Virtual power plant dynamic aggregation method and system based on Transform and AdaptMLP

Abstract

The invention provides a virtual power plant dynamic aggregation method and system based on Transform and AdaptMLP, and relates to the field of power system optimization scheduling, the method comprises the following steps: using a Transform encoder to extract global coupling characteristics between devices through a multi-head self-attention mechanism, using an AdaptMLP module to realize adaptive parameter adjustment through a gating mechanism and a multi-expert network, and using a virtual power plant dynamic aggregation method and a virtual power plant dynamic aggregation system based on the AdaptMLP. And a multi-agent reinforcement learning problem is solved in combination with an MAPPO algorithm. The method can adapt to equipment isomerism and dynamic change of the number of the fuel gas producers, supports immediate access of new fuel gas producers without retraining, and is suitable for resource aggregation and optimal scheduling of virtual power plants including multiple types of equipment such as energy storage equipment, electric boilers, electric refrigerators and interruptible loads.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power system optimization scheduling, and particularly relates to a virtual power plant dynamic aggregation method and system based on a Transformer and an AdaptMLP. BACKGROUND

[0002] A virtual power plant (VPP) provides auxiliary services in the electricity market by aggregating distributed energy resources. With the large-scale access of distributed photovoltaic, energy storage systems, and controllable loads, the number of producers and consumers and the dynamic changes in device configuration in the virtual power plant, new producers and consumers may access or exit the virtual power plant at any time, resulting in continuous adjustment of the system size and structure. The existing deep reinforcement learning-based methods usually use fully connected neural networks or recurrent neural networks, which have insufficient ability to process variable-length inputs (changes in the number of producers and consumers) and multi-device coupling feature extraction, and are difficult to adapt to new device access scenarios, requiring retraining. At the same time, there is a lack of adaptive network architecture for the dynamic aggregation scenario of the virtual power plant, and the existing methods have poor generalization ability when facing changes in device type, size, and operating environment. SUMMARY

[0003] The purpose of the present application is to overcome the shortcomings of the prior art and provide a virtual power plant dynamic aggregation method and system based on a Transformer and an AdaptMLP, which extracts global coupling features between devices through a Transformer encoder and realizes adaptive parameter adjustment through an AdaptMLP module, solving the efficient aggregation and scheduling problem of the virtual power plant under the condition of device heterogeneity and dynamic changes in size.

[0004] TECHNICAL SCHEME In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme: A virtual power plant dynamic aggregation method based on a Transformer and an AdaptMLP, comprising the following steps: Step 1, constructing a virtual power plant multi-agent system, including a VPP operator agent and a plurality of producer and consumer agents, each producer and consumer agent being configured with a plurality of distributed energy devices; Step 2, collecting device physical observation values of each producer and consumer to form an observation matrix , the device physical observation values including a state of charge of an energy storage, a temperature of an electric boiler, a temperature of a refrigeration machine, and a state of an interruptible load; Step 3, inputting the observation matrix into a Transformer encoder to extract coupling features between devices through a multi-head self-attention mechanism, and calculating a query matrix , a key matrix , and a value matrix The global interaction feature vector is obtained by attention weight calculation ; Step 4, constructing a context vector , including device type, capacity scale, ambient temperature and electricity price information; Step 5, inputting the Transformer output feature and the context vector into the AdaptMLP module to generate four groups of adaptive weights through a gating mechanism , and applying the weights to the outputs of the four expert networks to obtain adaptively adjusted decision features through weighted fusion; Step 6, outputting the scheduling actions of each device based on the decision features, including energy storage charging and discharging power, electric boiler power, refrigeration machine power and interruptible load reduction; Step 7, obtaining reward feedback by interacting with the environment, and updating the network parameters of the Transformer and the AdaptMLP; Step 8, repeating steps S2-S7 until the model converges.

[0005] As a further scheme of the application, the construction method of the device physical observation value in step S2 is: For the first producer and consumer containing energy storage system, electric boiler, electric refrigerator and interruptible load at time , the observation matrix includes: Energy storage system: state of charge ; Electric boiler: water tank temperature , normalized to ; Electric refrigerator: cold water temperature , normalized to ; Interruptible load: current state ; The dimension of the observation matrix is , where is the number of devices, is the feature dimension of each device.

[0006] As a further scheme of the application, the multi-head self-attention mechanism of the Transformer encoder in step S3 is specifically: Wherein: Wherein: is the key vector dimension, is the number of attention heads, , , , is the learnable weight matrix.

[0007] As a further scheme of the present application, the construction of the context vector in step S4 includes: Device type vector: represented by one-hot encoding, including energy storage, electric boiler, electric chiller, interruptible load; Device capacity: normalized rated power or capacity; Environmental information: outdoor temperature, solar radiation intensity; Market information: real-time electricity price, ancillary service price; The above information is spliced to form a context vector with a dimension of .

[0008] As a further scheme of the present application, the adaptive adjustment process of the AdaptMLP module in step S5 is specifically: The feature transformation formula is: The gating weight calculation formula is: The adaptive output formula is: Wherein: , is the weight and bias of the th expert network, is the learnable parameter of the gating network, is the context vector.

[0009] As a further scheme of the present application, the output constraint of the device scheduling action in step S6 is: Energy storage charging and discharging power: ; Electric boiler power: ; Electric chiller power: ; Interruptible load reduction: ; The action is mapped to the corresponding constraint range by the sigmoid activation function.

[0010] As a further scheme of the present application, the network parameter update in step S7 adopts the asynchronous advantage actor-critic algorithm MAPPO, which includes: Policy network: output probability distribution of device scheduling action; Value network: estimate the value function of current state-action pair; Advantage function: . Policy gradient update: maximize the truncated target function.

[0011] As a further scheme of the present application, the method further comprises a dynamic expansion step: When a new producer or consumer accesses the virtual power plant: The Transformer encoder automatically adapts to changes in input sequence length through a position encoding mechanism; The AdaptMLP module adjusts the context vector of the new producer or consumer Dynamically adjust the weights of the expert network; Without retraining the overall model, only fine-tune the gating network parameters.

[0012] Further, the present application also provides a virtual power plant dynamic aggregation system based on Transformer and AdaptMLP, comprising: Data acquisition module: acquire real-time operating state and physical parameters of producer or consumer devices; Feature extraction module: use Transformer encoder to extract coupling features between devices, including multi-head self-attention calculation unit, residual connection unit and layer normalization unit; Adaptive adjustment module: use AdaptMLP structure, including 4 parallel expert networks, gating mechanism unit and feature fusion unit; Decision output module: generate device scheduling action according to fused features; Parameter update module: update network parameters based on environmental feedback.

[0013] As a further scheme of the present application, the feature extraction module uses a multi-layer Transformer encoder stacking structure, each layer including a multi-head self-attention sublayer, a feedforward neural network sublayer, and a residual connection.

[0014] Advantages Compared with the prior art, the present application has the following advantages: 1) It can calculate the global dependence between devices in parallel, effectively capturing the complex coupling characteristics of electrical-thermal-cold multi-energy flow; 2) It can dynamically adjust network parameters according to device type, capacity, and environmental changes, without retraining to adapt to new producer or consumer access and device configuration changes, with strong generalization ability; 3) It can be applied to virtual power plants containing energy storage, electric boilers, electric chillers, interruptible loads and other types of devices, meeting the real-time scheduling needs of auxiliary service markets. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a whole flow chart of the method of the present application.

[0016] Figure 2 is a service bidding result chart for assisting services of different numbers of producers and consumers in an embodiment of the present application. DETAILED DESCRIPTION

[0017] The present application will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0018] The following detailed description is exemplary and is intended to provide further detailed description of the present application. Unless otherwise specified, all technical terms used in the present application have the same meanings as those generally understood by those skilled in the art. The terms used in the present application are only for the purpose of describing the specific embodiments and are not intended to limit the exemplary embodiments according to the present application.

[0019] As a preferred embodiment of the present application, a virtual power plant dynamic aggregation method based on Transformer and AdaptMLP includes the following steps: Step 1, constructing a virtual power plant multi-agent system, including a VPP operator agent and a plurality of producer and consumer agents, each of which is configured with a plurality of distributed energy equipment; Step 2, collecting the physical observation values of each producer and consumer to form an observation matrix , the physical observation values including the state of charge of energy storage, the temperature of electric boiler, the temperature of refrigeration machine and the state of interruptible load; Step 3, inputting the observation matrix into the Transformer encoder, extracting the coupling features between the equipment through the multi-head self-attention mechanism, calculating the query matrix , the key matrix and the value matrix , and obtaining the global interaction feature vector through attention weight calculation; Step 4, constructing a context vector containing equipment type, capacity scale, ambient temperature and electricity price information; Step 5, inputting the Transformer output feature and the context vector into the AdaptMLP module, generating four groups of adaptive weights through the gating mechanism, applying the weights to the outputs of the four expert networks, and obtaining the adaptively adjusted decision features through weighted fusion; Step 6, output the scheduling action of each device based on the decision feature, including energy storage charging and discharging power, electric boiler power, refrigeration machine power and interruptible load reduction; Step 7, interact with the environment to obtain reward feedback, and update the network parameters of the Transformer and AdaptMLP; Step 8, repeat steps S2-S7 until the model converges.

[0020] The construction of the device physical observation value in step S2 is as follows: For the first producer and consumer with energy storage system, electric boiler, electric refrigeration machine and interruptible load at time The observation matrix includes: Energy storage system: state of charge ; Electric boiler: water tank temperature , normalized to ; Electric refrigeration machine: cold water temperature , normalized to ; Interruptible load: current state ; The dimension of the observation matrix is , where is the number of devices, is the feature dimension of each device.

[0021] The multi-head self-attention mechanism of the Transformer encoder in step S3 is as follows: Where: Where: is the key vector dimension, is the number of attention heads, , , , is the learnable weight matrix.

[0022] The construction of the context vector in step S4 includes: Device type vector: one-hot encoding to represent energy storage, electric boiler, refrigeration machine and interruptible load; Device capacity: rated power or capacity after normalization processing; Environmental information: outdoor temperature, solar radiation intensity; Market information: real-time electricity price, ancillary service price The above information is spliced to form a context vector with a dimension of .

[0023] The adaptive adjustment process of the AdaptMLP module in step S5 is specifically: The feature transformation formula is: The gating weight calculation formula is: The adaptive output formula is: Wherein: , is the weight and bias of the th expert network, is the learnable parameter of the gating network, is the context vector.

[0024] The output constraint of the device scheduling action in step S6 is: Energy storage charging and discharging power: ; Electric boiler power: ; Electric refrigerator power: ; Interruptible load reduction: ; The action is mapped to the corresponding constraint range through the sigmoid activation function.

[0025] The network parameter update in step S7 adopts the asynchronous advantage actor-critic algorithm MAPPO, including: Policy network: output probability distribution of device scheduling action Evaluation network: estimate the value function of the current state-action pair Advantage function: ; Policy gradient update: maximize the truncated objective function.

[0026] The method further includes a dynamic expansion step: When a new producer or consumer accesses the virtual power plant: The Transformer encoder automatically adapts to changes in input sequence length through the position encoding mechanism The AdaptMLP module dynamically adjusts the weights of the expert networks according to the context vector of the new producer or consumer Without retraining the whole model, only fine-tune the gating network parameters.

[0027] Further, the application also provides a virtual power plant dynamic aggregation system based on the Transformer and AdaptMLP, comprising: a data acquisition module: acquiring real-time operation states and physical parameters of producer-consumer equipment; a feature extraction module: extracting coupling features between equipment by using a Transformer encoder, including a multi-head self-attention calculation unit, a residual connection unit, and a layer normalization unit; an adaptive adjustment module: using an AdaptMLP structure, including four parallel expert networks, a gating mechanism unit, and a feature fusion unit; a decision output module: generating equipment scheduling actions according to fused features; a parameter update module: updating network parameters based on environmental feedback.

[0028] The feature extraction module uses a multi-layer Transformer encoder stacking structure, each layer including a multi-head self-attention sublayer, a feedforward neural network sublayer, and a residual connection.

[0029] The application is further illustrated by an embodiment.

[0030] A plurality of producers and consumers are selected as dynamic aggregation objects of the VPP for example analysis, assuming that there are various flexible resources and multi-energy loads in the producers and consumers. The VPP needs to adapt to the characteristics of scale expansion and heterogeneous equipment update of the producers and consumers in actual operation. Table 1 shows the dynamic changes of the number of producers and consumers and the equipment configuration under different time periods.

[0031] Figure 2 shows the bidding results of the VPP participating in the grid peak shaving auxiliary service market under different producer-consumer aggregation scales, in which three typical days correspond to dynamic aggregation scenarios of 5, 10, and 20 producers and consumers. Positive values represent peak shaving auxiliary services, and negative values represent valley filling auxiliary services. By comparing the bidding results of the three typical days, it can be found that although the number of producers and consumers and the equipment configuration change significantly, the VPP can always meet the minimum bidding requirement of the grid ≥1000kWh.

[0032] From the time series distribution of bidding capacity, the bidding strategy of VPP presents obvious peak-valley characteristics. In the valley filling period (1:00-8:00 and 20:00-24:00), VPP increases electricity consumption or reduces power generation by coordinating producers and consumers to respond to the valley filling demand of the grid. In the peak shaving period (8:00-20:00), VPP organizes producers and consumers to reduce electricity consumption or increase power generation to provide peak shaving services for the grid. The number of producers and consumers increases from 5 to 20, and the bidding capacity of VPP presents a significant growth trend. This is because the complementary effect between heterogeneous resources is enhanced, such as the CHP unit added on Day 2 and the absorption chiller, which form a good synergy. In addition, under the P2P transaction mechanism, the energy mutual aid effect between producers and consumers makes the overall peak shaving capacity exceed the simple addition of individual capacity. When the aggregation scale expands, the uncertainty of individual producers and consumers is effectively dispersed, making VPP able to provide more stable auxiliary services.

Claims

1. A method for dynamic aggregation of virtual power plants based on Transformer and AdaptMLP, characterized in that, Includes the following steps: S1: Construct a virtual power plant multi-agent system, including a VPP operator agent and multiple prosumer agents, with each prosumer agent configured with a variety of distributed energy devices; S2: Collect physical observations of the equipment of each producer and consumer to form an observation matrix. The physical observations of the equipment include the energy storage state of charge, the electric boiler temperature, the chiller temperature, and the interruptible load status; S3: Observation matrix Inputting a Transformer encoder, the coupling features between devices are extracted using a multi-head self-attention mechanism: Calculate the query matrix Key matrix Sum matrix ; The global interaction feature vector is obtained by calculating attention weights. ; S4: Constructing Context Vectors This includes information on equipment type, capacity, ambient temperature, and electricity price; S5: Transformer output features With context vector Input the AdaptMLP module: N sets of adaptive weights are generated through a gating mechanism. ; Apply weights to the outputs of N expert networks; The adaptively adjusted decision features are obtained through weighted fusion. S6: Output the scheduling actions of each device based on decision characteristics, including energy storage charging and discharging power, electric boiler power, chiller power, and interruptible load reduction amount; S7: Interact with the environment to obtain reward feedback and update the network parameters of Transformer and AdaptMLP; S8: Repeat steps S2-S7 until the model converges.

2. The method for dynamic aggregation of virtual power plants based on Transformer and AdaptMLP according to claim 1, characterized in that, The method for constructing the physical observations of the equipment in step S2 is as follows: For the first type of load including energy storage systems, electric boilers, electric chillers, and interruptible loads Individual consumers at any time observation matrix include: Energy storage system: State of charge ; Electric boiler: water tank temperature Normalization ; Electric chiller: chilled water temperature Normalization ; Interruptible load: Current state ; The dimension of the observation matrix is ,in For the number of devices, The feature dimensions for each device.

3. The method for dynamic aggregation of virtual power plants based on Transformer and AdaptMLP according to claim 1, characterized in that, The multi-head self-attention mechanism of the Transformer encoder in step S3 is as follows: in: in: The dimension of the key vector. For the number of attention heads, This is a learnable weight matrix.

4. The method for dynamic aggregation of virtual power plants based on Transformer and AdaptMLP according to claim 1, characterized in that, Context vector in step S4 The construction includes: Equipment type vector: One-hot encoding is used to represent energy storage, electric boilers, chillers, and interruptible loads; Equipment capacity: Rated power or capacity after normalization; Environmental information: outdoor temperature, solar radiation intensity; Market information: real-time electricity prices, ancillary service prices; The above information is pieced together to form the dimension. The context vector.

5. The method for dynamic aggregation of virtual power plants based on Transformer and AdaptMLP according to claim 1, characterized in that, The adaptive adjustment process of the AdaptMLP module in step S5 is as follows: The feature transformation formula is: The formula for calculating the gating weight is: The adaptive output formula is: in: For the first The weights and biases of an expert network, These are the learnable parameters of the gated network. This is the context vector.

6. The method for dynamic aggregation of virtual power plants based on Transformer and AdaptMLP according to claim 1, characterized in that, The output constraint for the equipment scheduling action in step S6 is: Energy storage charging and discharging power: ; Electric boiler power: ; Electric chiller power: ; Interruptible load reduction: ; The action is mapped to the corresponding constraint range through the sigmoid activation function.

7. The method for dynamic aggregation of virtual power plants based on Transformer and AdaptMLP according to claim 1, characterized in that, The network parameter update in step S7 uses the Asynchronous Advantageous Actor-Critic Algorithm (MAPPO), which includes: Policy network: Outputs the probability distribution of device scheduling actions; Network evaluation: Estimating the value function of the current state-action pair; Advantage function: ; Policy gradient update: Maximize the truncated objective function.

8. The method for dynamic aggregation of virtual power plants based on Transformer and AdaptMLP according to claim 1, characterized in that, The method also includes a dynamic expansion step: When new consumers connect to the virtual power plant: The Transformer encoder automatically adapts to changes in the length of the input sequence through a positional encoding mechanism; The AdaptMLP module uses the context vector of the new pro-consumer. Dynamically adjust the weights of the expert network; There is no need to retrain the entire model; only the gating network parameters need to be fine-tuned.

9. A system for implementing the virtual power plant dynamic aggregation method based on Transformer and AdaptMLP as described in any one of claims 1-8, characterized in that, include: Data acquisition module: Collects real-time operating status and physical parameters of producer-consumer equipment; Feature extraction module: Employs a Transformer encoder to extract coupling features between devices, including a multi-head self-attention calculation unit, a residual connection unit, and a layer normalization unit; Adaptive adjustment module: Adopts the AdaptMLP structure, including 4 parallel expert networks, a gating mechanism unit and a feature fusion unit; Decision output module: Generates device scheduling actions based on fusion features; Parameter update module: Updates network parameters based on environmental feedback.

10. The system according to claim 9, characterized in that, The feature extraction module adopts a multi-layer Transformer encoder stacked structure, with each layer including a multi-head self-attention sub-layer, a feedforward neural network sub-layer, and residual connections.

Images