Artificial intelligence-based renewable energy power system operation optimization method and system

Abstract

This invention provides an artificial intelligence-based method and system for optimizing the operation of renewable energy power systems. The method constructs a hybrid deep learning model using convolutional neural networks, gated recurrent units, and Transformers, and utilizes multi-stage feature extraction and attention correction mechanisms to address the problem of insufficient prediction accuracy caused by the difficulty of capturing nonlinear spatiotemporal features in traditional prediction methods. By combining deep neural networks with reinforcement learning Actor-Critic networks to form a scheduling model, it achieves integrated solutions for optimal power flow and unit combination, solving the problems of progressively amplified prediction errors and excessively long iterative calculation times in scheduling. By deploying multiple agents to form a multi-agent deep reinforcement learning model, it addresses the problem of lag in centralized control response in low-inertia systems. Finally, by feeding back control commands to the data acquisition terminal to form a closed-loop optimization framework, it solves the problem of fragmented operation of various components.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, and in particular to a method and system for optimizing the operation of a renewable energy power system based on artificial intelligence. Background Technology

[0002] Traditional power systems rely on controllable generating units such as thermal, hydro, and nuclear power to maintain a real-time balance between power generation and consumption. When new energy sources such as wind or solar power are connected to the grid, frequent adjustments to traditional generating units are often necessary to smooth out fluctuations. If the adjustment capacity is insufficient, it can lead to wind and solar curtailment, meaning that new energy power generation is forced to stop due to insufficient system absorption capacity.

[0003] Existing technical solutions employ a separate forecasting, scheduling, and control model. First, a physical model of the weather station is used for independent forecasting. Then, the forecast results are input into traditional solvers based on the interior-point method for iterative calculation to obtain scheduling instructions. Finally, a single linear controller is used to regulate the frequency of the power grid.

[0004] The drawback of existing technologies is that renewable energy output is highly susceptible to environmental influences, exhibiting strong nonlinearity and random fluctuations. Existing technologies separate the four stages of forecasting, scheduling, control, and the market, with each stage relying on linear approximations or complex mathematical iterations. This results in extremely long computation times when dealing with complex grid topologies, sacrificing real-time performance. Furthermore, once an error occurs in the forecasting stage, this error is amplified at each stage, leading to delayed or even invalid control commands. Summary of the Invention

[0005] This invention provides an artificial intelligence-based method and system for optimizing the operation of renewable energy power systems, which addresses the shortcomings of existing technologies where errors occur in the prediction process, and these errors are amplified at each stage, leading to delays or even failures in the final control commands.

[0006] On the one hand, the present invention provides an artificial intelligence-based method for optimizing the operation of a renewable energy power system, characterized by comprising the following steps: Acquire multi-source operational data of the power system; The multi-source operating data is input into a hybrid deep learning model to predict the power generation of renewable energy and output the prediction results. The hybrid deep learning model is composed of a convolutional neural network, a gated recurrent unit, and a Transformer cascaded in sequence. The convolutional neural network is used to extract local fluctuation features, the gated recurrent unit is used to model long-period time-series dependencies, and the Transformer is used to correct the attention weights of the time-series features. The prediction results and grid topology parameters are input into the scheduling model to perform integrated scheduling of unit combination and optimal power flow, and generate scheduling instructions. The scheduling model includes a deep neural network for solving the optimal power flow and a reinforcement learning Actor-Critic network for solving the unit combination. The scheduling instructions and real-time operating status are input into a multi-agent deep reinforcement learning model to coordinate the control of multiple control objects on the source side, network side, load side, and storage side, and generate control instructions. The multi-agent deep reinforcement learning model includes multiple agents deployed on different control objects, and each agent makes collaborative decisions based on a decentralized architecture. The control commands are fed back to the acquisition terminal of the multi-source operating data to form a closed-loop optimization framework.

[0007] Optionally, the multi-source operating data is input into a hybrid deep learning model to predict renewable energy power generation and output the prediction results, including: Based on multi-source operational data, a multi-dimensional feature input tensor is constructed, which includes meteorological data and historical power generation data. The multidimensional feature input tensor is input into a convolutional neural network (CNN). A sliding window convolution operation is performed on the multidimensional feature input tensor through a one-dimensional convolution kernel to extract the local fluctuation features of renewable energy power generation and related influencing factors, and output a time series feature matrix. The time-series feature matrix is ​​input into a gated recurrent unit (GRU), and long-period time-series features are filtered and retained through update and reset gates, and time-dependent features are output. The time-dependent features are input into the Transformer network, and a multi-head self-attention mechanism is used to assign higher weights to key sample points of meteorological changes and power surges, thereby enhancing the core feature capture capability in extreme scenarios and outputting corrected prediction features. Using the mean squared error between the predicted and observed values ​​as the loss function, the network parameters of the hybrid deep learning model are iteratively optimized through the backpropagation algorithm to output the prediction results of future renewable energy power generation at multiple time scales.

[0008] Optionally, the prediction results and grid topology parameters are input into the scheduling model to perform integrated scheduling of unit combination and optimal power flow, generating scheduling instructions, including: A scheduling model is constructed that includes a DNN power flow mapping sub-model based on a deep neural network and a deep reinforcement learning scheduling sub-model based on a reinforcement learning Actor-Critic network. The node admittance matrix corresponding to the power grid topology parameters, the active load and reactive load corresponding to the prediction results are input into the DNN power flow mapping sub-model. The end-to-end mapping of the nonlinear power flow equation of the power grid is constructed through the forward propagation of the neural network, and the optimal power flow calculation results are output. At the same time, the safety constraint boundary of the unit output is determined. The prediction results and the safety constraint boundary are input into the deep reinforcement learning scheduling sub-model to construct a Markov decision process. The prediction results and the safety constraint boundary are used as the state space, the discrete start-stop state and continuous output adjustment value of the controllable generator set are used as the action space, and the reward function that comprehensively considers fuel cost, start-stop cost and over-limit penalty is used as the optimization objective to output scheduling instructions.

[0009] Optionally, the scheduling instructions and real-time operating status are input into a multi-agent deep reinforcement learning model to perform coordinated control of multiple control objects on the source side, network side, load side, and storage side, generating control instructions, including: Collect real-time power grid operation status data, and use a pre-trained convolutional neural network to denoise the real-time operation status data to extract the system state vector; A decentralized distributed control architecture is adopted, with independent control agents deployed for the source-side renewable energy generator sets, grid-connected converters, energy storage systems, and load-side flexible loads. The discrete control strategy is output by iterating the discrete control actions of energy storage charging and discharging mode switching and flexible load switching through a deep Q-network. The continuous control actions of generator power regulation, converter output control, and energy storage charging and discharging power regulation are iterated through a deep deterministic strategy gradient algorithm to output a continuous control strategy. Based on the discrete control strategy and the continuous control strategy, the control agent synchronously outputs control commands.

[0010] Optionally, it also includes: In the distributed power market, power generators, energy storage operators, and load aggregators are modeled as reinforcement learning bidding agents. The bidding agent is configured to operate within the price-volume curve, with its own profit maximization as the optimization objective. It iteratively optimizes its bidding strategy based on feedback from the marginal clearing results of the market environment.

[0011] Optionally, the control commands are fed back to the acquisition terminal of the multi-source operating data to form a closed-loop optimization framework, including: Collect system operation data after the control command is executed, execution deviation data of the scheduling command, and measured error data of the prediction results; The system operation data, the execution deviation data, and the measured error data are transmitted to the acquisition terminal of the multi-source operation data. The system operation data, execution deviation data, and measured error data are respectively input into the hybrid deep learning model, scheduling model, and multi-agent deep reinforcement learning model to update the network parameters of each model.

[0012] Optionally, it also includes: The node voltage amplitude, node voltage phase angle, and unit output safety constraint boundary output by the DNN power flow mapping sub-model are input into the state input layer of the deep reinforcement learning scheduling sub-model. The generator start-stop status and generator output adjustment value output by the deep reinforcement learning scheduling sub-model are input into the input tensor of the DNN power flow mapping sub-model.

[0013] Optionally, it also includes: Send the control command to the corresponding controllable object; Collect power grid operation status data after the control command is executed; The power grid operation status data are input to the corresponding control agents respectively; Based on the real-time operating status data, each control agent updates the network parameters of the deep Q-network and the network parameters of the deep deterministic policy gradient algorithm, respectively.

[0014] Optionally, the parameter update of the hybrid deep learning model, scheduling model, and multi-agent deep reinforcement learning model includes: Each data acquisition node and execution node generates encrypted gradient parameters for the corresponding model. The encryption gradient parameters are transmitted between each node; Each node updates the network parameters of the corresponding model based on the received encrypted gradient parameters.

[0015] On the other hand, the present invention also provides an artificial intelligence-based renewable energy power system operation optimization system, comprising: The data acquisition module is used to acquire multi-source operational data of the power system; The power prediction module is used to input the multi-source operating data into the hybrid deep learning model to predict the power generation of renewable energy and output the prediction results. The hybrid deep learning model is composed of a convolutional neural network, a gated recurrent unit and a Transformer cascaded in sequence. The convolutional neural network is used to extract local fluctuation features, the gated recurrent unit is used to model long-period time-series dependencies, and the Transformer is used to perform attention weight correction on the time-series features. The joint scheduling module is used to input the prediction results and grid topology parameters into the scheduling model to perform integrated scheduling of unit combination and optimal power flow, and generate scheduling instructions; the scheduling model includes a deep neural network for solving the optimal power flow and a reinforcement learning Actor-Critic network for solving the unit combination. The collaborative control module is used to input the scheduling instructions and real-time operating status into a multi-agent deep reinforcement learning model to perform collaborative control on multiple control objects on the source side, network side, load side, and storage side, and generate control instructions. The multi-agent deep reinforcement learning model includes multiple agents deployed on different control objects, and each agent makes collaborative decisions based on a decentralized architecture. The control instructions are fed back to the acquisition terminal of the multi-source operating data to form a closed-loop optimization framework.

[0016] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the artificial intelligence-based renewable energy power system operation optimization method as described above.

[0017] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the artificial intelligence-based renewable energy power system operation optimization method as described above.

[0018] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the artificial intelligence-based renewable energy power system operation optimization method as described above.

[0019] This invention provides an artificial intelligence-based method and system for optimizing the operation of renewable energy power systems. The method constructs a hybrid deep learning model by cascading convolutional neural networks, gated recurrent units, and Transformers. Utilizing its multi-stage feature extraction and attention correction mechanisms, it addresses the problem of insufficient prediction accuracy caused by the difficulty of capturing nonlinear spatiotemporal features in traditional prediction methods, thus improving prediction accuracy in extreme scenarios. By combining deep neural networks with reinforcement learning Actor-Critic networks to form a scheduling model, it achieves integrated solutions for optimal power flow and unit combination, solving the problems of progressively amplified prediction errors and excessively long iterative calculation times in traditional scheduling. This effectively eliminates error accumulation and compresses calculation time to the millisecond level. By deploying multiple agents based on decentralized architecture for collaborative decision-making to form a multi-agent deep reinforcement learning model, it solves the problem of lag in centralized control response in low-inertia systems, improving collaborative control efficiency and system dynamic stability. Finally, by feeding control commands back to the data acquisition terminal to form a closed-loop optimization framework, it solves the problem of fragmented operation in each link, achieving the beneficial effect of full-chain collaborative optimization. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the artificial intelligence-based renewable energy power system operation optimization method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system structure for optimizing the operation of a renewable energy power system based on artificial intelligence, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] Figure 1It is a schematic flowchart of an operation optimization method for a renewable energy power system based on artificial intelligence provided by an embodiment of the present invention.

[0024] As Figure 1 shown, the operation optimization method for a renewable energy power system based on artificial intelligence provided by an embodiment of the present invention mainly includes the following steps: 101. Obtain multi-source operation data of the power system.

[0025] Specifically, through the SCADA data acquisition and monitoring control system and the AIoT (Artificial Intelligence of Things) sensor network supporting the power system, multi-source operation data of the power system is collected at a high-frequency sampling rate. The collected multi-source operation data includes meteorological data, historical power generation power data of renewable energy, and historical load data of the power grid. The meteorological data includes meteorological variables such as wind speed, irradiance, and temperature that affect renewable energy power generation. At the same time, basic operation data of the power grid such as electrical parameters of power grid nodes and line operation states are collected. After cleaning and standardizing the collected multi-source operation data, removing outliers and missing values, the multi-source operation data is reconstructed into a multi-dimensional feature input tensor including batch size, historical time step, and feature dimension, forming multi-source operation data that can be directly called by subsequent models.

[0026] 102. Input the multi-source operation data into a hybrid deep learning model to predict the power generation power of renewable energy and output a prediction result.

[0027] Among them, the hybrid deep learning model is composed of a convolutional neural network, a gated recurrent unit, and a Transformer cascaded in sequence. For example, a one-dimensional convolutional neural network is used to perform sliding window convolution on the multi-dimensional feature tensor to extract local fluctuation features and output a time series feature matrix; the time series feature matrix is input into the gated recurrent unit to model long-term time series dependencies through update gates and reset gates and output time series dependency features; the time series dependency features are input into a Transformer network with 2-layer Transformer encoders, 8 heads of self-attention, and a feed-forward network dimension of 2048. High weights are assigned to meteorological mutations and power sudden change points through multi-head self-attention to complete feature correction; with the mean square error as the loss function, the model parameters are optimized by backpropagation iteration, and a prediction result is output.

[0028] Specifically, inputting the multi-source operation data into the hybrid deep learning model to predict the power generation power of renewable energy and outputting a prediction result includes: Based on the multi-source operation data, a multi-dimensional feature input tensor is constructed, and the multi-dimensional feature input tensor includes meteorological data and historical power generation data.

[0029] Among them, based on the preprocessed multi-source operation data of the power system, a multi-dimensional feature input tensor with a dimension of is constructed ,in, This represents the batch size, which is the number of samples processed simultaneously in a single batch. This represents the historical time step, i.e., the length of the historical time window selected for observation. The representative feature dimension includes meteorological data such as wind speed, irradiance, and temperature, as well as historical power generation data such as historical power generation of renewable energy. It integrates various heterogeneous feature data into a standardized tensor structure, providing a unified input format for feature extraction and analysis of subsequent models.

[0030] The multidimensional feature input tensor is input into the convolutional neural network (CNN). A sliding window convolution operation is performed on the multidimensional feature input tensor through a one-dimensional convolution kernel to extract the local fluctuation features of renewable energy power generation and related influencing factors, and output the time series feature matrix.

[0031] Among them, the constructed multidimensional feature input tensor Inputting a convolutional neural network (CNN), it performs a sliding window convolution operation on the input time series data using a one-dimensional convolution kernel, through the formula... Complete local feature mapping, where, Temporal feature matrix output by the convolutional layer Here is the weight matrix of the convolution kernel. Input tensors for multidimensional features, For bias vectors, The linear rectification activation function is used to extract the local fluctuation patterns of renewable energy power generation and related influencing factors such as meteorology and historical power generation over a very short period of time, ultimately outputting a time-series feature matrix containing local spatial characteristics. .

[0032] The time series feature matrix is ​​input into the gated recurrent unit (GRU), and the long-period time series features are filtered and retained through update gates and reset gates, and the time series dependent features are output.

[0033] Among them, the temporal feature matrix output by the convolutional neural network Input is gated recurrent unit (GRU), and the input passes through the GRU's update gate. The reset gate filters and retains long-term memories, while the update gate... The hidden layer state at the current moment is obtained by using GRU computation to control the degree of retention of historical information. The hidden layer state at the previous time step was The current candidate hidden layer state is This process models long-period time-series dependencies in the time-series feature matrix, effectively mitigating the gradient vanishing problem in long-term time-series prediction, filtering and retaining effective long-period time-series features, and finally outputting time-series dependency features.

[0034] Among them, the hidden layer state The calculation formula is: .

[0035] By inputting time-dependent features into the Transformer network, a multi-head self-attention mechanism is used to assign higher weights to key sample points for sudden weather changes and power fluctuations, thereby enhancing the ability to capture core features in extreme scenarios and outputting corrected prediction features.

[0036] The temporal dependency features output by the GRU are input into the Transformer network. At the end of the network, a multi-head self-attention mechanism is used to correct the attention weights of the temporal features. The attention calculation formula is as follows: .

[0037] in, These are the query matrix, key matrix, and value matrix, respectively. This is the transpose of the key matrix. Let be the dimension of the key matrix. Used to scale the dot product to prevent gradient vanishing. The normalized exponential function is used to enable the model to automatically identify key sample points of sudden weather changes and power fluctuations, and assign them higher attention weights to enhance the model's ability to capture core features in extreme scenarios. After correcting the time-dependent features, the corrected predicted features are output.

[0038] Using the mean squared error between predicted and observed values ​​as the loss function, the network parameters of the hybrid deep learning model are iteratively optimized through the backpropagation algorithm to output the prediction results of future renewable energy power generation at multiple time scales.

[0039] The loss function is the mean square error between the predicted value and the actual observed value.

[0040] The formula for calculating the loss function is as follows: .

[0041] in, The mean squared error loss value of the prediction model. The total number of samples, For the first The true observed value of each sample For the first The model prediction value for each sample. The summation operation is represented by the loss function, which is used to calculate the model prediction error. The error is then propagated from the model output layer to the input layer through the backpropagation algorithm. The network parameters of the convolutional neural network, gated recurrent unit, and Transformer network are iteratively optimized in turn to continuously reduce the model prediction loss. After the model converges, the prediction results of the power generation of renewable energy in the future at multiple time scales, such as ultra-short term and day-ahead, are output.

[0042] 103. Input the prediction results and grid topology parameters into the scheduling model to perform integrated scheduling of unit combination and optimal power flow, and generate scheduling instructions.

[0043] The scheduling model includes a deep neural network for solving the optimal power flow problem and a reinforcement learning Actor-Critic network for solving the unit combination problem.

[0044] Specifically, the prediction results and grid topology parameters are input into the scheduling model to perform integrated scheduling of unit combination and optimal power flow, generating scheduling instructions, including: A scheduling model is constructed that includes a DNN power flow mapping sub-model based on a deep neural network and a deep reinforcement learning scheduling sub-model based on an Actor-Critic network.

[0045] To achieve integrated scheduling of unit combination and optimal power flow, a scheduling model integrating two core sub-models is constructed. The scheduling model includes a DNN power flow mapping sub-model for solving optimal power flow and a deep reinforcement learning scheduling sub-model for solving unit combination. The DNN power flow mapping sub-model uses a deep neural network as its core architecture, abandoning the time-consuming iterative process of Jacobian matrix inversion in traditional optimal power flow calculation, and achieving end-to-end solution of power flow equations through neural network forward propagation. The deep reinforcement learning scheduling sub-model is built based on an Actor-Critic reinforcement learning network. It completes the optimization decision of unit combination through a collaborative approach of generating unit scheduling actions through the Actor network and evaluating the value of actions through the Critic network. The two sub-models cooperate with each other and share data to jointly achieve integrated solution of optimal power flow and unit combination.

[0046] The node admittance matrix corresponding to the power grid topology parameters and the active and reactive loads corresponding to the prediction results are input into the DNN power flow mapping sub-model. The end-to-end mapping of the nonlinear power flow equation of the power grid is constructed through the forward propagation of the neural network, and the optimal power flow calculation results are output. At the same time, the safety constraint boundary of the unit output is determined.

[0047] Specifically, the node admittance matrix corresponding to the grid topology parameters is extracted, and the active and reactive loads of the power system are determined based on the renewable energy power generation prediction results. The node admittance matrix, active load, and reactive load are used together as input tensors and input into the DNN power flow mapping sub-model. The DNN power flow mapping sub-model performs forward propagation calculations through a deep neural network, using the traditional nonlinear power balance equations of the power grid. Construct an end-to-end nonlinear mapping relationship.

[0048] in, For nodes Active power generation, For nodes Active load, and They are nodes , voltage amplitude, and They are nodes , The electrical conductivity and susceptance between them For nodes , The voltage phase angle difference between nodes is mapped to directly output the node voltage amplitude that meets the power system safety constraints. and voltage phase angle This is the optimal power flow calculation result. Based on this calculation result, the safety constraint boundary of the controllable generator output is defined and output, providing a constraint basis for subsequent unit combination optimization.

[0049] The prediction results and safety constraint boundaries are input into the deep reinforcement learning scheduling sub-model to construct a Markov decision process. The prediction results and safety constraint boundaries are used as the state space, the discrete start-stop state and continuous output adjustment value of the controllable generator set are used as the action space, and the reward function that comprehensively considers fuel cost, start-stop cost and over-limit penalty is used as the optimization objective to output scheduling instructions.

[0050] Specifically, the renewable energy power generation prediction results and the unit output safety constraint boundary output by the DNN power flow mapping sub-model are jointly input into the deep reinforcement learning scheduling sub-model. Based on this input, a Markov decision process is constructed for unit combination optimization. The above prediction results and safety constraint boundary are used as the state space of this process, and the discrete start-stop states of the controllable generator units and the continuous output adjustment values ​​are used as the action space.

[0051] Simultaneously, a reward function that comprehensively considers fuel cost, start-stop cost, and over-limit penalty is designed as the optimization objective. The reward function formula is: .

[0052] in, For the unit fuel costs, For the unit Start-up and shutdown costs For the unit exist The start / stop status at any given time. This is the penalty coefficient for exceeding the limit. The system voltage or line overload penalty term is used; the Actor network of the sub-model outputs unit scheduling actions based on the state space, and the Critic network evaluates the value of the action through the advantage function A(s,a) and feeds it back to the Actor network. The Actor network optimizes the action strategy by updating the network parameters through policy gradient until the reward function converges to the optimal value. Finally, it outputs the power system scheduling instructions that include unit start-up and shutdown instructions and output adjustment instructions, and completes the integrated scheduling decision of unit combination and optimal power flow.

[0053] The formula for updating the Actor network parameters using policy gradient is as follows: .

[0054] Among them, This is the advantage function calculated for the Critic network.

[0055] In addition, the DNN power flow mapping sub-model adopts a 4-layer fully connected DNN with ReLU as the activation function. The input node admittance matrix and active / reactive load are used to construct the end-to-end mapping of the nonlinear power flow equation through forward propagation. The output node voltage magnitude, phase angle and unit output safety constraint boundary are used. The Actor-Critic scheduling sub-model uses prediction results and safety constraints as its state space, unit start-up and shutdown and output adjustment as its action space, and fuel cost + start-up and shutdown cost + over-limit penalty as its reward function to output scheduling instructions. The two sub-models interact bidirectionally. The power flow results output by the DNN power flow mapping sub-model are input into the state layer of the Actor-Critic scheduling sub-model, and the unit decisions output by the Actor-Critic scheduling sub-model are input into the input tensor of the DNN power flow mapping sub-model, thus achieving integrated solution.

[0056] 104. Input scheduling instructions and real-time operating status into a multi-agent deep reinforcement learning model to perform coordinated control of multiple control objects on the source side, network side, load side, and storage side, and generate control instructions.

[0057] Among them, the multi-agent deep reinforcement learning model includes multiple agents deployed on different control objects, and each agent makes collaborative decisions based on a decentralized architecture.

[0058] Specifically, scheduling instructions and real-time operating status are input into a multi-agent deep reinforcement learning model to coordinate the control of multiple control objects on the source side, grid side, load side, and storage side, generating control instructions, including: Collect real-time power grid operation status data, and use a pre-trained convolutional neural network to denoise the real-time operation status data and extract the system state vector; A decentralized distributed control architecture is adopted, with independent control agents deployed for the source-side renewable energy generator sets, grid-connected converters, energy storage systems, and load-side flexible loads. The discrete control strategy is output by iterating the discrete control actions of energy storage charging and discharging mode switching and flexible load switching through a deep Q-network. The continuous control actions of generator power regulation, converter output control, and energy storage charging and discharging power regulation are iterated through a deep deterministic strategy gradient algorithm to output a continuous control strategy. Based on discrete control strategies and continuous control strategies, control commands are output synchronously by controlling the intelligent agent.

[0059] This includes collecting real-time status data during power grid operation, including system frequency deviation. Voltage deviation Real-time photovoltaic power generation Energy storage battery state of charge The collected real-time power grid operation status data is input into a pre-trained convolutional neural network. The noisy real-time operation status data is denoised by feature extraction and filtering operations of the convolutional neural network, removing system noise and interference signals from the data, and reconstructing and extracting the system state vector st that can truly reflect the power grid operation status.

[0060] The formula for calculating the system state vector st is: .

[0061] in Let be the system state vector. For frequency deviation, For voltage deviation, For photovoltaic power, This refers to the amount of electricity stored.

[0062] A decentralized distributed control architecture is established, which does not set up a centralized master control node. Each control unit makes independent decisions and achieves information exchange. For four different control objects, namely renewable energy generators on the source side, grid-connected converters on the grid side, energy storage systems on the storage side, and flexible loads on the load side, independent control agents are deployed one-to-one. Each control agent is only responsible for the control decision calculation of its corresponding control object. Each control agent obtains global grid operation information based on the extracted system state vector, ensuring the coordinated cooperation among multiple control objects.

[0063] For discrete control actions such as energy storage charging and discharging mode switching and flexible load switching, a deep Q-network is used to iteratively optimize the control strategy. The system state vector is used as the input of the deep Q-network, and the execution effect of discrete control actions is used as the evaluation criterion. The parameters of the deep Q-network are iteratively updated by minimizing the time difference error, which is the loss function of the deep Q-network. The strategy optimization of discrete control actions is completed through the iterative convergence of the loss function, and finally the discrete control strategy for energy storage charging and discharging and flexible load switching is adapted to the real-time operating state of the power grid is output.

[0064] The loss function of the deep Q-network is: .

[0065] For continuous control actions such as generator power regulation, grid-connected converter output control, and energy storage charging and discharging power regulation, a deep deterministic strategy gradient algorithm is used to iteratively optimize the control strategy. The system state vector is input into the decision network of the deep deterministic strategy gradient algorithm, and the algorithm directly outputs a continuous power reference value deltaPref. Based on the state feedback of the grid operation, the network parameters of the algorithm are iteratively adjusted. Through multiple rounds of action iteration and parameter optimization, the continuous control actions output by the algorithm are matched with the real-time operating state of the grid, and finally, a continuous power regulation control strategy for generators, grid-connected converters, and energy storage systems is generated.

[0066] The discrete control strategy output by the deep Q-network and the continuous control strategy output by the deep deterministic strategy gradient algorithm are respectively distributed to the corresponding source-side, network-side, load-side, and storage-side control agents. Based on the received control strategy, each control agent simultaneously performs decision calculations and generates matching control commands for its corresponding control object. Each control agent synchronously distributes the control commands to the corresponding execution unit, realizing the coordinated control execution of multiple control objects on the source-side, network-side, load-side, and storage-side.

[0067] 105. Feed back control commands to the acquisition terminal of multi-source operating data to form a closed-loop optimization framework.

[0068] Specifically, control commands are fed back to the multi-source operational data acquisition terminal to form a closed-loop optimization framework, including: Collect system operation data after the execution of control commands, execution deviation data of scheduling commands, and measured error data of prediction results.

[0069] The process involves collecting and executing full-scale power grid system operation data after control commands are issued to and executed by the execution units on the source, grid, load, and storage sides. This data includes real-time power grid operation parameters such as grid node voltage, frequency, actual output of each generating unit, energy storage state of charge, and actual power of flexible loads. Simultaneously, it calculates the execution deviation data of dispatch commands, which is the difference between the dispatch command values ​​(generator start-up, shutdown, output adjustment, etc.) output by the dispatch model and the actual action values ​​executed by each device. Furthermore, it calculates the measured error data of the prediction results, which is the difference between the predicted renewable energy power generation output by the hybrid deep learning model and the measured renewable energy power generation data actually collected by the power grid. This completes the comprehensive collection and accurate calculation of these three types of data.

[0070] The system operation data, execution deviation data, and measured error data are transmitted to the multi-source operation data acquisition terminal.

[0071] Specifically, the system operation data after the control commands are executed, the calculated scheduling command execution deviation data, and the actual error data of the prediction results are transmitted to the power system multi-source operation data acquisition terminal through the power system's data communication link. This realizes the back transmission of the three types of data to the data acquisition source, providing data support for subsequent model parameter updates and full-process optimization, and ensuring the orderly flow of data within the closed-loop framework.

[0072] The system operation data, execution deviation data, and measured error data are respectively input into the hybrid deep learning model, the scheduling model, and the multi-agent deep reinforcement learning model to update the network parameters of each model.

[0073] Specifically, the measured error data of the prediction results transmitted back to the acquisition end is input into the hybrid deep learning model. Based on this error data, the network parameters of the convolutional neural network, gated recurrent unit, and Transformer network in the model are iteratively updated through the backpropagation algorithm. The execution deviation data of the scheduling command is input into the scheduling model, and the deep neural network parameters of the DNN power flow mapping sub-model and the Actor-Critic network parameters of the deep reinforcement learning scheduling sub-model are corrected and updated respectively. The system operation data after the control command is executed is input into the multi-agent deep reinforcement learning model, and the network parameters of the deep Q network and the deep deterministic policy gradient algorithm in the model are optimized and updated. This achieves targeted adjustment of the parameters of each model, completes the closed-loop parameter iteration of the entire optimization framework, and allows each model to continuously adapt to the actual operating state of the power grid.

[0074] In some embodiments, the artificial intelligence-based renewable energy power system operation optimization method provided in this invention further includes: In the distributed power market, power generators, energy storage operators, and load aggregators are modeled as reinforcement learning bidding agents. The bidding agent is configured to operate within the price and volume curves of the bidding process, with the goal of maximizing its own profits. It iteratively optimizes its bidding strategy based on feedback from the marginal clearing results of the market environment.

[0075] Specifically, for each market participant in the distributed power market, such as power generators, energy storage operators, and load aggregators, corresponding reinforcement learning bidding agents are constructed. Each bidding agent independently maps the power trading behavior and decision-making needs of its corresponding market participant. It can autonomously generate power trading application decisions based on the real-time operating status of the distributed power market, its own operating costs, and power supply and demand. The reinforcement learning bidding agents form a game theory framework for the distributed power market, achieving accurate modeling and decision simulation of the trading behavior of a large number of heterogeneous participants in the market.

[0076] The action space for each reinforcement learning bidding agent is set as the bid volume-price curve. ,in, For the declared electricity volume of each market entity, This corresponds to the declared price.

[0077] Simultaneously, the optimization objective for each bidding agent is set as maximizing its own expected trading profit, and the formula for calculating the expected profit is: .

[0078] in, To ultimately unify and settle electricity prices, This refers to the actual amount of electricity won in the bid. This is the marginal cost function.

[0079] The central market environment of the distributed power market executes a marginal clearing algorithm on the bid volume and price curves of all bidding agents, generates and feeds back a unified clearing price and the actual winning bid volume of each bidding agent. After receiving the market clearing results, each bidding agent, with the goal of maximizing its own profit, iteratively adjusts the parameters of its internal decision network, continuously optimizes the generation strategy of the bid volume and price curves, and realizes dynamic optimization of the bidding strategy.

[0080] In some embodiments, the artificial intelligence-based renewable energy power system operation optimization method provided in this invention further includes: The node voltage magnitude, node voltage phase angle, and unit output safety constraint boundary output by the DNN power flow mapping sub-model are input into the state input layer of the deep reinforcement learning scheduling sub-model. The generator start-stop status and generator output adjustment value output by the deep reinforcement learning scheduling sub-model are input into the input tensor of the DNN power flow mapping sub-model.

[0081] Specifically, the grid node voltage amplitude is output by the DNN power flow mapping sub-model through end-to-end mapping operations. Node voltage phase angle The power output safety constraint boundary of the units, which is delineated based on the power flow calculation results, is used as the state parameter of the deep reinforcement learning scheduling sub-model and directly input into the state input layer of the scheduling sub-model. The power flow calculation results and safety constraints are integrated into the state space of the unit combination optimization, so that when the deep reinforcement learning scheduling sub-model solves the unit combination problem, it takes the actual power flow operation state of the power grid as the constraint basis, and ensures that the unit start-up and shutdown and power output adjustment decisions meet the safe operation requirements of the optimal power flow of the power grid.

[0082] The discrete start-stop states of controllable generator sets are optimized from the deep reinforcement learning scheduling sub-model. Continuous output adjustment value The input tensor of the DNN power flow mapping sub-model is integrated into the original input tensor, and together with the grid node admittance matrix, active load, and reactive load, it serves as the input data of the power flow mapping sub-model. This allows the DNN power flow mapping sub-model to incorporate the latest decision results of unit combination in real time when performing optimal power flow calculation, realizing mutual feedback of input data between optimal power flow calculation and unit combination optimization, and ensuring the coordinated adaptation of the results of the two types of calculations.

[0083] In some embodiments, the artificial intelligence-based renewable energy power system operation optimization method provided in this invention further includes: Send control commands to the corresponding controllable objects; Collect power grid operation status data after control commands are executed; The power grid operation status data are input to the corresponding control agents respectively; Each control agent updates the network parameters of the deep Q-network and the network parameters of the deep deterministic policy gradient algorithm based on real-time operating status data.

[0084] Specifically, the discrete and continuous control commands output by the multi-agent deep reinforcement learning model are distributed to the controllable execution objects corresponding to the source-side renewable energy generator sets, grid-connected converters, energy storage systems, and load-side flexible loads, according to the type and affiliation of the controlled objects. This ensures that various control commands are accurately matched with the corresponding control equipment and directly drive each controllable object to perform corresponding power regulation, mode switching, load switching, and other operations.

[0085] After each controllable object completes the execution of the control command, the real-time operating status data of the entire power grid is collected through the SCADA acquisition system and IoT sensors of the power system. The data includes core electrical parameters that reflect the grid regulation effect, such as grid system frequency, node voltage, actual output of each generator set, state of charge of energy storage batteries, and actual power consumption of flexible loads, so as to obtain the actual operating status of the power grid after the execution of the control command.

[0086] The collected power grid operation status data after the control commands are executed are classified according to the control objects of source, grid, load and storage, and transmitted to the corresponding control agents respectively. This allows each control agent to obtain the power grid status feedback data after the execution of the commands by the object it is responsible for regulating. At the same time, it ensures that each agent can share the global operation status information of the power grid, providing data basis for subsequent parameter updates of the control strategy.

[0087] The control agent responsible for discrete control calculates the time-series differential error of the deep Q-network based on the received grid operation status data. Using this error as the optimization basis, it iteratively updates the network parameters of the deep Q-network through backpropagation to optimize the discrete control strategies for energy storage charging and discharging mode switching and flexible load switching. The control agent responsible for continuous control iteratively adjusts the decision network and evaluation network parameters of the algorithm based on the received grid operation status data and the strategy evaluation results of the deep deterministic strategy gradient algorithm to optimize the continuous control strategies for generator power regulation, converter output control, and energy storage charging and discharging power regulation. This achieves targeted updates of algorithm parameters for each control agent and adaptive optimization of control strategies.

[0088] In some embodiments, parameter updates for hybrid deep learning models, scheduling models, and multi-agent deep reinforcement learning models include: Each data acquisition node and execution node generates encrypted gradient parameters for the corresponding model. Encrypted gradient parameters are transmitted between nodes; Each node updates the network parameters of the corresponding model based on the received encrypted gradient parameters.

[0089] Specifically, at the multi-source operation data acquisition nodes and the control execution nodes on the source side / grid side / load side / storage side of the power system, the gradient parameters of the hybrid deep learning model, scheduling model, and multi-agent deep reinforcement learning model are independently calculated based on the model training errors generated during their own operation. Each node only performs gradient solving for the corresponding model deployed locally, without involving the original operation data and model parameters of other nodes.

[0090] Subsequently, each node uses an encryption algorithm to encrypt the local model gradient parameters obtained from the solution, forming encrypted gradient parameters to prevent the leakage of core power grid operation and model configuration information during transmission.

[0091] Each data acquisition node and execution node uses a dedicated communication link for the power system to transmit encrypted gradient parameters between nodes. Each node only sends out the encrypted gradient parameters it generates, while receiving encrypted gradient parameters transmitted by other nodes. Throughout the transmission process, no node transmits the original power grid operation data, the original model parameters, or the unencrypted gradient information. Only the distributed sharing of encrypted gradient parameters is achieved.

[0092] Each data acquisition node and execution node decrypts the encrypted gradient parameters received from other nodes, and combines the locally solved and decrypted gradient parameters to perform aggregate calculations on the gradient information. Subsequently, based on the aggregated gradient parameters, each node iteratively updates the network parameters of the locally deployed hybrid deep learning model, scheduling model, and multi-agent deep reinforcement learning model, respectively, to achieve distributed collaborative optimization of the parameters of each model. While completing the model parameter update, the privacy and security of the original data and local model parameters of each node are guaranteed.

[0093] Based on the same inventive concept, this invention also protects an artificial intelligence-based renewable energy power system operation optimization system. The artificial intelligence-based renewable energy power system operation optimization system provided by this invention will be described below. The artificial intelligence-based renewable energy power system operation optimization system described below can be referred to in correspondence with the artificial intelligence-based renewable energy power system operation optimization method described above.

[0094] In some embodiments, such as Figure 2 As shown, this embodiment of the invention also provides an artificial intelligence-based renewable energy power system operation optimization system, comprising: Data acquisition module 210 is used to acquire multi-source operating data of the power system; The power prediction module 220 is used to input multi-source operating data into the hybrid deep learning model to predict the power generation of renewable energy and output the prediction results. The hybrid deep learning model is composed of a convolutional neural network, a gated recurrent unit and a Transformer cascaded in sequence. The convolutional neural network is used to extract local fluctuation features, the gated recurrent unit is used to model long-period time-series dependencies, and the Transformer is used to perform attention weight correction on the time-series features. The joint scheduling module 230 is used to input the prediction results and grid topology parameters into the scheduling model to perform integrated scheduling of unit combination and optimal power flow, and generate scheduling instructions. The scheduling model includes a deep neural network for solving the optimal power flow and a reinforcement learning Actor-Critic network for solving the unit combination. The collaborative control module 240 is used to input scheduling instructions and real-time operating status into a multi-agent deep reinforcement learning model to perform collaborative control on multiple control objects on the source side, network side, load side, and storage side, and generate control instructions. The multi-agent deep reinforcement learning model includes multiple agents deployed on different control objects, and each agent makes collaborative decisions based on a decentralized architecture. The control instructions are fed back to the multi-source operating data acquisition terminal to form a closed-loop optimization framework.

[0095] Figure 3 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention.

[0096] like Figure 3 As shown, the electronic device may include a processor 310, a communications interface 320, a memory 330, and a communication bus 340. The processor 310, communications interface 320, and memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions from the memory 330 to execute an artificial intelligence-based renewable energy power system operation optimization method.

[0097] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0098] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the artificial intelligence-based renewable energy power system operation optimization method provided by the above methods.

[0099] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the artificial intelligence-based renewable energy power system operation optimization method provided by the above methods.

[0100] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0101] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-based renewable energy power system operation optimization method, characterized in that, Includes the following steps: Acquire multi-source operational data of the power system; The multi-source operating data is input into a hybrid deep learning model to predict the power generation of renewable energy and output the prediction results. The hybrid deep learning model is composed of a convolutional neural network, a gated recurrent unit, and a Transformer cascaded in sequence. The convolutional neural network is used to extract local fluctuation features, the gated recurrent unit is used to model long-period time-series dependencies, and the Transformer is used to correct the attention weights of the time-series features. The prediction results and grid topology parameters are input into the scheduling model to perform integrated scheduling of unit combination and optimal power flow, and generate scheduling instructions. The scheduling model includes a deep neural network for solving the optimal power flow and a reinforcement learning Actor-Critic network for solving the unit combination. The scheduling instructions and real-time operating status are input into a multi-agent deep reinforcement learning model to coordinate the control of multiple control objects on the source side, network side, load side, and storage side, and generate control instructions. The multi-agent deep reinforcement learning model includes multiple agents deployed on different control objects, and each agent makes collaborative decisions based on a decentralized architecture. The control commands are fed back to the acquisition terminal of the multi-source operating data to form a closed-loop optimization framework.

2. The artificial intelligence-based renewable energy power system operation optimization method according to claim 1, characterized in that, The multi-source operational data is input into a hybrid deep learning model to predict renewable energy power generation, and the prediction results are output, including: Based on multi-source operational data, a multi-dimensional feature input tensor is constructed, which includes meteorological data and historical power generation data. The multidimensional feature input tensor is input into a convolutional neural network (CNN). A sliding window convolution operation is performed on the multidimensional feature input tensor through a one-dimensional convolution kernel to extract the local fluctuation features of renewable energy power generation and related influencing factors, and output a time series feature matrix. The time-series feature matrix is ​​input into a gated recurrent unit (GRU), and long-period time-series features are filtered and retained through update and reset gates, and time-dependent features are output. The time-dependent features are input into the Transformer network, and a multi-head self-attention mechanism is used to assign higher weights to key sample points of meteorological changes and power surges, thereby enhancing the core feature capture capability in extreme scenarios and outputting corrected prediction features. Using the mean squared error between the predicted and observed values ​​as the loss function, the network parameters of the hybrid deep learning model are iteratively optimized through the backpropagation algorithm to output the prediction results of future renewable energy power generation at multiple time scales. 3.The AI-based renewable energy power system operation optimization method of claim 1, wherein, The predicted results and grid topology parameters are input into the scheduling model to perform integrated scheduling of unit combination and optimal power flow, generating scheduling instructions, including: A scheduling model is constructed that includes a DNN power flow mapping sub-model based on a deep neural network and a deep reinforcement learning scheduling sub-model based on a reinforcement learning Actor-Critic network. The node admittance matrix corresponding to the power grid topology parameters, the active load and reactive load corresponding to the prediction results are input into the DNN power flow mapping sub-model. The end-to-end mapping of the nonlinear power flow equation of the power grid is constructed through the forward propagation of the neural network, and the optimal power flow calculation results are output. At the same time, the safety constraint boundary of the unit output is determined. The prediction results and the safety constraint boundary are input into the deep reinforcement learning scheduling sub-model to construct a Markov decision process. The prediction results and the safety constraint boundary are used as the state space, the discrete start-stop state and continuous output adjustment value of the controllable generator set are used as the action space, and the reward function that comprehensively considers fuel cost, start-stop cost and over-limit penalty is used as the optimization objective to output scheduling instructions.

4. The method for optimizing the operation of a renewable energy power system based on artificial intelligence according to claim 1, characterized in that, The scheduling instructions and real-time operating status are input into a multi-agent deep reinforcement learning model to perform coordinated control of multiple control objects on the source side, grid side, load side, and storage side, generating control instructions, including: Collect real-time power grid operation status data, and use a pre-trained convolutional neural network to denoise the real-time operation status data to extract the system state vector; A decentralized distributed control architecture is adopted, with independent control agents deployed for the source-side renewable energy generator sets, grid-connected converters, energy storage systems, and load-side flexible loads. The discrete control strategy is output by iterating the discrete control actions of energy storage charging and discharging mode switching and flexible load switching through a deep Q-network. The continuous control actions of generator power regulation, converter output control, and energy storage charging and discharging power regulation are iterated through a deep deterministic strategy gradient algorithm to output a continuous control strategy. Based on the discrete control strategy and the continuous control strategy, the control agent synchronously outputs control commands.

5. The method for optimizing the operation of a renewable energy power system based on artificial intelligence according to claim 1, characterized in that, Also includes: In the distributed power market, power generators, energy storage operators, and load aggregators are modeled as reinforcement learning bidding agents. The bidding agent is configured to operate within the price-volume curve, with its own profit maximization as the optimization objective. It iteratively optimizes its bidding strategy based on feedback from the marginal clearing results of the market environment.

6. The method for optimizing the operation of a renewable energy power system based on artificial intelligence according to claim 1, characterized in that, The control commands are fed back to the acquisition terminal of the multi-source operating data to form a closed-loop optimization framework, including: Collect system operation data after the control command is executed, execution deviation data of the scheduling command, and measured error data of the prediction results; The system operation data, the execution deviation data, and the measured error data are transmitted to the acquisition terminal of the multi-source operation data. The system operation data, execution deviation data, and measured error data are respectively input into the hybrid deep learning model, scheduling model, and multi-agent deep reinforcement learning model to update the network parameters of each model.

7. The method for optimizing the operation of a renewable energy power system based on artificial intelligence according to claim 3, characterized in that, Also includes: The node voltage magnitude, node voltage phase angle, and unit output safety constraint boundary output by the DNN power flow mapping sub-model are input into the state input layer of the deep reinforcement learning scheduling sub-model. The generator start-stop status and generator output adjustment value output by the deep reinforcement learning scheduling sub-model are input into the input tensor of the DNN power flow mapping sub-model.

8. The method for optimizing the operation of a renewable energy power system based on artificial intelligence according to claim 1, characterized in that, Also includes: Send the control command to the corresponding controllable object; Collect power grid operation status data after the control command is executed; The power grid operation status data are input to the corresponding control agents respectively; Based on the real-time operating status data, each control agent updates the network parameters of the deep Q-network and the network parameters of the deep deterministic policy gradient algorithm.

9. The method for optimizing the operation of a renewable energy power system based on artificial intelligence according to claim 1, characterized in that, The parameter updates for the hybrid deep learning model, scheduling model, and multi-agent deep reinforcement learning model include: Each data acquisition node and execution node generates encrypted gradient parameters for the corresponding model. The encryption gradient parameters are transmitted between each node; Each node updates the network parameters of the corresponding model based on the received encrypted gradient parameters.

10. An artificial intelligence-based renewable energy power system operation optimization system, characterized in that, include: The data acquisition module is used to acquire multi-source operational data of the power system; The power prediction module is used to input the multi-source operating data into the hybrid deep learning model to predict the power generation of renewable energy and output the prediction results. The hybrid deep learning model is composed of a convolutional neural network, a gated recurrent unit and a Transformer cascaded in sequence. The convolutional neural network is used to extract local fluctuation features, the gated recurrent unit is used to model long-period time-series dependencies, and the Transformer is used to perform attention weight correction on the time-series features. The joint scheduling module is used to input the prediction results and grid topology parameters into the scheduling model to perform integrated scheduling of unit combination and optimal power flow, and generate scheduling instructions; the scheduling model includes a deep neural network for solving the optimal power flow and a reinforcement learning Actor-Critic network for solving the unit combination. The collaborative control module is used to input the scheduling instructions and real-time operating status into a multi-agent deep reinforcement learning model to perform collaborative control on multiple control objects on the source side, network side, load side, and storage side, and generate control instructions. The multi-agent deep reinforcement learning model includes multiple agents deployed on different control objects, and each agent makes collaborative decisions based on a decentralized architecture. The control instructions are fed back to the acquisition terminal of the multi-source operating data to form a closed-loop optimization framework.

Images