Power system pre-decision method and device embedded with spatio-temporal information

Abstract

The invention relates to the technical field of power system stability control, and discloses a spatial-temporal information embedded power system pre-decision method and device. The method comprises the steps of obtaining power system data in real time; wherein the power system data comprises power grid operation data and power grid topological structure data; extracting time sequence features and spatial features of the power system data by using a transient stability evaluator, and evaluating a transient stability evaluation result of the power system; and when the transient stability evaluation result is instability, generating an optimal action by using a deep reinforcement learning agent based on the data of the power system, and controlling the power system according to the optimal action. According to the method, the time sequence characteristic and spatial characteristic information in the power system operation data can be effectively extracted, the accuracy of transient stability evaluation is improved, and the stability of the power system is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system stability control, and in particular to a power system pre-decision method and device embedded with spatiotemporal information. Background Art

[0002] With the increasing scale and complexity of power systems, its transient stability problem has become a key factor affecting the safe and stable operation of power systems. In order to avoid large-scale power outages caused by transient instability of power systems, it is urgent to develop effective transient stability assessment and control decision-making methods for power systems. Traditional transient stability assessment and control decision-making methods are model-driven time-domain simulation methods. Because their models contain nonlinear differential algebraic equations, the calculations are complex and the operation time is long, making it difficult to adapt to the needs of online stability monitoring and instability prevention and control decisions. Most of the existing deep reinforcement learning network structure designs only focus on the time series measurement data of the power system or the spatial feature information of the power grid topology for design. However, the transient process of the power system under large disturbances has obvious spatiotemporal characteristics. Accurately extracting key spatiotemporal features for transient stability assessment and generating control decision methods is still a problem worth studying.

[0003] At present, model-driven time-domain simulation needs to consider a large number of nonlinear differential constraints, which often leads to long solution time and poor convergence, and cannot meet the real-time requirements of current preventive control. At present, deep reinforcement learning methods do not take into account the impact of the spatial topological characteristics and temporal characteristics of the power grid on the transient stability of the power grid, resulting in low transient stability of the power grid. Summary of the invention

[0004] The present invention provides a method and device for pre-decision making of an electric power system embedded with time-space information, which can effectively extract the time series characteristics and spatial characteristic information in the operation data of the electric power system, improve the accuracy of transient stability assessment, and thus improve the stability of the electric power system.

[0005] In order to solve the above technical problems, the present invention provides a power system pre-decision method embedded with spatiotemporal information, comprising:

[0006] Acquire power system data in real time; wherein the power system data includes power grid operation data and power grid topology data;

[0007] Extracting the time series characteristics and spatial characteristics of the power system data using a transient stability evaluator, and evaluating the transient stability evaluation result of the power system;

[0008] When the transient stability assessment result is instability, based on the power system data, a deep reinforcement learning agent is used to generate an optimal action, and the power system is controlled according to the optimal action.

[0009] Further, the method of extracting the time series characteristics and spatial characteristics of the power system data by using a transient stability evaluator and evaluating the transient stability evaluation result of the power system is specifically as follows:

[0010] Extracting time series characteristics of power grid operation data;

[0011] Extract spatial features of power grid topology data;

[0012] splicing the temporal features and the spatial features to form feature data;

[0013] Nonlinear mapping is performed on the characteristic data to generate a transient stability assessment result of the power system.

[0014] Furthermore, the time series characteristics of the extracted power grid operation data are specifically:

[0015] Cut the power grid operation data to form odd sequences and even sequences;

[0016] exchanging information between the odd sequence and the even sequence through interactive learning to form an updated odd sequence and an updated even sequence;

[0017] The updated odd sequence and the updated even sequence are determined as timing features.

[0018] Furthermore, the information exchange between the odd sequence and the even sequence is performed through interactive learning to form an updated odd sequence and an updated even sequence, specifically:

[0019] Convolution calculation, exponential calculation and Hadamard product are performed on the odd sequence and the even sequence to form an intermediate odd sequence and an intermediate even sequence, specifically:

[0020]

[0021] In the formula, is an intermediate odd sequence; is the intermediate even sequence; ⊙ is the Hadamard product; exp(·) is the exponential function calculation; φ and ψ are one-dimensional convolution modules respectively; X odd is an odd sequence; X even is an even sequence;

[0022] Information is exchanged between the intermediate odd sequence and the intermediate even sequence to form an updated odd sequence and an updated even sequence, specifically:

[0023]

[0024] In the formula, X′ odd is the updated odd sequence; X′even is the updated even sequence; ρ and η are one-dimensional convolution modules respectively; is an intermediate odd sequence; is an intermediate even sequence.

[0025] Furthermore, the extraction of spatial features of the power grid topology data is specifically as follows:

[0026] Acquiring a system neighborhood matrix in the power grid topology data;

[0027] Based on the system neighborhood matrix, an attention mechanism is used to calculate the attention coefficient between each node in the power system and the corresponding neighbor node;

[0028] Normalize each attention coefficient through the Softmax function;

[0029] The normalized attention coefficient is used to aggregate the features of each node in the power system to form spatial features.

[0030] Furthermore, the nonlinear mapping process is performed on the characteristic data to generate a transient stability assessment result of the power system, specifically:

[0031] According to the stability association rule function of the transient stability evaluator, a transient stability evaluation result of the power system is generated based on the characteristic data, specifically:

[0032] [W0,W1]=ψ([X V ,X θ ,X PG ,X QG ],A)

[0033] Where [W0, W1] is the transient stability assessment result; ψ(·) is the stability association rule function of the transient stability assessor; [X V ,X θ ,X PG ,X QG ] is the characteristic data; among them, X V is the voltage amplitude at the grid node; X θ is the phase angle at the grid node; X PG is the active power at the grid node; X QG is the reactive power at the grid node; A is the system neighborhood matrix.

[0034] Furthermore, the training process of the deep reinforcement learning agent is specifically as follows:

[0035] Using historical power system data as training data for a deep reinforcement learning agent; wherein the historical power system data includes historical power grid operation data and historical power grid topology data;

[0036] Using the historical power system data, controlling the deep reinforcement learning agent to generate a plurality of first samples, and storing each of the first samples in an experience pool;

[0037] The deep reinforcement learning agent is trained using a plurality of first samples in the experience pool, and when a preset maximum number of training rounds is reached, it is determined that the training of the deep reinforcement learning agent is completed.

[0038] Furthermore, the deep reinforcement learning agent further includes:

[0039] Determining the active output adjustment of the generator as the action space of the deep reinforcement learning agent;

[0040] Among them, the upper and lower limits of the generator active output adjustment are specifically:

[0041]

[0042] In the formula, The upper and lower limits of the generator active output adjustment; α g is a preset constant, ranging from (0,1); is the minimum generator active output; is the maximum generator active output;

[0043] Among them, the active output of the generator after adjustment is specifically:

[0044]

[0045] In the formula, It is the active output after the active output of the generator is adjusted; It is the active output of the generator before active output adjustment; It is the upper and lower limit range of the generator active output adjustment; is the minimum generator active output; It is the maximum active output of the generator.

[0046] Furthermore, the deep reinforcement learning agent further includes:

[0047] The reward function of the deep reinforcement learning agent is specifically:

[0048]

[0049] R3=-α6W1

[0050] In the formula, r t is the reward function of the deep reinforcement learning agent; α1~α6 are preset constants; C gi Adjust the cost for the i-th generator; Pgi To adjust the output; is the original output of the i-th generator; Q Gi is the reactive power of the i-th generator; is the maximum reactive power of the i-th generator; is the minimum reactive power of the i-th generator; P Gs To balance the node output; Output for the maximum balance node; is the minimum balance node output; V i is the voltage at node i; V i max is the maximum allowable voltage at node i; V i min is the maximum allowable voltage of node i; W1 is the probability of instability.

[0051] The present invention provides a method for pre-decision-making of an electric power system embedded with time-space information, and acquires electric power system data in real time; wherein the electric power system data includes power grid operation data and power grid topology structure data; a transient stability evaluator is used to extract the time series characteristics and spatial characteristics of the electric power system data, and the transient stability evaluation result of the electric power system is evaluated; when the transient stability evaluation result is unstable, based on the electric power system data, a deep reinforcement learning agent is used to generate the optimal action, and the electric power system is controlled according to the optimal action. The present invention can effectively extract the time series characteristics and spatial characteristics information in the electric power system operation data, improve the accuracy of transient stability evaluation, thereby better providing the optimal strategy for the risk prevention and control of the electric power system, and further improving the stability of the electric power system.

[0052] Accordingly, the present invention provides a power system pre-decision device embedded with spatiotemporal information, comprising: a data acquisition module, a transient assessment module and a control module;

[0053] The data acquisition module is used to acquire power system data in real time; wherein the power system data includes power grid operation data and power grid topology data;

[0054] The transient assessment module is used to extract the time series characteristics and spatial characteristics of the power system data using the transient stability assessor, and evaluate the transient stability assessment result of the power system;

[0055] The control module is used to generate optimal actions based on the power system data using a deep reinforcement learning agent when the transient stability assessment result is instability, and control the power system according to the optimal actions.

[0056] The present invention provides a power system pre-decision device embedded with time and space information, which acquires power system data in real time based on the organic combination of modules; wherein the power system data includes power grid operation data and power grid topology structure data; a transient stability evaluator is used to extract the time series characteristics and spatial characteristics of the power system data, and the transient stability evaluation result of the power system is evaluated; when the transient stability evaluation result is unstable, based on the power system data, a deep reinforcement learning agent is used to generate the optimal action, and the power system is controlled according to the optimal action. The present invention can effectively extract the time series characteristics and spatial characteristic information in the power system operation data, improve the accuracy of transient stability evaluation, thereby better providing the optimal strategy for power system risk prevention and control, and thus improving the stability of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic flow chart of an embodiment of a method for pre-decision making of a power system embedded with spatiotemporal information provided by the present invention;

[0058] Figure 2 A schematic structural diagram of an embodiment of a power system pre-decision making device embedded with spatiotemporal information provided by the present invention. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0061] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0062] Example 1

[0063] See also Figure 1 , is a flow chart of an embodiment of a method for pre-decision making of a power system embedded with spatiotemporal information provided by the present invention, the method comprises steps 101 to 103, each of which is specifically as follows:

[0064] Step 101: Acquire power system data in real time; wherein the power system data includes power grid operation data and power grid topology data.

[0065] In the first embodiment of the present invention, the power system data acquired in real time is used to evaluate the transient stability of the new power system, including power grid operation data and power grid topology data. The power grid operation data X includes the voltage amplitude X at the power grid node. V 、Phase Angle X θ , Active power X PG and reactive power X QG , the power grid operation data X can be expressed as X = [X V ,X θ ,X PG ,X QG The power grid topology data includes the neighborhood matrix A of the power grid topology. The neighborhood matrix A is an N×N matrix, where N is the number of nodes. The elements A in the neighborhood matrix A are ij Represents the connection relationship between node i and node j. If there is a direct connection between node i and node j, then A is determined. ij The value of is 1; if there is no direct connection between node i and node j, then A is determined ij The value of is 0.

[0066] Step 102: extracting the time series characteristics and spatial characteristics of the power system data using a transient stability evaluator, and evaluating a transient stability evaluation result of the power system.

[0067] Furthermore, in the first embodiment of the present invention, a transient stability evaluator is used to extract the time series characteristics and spatial characteristics of the power system data, and to evaluate the transient stability evaluation result of the power system, specifically:

[0068] Extracting time series characteristics of power grid operation data;

[0069] Extract spatial features of power grid topology data;

[0070] splicing the temporal features and the spatial features to form feature data;

[0071] Nonlinear mapping is performed on the characteristic data to generate a transient stability assessment result of the power system.

[0072] In the first embodiment of the present invention, the transient stability assessor is built based on a hybrid neural network of Sample Convolution and Interaction Network (SCI) and Graph Attention Network (GAT). The number of neural network layers of the transient stability assessor is 5, including a data input layer, a spatiotemporal information extraction layer, a spatiotemporal splicing layer, a fully connected layer and an output layer. The real-time acquired power system data is transmitted to the data input layer, and the time series features and spatial features can be extracted respectively through the SCI layer and the GAT layer in the spatiotemporal information extraction layer. The time series features and spatial features are spliced ​​in the spatiotemporal splicing layer, and the spliced ​​feature data is nonlinearly mapped in the fully connected layer, and the transient stability assessment results of the power system are output using the output layer.

[0073] Furthermore, in the first embodiment of the present invention, the time series characteristics of the power grid operation data are extracted, specifically:

[0074] Cut the power grid operation data to form odd sequences and even sequences;

[0075] exchanging information between the odd sequence and the even sequence through interactive learning to form an updated odd sequence and an updated even sequence;

[0076] The updated odd sequence and the updated even sequence are determined as timing features.

[0077] Further, in the first embodiment of the present invention, the odd sequence and the even sequence are exchanged with each other through interactive learning to form an updated odd sequence and an updated even sequence, specifically:

[0078] Convolution calculation, exponential calculation and Hadamard product are performed on the odd sequence and the even sequence to form an intermediate odd sequence and an intermediate even sequence, specifically:

[0079]

[0080] In the formula, is an intermediate odd sequence; is the intermediate even sequence; ⊙ is the Hadamard product; exp(·) is the exponential function calculation; φ and ψ are one-dimensional convolution modules respectively; X odd is an odd sequence; X even is an even sequence;

[0081] Information is exchanged between the intermediate odd sequence and the intermediate even sequence to form an updated odd sequence and an updated even sequence, specifically:

[0082]

[0083] In the formula, X′ odd is the updated odd sequence; X′ even is the updated even sequence; ρ and η are one-dimensional convolution modules respectively; is an intermediate odd sequence; is an intermediate even sequence.

[0084] In the first embodiment of the present invention, the power grid operation data input to the transient stability assessor can be decomposed into two subsequences, namely, odd sequence X and odd sequence X, after the segmentation operation of the SCI layer in the spatiotemporal information extraction layer. odd Even sequence X even ; Odd sequence X odd Even sequence X even After convolution calculation, exponential calculation and Hadamard product, it is formed Intermediate odd sequence and intermediate even sequence Finally, information exchange between odd and even sequences is achieved through interactive learning to form updated odd sequences and even sequences, and the updated odd sequences and updated even sequences are determined as time series features.

[0085] Furthermore, in the first embodiment of the present invention, the spatial features of the power grid topology data are extracted, specifically:

[0086] Acquiring a system neighborhood matrix in the power grid topology data;

[0087] Based on the system neighborhood matrix, an attention mechanism is used to calculate the attention coefficient between each node in the power system and the corresponding neighbor node;

[0088] Normalize each attention coefficient through the Softmax function;

[0089] The normalized attention coefficient is used to aggregate the features of each node in the power system to form spatial features.

[0090] In the first embodiment of the present invention, while extracting the time series features from the time series data, the system neighborhood matrix is ​​obtained from the power grid topology data, and the system neighborhood matrix is ​​used as the input feature. The GAT layer in the spatiotemporal information extraction layer calculates the correlation between each node and its neighboring nodes through the attention mechanism, thereby aggregating the node features in the graph. For node i and its neighboring node j, the attention coefficient between them is:

[0091] e ij =α[concat(Wa i ,Wa j )],j∈N i

[0092] In the formula, e ijis the attention coefficient between node i and its neighbor node j; α is the calculation parameter of the shared attention mechanism; concat(·) is the concatenation operation; W is the shared weight parameter matrix; a i is the feature of node i; a j is the feature of node j; N i is the set of adjacent nodes of node i.

[0093] After obtaining the attention coefficients between nodes, the attention coefficients are normalized by the Softmax function, and the expression is:

[0094]

[0095] In the formula, α ij is the normalized attention coefficient between node i and its neighbor node j; LeakyReLU(·) is the nonlinear activation function; exp(·) is the exponential function; e ij is the attention coefficient between node i and its neighbor node j; e ih is the attention coefficient between node i and its neighbor node j; N i is the set of adjacent nodes of node i.

[0096] After obtaining the normalized attention coefficient, feature aggregation is performed on node i, and the expression is:

[0097]

[0098] In the formula, a′ i The new features of each node output by the GAT layer are integrated with domain information; (·) is the ReLU activation function; α ij is the normalized attention coefficient between node i and its neighbor node j; W is the shared weight parameter matrix; a j is the feature of node j; N i is the set of adjacent nodes of node i.

[0099] Furthermore, in the first embodiment of the present invention, the characteristic data is subjected to nonlinear mapping processing, and a transient stability assessment result of the power system is generated, specifically:

[0100] According to the stability association rule function of the transient stability evaluator, a transient stability evaluation result of the power system is generated based on the characteristic data, specifically:

[0101] [W0,W1]=ψ([X V ,X θ ,X PG ,X QG ],A)

[0102] Where [W0, W1] is the transient stability assessment result; ψ(·) is the stability association rule function of the transient stability assessor; [X V ,X θ ,X PG ,X QG ] is the characteristic data; among them, X V is the voltage amplitude at the grid node; X θ is the phase angle at the grid node; X PG is the active power at the grid node; X QG is the reactive power at the grid node; A is the system neighborhood matrix.

[0103] In the first embodiment of the present invention, after the extraction of time series features and spatial features is completed, the fully connected layer of the transient stability evaluator is used to perform nonlinear mapping processing, and the transient stability evaluation result of the power system is generated according to the stable association rule function of the transient stability evaluator, and the probability W0 of power grid stability and the probability W1 of power grid instability are outputted by the output layer of the transient stability evaluator. Wherein, W0+W1=1, when W0>W1, it is determined that the power system is stable; when W0<W1, it is determined that the power system is unstable, and the labels 0 and 1 represent the stability and instability of the power system respectively.

[0104] In the first embodiment of the present invention, the training process of the transient stability evaluator is: obtaining historical power system data of the power system, including historical power grid topology and historical power grid operation characteristics. Performing time domain simulation on the acquired historical power system data, and determining the transient stability of the power system according to the simulation results. Among them, the stability label is: A data set is constructed based on operating characteristic data, topological information and simulation annotation results, and the data records of all operating conditions are summarized to finally form a transient stability data knowledge base. The data in the transient stability data knowledge base is used as training data for training the transient stability evaluator. The transient stability evaluator uses the SCI layer and the GAT layer to extract the temporal characteristics and spatial characteristics information in the training data respectively. The training data is used as the input feature and the corresponding transient stability index is used as the output. The mapping relationship between the input and the transient stability of the system is established, and the transient stability evaluator is embedded in the environment to replace the transient constraints in the TSCOPF model, and the stability association rule function of the transient stability evaluator is generated.

[0105] The transient stability evaluator based on the SCI-GAT hybrid neural network of the present invention can extract the spatiotemporal characteristic information in the transient process and establish a mapping relationship between the input and the transient stability of the system to replace the nonlinear differential algebraic equation calculation and complicated transient stability verification in the traditional transient stability constrained optimal power flow model.

[0106] Step 103: When the transient stability assessment result is instability, based on the power system data, a deep reinforcement learning agent is used to generate an optimal action, and the power system is controlled according to the optimal action.

[0107] Furthermore, in the first embodiment of the present invention, the training process of the deep reinforcement learning agent is specifically as follows:

[0108] Using historical power system data as training data for a deep reinforcement learning agent; wherein the historical power system data includes historical power grid operation data and historical power grid topology data;

[0109] Using the historical power system data, controlling the deep reinforcement learning agent to generate a plurality of first samples, and storing each of the first samples in an experience pool;

[0110] The deep reinforcement learning agent is trained using a plurality of first samples in the experience pool, and when a preset maximum number of training rounds is reached, it is determined that the training of the deep reinforcement learning agent is completed.

[0111] Furthermore, in the first embodiment of the present invention, the deep reinforcement learning agent further includes:

[0112] Determining the active output adjustment of the generator as the action space of the deep reinforcement learning agent;

[0113] Among them, the upper and lower limits of the generator active output adjustment are specifically:

[0114]

[0115] In the formula, The upper and lower limits of the generator active output adjustment; α g is a preset constant, ranging from (0,1); is the minimum generator active output; is the maximum generator active output;

[0116] Among them, the active output of the generator after adjustment is specifically:

[0117]

[0118] In the formula, It is the active output after the active output of the generator is adjusted; It is the active output of the generator before active output adjustment; It is the upper and lower limit range of the generator active output adjustment; is the minimum generator active output; It is the maximum active output of the generator.

[0119] Furthermore, in the first embodiment of the present invention, the deep reinforcement learning agent further includes:

[0120] The reward function of the deep reinforcement learning agent is specifically:

[0121]

[0122] R3=-α6W1

[0123] In the formula, r t is the reward function of the deep reinforcement learning agent; α1~α6 are preset constants; C gi Adjust the cost for the i-th generator; P gi To adjust the output; is the original output of the i-th generator; Q Gi is the reactive power of the i-th generator; is the maximum reactive power of the i-th generator; is the minimum reactive power of the i-th generator; P Gs To balance the node output; Output for the maximum balance node; is the minimum balance node output; V i is the voltage at node i; V i max is the maximum allowable voltage at node i; V i min is the maximum allowable voltage of node i; W1 is the probability of instability.

[0124] In the first embodiment of the present invention, a deep reinforcement learning agent is constructed with the goal of safe operation of a new power system, wherein the deep reinforcement learning elements include state space, action space, reward function, action-valuation network (Actor-Critic) and target action-valuation network. The spatiotemporal information extraction layer of the transient stability evaluator is used as the input of the action network and valuation network of the deep reinforcement learning agent. The deep reinforcement learning agent gives actions by perceiving and learning the environment. After the environment accepts the action of the deep reinforcement learning agent, it obtains the state at the next moment and the reward at the next moment and performs the next round of state evaluation. According to this method, the deep reinforcement learning agent collects the "state-action-reward-next moment state" experience element by continuously interacting with the environment, and completes offline training by sampling the experience pool.

[0125] In the first embodiment of the present invention, the deep reinforcement learning agent is trained using the DDPG algorithm, and the spatiotemporal information extraction layer in the trained transient stability evaluator is embedded into the action and valuation network in the DDPG network structure. tThis is consistent with the input of the stability association rule function ψ(·) of the transient stability assessor, that is, s t ={X V ,X θ ,X PG ,X QG ,A}. The generator active output adjustment is determined as the action space of the deep reinforcement learning agent. The reward function of the deep reinforcement learning agent is the feedback of the environment to the deep reinforcement learning agent after the action, which is obtained by processing the TSCOPF model.

[0126] In the first embodiment of the present invention, in each interactive learning of the deep reinforcement learning agent, the deep reinforcement learning agent observes the environment state s t , given an action a t , the environment accepts action a t Backward deep reinforcement learning agent feedback reward r t And enter the next state s t+1 , the experienced environmental state s t 、Action a t , Reward t and the next state s t+1 The data is stored in the experience pool, and the policy network is trained by sampling from the experience pool until the maximum number of training rounds is reached, and the deep reinforcement learning agent training is completed.

[0127] The present invention uses the trained transient stability evaluator as an auxiliary link for the interaction between deep reinforcement learning and the environment, and uses the evaluator to replace the transient constraints in the TSCOPF model, which can greatly improve the computing efficiency and reduce the training cost of deep reinforcement learning.

[0128] In the first embodiment of the present invention, after the deep reinforcement learning agent is trained offline, it is loaded into the real-time operation stage, the current environmental state data of the real-time operation stage is transmitted to the constructed transient stability evaluator, and the transient stability evaluator is used to evaluate the transient stability of the current operating conditions; if the transient stability assessment result output by the transient stability evaluator indicates that there is a risk of instability, based on the trained deep reinforcement learning model, the optimal action is generated according to the current environmental state data, and the power system is controlled according to the optimal action, thereby realizing transient stability preventive control.

[0129] In summary, the first embodiment of the present invention provides a method for pre-decision making of an electric power system that embeds temporal and spatial information, and acquires electric power system data in real time; wherein the electric power system data includes power grid operation data and power grid topology data; a transient stability evaluator is used to extract the temporal characteristics and spatial characteristics of the electric power system data, and the transient stability evaluation result of the electric power system is evaluated; when the transient stability evaluation result is unstable, based on the electric power system data, a deep reinforcement learning agent is used to generate the optimal action, and the electric power system is controlled according to the optimal action. The present invention can effectively extract the temporal characteristics and spatial characteristics information in the electric power system operation data, improve the accuracy of the transient stability evaluation, thereby better providing the optimal strategy for the risk prevention and control of the electric power system, and thus improving the stability of the electric power system.

[0130] Example 2

[0131] See also Figure 2 , is a schematic structural diagram of an embodiment of a power system pre-decision-making device embedded with spatiotemporal information provided by the present invention, the device comprising a data acquisition module 201, a transient assessment module 202 and a control module 203;

[0132] The data acquisition module 201 is used to acquire power system data in real time; wherein the power system data includes power grid operation data and power grid topology data;

[0133] The transient assessment module 202 is used to extract the time series characteristics and spatial characteristics of the power system data using a transient stability assessor, and evaluate the transient stability assessment result of the power system;

[0134] The control module 203 is used to generate optimal actions based on the power system data using a deep reinforcement learning agent when the transient stability assessment result is instability, and control the power system according to the optimal actions.

[0135] Furthermore, in the second embodiment of the present invention, a transient stability evaluator is used to extract the time series characteristics and spatial characteristics of the power system data, and to evaluate the transient stability evaluation result of the power system, specifically:

[0136] Extracting time series characteristics of power grid operation data;

[0137] Extract spatial features of power grid topology data;

[0138] splicing the temporal features and the spatial features to form feature data;

[0139] Nonlinear mapping is performed on the characteristic data to generate a transient stability assessment result of the power system.

[0140] Furthermore, in the second embodiment of the present invention, the time series characteristics of the power grid operation data are extracted, specifically:

[0141] Cut the power grid operation data to form odd sequences and even sequences;

[0142] exchanging information between the odd sequence and the even sequence through interactive learning to form an updated odd sequence and an updated even sequence;

[0143] The updated odd sequence and the updated even sequence are determined as timing features.

[0144] Further, in the second embodiment of the present invention, the odd sequence and the even sequence are exchanged with each other through interactive learning to form an updated odd sequence and an updated even sequence, specifically:

[0145] Convolution calculation, exponential calculation and Hadamard product are performed on the odd sequence and the even sequence to form an intermediate odd sequence and an intermediate even sequence, specifically:

[0146]

[0147] In the formula, is an intermediate odd sequence; is the intermediate even sequence; ⊙ is the Hadamard product; exp(·) is the exponential function calculation; φ and ψ are one-dimensional convolution modules respectively; X odd is an odd sequence; X even is an even sequence;

[0148] Information is exchanged between the intermediate odd sequence and the intermediate even sequence to form an updated odd sequence and an updated even sequence, specifically:

[0149]

[0150] In the formula, X′ odd is the updated odd sequence; X′ even is the updated even sequence; ρ and η are one-dimensional convolution modules respectively; is an intermediate odd sequence; is an intermediate even sequence.

[0151] Furthermore, in the second embodiment of the present invention, the spatial features of the power grid topology data are extracted, specifically:

[0152] Acquiring a system neighborhood matrix in the power grid topology data;

[0153] Based on the system neighborhood matrix, an attention mechanism is used to calculate the attention coefficient between each node in the power system and the corresponding neighbor node;

[0154] Normalize each attention coefficient through the Softmax function;

[0155] The normalized attention coefficient is used to aggregate the features of each node in the power system to form spatial features.

[0156] Furthermore, in the second embodiment of the present invention, the characteristic data is subjected to nonlinear mapping processing, and a transient stability assessment result of the power system is generated, specifically:

[0157] According to the stability association rule function of the transient stability evaluator, a transient stability evaluation result of the power system is generated based on the characteristic data, specifically:

[0158] [W0,W1]=ψ([X V ,X θ ,X PG ,X QG ],A)

[0159] Where [W0, W1] is the transient stability assessment result; ψ(·) is the stability association rule function of the transient stability assessor; [X V ,X θ ,X PG ,X QG ] is the characteristic data; among them, X V is the voltage amplitude at the grid node; X θ is the phase angle at the grid node; X PG is the active power at the grid node; X QG is the reactive power at the grid node; A is the system neighborhood matrix.

[0160] Furthermore, in the second embodiment of the present invention, the training process of the deep reinforcement learning agent is specifically as follows:

[0161] Using historical power system data as training data for a deep reinforcement learning agent; wherein the historical power system data includes historical power grid operation data and historical power grid topology data;

[0162] Using the historical power system data, controlling the deep reinforcement learning agent to generate a plurality of first samples, and storing each of the first samples in an experience pool;

[0163] The deep reinforcement learning agent is trained using a plurality of first samples in the experience pool, and when a preset maximum number of training rounds is reached, it is determined that the training of the deep reinforcement learning agent is completed.

[0164] Furthermore, in the second embodiment of the present invention, the deep reinforcement learning agent further includes:

[0165] Determining the active output adjustment of the generator as the action space of the deep reinforcement learning agent;

[0166] Among them, the upper and lower limits of the generator active output adjustment are specifically:

[0167]

[0168] In the formula, The upper and lower limits of the generator active output adjustment; α g is a preset constant, ranging from (0,1); is the minimum generator active output; is the maximum generator active output;

[0169] Among them, the active output of the generator after adjustment is specifically:

[0170]

[0171] In the formula, It is the active output after the active output of the generator is adjusted; It is the active output of the generator before active output adjustment; It is the upper and lower limit range of the generator active output adjustment; is the minimum generator active output; It is the maximum active output of the generator.

[0172] Furthermore, in the second embodiment of the present invention, the deep reinforcement learning agent further includes:

[0173] The reward function of the deep reinforcement learning agent is specifically:

[0174]

[0175] R3=-α6W1

[0176] In the formula, rt is the reward function of the deep reinforcement learning agent; α1~α6 are preset constants; C gi Adjust the cost for the i-th generator; P gi To adjust the output; is the original output of the i-th generator; Q Gi is the reactive power of the i-th generator; is the maximum reactive power of the i-th generator; is the minimum reactive power of the i-th generator; P Gs To balance the node output; Output for the maximum balance node; is the minimum balance node output; V i is the voltage at node i; V i max is the maximum allowable voltage at node i; V i minis the maximum allowable voltage of node i; W1 is the probability of instability.

[0177] In summary, the second embodiment of the present invention provides a power system pre-decision device embedded with spatiotemporal information, which acquires power system data in real time based on the organic combination of modules; wherein the power system data includes power grid operation data and power grid topology data; a transient stability evaluator is used to extract the temporal characteristics and spatial characteristics of the power system data, and the transient stability evaluation result of the power system is evaluated; when the transient stability evaluation result is unstable, a deep reinforcement learning agent is used to generate the optimal action based on the power system data, and the power system is controlled according to the optimal action. The present invention can effectively extract the temporal characteristics and spatial characteristics information in the power system operation data, improve the accuracy of transient stability evaluation, thereby better providing the optimal strategy for power system risk prevention and control, and thus improving the stability of the power system.

[0178] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A power system pre-decision method embedded with spatiotemporal information, characterized in that: include: Acquire power system data in real time; wherein the power system data includes power grid operation data and power grid topology data; Extracting the time series characteristics and spatial characteristics of the power system data using a transient stability evaluator, and evaluating the transient stability evaluation result of the power system; When the transient stability assessment result is instability, based on the power system data, a deep reinforcement learning agent is used to generate an optimal action, and the power system is controlled according to the optimal action.

2. The power system pre-decision method embedded with spatiotemporal information according to claim 1, characterized in that: The method of extracting the time series characteristics and spatial characteristics of the power system data by using a transient stability evaluator and evaluating the transient stability evaluation result of the power system is specifically as follows: Extracting time series characteristics of power grid operation data; Extract spatial features of power grid topology data; splicing the temporal features and the spatial features to form feature data; Nonlinear mapping is performed on the characteristic data to generate a transient stability assessment result of the power system.

3. The power system pre-decision method embedded with spatiotemporal information according to claim 2, characterized in that: The time series characteristics of extracting the power grid operation data are specifically: Cut the power grid operation data to form odd sequences and even sequences; exchanging information between the odd sequence and the even sequence through interactive learning to form an updated odd sequence and an updated even sequence; The updated odd sequence and the updated even sequence are determined as timing features.

4. The power system pre-decision method embedded with spatiotemporal information according to claim 3, characterized in that: The information exchange between the odd sequence and the even sequence through interactive learning to form an updated odd sequence and an updated even sequence is specifically as follows: Convolution calculation, exponential calculation and Hadamard product are performed on the odd sequence and the even sequence to form an intermediate odd sequence and an intermediate even sequence, specifically: In the formula, is an intermediate odd sequence; is the intermediate even sequence; ⊙ is the Hadamard product; exp(·) is the exponential function calculation; φ and ψ are one-dimensional convolution modules respectively; X odd is an odd sequence; X even is an even sequence; Information is exchanged between the intermediate odd sequence and the intermediate even sequence to form an updated odd sequence and an updated even sequence, specifically: In the formula, X′ odd is the updated odd sequence; X′ even is the updated even sequence; ρ and η are one-dimensional convolution modules respectively; is an intermediate odd sequence; is an intermediate even sequence.

5. The power system pre-decision method embedded with spatiotemporal information according to claim 2, characterized in that: The extraction of spatial features of the power grid topology data is specifically as follows: Acquiring a system neighborhood matrix in the power grid topology data; Based on the system neighborhood matrix, an attention mechanism is used to calculate the attention coefficient between each node in the power system and the corresponding neighbor node; Normalize each attention coefficient through the Softmax function; The normalized attention coefficient is used to aggregate the features of each node in the power system to form spatial features.

6. The power system pre-decision method embedded with spatiotemporal information according to claim 2, characterized in that: The nonlinear mapping process is performed on the characteristic data to generate a transient stability assessment result of the power system, specifically: According to the stability association rule function of the transient stability evaluator, a transient stability evaluation result of the power system is generated based on the characteristic data, specifically: [W0,W1]=ψ([X V ,X θ ,X PG ,X QG ],A) Where [W0, W1] is the transient stability assessment result; ψ(·) is the stability association rule function of the transient stability assessor; [X V ,X θ ,X PG ,X QG ] is the characteristic data; where XV is the voltage amplitude at the grid node; X θ is the phase angle at the grid node; X PG is the active power at the grid node; X QG is the reactive power at the grid node; A is the system neighborhood matrix.

7. The power system pre-decision method embedded with spatiotemporal information according to claim 1, characterized in that: The training process of the deep reinforcement learning agent is specifically as follows: Using historical power system data as training data for a deep reinforcement learning agent; wherein the historical power system data includes historical power grid operation data and historical power grid topology data; Using the historical power system data, controlling the deep reinforcement learning agent to generate a plurality of first samples, and storing each of the first samples in an experience pool; The deep reinforcement learning agent is trained using a plurality of first samples in the experience pool, and when a preset maximum number of training rounds is reached, it is determined that the training of the deep reinforcement learning agent is completed.

8. The power system pre-decision method embedded with spatiotemporal information according to claim 7, characterized in that: The deep reinforcement learning agent further includes: Determining the active output adjustment of the generator as the action space of the deep reinforcement learning agent; Among them, the upper and lower limits of the generator active output adjustment are specifically: In the formula, The upper and lower limits of the generator active output adjustment; α g is a preset constant, ranging from (0,1); is the minimum generator active output; is the maximum generator active output; Among them, the active output of the generator after adjustment is specifically: In the formula, It is the active output after the active output of the generator is adjusted; It is the active output of the generator before active output adjustment; It is the upper and lower limit range of the generator active output adjustment; is the minimum generator active output; It is the maximum active output of the generator.

9. The power system pre-decision method embedded with spatiotemporal information according to claim 7, characterized in that: The deep reinforcement learning agent further includes: The reward function of the deep reinforcement learning agent is specifically: R3=-α6W1 In the formula, r t is the reward function of the deep reinforcement learning agent; α1~α6 are preset constants; C gi Adjust the cost for the i-th generator; P gi To adjust the output; is the original output of the i-th generator; Q Gi is the reactive power of the i-th generator; is the maximum reactive power of the i-th generator; is the minimum reactive power of the i-th generator; P Gs To balance the node output; Output for the maximum balance node; is the minimum balance node output; V i is the voltage at node i; V i max is the maximum allowable voltage at node i; V i min is the maximum allowable voltage of node i; W1 is the probability of instability.

10. A power system pre-decision device embedded with spatiotemporal information, characterized in that: include: Data acquisition module, transient assessment module and control module; The data acquisition module is used to acquire power system data in real time; wherein the power system data includes power grid operation data and power grid topology data; The transient assessment module is used to extract the time series characteristics and spatial characteristics of the power system data using the transient stability assessor, and evaluate the transient stability assessment result of the power system; The control module is used to generate optimal actions based on the power system data using a deep reinforcement learning agent when the transient stability assessment result is instability, and control the power system according to the optimal actions.

Images