Automatic power generation control method, device and system based on energy management system

Abstract

The invention belongs to the technical field of automatic power generation control, and particularly relates to an automatic power generation control method, device and system based on an energy management system. The method comprises the following steps: S1, acquiring real-time operation data of a power grid; s2, according to the real-time operation data of the power grid, through a pre-trained load prediction model and a pre-trained generating capacity prediction model, obtaining a predicted load and a predicted generating capacity in a period of time in the future; s3, according to the predicted load, the predicted generating capacity, the power grid real-time operation data and preset strategy generation parameters, a control strategy is generated through a strategy generation model, and the control strategy comprises output of each power generation device and / or energy storage device; and S4, executing the control strategy, and correcting strategy generation parameters in the strategy generation model through an optimization algorithm according to an execution result of the control strategy. The technical problem that the stability of a power system is affected due to the lack of feedback of a control strategy execution result in the prior art is solved.

Description

Technical Field

[0001] This invention belongs to the field of automatic power generation control technology, specifically relating to an automatic power generation control method, device, and system based on an energy management system. Background Technology

[0002] With the continuous growth of electricity demand and the large-scale integration of new energy sources into the grid, the operating environment of the power system is becoming increasingly complex. Traditional power generation control methods often rely on fixed strategies, which are not adaptable to rapidly changing grid conditions. For example, new energy power generation is intermittent and fluctuates, making it difficult for traditional control methods to accurately and promptly adjust the output of power generation equipment. This leads to fluctuations in grid frequency and voltage, affecting the stability of the power system and the quality of power supply.

[0003] A Chinese invention patent application with publication number CN120033837A and publication date of May 23, 2025, discloses a distribution network control system for distributed photovoltaic (PV) grid connection. This system includes a data acquisition layer, a data analysis layer, an intelligent strategy generation layer, and a control execution layer. The data acquisition layer collects real-time operating parameters of the PV system and the power grid. After preprocessing by the data analysis layer, the data is fed into a deep learning module in the intelligent strategy generation layer. The deep learning module uses this data and a built-in deep learning model to predict the future operating state of the PV system and the power grid. The strategy optimization module in the intelligent strategy generation layer generates and executes the optimal control strategy based on the prediction results. However, the overall control scheme lacks feedback on the execution results. Therefore, when the control strategy is not executed properly or significant changes occur in the system, the control strategy cannot be adjusted in a timely manner, thus affecting the stability of the power system and the quality of power supply. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic power generation control method, device, and system based on an energy management system, in order to solve the technical problem in the prior art of lacking feedback on the execution results of control strategies, which in turn affects the stability of the power system.

[0005] To solve the above-mentioned technical problems, the present invention provides a technical solution for an automatic power generation control method based on an energy management system: an automatic power generation control method based on an energy management system, the method comprising:

[0006] S1. Obtain real-time power grid operation data;

[0007] S2. Based on real-time power grid operation data, the predicted load and predicted power generation for a future period are obtained through pre-trained load prediction and power generation prediction models.

[0008] S3. Based on the predicted load, predicted power generation, real-time grid operation data and pre-set strategy generation parameters, a control strategy is generated through a strategy generation model. The control strategy includes the output of each power generation device and / or energy storage device.

[0009] S4. Execute the control strategy, and adjust the strategy generation parameters in the strategy generation model according to the execution result of the control strategy through an optimization algorithm;

[0010] The load forecasting model, power generation forecasting model, and strategy generation model are all constructed based on machine learning models.

[0011] The beneficial effects of the above technical solution are as follows: The technical solution of the automatic power generation control method based on an energy management system of the present invention belongs to an improved invention. Based on the prior art, the present invention introduces a feedback loop from the execution end (including various power generation devices and / or energy storage devices) to the intelligent control strategy generation module. In the feedback loop, based on the feedback information (i.e., the execution result), the strategy generation parameters in the strategy generation model are optimized through an optimization algorithm, enabling timely adjustment of the control strategy, thereby improving the stability of the power system and the power supply quality. The present invention solves the technical problem in the prior art of lacking feedback on the execution result of the control strategy, which affects the stability of the power system.

[0012] Furthermore, the optimization objectives of the optimization algorithm include one, two, or three of the following objectives: a stability objective to ensure grid stability, an economic objective to minimize economic costs, and an environmental objective to minimize carbon emissions.

[0013] Furthermore, the strategy generation parameters include one, two, or a combination of three or more of the following: the output adjustment priority of each generator set, the energy storage charging and discharging power threshold, the AGC control dead zone, the power allocation ratio coefficient, and the safety margin parameters.

[0014] Furthermore, the machine learning model on which the policy generation model is based is a reinforcement learning model; the policy generation parameters include the weights of each reward item.

[0015] Furthermore, S2 also includes: inputting the current region's power grid topology parameters into a pre-trained state assessment model to obtain power grid assessment indicators;

[0016] The power grid evaluation index includes the characteristic value perturbation of power transmission between nodes; the state evaluation model is constructed based on a machine learning model, and the machine learning model on which the state evaluation model is based is a graph convolutional network.

[0017] S4 also includes: adjusting the parameters of the state assessment model according to the changes in the power grid assessment indicators.

[0018] Furthermore, S2 also includes: performing state assessment calculations based on the current power grid topology parameters of the region to obtain power grid assessment indicators;

[0019] The power grid assessment indicators include one, two, or three of the following: node vulnerability score, system damping ratio, and power transmission vulnerability matrix.

[0020] S4 also includes: correcting the parameters calculated for the state assessment based on changes in the power grid assessment indicators.

[0021] Furthermore, the reinforcement learning model is a DQN network; the loss function L(θ) for training the DQN network is:

[0022]

[0023] Where θ represents the current network parameters of the DQN network; θ - γ represents the target network parameters of the DQN network; r is the immediate reward obtained after taking action a in state s; γ is the discount factor. Q(s,a;θ) is the maximum value estimate of the target network for all possible actions in the next state s′; Q(s,a;θ) is the current Q network's value estimate for taking action a in state s; E[·] represents the expected value.

[0024] Furthermore, the reward function in the reinforcement learning model includes one, two, or more of the following: stability penalty, economic reward, and environmental reward; the stability penalty is less than 0, and the more the system deviates from the steady state, the larger the absolute value of the stability penalty; the lower the economic cost, the higher the economic reward; and the lower the carbon emissions, the higher the environmental reward.

[0025] Furthermore, the machine learning model on which the load forecasting model and / or power generation forecasting model is based includes an LSTM network, a CNN network, and an output layer; the LSTM network is used to extract the temporal features of the input data; the CNN network is used to extract the spatial features of the input data; the output layer is used to concatenate the features output by the LSTM network and the CNN network, and the concatenated features are processed by a fully connected layer to output the prediction result.

[0026] Furthermore, S4 also includes: when a grid fault or a sudden change in the penetration rate of new energy sources is detected, it is considered that the grid environment has changed, and the model parameters of each machine learning model are updated through online transfer learning so that the updated machine learning models can adapt to the new grid environment.

[0027] Furthermore, methods for updating the model parameters of various machine learning models through online transfer learning include:

[0028] θ new=(1-λ)θ old +λθ adapt

[0029] Where, θ old The original model parameters for the machine learning model; θ adapt The parameters obtained by the machine learning model to adapt to the new environment; θ new λ represents the updated machine learning model parameters; λ is the transfer weight, a coefficient between 0 and 1.

[0030] Furthermore, the training process for all machine learning models includes:

[0031] After training local machine learning models using data from the current region, all parameters of the trained machine learning models are sent to the central server so that the central server can aggregate all parameters of the machine learning models from each region to obtain global model parameters.

[0032] Each machine learning model is updated based on the global model parameters issued by the central server.

[0033] Furthermore, the global model parameters are obtained in the following manner:

[0034]

[0035] Where, θ t+1 n represents the global model parameters for a given machine learning model. k is the number of training samples for the machine learning model in the k-th region; n is the sum of the number of training samples for the machine learning models in all regions; For n based on the k-th region k The parameters of the machine learning model are obtained by training with 10 training samples; K is the number of all regions.

[0036] Furthermore, the optimization algorithm is a Bayesian optimization algorithm.

[0037] The present invention also provides a technical solution for an automatic power generation control device based on an energy management system: an automatic power generation control device based on an energy management system includes a processor, the processor being used to execute a computer program to implement the steps of the automatic power generation control method based on an energy management system as described above.

[0038] This invention also provides a technical solution for an automatic power generation control system based on an energy management system: an automatic power generation control system based on an energy management system, comprising:

[0039] Data acquisition module: used to acquire real-time operating data of the power grid;

[0040] Predictive analysis module: Based on real-time power grid operation data, it obtains the predicted load and predicted power generation for a future period of time through pre-trained load prediction and power generation prediction models;

[0041] Intelligent control strategy generation module: Based on the predicted load, predicted power generation, real-time grid operation data and pre-set strategy generation parameters, the module generates a control strategy through a strategy generation model. The control strategy includes the output of each power generation device and / or energy storage device.

[0042] The load forecasting model, power generation forecasting model, and strategy generation model are all constructed based on machine learning models.

[0043] Adaptive adjustment and feedback module: Based on the execution results of the control strategy, the strategy generation parameters in the strategy generation model are corrected through an optimization algorithm.

[0044] The beneficial effects of the above technical solution are as follows: The technical solution of the automatic power generation control system based on the energy management system of the present invention is an improved invention. Based on the prior art, the present invention introduces a feedback loop from the execution end (including various power generation devices and / or energy storage devices) to the intelligent control strategy generation module. In the feedback loop, based on the feedback information (i.e., the execution result), the strategy generation parameters in the strategy generation model are optimized through an optimization algorithm, enabling timely adjustment of the control strategy, thereby improving the stability of the power system and the power supply quality. The present invention solves the technical problem in the prior art of lacking feedback on the execution result of the control strategy, which affects the stability of the power system.

[0045] Furthermore, it also includes a state assessment module: inputting the current region's power grid topology parameters into a pre-trained state assessment model to obtain power grid assessment indicators;

[0046] The power grid evaluation index includes the characteristic value perturbation of power transmission between nodes; the state evaluation model is constructed based on a machine learning model, and the machine learning model on which the state evaluation model is based is a graph convolutional network.

[0047] The adaptive adjustment and feedback module is also used to adjust the parameters of the state assessment model according to the changes in the power grid assessment indicators. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the data acquisition module architecture in an embodiment of the automatic power generation control system based on the energy management system of the present invention.

[0049] Figure 2 This is a flowchart of the state assessment and predictive analysis module of the automatic power generation control system based on the energy management system of the present invention.

[0050] Figure 3 This is a schematic diagram of the intelligent control strategy generation module architecture in an implementation of the automatic power generation control system based on the energy management system of the present invention.

[0051] Figure 4 This is a schematic diagram illustrating the working principle of the adaptive adjustment and feedback module in the implementation of the automatic power generation control system based on the energy management system of the present invention.

[0052] Figure 5 This is a schematic diagram illustrating the execution of the control strategy in an implementation of the automatic power generation control system based on an energy management system according to the present invention. Detailed Implementation

[0053] This invention, based on existing technologies, introduces a feedback loop from the execution end (including various power generation devices and / or energy storage devices) to the intelligent control strategy generation module. Within this feedback loop, the strategy generation parameters in the strategy generation model are optimized using an optimization algorithm based on feedback information (i.e., execution results). This enables timely adjustment of the control strategy, thereby improving the stability and power supply quality of the power system. This invention solves the technical problem in existing technologies of lacking feedback on the execution results of control strategies, which consequently affects the stability of the power system.

[0054] Implementation methods of automatic power generation control systems based on energy management systems:

[0055] An automatic power generation control system based on an energy management system includes a data acquisition module, a state assessment and predictive analysis module, an intelligent control strategy generation module, and an adaptive adjustment and feedback module.

[0056] Data acquisition module: such as Figure 1 As shown, the data acquisition module includes a real-time data acquisition unit and a real-time operational anomaly detection submodule. The real-time data acquisition unit is used to acquire real-time power grid operation data, including power grid frequency and voltage. To ensure the reliability of the input data, the real-time operational anomaly detection submodule performs anomaly detection on the data acquired by the real-time data acquisition unit. Once anomaly data is detected, it is marked or removed to avoid the anomaly data affecting subsequent predictive analysis or control. Specifically, the method for processing anomaly data can refer to existing technologies, and will not be described in detail in this embodiment.

[0057] Specifically, in this embodiment, the real-time anomaly detection submodule identifies anomalous data using the isolated forest algorithm: an anomaly score is calculated using the following formula, and when the anomaly score exceeds a preset score threshold, it is determined to be an anomaly:

[0058]

[0059] Where E(h(x)) represents the mean path length of sample x in the isolated tree; n is the number of samples; c(n) is the average path length of the binary tree, calculated according to the following formula:

[0060]

[0061] Where H(n-1) is the harmonic number,

[0062] In this embodiment, the score threshold is 0.6, that is, when s(x) > 0.6, it is judged as abnormal data.

[0063] The Isolation Forest algorithm includes a tree-building phase: the input multidimensional data stream samples are recursively and randomly partitioned. This process continues until all samples are isolated or a pre-defined tree height is reached, ultimately generating a tree T composed of multiple trees. i An isolated forest F = {T1,T2,…,T} consisting of (i = 1, 2,…,t) t For example, suppose there is a multidimensional data stream sample containing power grid load data at different points in time. The algorithm will randomly select a feature dimension and a split point on that dimension to divide the sample into two parts, and then continue to perform similar recursive splitting operations on these two parts of the sample.

[0064] Anomaly score calculation: For a new input sample x i It is necessary to calculate the path length h(x) of each tree in the isolated forest. i Then, the anomaly score s(x,n) is calculated according to the formula above. Here, c(n) is a tree average path length correction factor related to the sample size n. When the calculated anomaly score s(x) i When n is greater than 0.6, the data point x is determined to be... i This is considered outlier data. For example, if the path length of a new sample in each tree, after calculation and correction, yields an outlier score greater than 0.6, it indicates that the sample differs significantly in distribution from other normal samples and may be outlier data.

[0065] State assessment and predictive analysis module: such as Figure 2 As shown, the state assessment and predictive analysis module is the core of the entire system, performing in-depth analysis and processing of the data collected by the data acquisition module. It primarily performs two tasks:

[0066] Status assessment: This evaluates the current operating status of the power grid to determine its stability and identify potential risks. For example, it analyzes whether the voltage at each node is within normal range and whether there is a risk of line overload. This part combines real-time and historical power grid data, using a Graph Neural Network (GNN) to model the power grid topology. GNNs enable in-depth analysis of the relationships between nodes in the power grid, dynamically assess the stability of power transmission between nodes, and accurately identify weak points and potential risks in the power grid.

[0067] Construct a power grid topology graph G = (V, E), where nodes v ∈ V represent power generation equipment, substations, etc., and edges e ∈ E represent transmission lines. Perform state evaluation on the above power grid topology using a Graph Convolution Network (GCN). The input parameters of the GCN include the node feature matrix X ∈ R. |V|×d and adjacency matrix A∈R |V|×|V| Node feature matrix X∈R |V|×d This represents the real-time operating parameters such as power and voltage of each node in the power grid, where d is the dimension of the real-time operating parameters; the adjacency matrix A∈R |V|×|V| Indicates the connection relationship of the supports at each node in the power grid, A ij =1 indicates that there is a connection between nodes i and j.

[0068] A GCN consists of multiple sequentially connected GCN layers, and the output feature H of the l-th GCN layer is... (l+1) Update according to the following formula:

[0069]

[0070] in, for The degree matrix, I is the identity matrix; W (l) Let σ be the learnable weight matrix, and σ be the activation function (e.g., ReLU, tanh). Node features H (1) This represents the operating status of various devices (generators, substations, etc.) in the power grid (i.e., X∈R above). |V|×d ), H (l) The input features are those of the l-th GCN layer. Through propagation via GCN layers, each node can aggregate information from its neighbors, simulating the propagation of electricity in the network. Multi-layer GCNs can capture long-distance dependencies in the power grid, such as the impact of long-distance transmission lines on local nodes.

[0071] The GCN assesses the current operating status of the power grid and outputs various evaluation indicators. In this embodiment, these evaluation indicators include: the characteristic perturbation Δλ of power transfer between nodes, calculated by...i Assess power grid stability:

[0072]

[0073] The above formula is used to evaluate the stability Δλ of the power grid. i λ represents the i-th eigenvalue of the system state matrix A. i The change in ΔA represents the disturbance of the system state matrix A, v i Let λ be the i-th eigenvalue of the system state matrix A. i The corresponding right eigenvector, u i Let λ be the i-th eigenvalue of the system state matrix A. i The left eigenvector of the eigenvalue. By calculating the change in eigenvalues, the stability of the power grid under disturbance can be assessed. If the change in eigenvalues ​​causes its real part to approach or be greater than 0, it indicates that the power grid may become unstable. Δλ i The larger the absolute value, the more sensitive the node is to disturbances, and the worse the grid stability.

[0074] It should be noted that the right eigenvector v i Satisfy: λ i v i =Av i Left eigenvector u i satisfy: Indicate u i The transpose of .

[0075] In addition to extracting evaluation metrics through GCN, this implementation method also directly extracts evaluation metrics from the power grid topology diagram G=(V,E), including: node vulnerability score, system damping ratio, and power transmission vulnerability matrix, as detailed below:

[0076] (1) Node vulnerability score S v Node vulnerability scoring quantifies the vulnerability of a single node (such as a generator or substation) within the power grid. A higher score indicates that the node is more susceptible to failure during grid faults or disturbances, potentially triggering a cascading effect.

[0077] The node vulnerability score comprehensively considers the node's load rate, connection importance, and the scope of failure impact. The calculation formula is as follows:

[0078]

[0079] Among them, P v P represents the actual transmission power of node v. v,max k represents the maximum permissible transmission power (rated capacity) of node v. v Let v be the degree centrality of node v (the number of nodes directly connected to it); ∑i∈V k i / |V| represents the average degree centrality of all nodes in the network, where |V| is the total number of nodes in the network; N v ω1, ω2, and ω3 are the number of downstream nodes affected if node v fails; ω1, ω2, and ω3 are weighting coefficients adjusted according to the characteristics of the power grid, and ω1 + ω2 + ω3 = 1.

[0080] Physical meaning: S v The value range is ([0,1]). The larger the value, the more fragile the node, the higher the load rate, the more critical the connection, and the wider the scope of the failure impact.

[0081] (2) System damping ratio ξ: The system damping ratio reflects the attenuation capability of grid oscillations (such as low-frequency oscillations). Based on the generator rotor motion equation, the calculation formula is as follows:

[0082]

[0083] D is the system damping coefficient (considering generator damping, load damping, etc., unit: N·m·s / rad); J is the generator rotor moment of inertia (unit: kg·m). 2 K is the synchronization power coefficient (reflecting the synchronization capability between the generator and the power grid, unit: N·m / rad).

[0084] (3) Power transmission vulnerability matrix M: describes the vulnerability of transmission lines between nodes. The larger the element value in the matrix, the more likely the corresponding line is to experience a decrease in transmission efficiency or interruption due to overload, voltage fluctuations and other problems during power transmission.

[0085] The power transmission vulnerability matrix M is an N×N matrix (N is the number of nodes, i.e., |V|), where the elements M i,j The vulnerability of the transmission line from node i to node j is represented by the following formula:

[0086]

[0087] Among them, P i,j This represents the actual transmission power of the line (ij) between node i and node j; ΔU is the rated transmission power of line (ij); i,j =|U i -U j | represents the voltage difference between the two ends of the line (ij); This indicates the rated voltage of the line; α is the weighting coefficient, 0 < α < 1, such as α = 0.6, which emphasizes the power load rate.

[0088] Physical meaning: M i,j∈[0,1], the larger the value, the more fragile the line (ij) is (the closer the load is to the rated value or the larger the voltage difference, the more likely it is to overload or voltage collapse). Predictive Analysis: The predictive analysis module predicts the electricity demand and power generation for a future period based on the data collected by the data acquisition module, providing important support for subsequent control decisions. For example, it predicts the peak electricity consumption period and solar power generation for tomorrow based on historical data and weather forecasts. A multimodal deep learning model is introduced, integrating the time-series prediction capability of Long Short-Term Memory (LSTM) network with the spatial feature extraction capability of Convolutional Neural Network (CNN). This model can comprehensively consider time-series information and spatial distribution characteristics to more accurately predict the load demand (i.e., electricity demand) and new energy power generation for the next 15 minutes to 24 hours, providing a more reliable basis for power generation control.

[0089] Specifically, the structure of the multimodal model that integrates LSTM and CNN is as follows:

[0090] First, LSTM is used to extract time-series features from the time-series data X (historical power load data for load forecasting or historical power generation data for power generation forecasting):

[0091] i t =σ(W ii x t +W hi h t-1 +b i )

[0092] f t =σ(W if x t +W hf h t-1 +b f )

[0093] o t =σ(W io x t +W ho h t-1 +b o )

[0094] c t =f t ⊙c t-1 +i t ⊙tanh(W ic x t +W hc h t-1 +b c )

[0095] h t =o t ⊙tanh(c t )

[0096] Among them, i t ,f t ,o t These are the input gate, forget gate, and output gate of the LSTM; input gate i t Determine the current input x t How much information is added to the cell state, σ is the activation function (usually the sigmoid function), which maps the weighted input to between 0 and 1, W ii and W hi It is the weight matrix of the input gate, b i It is the bias term of the input gate; the forget gate f t The sigmoid function, W, is also used to determine how much old information to discard from the cell state. if W hf and b f These are the weight matrix and bias term of the forget gate; the output gate o t Which part of the cell state will be used as the output of the current LSTM unit, W io W ho and b o These are the weight matrix and bias terms of the output gate; c t For cell state updates, this is based on the forgetting gate f t Determine how much of the old cell state c to retain t-1 In addition, by input gate i t New information for control (input processed by the tanh function), ⊙ represents element-wise multiplication, W ic W hc and b c These are the weight matrix and bias terms during cell state updates; h t Update the hidden state, the final hidden state h t By output gate o t and the cell state after tanh treatment c t We obtain h by element-wise multiplication. t It is typically used as the output of the LSTM unit at the current time step.

[0097] For example, in an LSTM model for predicting electricity load, the historical electricity load data input at each time step t is x. t The model calculates the input gate, forget gate, and output gate using the above formulas, thereby updating the cell state and hidden state to capture long-term dependencies in power load data and make accurate predictions.

[0098] Then, spatial features are extracted from the input data X using a CNN:

[0099] y = σ(W*x + b)

[0100] Where * represents the convolution operation, W is the convolution kernel, b is the bias, and x is the input data.

[0101] Finally, the temporal features output by the LSTM are concatenated with the spatial features output by the CNN to form a new feature vector, which is then passed through a fully connected layer to output the final prediction result.

[0102]

[0103] in, This represents the final output forecast result (predicted load or predicted power generation); [h T c T [z] is a new feature vector formed by concatenating the LSTM output with the CNN features, where h T and c T The hidden state h at time step T of the LSTM output are respectively t and cell state c t z represents the spatial features output by the CNN; W f and b f For fully connected layer parameters, W f This is the weight matrix of the fully connected layer, which determines the degree of influence of each element in the concatenated feature vector on the prediction result. f It is the bias term of the fully connected layer, which plays a role in adjusting the prediction results.

[0104] Intelligent control strategy generation module: Based on the results of the state assessment and predictive analysis module, the intelligent control strategy generation module generates the optimal power generation control strategy.

[0105] To generate the optimal control strategy, this module uses a technique called Deep Reinforcement Learning (DRL). Deep Reinforcement Learning acts like an intelligent decision-maker, finding the optimal strategy through continuous trial and learning. Based on predicted electricity demand and renewable energy generation, it determines which power generation equipment should increase output and which should decrease output to ensure stable grid operation and economic efficiency.

[0106] A control strategy generator based on the DRL framework was designed, which is interactively trained with the power grid environment. During the training process, frequency deviation, economic cost, and carbon emissions are used as reward and penalty indicators to dynamically generate the optimal power generation equipment output allocation strategy, thereby achieving efficient, economical, and environmentally friendly operation of the power system. Details are as follows:

[0107] Constructing a Markov decision process model<S,A,P,R,γ> Wherein, the state space S = {s} t The system includes data acquisition modules that obtain current grid operating status parameters such as system frequency and current output of each generating unit, as well as prediction results output by status assessment and prediction analysis modules such as load forecasting and power generation forecasting. This information comprehensively describes the operating status of the power system at a certain moment and serves as the basis for subsequent decision-making.

[0108] Action space A = {a t} represents the output adjustment of each power generation device, i.e., based on the current state s t The system can make actions to change the output of each power generation device. t A set of.

[0109] The reward function R is:

[0110] R(s t ,a t )=w1R f +w2R c +w3R e

[0111] This function defines the state s. t Take action a t The reward value obtained afterwards. Among them, R f This is a frequency deviation penalty, meaning that if the system frequency deviates from the ideal value, it will receive a negative reward value to prompt the system to adjust and return the frequency to normal; R c The economic cost reflects the economic impact of taking the action; the lower the economic cost, the higher the reward. e The reward represents carbon emission indicators, taking environmental factors into account; the lower the carbon emissions, the higher the reward. w1, w2, and w3 are weighting coefficients used to adjust the relative importance of these three factors in the reward function. These weights can be flexibly adjusted according to actual needs and objectives to balance requirements such as frequency stability, economic costs, and environmental protection.

[0112] The transition probability P represents the probability of transitioning from a state s to a given state s. t By taking a certain action a t Transition to another state t+1 The probability, i.e., P(s) t+1 |s t ,a t ), which describes the randomness of the system's state transitions.

[0113] The discount factor γ is a number between 0 and 1 used to measure the importance of future rewards. The closer γ is to 1, the more important future rewards are; the closer γ is to 0, the more important immediate rewards are.

[0114] Parameter update rules for Deep Q Networks (DQN):

[0115]

[0116] This is the loss function used to train DQN. Here, θ represents the current network parameters, determining the structure and function of the Q-network. Adjusting θ continuously helps the Q-network better estimate the Q-value. - Here, θ represents the target network parameters. The target network is used to provide a relatively stable target value to help the current network learn better. r is the immediate reward obtained after taking action a in state s, i.e., the value of the reward function R(s,a). Q(s,a; θ) is the current Q-network's estimate of the value of taking action a in state s; The target network (i.e., the network parameter is θ) - The Q-network is used to estimate the maximum Q-value of all possible actions in the next state s′; γ is a discount factor.

[0117] The above update rule means that by minimizing the current Q-network estimate (i.e., Q(s,a;θ)) and the target value (i.e., the next-state maximum Q-value estimate after the immediate reward r plus the discount, i.e., r + γmax), the network can achieve the desired update. a′ Q(s′,a′;θ - The mean square error between the two states is used to continuously optimize the parameters θ of the Q network, enabling the Q network to more accurately evaluate the value of taking different actions under different states, thereby finding the optimal control strategy.

[0118] Assume the current state is s t The system frequency is slightly higher than normal, and the output of generator A is P. A,t The load forecast value is L t Based on the current state, the system can choose action a. t This could involve reducing the output of generator A. If this action returns the system frequency to normal, the economic cost is reasonable, and carbon emissions are within acceptable limits, then the reward function R(s) will... t ,a t This will result in a relatively high reward value. By continuously repeating this process and using the deep Q-network update rules to adjust the network parameters, the system can gradually learn to take the optimal action under various conditions, thereby achieving stable, economical, and environmentally friendly operation of the power system.

[0119] After the Q-network is trained, during decision-making, it uses the current state s... t The action 'a' corresponding to the maximum Q value is used as the control strategy output by the intelligent control strategy generation module.

[0120] like Figure 3As shown, the intelligent control strategy generation module in this embodiment introduces a federated learning mechanism, allowing multiple regional EMS (Energy Management System) systems to share model parameters without sharing raw data. Through federated learning, the power grid systems in each region can jointly improve the accuracy and effectiveness of the control strategy. This fully utilizes the local data and knowledge of each region, enhances global collaborative control capabilities, and makes the control of the entire power system more coordinated and optimized.

[0121] The models here include not only the DRL model, but also all the aforementioned machine learning models, such as the state assessment model, load forecasting model, and power generation forecasting model. While the model structures are the same across different regions, their grid topologies, load characteristics, and renewable energy ratios differ. Local models are trained based on their respective data, and their parameters adapt to local characteristics. For example, in regions with a high proportion of wind power, the forecasting model parameters will place greater emphasis on the weight of wind speed data.

[0122] Suppose there are K EMS systems in different regions, each with its own local data. For the k-th system, its local parameter update rule is as follows:

[0123]

[0124] here These are the model parameters of the k-th system at time t, and α is the learning rate, used to control the step size for each parameter update. Represents the loss function L k Regarding the model parameter θ k The gradient at time t is calculated using this formula. The k-th system then calculates the gradient based on its local data and updates its local model parameters accordingly.

[0125] The global aggregation rule is:

[0126]

[0127] Where n k It is the sample size of the k-th system. It is the sum of the sample sizes of all K systems. This formula indicates that in each training round, the EMS systems in each region will update the model parameters locally. According to its sample size n k The global model parameters θ are obtained by weighting the proportions of the samples in the total sample size n. t+1 This approach leverages the characteristics of local data in each region while simultaneously updating the global model through aggregation, enabling multi-region EMS systems to enhance global collaborative control capabilities without sharing raw data.

[0128] For example, suppose there are three EMS systems (K=3) in three regions. The first region system has a sample size of n1=100, the second region system has a sample size of n2=200, and the third region system has a sample size of n3=300. Then the total sample size n=100+200+300=600. In a certain round of training, the three region systems obtain updated parameters according to their local update rules. Then, according to the global aggregation rules, the global model parameters...

[0129] In federated learning, all learnable parameters of the model are shared (such as the weight matrix W and bias terms b in a neural network). Specifically: during global aggregation, each region shares all parameters of its local model (such as the weight matrix W and bias terms b in a GNN). (l) ), LSTM in (W ii W hi The parameters are weighted and aggregated according to the sample size to generate new global parameters. The parameter update logic is as follows: each region trains the model with local data and updates all local parameters; all local parameters are uploaded to the central server (or through a distributed protocol); the server calculates the new global parameters according to the global aggregation rules and then distributes them to each region.

[0130] Adaptive adjustment and feedback module: such as Figure 5 As shown, the main function of the adaptive adjustment and feedback module is to monitor the execution effect of the control strategy output by the intelligent control strategy generation module in real time and feed the monitoring results back to the intelligent control strategy generation module. If the execution effect is found to be unsatisfactory, the intelligent control strategy generation module will adjust the control strategy based on the feedback information. It acts like a supervisor, constantly monitoring the power generation control process to ensure that everything proceeds according to plan. If the actual situation is found to be inconsistent with the prediction during execution, such as a sudden increase in electricity demand or a sudden decrease in renewable energy generation, this module will immediately take measures to adjust, including online transfer learning and Bayesian optimization algorithms.

[0131] Online transfer learning: such as Figure 4 As shown, when a fault is detected in the renewable energy grid or a sudden change in renewable energy penetration is detected, online transfer learning technology is used to quickly transfer existing knowledge and experience to the new scenario. This reduces the lag time for strategy adjustments, enabling the system to quickly adapt to new operating conditions and ensuring the stable operation of the power system.

[0132] This implementation method employs a transfer learning approach based on model parameters. When an environmental change is detected, knowledge transfer is performed using the following formula:

[0133] θ new =(1-λ)θ old +λθ adapt

[0134] Where, θ old These are the original model parameters, referring to the parameters used by the model before the environmental change. These parameters are obtained based on previous training data and tasks, representing the knowledge the model learned in the original environment; θ adapt The parameters for adapting to the new environment are those learned or adjusted for the new environment, reflecting the changes the model needs to make to better adapt to the new situation; λ is the transfer weight, a coefficient between 0 and 1, which determines the original model parameters θ. old and the parameter θ for adapting to the new environment adapt In generating new model parameters θ new The relative importance of λ. If λ is close to 0, the new model parameters will mainly depend on the original model parameters; if λ is close to 1, the new model parameters will focus more on adapting to the new environment. By adjusting λ, the model's ability to retain original knowledge and adapt to the new environment can be flexibly balanced.

[0135] For example, in an automatic generation control model of a power system, the above formula can be used when a sudden change in the penetration rate of new energy sources is detected. Assume the original model parameter θ... old It was obtained based on training when the penetration rate of new energy sources was relatively stable, and θ adapt These are adaptive parameters trained on new data following a sudden change in the penetration rate of new energy sources. At this point, by setting an appropriate λ value, such as 0.3, the new model parameters θ... new In this model, 70% of the parameters come from the original model, and 30% come from parameters adapted to the new environment. This allows the model to utilize existing knowledge while adapting to the new environment following the sudden change in the penetration rate of new energy sources.

[0136] Bayesian optimization algorithm: By embedding the Bayesian optimization algorithm in the feedback loop, the policy generation parameters in the policy generation model are continuously updated and optimized, thereby adjusting the control policy to ensure that the system still has strong robustness in uncertain environments and reduce the system performance fluctuations caused by external interference and uncertain factors.

[0137] The feedback loop refers to the complete closed loop from the execution result of the control strategy to the optimization of model parameters, including:

[0138] Execution end: Control actions such as power generation equipment output adjustment and energy storage charging and discharging;

[0139] Monitoring terminal: The data acquisition module acquires real-time power grid operation data;

[0140] Optimization side: The adaptive adjustment and feedback module analyzes data and updates the strategy.

[0141] The feedback included:

[0142] Control effect data: deviation between actual power output of power generation equipment and target value, grid frequency / voltage fluctuation amplitude, renewable energy consumption, etc., to evaluate whether the strategy has achieved the expected effect, such as "the deviation of energy storage charging and discharging power exceeds 5%".

[0143] Abnormal event logs: identification strategy vulnerabilities such as system frequency exceeding limits, power imbalance alarms, and equipment failures, such as "when wind power drops suddenly, insufficient backup capacity leads to frequency drop".

[0144] Prediction error indicators include the deviation between the predicted and actual load values, changes in the confidence interval of new energy power generation predictions, and other parameters of the optimized prediction model. For example, "the load prediction error last week was 10% higher on average during the evening peak period."

[0145] Stability indicators include changes in node vulnerability scores, fluctuations in system damping ratios, and updates to the power transmission vulnerability matrix. Adjusting control strategies can enhance system stability, such as "reducing transmission power when the vulnerability score of a critical line increases."

[0146] Node vulnerability score change ΔS v The formula used to quantify the dynamic changes in node vulnerability is:

[0147] ΔS v =S v (t)-S v (t0)

[0148] S v (t) represents the node vulnerability score at time t; S v (t0) represents the node vulnerability score at the reference time t0 (e.g., when the system is in steady state).

[0149] Physical meaning: ΔS v >0 indicates increased node vulnerability (deteriorated stability); ΔS v A value greater than 0 indicates a decrease in vulnerability (improved stability).

[0150] The system damping ratio fluctuation Δξ is used to quantify the dynamic change of the damping ratio, and the calculation formula is as follows:

[0151] Δξ=|ξ(t)-ξ ref |

[0152] ξ(t) represents the system damping ratio at the current time t; ref This indicates the reference damping ratio (such as the design value or historical steady-state average).

[0153] Physical meaning: The larger Δξ is, the further the damping ratio deviates from the ideal state, and the more significant the fluctuation in the power grid's anti-oscillation capability.

[0154] Based on the feedback information mentioned above, such as: the state assessment module corrects the power grid topology modeling parameters, such as updating the calculation weights of node vulnerability scores; the predictive analysis module optimizes the LSTM / CNN model parameters, such as adjusting the influence coefficient of temperature factors in load forecasting; and the control strategy generation module updates the reward and penalty functions of deep reinforcement learning, such as adjusting the weight value of "frequency deviation penalty".

[0155] For the state assessment module, with GCN as the core, the main optimizations are made to parameters related to power grid topology modeling and node state assessment, as follows:

[0156] The optimized parameters include: (1) the learning parameters of the GCN model, such as the weight matrix (W) of each GCN layer. (l) ): Used for the aggregation and transformation of node features, determining the influence weight of adjacent node features on the current node (e.g., the power influence weight of transmission lines on substation nodes). Non-learnable hyperparameters in the GCN model, such as activation function parameters: the threshold of the ReLU function (controlling the degree of nonlinearity of feature mapping), etc.

[0157] (2) Calculation weight of node vulnerability score The weight coefficients (ω1, ω2, ω3) in the node vulnerability score formula: respectively control the contribution ratio of "node load rate", "connection importance" and "fault impact range" to the final score.

[0158] (3) Weighting coefficients of the power transmission vulnerability matrix: elements of the power transmission vulnerability matrix (M i,j In the calculation formula, α is used to balance the impact of "line load rate" and "voltage difference" on line vulnerability (for example, α can be reduced for high-voltage lines to decrease the weight of load rate).

[0159] The optimization methods include: monitoring power grid operation data (such as the cascading effects after node failures); if it is found that the node vulnerability score deviates significantly from the actual failure impact (such as nodes with low scores failing first), then the weights of (ω1, ω2, ω3) are corrected using a Bayesian optimization algorithm (such as taking "minimizing the error between the score and the actual vulnerability" as the optimization objective).

[0160] For the predictive analytics module, the main optimizations are related to the accuracy of load / power generation forecasts, as follows:

[0161] The optimized parameters include: (1) the weight matrices of the LSTM network input gate, forget gate, and output gate (e.g., ({W)). ii},{W hi},{W if}) and bias terms (such as (b) i ,b f ): Weights for extracting time-series features (such as trends and periodicity of historical load). Weight matrix (W) for cell state updates.ic W hc (2) CNN network parameters: Convolution kernel weights (W) and biases (b): Control the extraction accuracy of spatial features (such as the spatial correlation of loads in different regions and the influence of geographical factors on power generation). Pooling layer parameters: Such as the pooling window size, which affects the aggregation granularity of spatial features (such as the distinction between urban and suburban loads). (3) Feature fusion and output layer parameters: Temporal features (LSTM output (h) T ,c T The concatenation weights of temporal features (CNN output ({z})) and spatial features (CNN output ({z})) are adjusted to optimize the contribution ratio of the two types of features to the final prediction result (e.g., increasing the weight of temporal features during the winter heating season and increasing the weight of spatial features during the summer). The influence coefficients of environmental factors such as temperature and humidity are quantified in load forecasting to quantify the correlation between environmental factors and load (e.g., the coefficient of load increase for every 1°C increase in summer temperature).

[0162] Optimization methods include: (1) Backpropagation optimization based on loss function: using prediction error (such as the mean square error between predicted load and actual load) as the basis for optimization. To achieve this, the weight matrix and bias terms of the LSTM / CNN are updated using gradient descent (e.g., ...). (α is the learning rate) to minimize the prediction bias. (2) Bayesian optimization based on prediction error feedback: monitor the prediction error (e.g., the load prediction is 10% higher during the evening peak period), adjust the influence coefficient of environmental factors through the Bayesian optimization algorithm (e.g., increase the weight of temperature factors in the summer evening peak period), or adjust the LSTM forget gate parameter (enhance the memory of recent load changes) to reduce the prediction bias in a specific period.

[0163] For the control strategy generation module, the Bayesian optimization algorithm optimizes some parameters in the state space of the intelligent control strategy generation module, including: power generation scheduling parameters: the output adjustment priority of each generator set, the charging and discharging power threshold of each energy storage, etc.; strategy logic parameters: AGC control dead zone size, power allocation ratio coefficient, etc.; safety margin parameters, etc.

[0164] The optimization objectives consist of three parts: stability objectives, economic objectives, and environmental objectives.

[0165] Stability objective: Minimize grid frequency deviation and voltage fluctuation amplitude, such as minimizing the root mean square error (RMSE) between the actual frequency value and the target value.

[0166] minRMSE(f actual -f target )

[0167] Among them, f actual f is the actual frequency value.target This is the target frequency value.

[0168] Economic objectives: Minimize power generation costs and grid losses, such as:

[0169]

[0170] Among them, (a i ,b i ,c i P represents the generator cost coefficient. loss For network loss; P i Let P be the power of the i-th generating device, N be the number of generating devices, and λ be the network loss cost conversion factor, used to convert the network loss P... loss This is converted into an economic cost comparable to the cost of generating electricity.

[0171] Environmental goals: Minimize carbon emissions, such as:

[0172]

[0173] Among them, e i Carbon emissions per unit power of the generating unit; P i Let be the power of the i-th power generation device.

[0174] The specific process of the Bayesian optimization algorithm is as follows:

[0175] Since directly calculating f(θ) (which requires actual adjustment of parameter θ and evaluation of grid operation) is time-consuming and may affect system stability, the distribution of f(θ) can be approximated by a small number of samples using a Gaussian process, which can efficiently predict the optimization effect of any θ.

[0176] Construct a Gaussian process model for the objective function f(θ):

[0177]

[0178] Where m(θ) is the mean function and k(θ,θ′) is the kernel function, used to quantify the similarity between different parameters (i.e., θ and θ′): θ is the parameter to be optimized, and θ′ is other candidate parameters in the parameter search space compared with θ. If the values ​​of the strategy generation parameters corresponding to θ and θ′ are close (e.g., the output adjustment priorities of two generator sets are small), then the value of the kernel function k(θ,θ′) is large, indicating that the two have a high correlation in their impact on the objective function.

[0179] The objective function f(θ) follows a Gaussian process. It is determined by the mean function m(θ) and the kernel function k(θ,θ′). The mean function m(θ) gives the average predicted value of the function f(θ) at θ; the kernel function k(θ,θ′) describes the correlation between the two points θ and θ′, which determines the covariance structure of the Gaussian process and reflects the degree of similarity of the function values ​​between different input points.

[0180] The next evaluation point is selected by using the acquisition function α(θ), and the value of θ corresponding to the maximum value of α(θ) is used as the optimized parameter, i.e., θ t+1 :

[0181]

[0182] The commonly used data acquisition function α(θ) is the one that we hope to improve:

[0183] α(θ)=EI(θ)=E{max[f(θ)-f(θ + ),0]}

[0184] Where, θ + Let θ be the currently optimal parameter to be optimized. This formula means that, based on the current optimal solution, we calculate the expected improvement that a new point θ might bring. By maximizing this expected improvement, we can find the next potentially better evaluation point to progressively optimize the objective function.

[0185] The acquisition function (desired improvement EI) is a "search strategy" based on the Gaussian process model: its core is to find the θ that minimizes f(θ) to the greatest extent in the unknown parameter space. By maximizing EI, we can both explore new parameters that may bring significant improvements (raising the upper limit of optimization) and utilize known optimal regions (avoiding invalid attempts), ultimately achieving synergistic optimization of the objective.

[0186] Example: 1. Data acquisition module:

[0187] Data input processing: Real-time collection of data such as grid connection frequency, power generation equipment output, load demand, and renewable energy generation forecasts, for example:

[0188] The frequency sensor collects the grid connection point frequency f once per second, forming a sequence [f1, f2, ..., ft];

[0189] The anomaly detection submodule filters out bad data using the Isolation Forest algorithm, for example:

[0190] When the output data of a generator changes by 20% and contradicts the data of adjacent equipment, it is marked as abnormal and removed.

[0191] b. Data output application: Provide standardized input data to AGC logic.

[0192] 2. Status Assessment and Predictive Analysis Module:

[0193] Application of status assessment in AGC:

[0194] Graph neural networks (GNNs) can be used to analyze the stability of power grid topology. For example, the vulnerability score S_v of critical nodes (such as hub substations) can be calculated. When S_v > 0.8, AGC will prioritize adjusting the output of generating equipment near the node to avoid cascading failures.

[0195] Applications of predictive analytics in AGC:

[0196] Model predicts load for the next 15 minutes and new energy power generation For example: If the peak load at 8 PM tonight is predicted to be 500MW (confidence range ±10MW) and wind power output is 100MW (confidence range ±20MW), the AGC will reserve 500-100+30=430MW of adjustable capacity in advance (30MW as a reserve). The prediction error is used to adjust the AGC control dead zone. For example, when the prediction error exceeds 5%, the AGC control dead zone is increased from 0.2Hz to 0.3Hz to avoid frequent adjustments caused by inaccurate predictions.

[0197] 3. Intelligent control strategy generation module:

[0198] Policy generation based on deep reinforcement learning (DRL): Constructing a Markov decision process (MDP) for AGC: State S: [current frequency, unit output, load forecast, renewable energy forecast, system damping ratio]; Action A: [unit output adjustment, energy storage charging and discharging power]; Reward R: -|frequency deviation| - generation cost - carbon emissions, for example:

[0199]

[0200] The DRL model outputs optimal control commands. For example, when the frequency f = 49.8 Hz (below 50 Hz) and the load forecast increases, the DRL generates commands: increase the gas turbine capacity by 10 MW, discharge the energy storage by 5 MW, and prioritize stabilizing the frequency.

[0201] Federated learning optimizes global strategies: AGC systems in different regions share DRL model parameters. For example, a power grid in a northern region found that the strategy of "adjusting energy storage before starting gas turbines when heating load increases sharply in winter" is more efficient. Through federated learning, this empirical parameter (such as increasing the priority weight of energy storage from 0.3 to 0.6) is shared with other regions.

[0202] 4. Adaptive Adjustment and Feedback Module: AGC's "Optimization Engine" Online Transfer Learning Correction Strategy: When the penetration rate of new energy exceeds 30% (new scenario), transfer strategy parameters of similar scenarios, for example: transfer energy storage charging and discharging parameters from the historical strategy library of "high wind power ratio", and adjust the energy storage response speed parameter from 2s to 1s to adapt to the high volatility characteristics of new energy.

[0203] Mathematical implementation:

[0204] θ new =0.7θ high_wind +0.3θ current

[0205] Bayesian optimization for dynamic parameter tuning: Optimizing AGC control parameters, such as adjusting the power distribution ratio coefficient.

[0206] Optimization objective: Minimize frequency recovery time + power generation cost using a Bayesian optimization process;

[0207] First, we tested the scaling factors of 0.5, 0.6, and 0.7. We found that the frequency recovery time was 8 seconds and the cost increased by 5% when the scaling factor was 0.6, which is the current optimal solution.

[0208] Real-time feedback correction: If the actual frequency recovery time after execution is 10s (2s longer than the prediction), the Bayesian optimization update model will be performed, and the coefficients of 0.65 and 0.7 will be tested next time.

[0209] Implementation methods of automatic power generation control method based on energy management system:

[0210] An automatic power generation control method based on an energy management system includes: S1, acquiring real-time power grid operation data; S2, obtaining predicted load and predicted power generation for a future period based on the real-time power grid operation data using a pre-trained load prediction model and power generation prediction model; S3, generating a control strategy using a strategy generation model based on the predicted load, predicted power generation, real-time power grid operation data, and pre-set strategy generation parameters, wherein the control strategy includes the output of each power generation device and / or energy storage device; S4, executing the control strategy, and correcting the strategy generation parameters in the strategy generation model using an optimization algorithm based on the execution result of the control strategy; the load prediction model, power generation prediction model, and strategy generation model are all constructed based on machine learning models. The specific automatic power generation control method based on an energy management system has been described in sufficient detail in the above-described implementation of an automatic power generation control system based on an energy management system, and will not be repeated here.

[0211] Implementation methods for automatic power generation control devices based on energy management systems:

[0212] An automatic power generation control device based on an energy management system includes a processor for executing a computer program to implement the steps of the automatic power generation control method based on an energy management system as described above. The specific automatic power generation control method based on an energy management system has been described in sufficient detail in the above-described embodiments of the automatic power generation control system based on an energy management system, and will not be repeated here.

[0213] Specifically, a processor can be a CPU, or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. A processor can also be a processor that supports the Advanced Reduced Instruction Set Machine (ARM) architecture.

[0214] The present invention has the following advantages:

[0215] Enhanced data reliability: By running the anomaly detection submodule and the isolated forest algorithm in real time, the reliability of the input data is effectively ensured, providing a solid data foundation for subsequent analysis and control.

[0216] Accurate assessment and prediction: By using graph neural networks for power grid topology modeling and multimodal deep learning models for load and renewable energy generation prediction, the accuracy of state assessment and prediction precision are improved, making power generation control more forward-looking.

[0217] Intelligent strategy generation: Based on a deep reinforcement learning-based control strategy generator and federated learning mechanism, it can dynamically generate the optimal power allocation strategy and improve global collaborative control capabilities, thereby achieving efficient and economical operation of the power system.

[0218] Rapid Adaptation and Robustness: Online transfer learning technology enables the system to quickly adapt to new scenarios such as sudden changes in the penetration rate of new energy sources and grid failures, while Bayesian optimization algorithms ensure the robustness of the system in uncertain environments and reduce system performance fluctuations.

[0219] Enhancing automation and intelligence: The system is highly automated and intelligent, reducing reliance on manual operation, lowering the risk of human error, and providing more intelligent decision support, thereby improving the management efficiency of the power system.

[0220] In summary, the intelligent adaptive automatic generation control method and system based on EMS, with its advanced technology and innovative design, provides strong technical support for the stable, efficient and economical operation of power systems.

[0221] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still make modifications to the technical solutions described in the foregoing embodiments without creative effort, or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic generation control method based on an energy management system, characterized by, The method comprises: S1, acquiring real-time operation data of a power grid; S2, obtaining predicted load and predicted power generation of a future period of time according to the real-time operation data of the power grid through a pre-trained load prediction model and a power generation prediction model; S3, generating a control strategy through a strategy generation model according to the predicted load, the predicted power generation, the real-time operation data of the power grid and pre-set strategy generation parameters, the control strategy comprising output of each power generation device and / or energy storage device; S4, executing the control strategy and correcting the strategy generation parameters in the strategy generation model through an optimization algorithm according to an execution result of the control strategy; The load prediction model, the power generation prediction model and the strategy generation model are all constructed based on a machine learning model.

2. The energy management system based automatic generation control method of claim 1, wherein, The optimization target of the optimization algorithm comprises one or two or three of a stability target for guaranteeing stability of the power grid, an economic target for minimizing economic cost and an environmental protection target for minimizing carbon emission.

3. The energy management system based automatic generation control method of claim 1, wherein, The strategy generation parameters comprise one or two or three or a combination of more than three of output adjustment priority of each power generation unit, energy storage charging and discharging power threshold, AGC control dead zone, power distribution proportion coefficient and safety margin parameter.

4. The energy management system based automatic generation control method according to claim 1 or 3, characterized by, The machine learning model based on which the strategy generation model is based on is a reinforcement learning model; the strategy generation parameters comprise weight of each reward item.

5. The energy management system based automatic generation control method of claim 1, wherein, S2 further comprises: inputting power grid topology parameters of a current area to a pre-trained state evaluation model to obtain a power grid evaluation index; The power grid evaluation index comprises eigenvalue perturbation of power transmission between nodes; the state evaluation model is constructed based on a machine learning model, and the machine learning model based on which the state evaluation model is based on is a graph convolution network; S4 further comprises: adjusting parameters of the state evaluation model according to change of the power grid evaluation index.

6. The energy management system based automatic generation control method of claim 1, wherein, S2 further comprises: performing state evaluation calculation according to power grid topology parameters of a current area to obtain a power grid evaluation index; The power grid evaluation index comprises one or two or three of node vulnerability score, system damping ratio and power transmission vulnerability matrix; S4 further comprises: correcting parameters of the state evaluation calculation according to change of the power grid evaluation index.

7. The energy management system based automatic generation control method of claim 4, wherein, The reinforcement learning model is a DQN network; a loss function L(θ) for training the DQN network is: where θ is the current network parameter of the DQN network; θ - is the target network parameter of the DQN network; r is the immediate reward obtained after taking action a in state s; γ is the discount factor; is the maximum value estimate of all possible actions by the target network for the next state s ′ is the value estimate of the current Q network for taking action a in state s; E[·] denotes the calculation of expectation.

8. The energy management system based automatic generation control method of claim 4, wherein, The reward function in the reinforcement learning model comprises one or two or three or a combination of more than three of stability penalty, economic reward and environmental protection reward; the stability penalty is less than 0, and the more the system deviates from a steady state, the greater the absolute value of the stability penalty is; The lower the economic cost, the higher the economic reward is; The lower the carbon emission, the higher the environmental protection reward is.

9. The energy management system based automatic generation control method of claim 1, wherein, The machine learning model based on which the load prediction model and / or the power generation prediction model is based on comprises an LSTM network, a CNN network and an output layer; the LSTM network is used to extract time sequence features of input data; the CNN network is used to extract spatial features of input data; The output layer is used to splice features output by the LSTM network and the CNN network, and the spliced features are output after being processed by a fully connected layer.

10. The energy management system based automatic generation control method as claimed in claim 1 or 5, wherein, S4 further comprises: when it is detected that the power grid fails or the new energy penetration rate suddenly changes, it is considered that the power grid environment changes, and then the model parameters of each machine learning model are updated through online transfer learning, so that the updated each machine learning model can adapt to the new power grid environment.

11. The energy management system based automatic generation control method of claim 10, wherein, The way of updating the model parameters of each machine learning model through online transfer learning comprises: θ new = (1 - λ)θ old + λθ adapt wherein θ old is the original model parameter of the machine learning model; θ adapt is the parameter of the machine learning model adjusted for the new environment; θ new is the updated machine learning model parameter; and λ is the transfer weight, which is a coefficient between 0 and 1.

12. The energy management system based automatic generation control method as claimed in claim 1 or 5, wherein, The training process of all machine learning models comprises: After training each machine learning model locally using the data of the current region, all parameters of the trained machine learning model are uploaded to the central server, so that the central server can aggregate global model parameters according to all parameters of the machine learning model of each region; Each machine learning model is updated according to the global model parameters issued by the central server.

13. The energy management system based automatic generation control method of claim 12, wherein, The global model parameters are obtained according to the following way: wherein θ t+1 is a global model parameter of a certain machine learning model; n k is the number of training samples of the machine learning model of the kth region; n is the sum of the number of training samples of the machine learning model of all regions; is the parameter of the machine learning model trained according to n k training samples of the kth region; K is the number of all regions.

14. An automatic generation control device based on an energy management system, comprising a processor, characterized in that, The processor is configured to execute a computer program to implement the steps of the automatic generation control method based on the energy management system according to any one of claims 1-13.

15. An automatic generation control system based on an energy management system, characterized by, Comprise: A data acquisition module for acquiring real-time operation data of the power grid; A prediction analysis module for obtaining predicted load and predicted power generation of a future period of time according to the real-time operation data of the power grid through a pre-trained load prediction model and a power generation prediction model; An intelligent control strategy generation module for generating a control strategy through a strategy generation model according to the predicted load, the predicted power generation, the real-time operation data of the power grid, and pre-set strategy generation parameters, the control strategy comprising output of each power generation device and / or energy storage device; The load prediction model, the power generation prediction model, and the strategy generation model are all constructed based on machine learning models; An adaptive adjustment and feedback module for correcting strategy generation parameters in the strategy generation model through an optimization algorithm according to an execution result of the control strategy.

16. The energy management system based automatic generation control system of claim 15, wherein, Further comprising a state evaluation module for inputting power grid topology parameters of the current region to a pre-trained state evaluation model to obtain power grid evaluation indexes; The power grid evaluation indexes comprise eigenvalue perturbation of power transmission between nodes, the state evaluation model is constructed based on a machine learning model, and the machine learning model on which the state evaluation model is based is a graph convolution network; The adaptive adjustment and feedback module is further configured to adjust parameters of the state evaluation model according to changes in the power grid evaluation indexes.

Images