Virtual power plant intelligent scheduling and optimizing system and method based on multi-source data fusion

Abstract

The invention relates to a virtual power plant intelligent scheduling and optimizing system and method based on multi-source data fusion, and belongs to the field of power grid technology and artificial intelligence application. The system comprises a memory and a processor, multi-source data such as power grid market price, renewable energy prediction, load demand and energy storage state are acquired in real time, and a pre-trained deep reinforcement learning model is utilized to generate a scheduling instruction. The model takes maximization of a multi-target reward function as a guide, covers income, resource utilization rate and carbon emission targets, and optimizes each target weight in real time through a dynamic weight adjustment mechanism to balance conflicts (for example, the income weight is improved when the income is insufficient, and the carbon target weight is enhanced when the carbon emission exceeds the standard). The core technology adopts a space-time fusion network architecture (CNN extracts spatial features, and LSTM captures time sequence dependence), lightweight deployment is realized in combination with a knowledge distillation technology, and decision delay is less than 200 milliseconds. The method comprises the steps of data acquisition, DRL model decision making, reward calculation, weight dynamic adjustment and instruction issuing.

Description

Technical Field

[0001] This invention relates to the fields of power grid technology and artificial intelligence applications, and in particular to a virtual power plant intelligent dispatching and optimization system and method based on multi-source data fusion. Background Technology

[0002] As the global energy structure transitions towards low-carbon and clean energy, the penetration rate of renewable energy sources, such as photovoltaics and wind power, in the power grid is increasing. These distributed energy resources (DERs) are characterized by intermittency, volatility, and uncertainty, posing significant challenges to the stable operation of the power grid and power balance. Virtual power plants (VPPs), as an advanced energy management technology, have emerged to address this challenge. Through advanced communication and control technologies, they aggregate geographically dispersed distributed generator units, controllable loads, energy storage systems, and other DERs, allowing them to participate as a special power plant entity in electricity market transactions and grid ancillary services, thereby achieving efficient utilization and unified management of distributed energy resources.

[0003] Traditional virtual power plant scheduling and optimization methods mainly rely on mathematical programming models, such as mixed-integer linear programming (MILP). While these methods are theoretically sound, they have several limitations in practical applications:

[0004] High computational complexity and low decision-making efficiency: As the number of DERs aggregated by VPP increases (e.g., reaching thousands or even tens of thousands of nodes), the variables and constraints of the mathematical model grow exponentially, and the solution time increases dramatically, often requiring tens of minutes or even longer, making it difficult to meet the needs of the electricity market (such as a real-time market at the 15-minute or 5-minute level) for rapid decision-making.

[0005] Limitations of Single-Objective Optimization: Traditional methods mostly take the lowest operating cost or highest return as the sole optimization objective. However, the operation of modern VPPs needs to consider multiple dimensions, including economic benefits, grid friendliness (such as reducing wind and solar curtailment), asset health (such as energy storage lifespan), and environmental responsibility (such as reducing carbon emissions). Single-objective optimization cannot achieve an effective balance among these conflicting objectives.

[0006] Poor adaptability to uncertainty: Planning-based methods typically rely on accurate forecasts of future electricity prices, load, and renewable energy output. When market conditions fluctuate drastically or forecasts deviate significantly, the robustness of the optimization results and actual returns are greatly reduced.

[0007] In recent years, although some research has attempted to apply machine learning or heuristic algorithms to VPP scheduling, these methods or models are relatively simple or rely heavily on human experience and rule adjustments, lacking the ability to learn autonomously and adapt dynamically in complex and ever-changing market environments, resulting in limited performance improvements. Therefore, there is an urgent need for a novel technical solution that can overcome the bottlenecks of existing technologies and achieve fast, intelligent, and multi-objective collaborative optimization scheduling for large-scale VPPs. Summary of the Invention

[0008] In view of the above situation, it is necessary to provide a virtual power plant intelligent scheduling and optimization system based on multi-source data fusion to solve at least one of the above problems, characterized in that it includes:

[0009] Memory, used to store instructions;

[0010] and a processor coupled to the memory, the processor being configured to execute the instructions to: acquire multi-source data in real time, the multi-source data including at least: grid market prices, renewable energy output forecasts, load demand forecasts, and energy storage system status;

[0011] Based on the multi-source data, a pre-trained deep reinforcement learning model is used to generate scheduling instructions; the deep reinforcement learning model is trained with the goal of maximizing a multi-objective reward function that includes multiple optimization objective functions, the optimization objective functions including at least: revenue objective, resource utilization objective and carbon emission objective;

[0012] Based on a preset dynamic weight adjustment mechanism, the weights of the multiple optimization objective functions are adjusted in real time during the scheduling process.

[0013] The dispatch command is issued to one or more distributed energy sources in the virtual power plant.

[0014] Preferably, the deep reinforcement learning model includes a spatiotemporal fusion network architecture, which comprises:

[0015] A convolutional neural network module is used to extract spatial features from the multi-source data;

[0016] And a long short-term memory network module for extracting time-series dependencies from the multi-source data.

[0017] Preferably, the processor is configured to execute the dynamic weight adjustment mechanism, further comprising: during the scheduling decision process, detecting whether there is a conflict among the multiple optimization objectives through Pareto front analysis;

[0018] When a conflict is detected, the weights of each optimization objective are dynamically adjusted according to preset rules to balance the overall scheduling performance.

[0019] Preferably, the preset rules include:

[0020] When the real-time revenue is lower than the predetermined average revenue threshold, the weight of the revenue target is increased;

[0021] When the resource utilization rate exceeds a predetermined upper limit threshold, the weight of the resource utilization rate target is reduced.

[0022] And when the carbon emission intensity of a unit transaction is higher than a predetermined carbon emission benchmark, the weight of the carbon emission target is increased.

[0023] Preferably, the deep reinforcement learning model is a lightweight model generated by knowledge distillation; the knowledge distillation technique uses a function that includes Kullback-Leibler divergence loss and mean squared error loss.

[0024] In this invention, another solution is also disclosed: a computer implementation method for intelligent scheduling and optimization of virtual power plants based on multi-source data fusion, comprising the following steps:

[0025] Acquire multi-source input data, which includes at least real-time electricity prices, distributed energy status, and load forecasts;

[0026] The multi-source input data is fed into a deep reinforcement learning model to generate scheduling actions for multiple optimization objectives;

[0027] The reward signal for the current scheduling action is calculated based on a reward function that is weighted by a revenue item, a resource utilization item, and a carbon emission item.

[0028] In response to the reward signal and changes in system state, the weights of the corresponding terms in the reward function are dynamically adjusted;

[0029] Based on the scheduling action, specific scheduling instructions for resources within the virtual power plant are generated and output.

[0030] Preferably, the step of dynamically adjusting the weights further includes:

[0031] When the real-time revenue is lower than a first predetermined threshold, the weight of the revenue item is increased;

[0032] When the resource utilization rate exceeds the second predetermined threshold, the weight of the resource utilization rate item is reduced.

[0033] And when the carbon emission intensity exceeds a third predetermined threshold, the weight of the carbon emission item is increased.

[0034] Preferably, the method further includes the step of training the deep reinforcement learning model:

[0035] The model policy network is initialized by offline pre-training using historical market and weather datasets; and then online learning is performed at a preset time frequency during actual system operation.

[0036] Preferably, in the step of generating scheduling actions, the deep reinforcement learning model processes the input data in the following manner:

[0037] Using convolutional neural networks to extract spatial correlation features among distributed energy resources in different geographical locations;

[0038] And utilize Long Short-Term Memory (LSTM) networks to capture the dynamic characteristics of data such as electricity prices, load, and weather over time.

[0039] Compared with the prior art, the system and method proposed in this invention have one or more of the following significant advantages:

[0040] Breakthrough improvement in decision-making efficiency: By adopting a lightweight deep reinforcement learning model, this invention reduces the decision-making time for complex scheduling problems involving more than 10,000 distributed energy nodes from tens of minutes or more in traditional methods to less than 200 milliseconds, fully meeting the stringent requirements of the real-time electricity market for high-frequency, low-latency decision-making.

[0041] Significantly enhanced economic benefits: Through multi-objective collaborative optimization and precise response to market price fluctuations, this invention can maximize the operating revenue of the VPP while ensuring grid security. In backtesting with real market data, compared with the traditional MILP method, the solution of this invention can increase cumulative revenue by more than 23%.

[0042] Significantly optimized resource utilization: By finely managing the charging and discharging of energy storage and coordinating the output of various DERs, this invention can increase the average utilization rate of distributed energy from about 68% to over 89%, significantly reducing wind and solar curtailment caused by grid congestion or inaccurate forecasts.

[0043] Multi-objective synergy and dynamic adaptation: The innovative dynamic weight adjustment mechanism enables the system to intelligently balance multiple objectives such as revenue, utilization rate, and carbon emissions. Even in extreme scenarios where electricity price fluctuations reach ±50%, the system can still control revenue fluctuations within ±3%, demonstrating strong robustness and market adaptability.

[0044] Significant environmental benefits: This invention takes carbon emissions as a core optimization objective and guides the scheduling behavior of VPPs towards cleaner and lower-carbon practices through a carbon penalty mechanism. After implementing this invention, approximately 12.6 kg of carbon dioxide emissions can be reduced per megawatt-hour (MWh) of trading. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the overall architecture of a virtual power plant intelligent scheduling and optimization system according to an embodiment of the present invention.

[0046] Figure 2 This is an overall flowchart of a virtual power plant intelligent scheduling and optimization method according to an embodiment of the present invention.

[0047] Figure 3 This is a detailed flowchart of the training and decision-making of a deep reinforcement learning model according to an embodiment of the present invention, which focuses on the mechanism of target conflict detection and dynamic weight adjustment.

[0048] Figure 4 This is a schematic diagram of the spatiotemporal fusion network (CNN-LSTM) used in an embodiment of the present invention.

[0049] Figure 5 This is a schematic diagram illustrating the principle of knowledge distillation technology employed according to an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0051] In the description of this invention, unless otherwise stated, "a plurality of" means two or more; the terms "center," "longitudinal," "lateral," "upper," "lower," "left," "right," "inner," "outer," "front end," "rear end," "head," "tail," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0052] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0054] Example 1: System Architecture

[0055] Please see Figure 1 This demonstrates the overall architecture of the virtual power plant intelligent scheduling and optimization system according to an embodiment of the present invention. The system 100 can be deployed on a cloud server or an edge computing gateway, and mainly includes three core layers: a data interaction layer 110, an intelligent decision-making layer 120, and a scheduling execution layer 130.

[0056] The data interaction layer 110 is responsible for exchanging data with the outside world. It contains multiple data interface modules:

[0057] Electricity Market Interface Module 111: Connects with the platforms of electricity market operators (such as CAISO and PJM in the United States, or EPEX SPOT in Europe) through application programming interfaces (APIs) to obtain real-time electricity price information, bidding information, etc. in the day-ahead market, real-time market, and ancillary services market.

[0058] DER Data Interface Module 112: Connects to the local controllers of various distributed energy sources aggregated by VPP via Internet of Things (IoT) protocols (such as MQTT) or industrial bus protocols (such as Modbus TCP). It is responsible for collecting data such as SOC, SOH (health status), and charging / discharging power of the energy storage system; real-time power generation and inverter status of the photovoltaic array; and operating status and adjustable potential of controllable loads (such as electric vehicle charging piles and smart air conditioners).

[0059] Weather forecasting interface module 113: Connects to third-party commercial weather services or public weather data sources to obtain high-resolution future weather forecasts, including light intensity, wind speed, temperature, etc., providing input for renewable energy output forecasting.

[0060] These raw data from different sources, after being cleaned, aligned, and preliminarily processed by the data interaction layer 110, are uniformly sent to the state-aware module 114. The state-aware module 114 performs deep fusion and feature engineering on the multi-source data. For example, it combines weather forecasts and historical power generation data from photovoltaic power plants, using physical models or machine learning models to generate power prediction curves with confidence intervals for the next 24 hours or even longer. Finally, it integrates all key information into a standardized, high-dimensional "state vector S". t This serves as the basis for intelligent decision-making.

[0061] The intelligent decision-making layer 120 is the "brain" of the entire system, and its core is the DRL agent 121. This agent runs on the processor and specifically implements the core algorithm of this invention.

[0062] The DRL agent 121 internally encapsulates a pre-trained deep reinforcement learning model. In this embodiment, the model employs the Multi-Objective Proximal Policy Optimization (MO-PPO) algorithm. As an advanced DRL algorithm, PPO ensures training stability and sample utilization efficiency by introducing importance sampling and pruning of the objective function. The "multi-objective" extension enables it to handle our defined complex reward functions.

[0063] The multi-objective decision engine 122 is a crucial component of the DRL agent 121, responsible for implementing the dynamic weight adjustment mechanism described above. At each decision step, it analyzes the performance of multiple objectives, including current revenue, utilization rate, and carbon emissions, and then... Figure 3 The logic shown determines whether there is an optimization conflict, and then dynamically adjusts the multi-objective reward function R. t The weights w of each item rev w util w carb .

[0064] Reward function R t The specific form can be defined as:

[0065]

[0066] in:

[0067] It is the profit at the current moment. It is a normalized maximum return reference value.

[0068] U t It is the resource utilization rate at the current moment, U max This is the maximum utilization rate (usually 100%).

[0069] C t It is the carbon emission intensity at the current moment, C base It is a carbon emission benchmark.

[0070] w rev ,w util ,w carb The weights are dynamically adjusted by the multi-objective decision engine 122.

[0071] When DRL agent 121 receives the state vector S t Then, its internal policy network will output a "scheduling action vector a".t Each element of this vector corresponds to a scheduling setting value for a controllable resource, such as the charging and discharging power of an energy storage system, or the active / reactive power adjustment commands of a photovoltaic inverter.

[0072] The scheduling execution layer 130 is responsible for transforming the abstract instructions generated by the intelligent decision-making layer 120 into actual operations in the physical world.

[0073] The scheduling instruction parsing module 131 receives the action vector a. t It then translates these into control commands that conform to the specific device's communication protocol.

[0074] Subsequently, these instructions are precisely distributed to various execution units via the data interaction layer 110, including: energy storage system 132, photovoltaic inverter 133, and controllable load 134. For example, an instruction might be "set the charging power of the energy storage system with ID BESS_001 to 50kW".

[0075] Example 2: Method Flow and Core Algorithm

[0076] Please combine Figure 2 and Figure 3 To understand the method flow of the present invention. Figure 2 It demonstrates the overall steps of the method, while Figure 3 It provides an in-depth look at the decision-making and learning cycle within the DRL intelligent body.

[0077] Overall process (reference) Figure 2 )

[0078] Step S201: Initialization. When the system starts, it loads the pre-trained DRL model parameters and initializes the experience replay pool. The experience replay pool is used to store historical sequences (states, actions, rewards, next states) for model training.

[0079] Step S202: State Acquisition and Perception. The system periodically (e.g., every 5 minutes) collects multi-source data through the data interaction layer 110, and the state perception module 114 fuses this data into the current state vector S. t .

[0080] Step S203: Intelligent Decision Making. The state vector S... t The input is fed into the policy network of DRL agent 121. The policy network performs forward propagation calculations and outputs an optimal scheduling action vector a. t .

[0081] Step S204: Instruction Issuance and Execution. The scheduling instruction parsing module 131 will execute instruction a. t The instructions are translated into specific commands and sent to the various DERs in execution layer 130. The DERs execute the commands, and the overall state of the VPP changes.

[0082] Step S205: Environmental Feedback and Reward Calculation. At the start of the next time step, the system observes the execution of action a. t The new state S that is generated later t+1 And calculate the actual reward R obtained at time t. t This reward is calculated based on the actual market transaction price, the actual output of DERs, and carbon emissions, according to the aforementioned multi-objective reward function formula.

[0083] Step S206: Online Model Learning and Update. The system will update the new experience tuples (S... t ,a t ,R t ,S t+1 The data is stored in the experience replay pool. Simultaneously, the system samples a batch of historical data from the replay pool to iteratively update the network parameters of the DRL model. This online learning step ensures the model can continuously adapt to environmental changes. Details of this step are provided in [link to article / document / etc.]. Figure 3 The details are as follows. After that, the process returns to step S202, forming a continuously running closed loop.

[0084] Detailed process of DRL model training and decision-making (reference) Figure 3 )

[0085] This is a detailed explanation of step S206 above, and it is also one of the core innovations of this invention.

[0086] Step S301: Initialize the experience replay pool.

[0087] Step S302: Sample batch data. Randomly select a batch of data from the experience replay pool.

[0088] Step S303: Calculate the multi-objective reward R t For each sample in the sampled data, calculate its multi-objective reward.

[0089] Step S304: Forward propagation of the policy network.

[0090] Step S305: Calculate the policy gradient. Based on the PPO algorithm, calculate the gradient used to update the policy network. This gradient indicates the direction in which the parameters should be adjusted to produce actions that yield higher rewards.

[0091] Step S306: Target Conflict Detection. This is a key innovative step. Before applying gradient updates, the system invokes the multi-objective decision engine 122 to analyze the reward distribution and Pareto frontier of the current batch of data. If it is found that the gradient calculated to improve one objective (such as profit) would severely harm another objective (such as causing a surge in carbon emissions), it is determined that "a conflict exists".

[0092] Decision branches:

[0093] If a conflict exists:

[0094] Step S307: Perform gradient pruning. To prevent the gradient of a certain objective from becoming too large and dominating the entire update process, the calculated policy gradient is pruned to limit its norm.

[0095] Step S308: Dynamic weight adjustment. According to the preset rules described in claim 4, adjust the weights of each item in the multi-objective reward function. For example, if a loss of profit is detected, increase w. rev This adjustment will affect the reward and gradient calculations in the next round, thus guiding the model's learning direction.

[0096] If no conflict exists (No):

[0097] Step S309: Normal policy gradient update. Update directly using the calculated gradient.

[0098] Step S310: Update the policy network parameters θ. Based on the selected gradient update policy, update the policy network parameters θ using an optimizer (such as Adam).

[0099] Step S311: Target network soft update. To ensure training stability, there is usually a target network with the same structure but slower parameter updates. This step performs a soft update on its parameters. Then, the process loops back to step S302.

[0100] Example 3: Detailed Description of Core Technology Modules

[0101] 1. Spatiotemporal Fusion Network Architecture (Reference) Figure 4 )

[0102] Please see Figure 4 This demonstrates the optimal structure of the policy network within the DRL agent. The input state vector S... t It contains multiple data channels, some of which are geographically related (such as the output of multiple photovoltaic power plants), and others are purely time-series data (such as market electricity prices).

[0103] CNN Module 401: For geolocation-related multidimensional data, it is first processed through one or more one-dimensional convolutional layers (Conv1D). The convolutional kernels can effectively capture the local correlations between adjacent DERs, for example, the passing of a cloud causing a simultaneous decrease in photovoltaic output in a region. This extracts "spatial features".

[0104] LSTM Module 402: The spatial features extracted by the CNN are concatenated with other pure time series data (such as electricity prices) and then fed into the LSTM module. The gating mechanism inside the LSTM unit (input gate, forget gate, output gate) enables it to effectively learn and remember long-term dependencies in time series data, such as the intraday periodicity of electricity prices and the pattern of day-to-day / real-time price difference.

[0105] Fully Connected Layer (FC) 403: The output of the LSTM is ultimately mapped to the final action space through several fully connected layers, outputting the scheduling action vector a. t .

[0106] This hybrid architecture of CNN+LSTM enables the model to understand both "where" (space) and "when" (time) events, thereby making more accurate and predictive decisions.

[0107] 2. Knowledge distillation technology (reference) Figure 5 )

[0108] Please see Figure 5 It describes the process of knowledge distillation, which is key to ensuring that the system can make decisions at the millisecond level.

[0109] Teacher Model 501: This is a large, complex model, such as a spatiotemporal fusion network with more layers and more neurons. It is trained offline using massive amounts of historical data in the cloud with powerful computing resources (such as GPU clusters) to achieve the highest possible decision accuracy.

[0110] Student Model 502: This is a lightweight network with a simple structure and a very small number of parameters. Its network layer count and width are much smaller than those of the teacher model, resulting in extremely fast computation speed, making it suitable for deployment on resource-constrained edge computing devices.

[0111] Knowledge transfer process 503: The goal of training the student model is not only to learn rewards in the real environment, but more importantly, to imitate the "behavior" and "thinking style" of the teacher model. This is achieved through a special loss function L. distill Achieved:

[0112] L distill =α·L KL (y′ teacher ,y′student )+β·L MSE (f teacher ,f student )

[0113] The first term is the KL divergence loss, where y′ teacher and y′ student These are the logits of the output layers (before softmax) of the teacher and student models, respectively. Minimizing the KL divergence means making the output probability distribution of the student model as consistent as possible with that of the teacher model, i.e., "behavioral imitation".

[0114] The second term is the mean squared error loss, where f teacher and f student This is a feature map representing a certain intermediate layer in the teacher and student models. Minimizing this loss means forcing the student model to learn how the teacher model extracts and organizes information from the raw data, i.e., "mental imitation".

[0115] α and β are hyperparameters that balance the two losses.

[0116] Through this process, the "wisdom" of the teacher model is "distilled" and compressed into the student model. Experiments have shown that the number of parameters in the student model can be reduced from over 100 million to approximately 1.2 million, while the inference speed increases by more than 1,000 times and the loss of decision accuracy is minimal, thus perfectly solving the efficiency bottleneck of real-time decision-making.

[0117] In summary, this invention constructs a complete and high-performance intelligent dispatching and optimization solution for virtual power plants through systematic architecture design, innovative multi-objective DRL algorithm, unique dynamic weight adjustment mechanism, and efficient knowledge distillation technology. It is not only theoretically advanced but also achieves generational advantages over existing technologies in key performance indicators, providing a practical technical path for grid dispatching and energy management under future high-proportion renewable energy integration.

[0118] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A virtual power plant intelligent scheduling and optimization system based on multi-source data fusion, characterized in that, include: Memory, used to store instructions; and a processor coupled to the memory, the processor being configured to execute the instructions to: acquire multi-source data in real time, the multi-source data including at least: grid market prices, renewable energy output forecasts, load demand forecasts, and energy storage system status; Based on the multi-source data, a pre-trained deep reinforcement learning model is used to generate scheduling instructions; the deep reinforcement learning model is trained with the goal of maximizing a multi-objective reward function that includes multiple optimization objective functions, the optimization objective functions including at least: revenue objective, resource utilization objective and carbon emission objective; Based on a preset dynamic weight adjustment mechanism, the weights of the multiple optimization objective functions are adjusted in real time during the scheduling process. The dispatch command is issued to one or more distributed energy sources in the virtual power plant.

2. The system according to claim 1, characterized in that, The deep reinforcement learning model includes a spatiotemporal fusion network architecture, which comprises: A convolutional neural network module is used to extract spatial features from the multi-source data; And a long short-term memory network module for extracting time-series dependencies from the multi-source data.

3. The system according to claim 1, characterized in that, The processor is configured to execute the dynamic weight adjustment mechanism, further comprising: during the scheduling decision process, detecting whether there is a conflict among the multiple optimization objectives through Pareto front analysis; When a conflict is detected, the weights of each optimization objective are dynamically adjusted according to preset rules to balance the overall scheduling performance.

4. The system according to claim 3, characterized in that, The preset rules include: When the real-time revenue is lower than the predetermined average revenue threshold, the weight of the revenue target is increased; When the resource utilization rate exceeds a predetermined upper limit threshold, the weight of the resource utilization rate target is reduced. And when the carbon emission intensity of a unit transaction is higher than a predetermined carbon emission benchmark, the weight of the carbon emission target is increased.

5. The system according to claim 1, characterized in that, The deep reinforcement learning model is a lightweight model generated through knowledge distillation; the knowledge distillation technique uses a function that includes Kullback-Leibler divergence loss and mean squared error loss.

6. A computer-based implementation method for intelligent scheduling and optimization of virtual power plants based on multi-source data fusion, characterized in that, Includes the following steps: Acquire multi-source input data, which includes at least real-time electricity prices, distributed energy status, and load forecasts; The multi-source input data is fed into a deep reinforcement learning model to generate scheduling actions for multiple optimization objectives; The reward signal for the current scheduling action is calculated based on a reward function that is weighted by a revenue item, a resource utilization item, and a carbon emission item. In response to the reward signal and changes in system state, the weights of the corresponding terms in the reward function are dynamically adjusted; Based on the scheduling action, specific scheduling instructions for resources within the virtual power plant are generated and output.

7. The method according to claim 6, characterized in that, The step of dynamically adjusting the weights further includes: When the real-time revenue is lower than a first predetermined threshold, the weight of the revenue item is increased; When the resource utilization rate exceeds the second predetermined threshold, the weight of the resource utilization rate item is reduced. And when the carbon emission intensity exceeds a third predetermined threshold, the weight of the carbon emission item is increased.

8. The method according to claim 6, characterized in that, The method also includes the step of training the deep reinforcement learning model: The model policy network is initialized by offline pre-training using historical market and weather datasets; and then online learning is performed at a preset time frequency during actual system operation.

9. The method according to claim 6, characterized in that, In the step of generating scheduling actions, the deep reinforcement learning model processes the input data in the following manner: Using convolutional neural networks to extract spatial correlation features among distributed energy resources in different geographical locations; And utilize Long Short-Term Memory (LSTM) networks to capture the dynamic characteristics of data such as electricity prices, load, and weather over time.

Images