method based on graph attention networks and sequence models as described above.

[0044] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of a low-carbon power generation scheduling method based on graph attention network and sequence model as described above.

[0045] One or more technical solutions proposed in this application have at least the following technical effects:

[0046] By incorporating multiple energy types into the carbon emission assessment calculation formula and considering three different power generation constraints, the carbon emission characteristics of multi-energy systems can be more comprehensively reflected. Through the construction of a bipartite generation scheduling graph, the structural features of the graph are extracted using the GAT model, and the dependency features between sequential data are extracted using the LSTM model, effectively capturing the dependencies between the carbon emission calculation formula, the power generation constraint sequence, and the carbon emission decision variables. By introducing a self-supervised method into low-carbon generation scheduling, alternating between graph attention network learning and long short-term memory network learning, the model's generalization ability in handling complex generation scheduling constraints is improved. Attached Figure Description

[0047] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart illustrating an embodiment of a low-carbon power generation scheduling method based on graph attention networks and sequence models according to this application.

[0050] Figure 2 This application presents a two-part method for low-carbon power generation scheduling based on graph attention networks and sequence models, specifically focusing on power generation scheduling. Figure 1 Schematic diagram of the embodiment;

[0051] Figure 3 This is a flowchart illustrating an embodiment of step S400 of a low-carbon power generation scheduling method based on graph attention networks and sequence models according to this application.

[0052] Figure 4 This is a schematic diagram of an embodiment of a two-level message passing method using a bipartite graph attention network for low-carbon power generation scheduling based on graph attention networks and sequence models, as described in this application.

[0053] Figure 5 This is a flowchart illustrating another embodiment of step S400 of a low-carbon power generation scheduling method based on graph attention networks and sequence models according to this application.

[0054] Figure 6 This is a framework diagram of an embodiment of an LSTM model for a low-carbon power generation scheduling method based on graph attention networks and sequence models according to this application;

[0055] Figure 7 This is an example of a low-carbon power generation scheduling method based on graph attention networks and sequence models, with a table of carbon emission coefficients for each generator unit.

[0056] Figure 8 This paper presents an example of a low-carbon power generation scheduling method based on graph attention networks and sequence models, which uses a feature vector table for message passing from carbon emission decision variables to power generation constraints.

[0057] Figure 9 This application presents an example of a low-carbon power generation scheduling method based on graph attention networks and sequence models, which includes a feature vector table for message passing from the power generation constraint side to the carbon emission decision variable side.

[0058] Figure 10 This is a schematic diagram of the equipment structure of the hardware operating environment involved in a low-carbon power generation scheduling method based on graph attention network and sequence model in an embodiment of this application.

[0059] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0060] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0061] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0062] Based on this, embodiments of this application provide a low-carbon power generation scheduling method based on graph attention networks and sequence models, referring to... Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the low-carbon power generation scheduling method based on graph attention networks and sequence models in this application.

[0063] In this embodiment, the low-carbon power generation scheduling method based on graph attention networks and sequence models includes steps S100 to S600:

[0064] Step S100: Obtain the power generation of generator sets of each power generation type in the power generation system, and obtain the power generation limit of each generator set, the power generation limit of each power plant, and the power generation limit of the power generation terminal in the power generation system.

[0065] In this embodiment, depending on the energy source, the power generation type includes, but is not limited to, fossil fuel power generation, nuclear power generation, and renewable energy power generation. Fossil fuel power generation mainly refers to thermal power generation, while renewable energy power generation includes solar energy, wind energy, hydropower, geothermal energy, biomass energy, and ocean energy power generation, including tidal energy, wave energy, and ocean current energy.

[0066] Step S200: The power generation of each generator set is used as the carbon emission decision variable in the carbon emission assessment calculation formula, and corresponding power generation constraints are constructed based on the power generation limit of the generator set, the power generation limit of the power plant and the power generation limit at the power generation end. The power generation constraints include a first power generation constraint, a second power generation constraint and a third power generation constraint.

[0067] In this embodiment, since the carbon emission characteristics of different power generation methods are different, and the power generation of generator sets directly affects the carbon emission, the present invention uses the power generation of various generator sets as the carbon emission decision variable.

[0068] For example, using P A,a,t Let P represent the actual power generation of the a-th thermal power generating unit during a certain time period t; B,b,t Let P represent the actual power generation of the b-th nuclear power generator unit during a certain time period t; C,c,t Let P represent the actual power generation of the c-th solar generator unit during a certain time period t; D,d,t Let P represent the actual power generation of the d-th wind turbine generator during a certain time period t; M,e,t Let P represent the actual power generation of the e-th hydroelectric generator unit during a certain time period t; N,f,t P represents the actual power generation of the f-th geothermal generator unit during a certain time period t;Q,g,t Let P represent the actual power generation of the g-th ocean energy generator unit during a certain time period t; R,h,t This represents the actual power generation of the h-th biomass generator unit during a certain time period t.

[0069] Understandably, the above P A,a,t P B,b,t P C,c,t P D,d,t P M,e,t P N,f,t P Q,g,t and P R,h,t The units are all kWh.

[0070] It should be noted that the above carbon emission assessment calculation formula includes, but is not limited to, the sum of the carbon emission calculation formulas for thermal power generation, renewable energy power generation, and nuclear power generation.

[0071] Optionally, the power generation of thermal power generating units is used as the carbon emission decision variable in the thermal power generation carbon emission calculation formula; the power generation of nuclear power generating units is used as the carbon emission decision variable in the nuclear power generation carbon emission calculation formula; and the power generation of renewable energy generating units is used as the carbon emission decision variable in the renewable energy generation carbon emission calculation formula.

[0072] As an optional implementation method, the carbon emission calculation formula for thermal power generation is defined as follows:

[0073] in, This is the carbon emission of the a-th thermal power generating unit, expressed in grams (g); α a β a γ a δ a and ε a These are the carbon emission coefficients unique to the a-th thermal power generating unit, expressed in grams per kilowatt-hour (g / kWh).

[0074] Understandably, under low load conditions, thermal power generating units have relatively high carbon emission intensity per unit of electricity generated due to reduced combustion efficiency; conversely, under high load conditions, due to the positive effects of economies of scale, the carbon emission intensity per unit of electricity generated actually decreases. In other words, during the operation of thermal power generating units, carbon emissions per unit of electricity generated do not increase linearly with increasing load. Therefore, this non-linear relationship can be described by a quadratic function. Furthermore, under unconventional operating conditions such as deep peak shaving, the carbon emissions of thermal power generating units may change significantly. For example, in older thermal power units, when the power output approaches the unit's upper limit, carbon emissions no longer increase linearly or gradually.

[0075] Therefore, by introducing an exponent to better capture the nonlinear characteristics of thermal power generating units under unconventional operating conditions, and by superimposing two different forms of functions to better match the actual operating data of thermal power generating units, the accuracy of carbon emission estimation can be improved.

[0076] As another alternative implementation method, the carbon emission calculation formula for solar power generation is defined as:

[0077] in, This is the carbon emission of the cth solar generator unit, in grams (g), θ c The carbon emission factor is unique to the cth solar generator unit, expressed in grams per kilowatt-hour (g / kWh).

[0078] Understandably, renewable energy and nuclear power generation do not directly rely on fuel combustion to generate energy, but their manufacturing, transportation, and maintenance processes throughout their lifecycle still produce carbon emissions. Taking solar power generation as an example, the manufacture of solar panels and other equipment, the construction of power generation facilities involving heavy machinery and transport vehicles, and the transportation and repair of equipment during the maintenance phase all potentially generate carbon emissions. Therefore, the carbon emissions from renewable energy and nuclear power generation are directly proportional to the actual amount of electricity generated; that is, the higher the electricity generation, the higher the overall carbon emissions.

[0079] Therefore, the carbon emission calculation formulas for renewable energy power generation and nuclear power generation are similar to those for solar power generation, and can be adjusted using corresponding carbon emission coefficients.

[0080] For example, suppose the carbon emission assessment is calculated as follows:

[0081]

[0082] Among them, a v b v c v d v e v f v g v and h v The number of generating units for thermal power, nuclear power, solar power, wind power, hydropower, geothermal power, ocean energy, and biomass power, respectively. b μ d τ e , ω g and ζ hThese are the unique carbon emission coefficients for each nuclear, wind, hydro, geothermal, ocean, and biomass power generation unit, expressed in grams per kilowatt-hour (g / kWh).

[0083] In this embodiment, the power generation of generator set i during time period t is represented as P. U,i,t (1≤i≤n). Where U∈{A,B,C,D,M,N,Q,R}, representing the energy type; n=a v +b v +c v +d v +e v +f v +g v +h v , representing the total number of all generator sets in the aforementioned power generation system.

[0084] The total power generation P of power plant k in time period t U,k,t It consists of the sum of the outputs of all its generator sets, that is:

[0085]

[0086] Where r represents the number of power plants, k (1≤k≤r) represents the specific power plant number, and J k ={1,2,……,s} represents the set of generator units in power plant k, and s represents the number of generator units in power plant k.

[0087] Optionally, the first power generation constraint is constructed based on the power generation limit of the generator set, i.e., the power generation output constraint of the generator set. This can be expressed as, in, and These are the minimum and maximum power generation of the i-th generator unit within time period t, respectively. The purpose of this is to ensure that the power generation P of generator unit i is... U,i,t Within the range of the maximum and minimum generating capacity of generator set i.

[0088] The second power generation constraint is constructed based on the power generation limit of the power plant, namely the total power generation output constraint of the power plant. This can be expressed as, in, and These represent the minimum and maximum total power generation of the k-th power plant within time period t, respectively. The purpose of this is to ensure that the total power generation P of power plant k is [not specified]. U,k,t Within the range of the maximum and minimum power generation capacity of power plant k.

[0089] The third generation constraint, namely the power balance constraint, is constructed based on the generation limit at the generation end. This can be expressed as, in, This is the minimum total power generation that all generating units must achieve within a certain time period t. The purpose of this is to ensure that, within any given time period, the total power generation of all generating units at least meets the electricity demand.

[0090] Step S300: Construct a bipartite graph for power generation scheduling based on the carbon emission decision variables and the power generation constraints, and use the carbon emission assessment formula and the power generation constraints as sequence data.

[0091] Optionally, step S300 includes constructing a low-carbon power generation scheduling problem by combining the carbon emission decision variables and the power generation constraints; then, encoding the low-carbon power generation scheduling problem to obtain a bipartite graph of power generation scheduling with nodes and edges. It should be noted that the nodes in the above-mentioned bipartite graph of power generation scheduling include carbon emission decision variable nodes and power generation constraint nodes; the edges in the bipartite graph of power generation scheduling refer to the edges formed by connecting carbon emission decision variable nodes and power generation constraint nodes. Furthermore, a target vector is formed based on the carbon emission assessment calculation formula, and a constraint vector is formed based on the first power generation constraint, the second power generation constraint, and the third power generation constraint. It can be understood that the above-mentioned sequence data includes the target vector and the constraint vector.

[0092] In this embodiment, the carbon emission assessment calculation formula F(P) is used. U,i,t Generator output constraints Total power generation output constraints of power plants and power balance constraints Constructing a low-carbon power generation dispatch problem Where, p * =(P U,1,t ,P U,2,t ,……,P U,i,t Understandably, the goal of solving this optimization problem is to minimize total carbon emissions while satisfying the constraints; in other words, to achieve the goal of satisfying the constraints. Under the premise of minimizing F X (p * ).in, The matrix represents the parameters for determining an instance of the low-carbon power generation dispatch problem. This indicates low-carbon power generation dispatch constraints. |·| represents the number of elements in the set, and the instance parameter determines the function F. X and The specific form.

[0093] Specifically, the output power of the generator set is first constrained. and the total power generation output constraint of the power plant Each interval constraint in the equation is decomposed into two independent one-sided inequalities. These decomposed one-sided inequalities are then combined with the power balance constraint. It is uniformly represented as less than or equal to 0. The carbon emission assessment calculation formula F(P) is used. U,i,t Construct the target vector x0, and construct the constraint vectors x1, x2, ..., x0 for each of the above one-sided inequalities. m-1 ,x m Each vector has a dimension of n+1, where n is the number of carbon emission decision variables, and the extra 1 is a constant term. For the carbon emission assessment calculation formula F(P)... U,i,t For each one-sided inequality, if the weight of a certain carbon emission decision variable is not 0, then the corresponding vector is matched with that weight. In particular, when the power generation of a thermal power generating unit is used as the carbon emission decision variable, the corresponding vector position is filled with the sum of the carbon emission weights specific to that thermal power generating unit.

[0094] Let the objective vector x0 and the constraint vectors x1, x2, ..., x be... m-1 ,x m The parameters are concatenated into a matrix X, containing the instantiation parameters for a specific low-carbon power generation scheduling problem within a certain time period t. By collecting the instantiation parameter matrices for a certain region over T' time periods, a complete low-carbon power generation dataset is constructed. Z is divided into training and test sets in an 8:2 ratio.

[0095] Optionally, Z is normalized using min-max scaling to transform the original electrocarbon data in the electrocarbon dataset Z to values ​​between 0 and 1. This ensures that the electrocarbon data is processed on a uniform scale and improves the model's stability and generalization ability. This is based on the maximum value in each instantiated parameter matrix. and minimum value Normalize each value s in the instantiation parameter matrix to obtain s', i.e.

[0096] Further, please refer to Figure 2 In constructing the aforementioned low-carbon power generation dispatch problem Next, the issue of low-carbon power generation scheduling will be addressed. Encoding as a bipartite graph of generation scheduling with nodes and edges in, and It is two disjoint sets of nodes. This represents the entire set of nodes.

[0097] Optionally, first divide the nodes of the bipartite graph into two non - overlapping sets: carbon emission decision variables and power generation constraints. That is, the bipartite graph of power generation scheduling The node set V of includes the set of carbon emission decision variables and the set of power generation constraint nodes and

[0098] where, the bipartite graph of power generation scheduling On one side, it contains n carbon emission decision variable nodes. In the feature vector of the j - th (0 ≤ j < n) carbon emission decision variable node, the value at the j - th (0 ≤ j < n + 1) position is denoted as x0[j], and all other positions are 0. Denote as the carbon emission decision variable node feature matrix. On the other side of the bipartite graph of power generation scheduling it contains m power generation constraint nodes. The feature vector of the i - th (0 < i ≤ m) power generation constraint node is Denote as the power generation constraint feature matrix.

[0099] Then, create an edge between the j - th carbon emission decision variable node and the i - th power generation constraint node. If the j - th carbon emission decision variable has a non - zero weight in the i - th power generation constraint, create an edge between the nodes and

[0100] The purpose of this is to highlight the interaction between power generation constraints and carbon emission decision variables in the low - carbon power generation scheduling problem, and facilitate the extraction of key features through the message - passing mechanism of the graph structure.

[0101] Step S400: Based on the graph attention network, extract the structural features of the bipartite graph of power generation scheduling, and based on the long short - term memory network, extract the dependency relationship features between the sequence data.

[0102] In this embodiment, a bipartite graph attention network (Graph Attention Network, GAT) is used for two - level message passing to extract the structural features of power generation scheduling. On the one hand, it improves the effectiveness of capturing the complex relationship between carbon emission decision variables and power generation constraints, and on the other hand, it improves the prediction ability and optimization effect of the model.

[0103] Optionally, refer to Figures 3 to 4 , the steps of extracting the structural features of the bipartite graph of power generation scheduling based on the graph attention network include steps S411 to S413: ​

[0104] Step S411: Based on the first-level message transmission from the carbon emission decision variable node to the power generation constraint node, capture the impact of the carbon emission decision variable on the power generation constraint;

[0105] Step S412: Based on the second-level message transmission from the power generation constraint node to the carbon emission decision variable node, determine the reaction of the power generation constraint on the carbon emission decision variable;

[0106] Step S413: Based on the impact of the carbon emission decision variables on power generation constraints and the reaction of the power generation constraints on carbon emission decision variables, obtain the structural characteristics of the power generation dispatch bipartite graph.

[0107] Specifically, each power generation constraint node is updated based on information from the carbon emission decision variable nodes directly connected to it, according to the formula... The first-level message transmission from the carbon emission decision variable node to the power generation constraint node captures the impact of carbon emission decision variables on power generation constraints. Let i be a neighbor of the i-th node. The weight matrix for the nodes of carbon emission decision variables. Let σ be the attention coefficient between node i (constrained by power generation) and node j (decision variable for carbon emissions), and let σ be the activation function. According to the formula... The second-level message passing from the power generation constraint node to the carbon emission decision variable node is executed to determine the reaction of power generation constraints to carbon emission decision variables. This represents the weight matrix of the nodes constrained by power generation.

[0108] It should be noted that graph attention networks can comprehensively consider the impact of carbon emission decision variables on power generation constraints and the feedback effect of power generation constraints on carbon emission decision variables. They can learn how to adjust and optimize each carbon emission decision variable while satisfying power generation constraints, and obtain updated carbon emission decision variables.

[0109] In this embodiment, the aforementioned attention coefficient The ability to capture which carbon emission decision variables have a greater impact on a certain power generation constraint, or which power generation constraints have a stronger constraint on a certain carbon emission decision variable, is learned by the graph attention network. The training process aims to minimize the network's loss function, including multiple iterative calculations of input features and parameter optimization, ensuring that the attention coefficients accurately reflect the relationships between nodes. The formula for calculating the attention coefficients is:

[0110]

[0111] in, A learnable parameter vector, defined as a set of weights for the graph attention network, typically initialized with random values; ∥ represents the concatenation operation, and ψ represents the activation function. Attention coefficients. Definition and Similarly, only the set of carbon emission decision variables is exchanged. With power generation constraints And swap the positions of the i-th and j-th nodes.

[0112] Furthermore, after obtaining the updated carbon emission decision variables, the node features of these updated carbon emission decision variables are concatenated into a matrix H. This matrix H is then passed to a multi-layer perceptron containing only one hidden layer to obtain the predicted values ​​of the carbon emission decision variables. Expressed as a formula: in, W is the activation function. (1) Let b be the weight matrix from the input layer to the hidden layer. (1) W is the bias vector of the hidden layer. (2) Let b be the weight matrix from the hidden layer to the output layer. (2) This is the bias vector for the output layer.

[0113] In this embodiment, a Long Short-Term Memory (LSTM) network is used. To extract the dependency features between the data sequences mentioned above, it should be noted that these dependency features refer to the dependency between the carbon emission assessment calculation formula and the power generation constraints. The purpose of this is to address the low-carbon power generation dispatch problem. In this study, the complex relationship that may exist between the carbon emission assessment calculation formula and the power generation constraints, spanning multiple power generation constraints, is effectively captured.

[0114] Optionally, please refer to Figures 5 to 6 The steps, based on a long short-term memory network, involve extracting dependency features between the sequence data, including steps S421 to S423:

[0115] Step S421: Pass the target vector and the constraint vector to the long short-term memory network to determine the candidate memory state and the gating value of the output gate;

[0116] Step S422: Obtain the hidden state based on the candidate memory state and the gating value of the output gate;

[0117] Step S423: Determine the dependency features based on the hidden state.

[0118] Specifically, the target vector x0 and the constraint vectors x1, x2, ……, x m-1 , x m are used as the inputs of the LSTM. Understandably, the sequence length is m + 1, and the dimension of each sequence vector is n + 1. The formula for the LSTM to extract features from each sequence vector includes:

[0119] p t = Θ(W p · [h t-1 , x t + b p )

[0120] q t = Θ(W q · [h t-1 , x t + b q )

[0121]

[0122] where x t (0 ≤ t < m + 1) represents the t-th sequence; p t is the output of the forget gate; W p is the weight matrix of the forget gate; Θ is the Sigmoid activation function, representing the proportion of discarding or retaining; tanh(·) is the activation function; q t is the gating value of the input gate, used to determine which information will be written into the cell state; is the candidate memory state, used to generate new information that may be updated to the cell state; is the current cell state updated by combining the output of the forget gate and the output of the input gate; is the gating value of the output gate; h t is the hidden state.

[0123] By updating the hidden state h t of the LSTM, the hidden state of each sequence step will contain the corresponding output information. Finally, the vector h m obtained at the last sequence step converges all the long-term and short-term dependency relationship features. Therefore, the dual variable predicted by the LSTM is λ = h m .

[0124] Step S500: Transmit the structural features to the first self-supervised loss function for training the graph attention network to perform graph attention network learning, and transmit the dependency relationship features to the second self-supervised loss function for training the long short-term memory network to perform long short-term memory network learning to obtain the target model.

[0125] In this embodiment, before proceeding to step S500, a first self-supervised loss function and a second self-supervised loss function need to be constructed. The first self-supervised loss function can be obtained by incorporating the aforementioned power generation constraint as a penalty term into the carbon emission assessment calculation formula; the second self-supervised loss function can be constructed by determining the minimum difference between the current dual estimate and the predicted dual variable, and based on this difference.

[0126] Optionally, by introducing Lagrange multipliers to construct a Lagrange function, the power generation constraint can be incorporated into the carbon emission assessment calculation formula as a penalty term, thereby addressing the aforementioned low-carbon power generation dispatch problem. The generation constraints in the problem are relaxed. It should be noted that the Lagrange multiplier is the dual variable λ mentioned above. Understandably, the relaxation process introduces the cost of violating the generation constraints into the carbon emission assessment formula, thereby addressing the low-carbon generation dispatch problem. This is transformed into a low-carbon power generation scheduling problem with a penalty term and no constraints.

[0127] Therefore, the first self-supervised function used for GAT learning can be defined as:

[0128]

[0129] Where ρ (0 < ρ < 1000) is the penalty factor.

[0130] The purpose of this is to extend the loss function of the GAT model by using the augmented Lagrangian method, transforming the constrained problem into an unconstrained problem, so that in the low-carbon power generation scheduling problem, the carbon emission decision variables approximated by the GAT model meet the power generation constraints, thereby improving the flexibility and efficiency of the GAT model in the optimization process.

[0131] Alternatively, by minimizing the current dual estimate The LSTM model is trained by constructing a loss function based on the difference between the predicted dual variable λ and the LSTM model. The second loss function can be expressed as: Where ||·|| represents L p Norm.

[0132] Furthermore, in one embodiment, when the violation of the power generation constraint is severe, the aforementioned penalty factor ρ is introduced to penalize the violation when updating the dual variable λ, and to adjust the step size to some extent. The degree of violation of the power generation constraint is represented by the violation value. This means, through the formula Calculate violation value This allows for monitoring of the solution in the current iteration, thereby determining whether the solution satisfies all constraints. Where, ||·|| ∞ This represents the infinity norm, which is defined as the element with the largest absolute value in a vector.

[0133] Specifically, when a violation value is detected to be too large, the penalty factor ρ is increased, so that solutions that violate the power generation constraint are subject to greater penalties in subsequent iterations, thereby adjusting the optimization process for each solution that satisfies the constraint. Furthermore, solutions that violate the power generation constraint, if the violation value Greater than the tolerance value Right now Then increase the penalty factor ρ, that is Where Γ∈(0,1) is used to determine the tolerance for whether to update the penalty factor ρ. It is the updating multiplier, ρ max It is the upper limit protection value of ρ.

[0134] In this embodiment, the predicted carbon emission decision variables are obtained by passing the aforementioned structural features to a graph attention network. Then, the predicted carbon emission decision variables are passed to the first self-supervised loss function, and the model parameters of the graph attention network are optimized using gradient descent to obtain an updated graph attention network, thus achieving graph attention network learning. Additionally, the predicted dual variable λ is obtained by passing the aforementioned dependency features to a long short-term memory network. Then, the predicted dual variable is passed to the second self-supervised loss function, and the model parameters of the long short-term memory network are updated using gradient descent to obtain an updated long short-term memory network, thus achieving long short-term memory network learning. Finally, the target model is obtained based on the updated graph attention network and the updated long short-term memory network.

[0135] In this embodiment, an end-to-end self-supervised method is used to solve the low-carbon power generation scheduling problem. It should be noted that before performing graph attention network learning and long short-term memory network learning, it is necessary to first analyze λ, ρ, and... Γ、ρ max Initialization is performed to set up constraint relaxation, penalty term adjustment, and initial model weight settings. During the entire training process, the global training counter is first set. This process alternately executes graph attention network learning, long short-term memory network learning, and penalty coefficient updates until the set maximum number of global training iterations is met.

[0136] Optionally, during each round of global training, in the graph attention network learning phase, an initial value is set for the graph attention network learning counter. The instance parameter matrix X of the low-carbon power generation dispatch problem is input into the GAT model, and the construction of the power generation dispatch bipartite graph and the extraction of the structural features of the power generation dispatch bipartite graph are performed sequentially to obtain the predicted carbon emission decision variable p. * Next, the first loss function described above is used. Calculate the loss value and optimize the GAT model parameters using gradient descent. Increment the learning counter by 1 after each optimization until a predetermined number of iterations are reached. The learning phase of the graph attention network is then completed.

[0137] During the learning phase of the Long Short-Term Memory (LSTM) network, the current LSTM model is first copied to a static LSTM model. middle, Used to provide the current dual estimate Set the sequence learning counter The parameter matrix X is input into the LSTM model to extract the dependency features between the sequence data, obtaining the predicted dual variable λ. Then, a second loss function is applied. Calculate the loss value and update the parameters of the LSTM model using gradient descent. Similarly, after each optimization, the learning counter is incremented by 1 until a predetermined number of iterations are reached. The sequence learning phase is then completed.

[0138] After each round of global training completes the learning of the graph attention network and the long short-term memory network, the penalty factor ρ is updated. That is, according to the formula... Calculate violation value According to the formula Update the penalty factor ρ. At this point, the global training counter... Increase by 1, and repeat the above process of attention network learning, long short-term memory network learning, and penalty factor update until the maximum number of global training iterations is reached. After that, the entire training process ends and the target model, i.e., the trained model, is obtained. Model.

[0139] Step S600: Input test set data into the target model to obtain predicted power generation data for each generator set.

[0140] In this embodiment, the test set data is input into the target model for forward propagation calculation; then, the data output by the target model is subjected to inverse normalization processing to obtain the predicted power generation data of each generator set.

[0141] In the technical solution provided in this embodiment, by incorporating multiple energy types into the carbon emission assessment calculation formula and considering three different power generation constraints, the carbon emission characteristics of a multi-energy system can be more comprehensively reflected. By constructing a bipartite graph for power generation scheduling, the structural features of the graph are extracted using a GAT model, and the dependency features between sequential data are extracted using an LSTM model, effectively capturing the dependencies between the carbon emission calculation formula, the power generation constraint sequence, and the carbon emission decision variables. By introducing a self-supervised method into low-carbon power generation scheduling, alternating between graph attention network learning and long short-term memory network learning, the model's generalization ability in handling complex power generation scheduling constraint problems is improved.

[0142] For example, carbon emission data for 100 generator sets in a certain region are collected over 50 time periods. Then, using the carbon emission coefficients of each of the 100 generator sets, the following formula is used:

[0143]

[0144] To establish a formula for carbon emission assessment, 100 generating units were selected, including thermal, renewable energy, and nuclear power units. The carbon emission coefficients for some generating units in the first time period are as follows: Figure 7 As shown, where z i It is the label of the i-th unit.

[0145] Then, a low-carbon power generation dispatch constraint set is constructed by using the power output constraints of each generator unit, the total power output constraints of each power plant, and the power balance constraints over 50 time periods. i represents different time periods (1≤i≤50).

[0146] Then, all low-carbon power generation scheduling constraints are transformed into formulas. The form of the inequality constraint requires that the right-hand side of all inequalities be less than or equal to zero. This means decomposing the bilateral inequality into two one-sided inequalities. Taking the first time period's partial constraint data as an example, 10 ≤ P... U,1,1 ≤20,5≤P U,2,1 ≤30,1000≤P U,1,1 +P U,2,1 +……+P U,7,1 ≤2000,1000≤P U,8,1 +P U,9,1 +……+P U,13,1 ≤2000 (unit is kWh), converted into the form of inequality constraint: 10-P U,1,1 ≤0,P U,1,1 -20≤0,5-P U,2,1 ≤0,P U,2,1 -30≤0,1000-PU,1,1 +P U,2,1 +……+P U,7,1 ≤0,P U,1,1 +P U,2,1 +……+P U,7,1 -2000≤0,1000 - P U,8,1 +P U,9,1 +……+P U,13,1 ≤0,P U,8,1 +P U,9,1 +……+P U,13,1 -2000≤0。

[0147] Combined with the unilateral inequality of low - carbon power generation scheduling constraints and the carbon emission assessment calculation formula obtained after decomposition, a low - carbon power generation scheduling problem is constructed. The carbon emission assessment calculation formula is constructed into a target vector Each unilateral inequality of the power generation constraints is respectively constructed into a constraint vector The target vector and the constraint vector are spliced to form an instantiated parameter matrix X (i) . Similar processing is performed on the electricity - carbon data for 50 time periods to form a complete electricity - carbon data set The power data set is divided into a training set and a test set Next, using the formula to perform normalization on the electricity - carbon data set to convert the power data into values between 0 and 1.

[0148] Then, the construction of the bipartite graph of power generation scheduling is carried out, taking the low - carbon power generation scheduling of the first time period as an example. It is constructed into a trainable constrained power generation scheduling bipartite graph The node set where represents the set of carbon emission decision variables, represents the set of power generation constraint nodes. One side of the power generation scheduling bipartite graph contains 100 variable nodes. The feature vector of the j - th (0 ≤ j < 100) carbon emission decision variable node The value at the j - th (0 ≤ j < 101) position in is x0[j], and all other positions are 0, thus representing the feature of each carbon emission decision variable node to obtain the carbon emission decision variable node feature matrix The other side contains 218 constraint nodes. The feature vector of the i - th (0 < i ≤ 218) power generation constraint node is Using This represents the power generation constraint feature matrix. Next, a bipartite graph attention network is used for two-level message passing. First, using the formula:

[0149]

[0150] To calculate the attention coefficient when transmitting messages from the carbon emission decision variable side to the power generation constraint side. Then use the calculated attention coefficient According to the formula The message passing from the carbon emission decision variable side to the power generation constraint side is calculated to obtain a new feature representation for each power generation constraint node. right The matrix is ​​obtained after one update. Partial results after message passing are as follows Figure 8 As shown.

[0151] Through the formula:

[0152]

[0153] Calculate the attention coefficient when relaying messages from the power generation constraint side to the carbon emission decision variable side. Using the calculated attention coefficient, according to the formula The message passing from the power generation constraint side to the carbon emission decision variable side is calculated to obtain a new feature representation for each carbon emission decision variable. right The matrix is ​​obtained after one update. The results of message passing are as follows Figure 9 As shown.

[0154] According to the formula The updated carbon emission decision variable node feature matrix The input is fed into a multilayer perceptron to predict the carbon emission decision variable value.

[0155] Then, the dependency features between the sequence data are extracted to analyze the low-carbon power generation scheduling for the first time period. For example, the dependency between the carbon emission assessment calculation formula and the power generation constraint is extracted. The total sequence length is 219, and the dimension of each sequence vector is 101. According to the formula:

[0156] p t =Θ(W p ·[h t-1 ,x t ]+b p )

[0157] q t =Θ(W q ·[ht-1 ,x t ]+b q )

[0158]

[0159] target vector and constraint vector The input is progressively fed into the LSTM. At each sequence step, the LSTM updates the current hidden state based on the current input vector and the hidden state of the previous step, thus calculating the hidden states for all sequence steps. Select hidden layer state As a dual variable representation of LSTM prediction.

[0160] Then, after defining the first loss function, the second loss function, and the update rule for the penalty factor when violating power generation constraints, data prediction of power generation for each generator unit is performed. First, λ is initialized to 0 to ensure unbiased estimation; ρ is set to 0.5 to balance the constraint relaxation in the early stage, using empirical values. To control the growth rate of the penalty factor, Γ = 0.5 limits over-adjustment, ρ max =1000 To avoid excessively large penalty factors leading to numerical instability, a maximum number of global training iterations is set. Maximum counter count during graph attention network learning Maximum counter count during Long Short-Term Memory network learning training set Z 1 The entire training process is executed as input, with the global training internally alternating between learning the graph attention network, learning the long short-term memory network, and updating the penalty factor, thus updating the parameters of the GAT model and the LSTM model. After the update is complete, the test set Z is used. 2 The low-carbon power generation scheduling problem in 10 time periods is predicted using a trained model, and the predicted actual power generation of each generator unit is obtained after inverse normalization.

[0161] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the low-carbon power generation scheduling method based on graph attention networks and sequence models in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0162] This application provides a low-carbon power generation scheduling device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the low-carbon power generation scheduling method based on graph attention network and sequence model in the above embodiment 1.

[0163] The following is for reference. Figure 10 The diagram illustrates a structural schematic suitable for implementing the low-carbon power generation dispatching equipment of the embodiments of this application. The low-carbon power generation dispatching equipment in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), vehicle terminals (e.g., vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 10 The low-carbon power generation dispatching equipment shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0164] like Figure 10 As shown, the low-carbon power generation dispatching equipment may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the low-carbon power generation dispatching equipment to communicate wirelessly or wiredly with other equipment to exchange data. Although the figure shows low-carbon power generation dispatching equipment with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0165] In particular, according to the embodiments disclosed in the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program comprising program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0166] The low-carbon power generation scheduling device provided in this application employs a low-carbon power generation scheduling method based on graph attention networks and sequence models as described in the above embodiments. This method addresses the technical problems of commonly used low-carbon power generation scheduling methods, which not only have high computational costs but also require significant improvement in convergence and efficiency under high-dimensional constrained spaces, and cannot guarantee finding the global optimal solution. Compared with existing technologies, the beneficial effects of the low-carbon power generation scheduling device provided in this application are the same as those of the low-carbon power generation scheduling method based on graph attention networks and sequence models provided in the above embodiments. Furthermore, other technical features of this low-carbon power generation scheduling device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0167] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0168] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0169] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute a low-carbon power generation scheduling method based on graph attention network and sequence model in the above embodiments.

[0170] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0171] The aforementioned computer-readable storage medium may be included in the low-carbon power generation dispatching equipment; or it may exist independently and not be assembled into the low-carbon power generation dispatching equipment.

[0172] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the low-carbon power generation dispatching equipment, cause the low-carbon power generation dispatching equipment to:

[0173] The power generation of generator sets of each power generation type in the power generation system is obtained, as well as the power generation limit of each generator set in the power generation system, the power generation limit of each power plant, and the power generation limit of the power generation terminal.

[0174] The power generation of each generator set is used as the carbon emission decision variable in the carbon emission assessment calculation formula, and corresponding first power generation constraint, second power generation constraint and third power generation constraint are constructed based on the power generation limit of the generator set, the power generation limit of the power plant and the power generation limit at the power generation end.

[0175] A bipartite graph for power generation scheduling is constructed based on the carbon emission decision variables and the power generation constraints, and the carbon emission assessment formula and the power generation constraints are used as sequence data.

[0176] Based on graph attention networks, the structural features of the power generation scheduling bipartite graph are extracted, and based on long short-term memory networks, the dependency features between the sequence data are extracted.

[0177] The structural features are passed to the first self-supervised loss function for training the graph attention network to perform graph attention network learning, and the dependency features are passed to the second self-supervised loss function for training the long short-term memory network to perform long short-term memory network learning, so as to obtain the target model;

[0178] Input the test set data into the target model to obtain the predicted power generation data of each generator set.

[0179] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0180] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0181] The modules described in the embodiments of the present application may be implemented in software or hardware, wherein the name of a module does not necessarily limit the unit itself.

[0182] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned low-carbon power generation scheduling method based on graph attention networks and sequence models. This addresses the technical problems of commonly used low-carbon power generation scheduling methods, which not only have high computational costs but also require significant improvement in convergence and efficiency under high-dimensional constraints, and cannot guarantee finding a globally optimal solution. Compared with existing technologies, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the low-carbon power generation scheduling method based on graph attention networks and sequence models provided in the above embodiments, and will not be elaborated upon here.

[0183] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of a low-carbon power generation scheduling method based on graph attention networks and sequence models as described above.

[0184] The computer program product provided in this application can solve the technical problems that commonly used low-carbon power generation scheduling methods are not only computationally expensive, but also have low convergence and efficiency under high-dimensional constraint spaces, and cannot guarantee finding the global optimal solution. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the low-carbon power generation scheduling method based on graph attention network and sequence model provided in the above embodiments, and will not be repeated here.

[0185] The above description is only part of the embodiments of the present application and does not limit the patent scope of the present application. All equivalent structural transformations made by using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A low-carbon power generation scheduling method based on graph attention networks and sequence models, characterized in that, The method includes: The power generation of generator sets of each power generation type in the power generation system is obtained, as well as the power generation limit of each generator set in the power generation system, the power generation limit of each power plant, and the power generation limit of the power generation terminal. The power generation of each generator set is used as the carbon emission decision variable in the carbon emission assessment calculation formula, and corresponding power generation constraints are constructed based on the power generation limit of the generator set, the power generation limit of the power plant and the power generation limit at the power generation end. The power generation constraints include a first power generation constraint, a second power generation constraint and a third power generation constraint. A bipartite graph for power generation scheduling is constructed based on the carbon emission decision variables and the power generation constraints, and the carbon emission assessment formula and the power generation constraints are used as sequence data. Based on graph attention networks, the structural features of the power generation scheduling bipartite graph are extracted, and based on long short-term memory networks, the dependency features between the sequence data are extracted. The structural features are passed to the first self-supervised loss function for training the graph attention network to perform graph attention network learning, and the dependency features are passed to the second self-supervised loss function for training the long short-term memory network to perform long short-term memory network learning, so as to obtain the target model; Input the test set data into the target model to obtain the predicted power generation data of each generator set; The steps of constructing a bipartite graph for power generation scheduling based on the carbon emission decision variables and the power generation constraints, and of using the carbon emission assessment formula and the power generation constraints as sequential data, include: The low-carbon power generation scheduling problem is constructed by combining the carbon emission assessment formula and the power generation constraints. The low-carbon power generation scheduling problem is encoded to obtain the power generation scheduling bipartite graph, which includes carbon emission decision variable nodes, power generation constraint nodes, and edges connecting the carbon emission decision variable nodes and the power generation constraint nodes. A target vector is formed based on the carbon emission assessment formula, and a constraint vector is formed based on the first power generation constraint, the second power generation constraint, and the third power generation constraint. The sequence data includes the target vector and the constraint vector. The step of extracting the structural features of the generation scheduling bipartite graph based on the graph attention network includes: Based on the first-level message transmission from the carbon emission decision variable node to the power generation constraint node, the impact of the carbon emission decision variable on the power generation constraint is captured; Based on the second-level message transmission from the power generation constraint node to the carbon emission decision variable node, the reaction of the power generation constraint on the carbon emission decision variable is determined. Based on the impact of the carbon emission decision variables on power generation constraints and the reaction of power generation constraints on carbon emission decision variables, the structural characteristics of the power generation dispatch bipartite graph are obtained.

2. The method as described in claim 1, characterized in that, The step of extracting dependency features between the sequence data based on the Long Short-Term Memory network includes: The target vector and the constraint vector are passed to the long short-term memory network to determine the candidate memory state and the gating value of the output gate; The hidden state is obtained based on the candidate memory state and the gating value of the output gate; The dependency features are determined based on the hidden state.

3. The method as described in claim 1, characterized in that, Before the steps of passing the structural features to the first self-supervised loss function for training the graph attention network and performing graph attention network learning, and passing the dependency features to the second self-supervised loss function for training the long short-term memory network and performing long short-term memory network learning, the method further includes: The power generation constraint is incorporated as a penalty term into the carbon emission assessment formula to obtain the first self-supervised loss function; and... Determine how to minimize the difference between the current dual estimate and the predicted dual variable, and construct the second self-supervised loss function based on the difference.

4. The method as described in claim 3, characterized in that, The steps of passing the structural features to the first self-supervised loss function for training the graph attention network, and passing the dependency features to the second self-supervised loss function for training the long short-term memory network, to obtain the target model, include: The structural features are passed to the graph attention network to obtain the predicted carbon emission decision variables; The predicted carbon emission decision variables are passed to the first self-supervised loss function, and the model parameters of the graph attention network are optimized using the gradient descent method to obtain an updated graph attention network; and, The dependency features are passed to the long short-term memory network to obtain the predicted dual variable; The predicted dual variable is passed to the second self-supervised loss function, and the model parameters of the long short-term memory network are updated by gradient to obtain the updated long short-term memory network. The target model is obtained based on the updated graph attention network and the updated long short-term memory network.

5. The method as described in claim 3, characterized in that, The step of inputting the input test set data into the target model to obtain the predicted power generation data of each generator set includes: Input the test set data into the target model for forward propagation calculation; The data output by the target model is subjected to inverse normalization to obtain the predicted power generation data of each generator set.

6. The method according to any one of claims 1 to 5, characterized in that, The power generation types include fossil fuel power generation, nuclear power generation, and renewable energy power generation; the carbon emission assessment formula consists of carbon emission calculation formulas for fossil fuel power generation, nuclear power generation, and renewable energy power generation; the step of using the power generation of each of the aforementioned generator sets as the carbon emission decision variable in the carbon emission assessment calculation formula includes: The power generation of the fossil fuel power generation generator set is used as the carbon emission decision variable in the calculation formula of the carbon emission of fossil fuel power generation; The power generation of the nuclear power generation generator set is used as the carbon emission decision variable in the calculation formula of the nuclear power generation carbon emissions; The power generation of the renewable energy generator sets is used as the carbon emission decision variable in the calculation formula for the carbon emissions of renewable energy generation.

7. A low-carbon power generation dispatching device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of a low-carbon power generation scheduling method based on graph attention networks and sequence models as described in any one of claims 1 to 6.

8. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of a low-carbon power generation scheduling method based on graph attention network and sequence model as described in any one of claims 1 to 6.

Images