Park-level adjustable load-oriented resource regulation and control method, device and equipment and medium

Abstract

The invention belongs to the technical field of power grid regulation and control, and particularly relates to a park-level adjustable load-oriented resource regulation and control method and system, and the method comprises the steps: formulating an initial energy scheduling scheme according to the predicted energy demand in combination with the output characteristics of distributed energy in a park and the load condition of a power grid; and training a load regulation and control strategy model by adopting a reinforcement learning algorithm based on the initial energy scheduling scheme, and generating a personalized load regulation and control instruction for each energy consumption unit. According to the scheme, the real-time energy consumption data of the plurality of energy consumption units in the park, the output characteristics of the distributed energy and the power grid load condition can be comprehensively considered, a scientific and reasonable energy scheduling and load regulation and control strategy is formulated, and efficient utilization and green low-carbon development of park energy are realized. Meanwhile, the method also needs to have the characteristics of data sharing, accurate prediction, intelligent scheduling, rapid response and the like so as to adapt to complex and changeable energy requirements and environmental protection requirements of the park.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid regulation and control, and specifically relates to a resource regulation method, device, equipment and medium for park-level adjustable loads. Background Art

[0002] With the rapid development of energy internet and smart grid, intelligent management of park-level energy systems has become a research hotspot. Among them, effective mining and precise regulation of adjustable load resources are the key to achieving efficient operation of park energy systems and improving energy utilization. However, the current resource mining and regulation of park-level adjustable loads still faces many challenges: Diversity of load characteristics: The park includes various electricity users such as industry, commerce, and residents, and the load characteristics vary significantly. The traditional single load model is difficult to accurately describe the overall energy consumption behavior of the park.

[0003] Data acquisition and privacy protection: It is often difficult to fully obtain detailed energy consumption data of each energy-consuming unit in the park, and there are data privacy protection issues, which restricts the application of traditional centralized modeling methods.

[0004] Personalization and coordination of control strategies: Different energy-consuming units have different acceptance and response capabilities to load control. How to achieve the overall control goals of the park while ensuring personalized needs is an urgent problem to be solved.

[0005] Adaptability to dynamic environments: The park’s energy system is subject to the influence of dynamic factors such as weather changes and changes in energy consumption patterns. Traditional static control strategies are difficult to adapt to this complex and changing environment.

[0006] Computing efficiency and real-time performance: Park-level load control involves a large number of energy-consuming units. How to achieve rapid response and real-time control with limited computing resources is a major challenge in practical applications.

[0007] Explainability and credibility: Although some existing control methods based on black box models have superior performance, they lack explainability and are difficult to gain the trust of users and managers.

[0008] Evaluation of control effects: There is a lack of unified and objective evaluation criteria to measure the effects of load control, especially when considering multiple objectives such as user comfort and system stability. Summary of the invention

[0009] The embodiment of the present invention provides a resource control method and system for campus-level adjustable loads to at least solve one of the problems in the above-mentioned prior art.

[0010] In order to achieve the above object, the present invention adopts the following technical solution: In a first aspect, the present invention provides a resource control method for a park-level adjustable load, comprising: Obtain historical energy consumption data and real-time energy consumption data of multiple energy consumption units in the park, the energy consumption units including buildings, industrial equipment and new energy facilities; based on the historical energy consumption data, use a deep learning algorithm to build an energy consumption prediction model, the deep learning algorithm includes a long short-term memory network and an attention mechanism; use the energy consumption prediction model and the real-time energy consumption data to predict the energy consumption demand of the park in the future period; According to the predicted energy demand, combined with the output characteristics of distributed energy in the park and the load of the power grid, an initial energy scheduling plan is formulated; based on the initial energy scheduling plan, a reinforcement learning algorithm is used to train a load control strategy model, and the reinforcement learning algorithm includes a deep Q network and a policy gradient method; using the load control strategy model, a personalized load control instruction for each energy unit is generated; and the personalized load control instruction is sent to the intelligent control terminal of each energy unit; Monitor the energy consumption status of each energy-consuming unit after executing the load control instruction in real time; calculate the deviation between the actual load control effect and the expected target based on the energy consumption status; according to the deviation, use the federated learning method to update the energy consumption prediction model and the load control strategy model, wherein each energy-consuming unit trains the model locally and encrypts the model parameters and uploads them to the central server for aggregation; iteratively optimize the energy scheduling plan and load control instructions until the predetermined control target is met or the maximum number of iterations is reached.

[0011] In an optional solution, based on the historical energy consumption data, a deep learning algorithm is used to build an energy consumption prediction model, and the deep learning algorithm includes a long short-term memory network and an attention mechanism including: The energy consumption data of the park is preprocessed, the original timestamp is converted into periodic features and one-hot encoded, the continuous features are standardized, and sliding window samples are constructed; the preprocessed energy consumption data is extracted with multi-scale time features using multiple one-dimensional convolutional layers with different convolution kernel sizes, the extracted features are concatenated in the channel dimension, and the dimension is reduced through one-dimensional convolution; the reduced-dimensional features are input into the bidirectional long short-term memory network encoder to generate the encoder output; The encoder output is decoded using a unidirectional long short-term memory network decoder. During the decoding process, the prediction result of the previous step is used as the input of the current step. At each time step of the decoder, the attention weight between the current hidden state of the decoder and the encoder output is calculated. Feed the decoder output into a multi-head attention layer, which contains multiple attention heads. Each attention head independently calculates the query matrix, key matrix, and value matrix, and uses the scaled dot product attention mechanism to calculate the attention score. Concatenate the outputs of multiple attention heads and map them back to the original dimension through a linear layer. Add residual connections after the convolutional layer, the LSTM layer, and the attention layer, and apply layer normalization after each residual connection; use the fully connected layer and activation function to process the output of the attention layer to generate the final energy consumption prediction result; The model was trained using a loss function containing root mean square error and L1 regularization term, combined with the Adam optimizer and cosine annealing learning rate strategy. The early stopping strategy was applied during the training process, and the training process was dynamically adjusted according to the performance of the validation set. Multiple models with different random initializations were trained, and the prediction results of these models were integrated through the averaging method to obtain the final park energy consumption prediction results.

[0012] In the optional scheme, according to the predicted energy demand, combined with the output characteristics of distributed energy in the park and the grid load, the initial energy dispatch plan is formulated, including: Construct a multi-objective optimization model, taking minimization of operating costs, maximization of renewable energy utilization and minimization of carbon emissions as optimization goals, while setting power balance constraints, thermal balance constraints, energy storage system constraints and distributed energy output constraints; An improved multi-objective particle swarm optimization algorithm is designed. The algorithm includes adaptive inertia weight, crowding distance sorting mechanism and ε-constraint method. The adaptive inertia weight is dynamically adjusted with the number of iterations. The crowding distance sorting mechanism is used to maintain the diversity of non-dominated solution sets. The ε-constraint method is used to deal with complex constraints. Integrate deep learning energy demand prediction model and renewable energy output prediction model to generate multiple energy demand scenarios and renewable energy output probability distribution; Using the improved multi-objective particle swarm optimization algorithm, a Pareto optimal solution set is generated based on the multi-objective optimization model, energy demand scenarios and renewable energy output probability distribution; Performing cluster analysis on the Pareto optimal solution set, and selecting the central solution of each cluster as a representative scheduling solution; Monte Carlo simulation is performed on each representative scheduling scheme to calculate the expected performance and risk metrics of the scheme under different scenarios and evaluate the robustness of the scheme; Based on the robustness evaluation results, multiple alternative scheduling schemes are screened out, and a 24-hour scheduling schedule is generated that includes the output of each power generation unit, energy storage charging and discharging plan, and controllable load scheduling; Conduct sensitivity analysis on the selected alternative scheduling schemes to evaluate the impact of key parameters on the schemes; Based on the sensitivity analysis results, the optimal initial energy scheduling scheme is selected, which can maintain good performance under different conditions; The selected initial energy scheduling scheme is output as the basis for generating subsequent load control strategies.

[0013] In an optional solution, based on the initial energy scheduling plan, a reinforcement learning algorithm is used to train a load control strategy model, and the reinforcement learning algorithm includes a deep Q network and a policy gradient method including: Construct a campus load control environment model, which includes state space, action space and reward function. The state space includes time characteristics, load status, energy price, weather conditions and energy supply information. The action space is the discretized load adjustment amount. The reward function comprehensively considers energy cost savings, user comfort and grid stability. Design a multi-agent collaborative learning framework, assign an independent deep Q-network agent to each adjustable energy unit, and set up a central coordinator to be responsible for information aggregation and global reward distribution; Construct a dual deep Q network structure, including a main network for selecting actions and updating, and a target network for calculating the target Q value. Both networks have the same multi-layer neural network structure. Integrated Actor-Critic architecture, where the Actor network outputs the action probability distribution, the Critic network estimates the state value, and calculates the advantage function for policy gradient update; Design a priority experience replay mechanism, including priority calculation based on temporal difference error, priority-based sampling probability calculation, and importance sampling weight calculation; Implement a communication mechanism among multiple agents, use the attention mechanism to aggregate key information of other agents, and regularly broadcast global status information through a central coordinator; Design a hybrid reward allocation scheme that combines local rewards for individual energy-consuming units with global rewards for the entire park; The course learning strategy is adopted to divide the training process into three stages: load control of a single energy-consuming unit, coordinated control of multiple similar energy-consuming units, and comprehensive control of multiple types of energy-consuming units in the entire park. During the training process, the ε-greedy strategy is used for action selection; Regularly evaluate model performance, including cumulative rewards, energy cost savings, user comfort satisfaction, and load volatility, and conduct load forecast error sensitivity analysis, energy price fluctuation adaptability testing, and performance evaluation under extreme weather conditions; The trained model is subjected to knowledge distillation, quantization, and ONNX conversion to obtain a compressed model suitable for deployment.

[0014] In an optional solution, the compressed model is used to generate a load control strategy based on the real-time park status, and output load adjustment instructions for each energy-consuming unit.

[0015] In an optional solution, based on the deviation, a federated learning method is used to update the energy consumption prediction model and the load regulation strategy model, wherein each energy consumption unit trains the model locally and encrypts the model parameters and uploads them to a central server for aggregation, including: Initializing global model parameters and encryption keys, wherein the global model parameters include energy consumption prediction model parameters and load regulation strategy model parameters, and the encryption keys include a public key and a private key; Distributing the global model parameters and the public key to multiple energy consuming units; Each of the energy-consuming units collects local data, preprocesses the local data, and calculates a data quality score; each of the energy-consuming units uses the local data to update local model parameters and calculates a prediction accuracy score; each of the energy-consuming units adds differential privacy noise to the updated local model parameters; each of the energy-consuming units uses the public key to encrypt the local model parameters after adding noise; each of the energy-consuming units uploads the encrypted local model parameters to a central server; The central server calculates the contribution weight of each of the energy-consuming units, where the contribution weight is based on the data quality score and the prediction accuracy score; the central server performs weighted aggregation on the encrypted local model parameters; the central server uses the private key to decrypt the aggregation result to obtain updated global model parameters.

[0016] In an optional solution, the method further comprises: Constructing a knowledge distillation model, and distributing the updated global model parameters and the knowledge distillation model to each of the energy consuming units; Each of the energy-consuming units calculates the similarity between the local data distribution and the global distribution; each of the energy-consuming units dynamically adjusts the knowledge distillation weight according to the similarity; each of the energy-consuming units defines a new local loss function based on the original task loss and the knowledge distillation loss; each of the energy-consuming units uses the new local loss function to perform local model fine-tuning; The central server evaluates the performance of the updated global model parameters on a validation dataset; When the performance degradation exceeds a preset threshold, a rollback mechanism is triggered to roll back the global model parameters to the last stable version and temporarily freeze the energy-consuming units that cause the performance degradation; Repeat the above steps until the preset termination condition is met.

[0017] In a second aspect, the present invention provides a resource control device for a park-level adjustable load, comprising: The first unit is used to obtain historical energy consumption data and real-time energy consumption data of multiple energy consumption units in the park, wherein the energy consumption units include buildings, industrial equipment and new energy facilities; based on the historical energy consumption data, an energy consumption prediction model is constructed using a deep learning algorithm, wherein the deep learning algorithm includes a long short-term memory network and an attention mechanism; and the energy consumption demand of the park in the future period is predicted using the energy consumption prediction model and the real-time energy consumption data; The second unit is used to formulate an initial energy dispatching plan according to the predicted energy demand, combined with the output characteristics of distributed energy in the park and the load of the power grid; based on the initial energy dispatching plan, a reinforcement learning algorithm is used to train a load control strategy model, and the reinforcement learning algorithm includes a deep Q network and a policy gradient method; using the load control strategy model, a personalized load control instruction for each energy unit is generated; and the personalized load control instruction is sent to the intelligent control terminal of each energy unit; The third unit is used to monitor the energy consumption status of each energy-consuming unit after executing the load control instruction in real time; based on the energy consumption status, calculate the deviation between the actual load control effect and the expected target; according to the deviation, use the federated learning method to update the energy consumption prediction model and the load control strategy model, wherein each energy-consuming unit trains the model locally and encrypts the model parameters and uploads them to the central server for aggregation; iteratively optimize the energy scheduling plan and load control instructions until the predetermined control target is met or the maximum number of iterations is reached.

[0018] According to a third aspect of the present invention, an electronic device is provided, comprising a processor and a memory, wherein the processor is configured to execute a computer program stored in the memory to implement the resource control method as described above.

[0019] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by a processor, the resource control method as described above is implemented.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: Improve prediction accuracy: By combining the long short-term memory network and the deep learning algorithm of the attention mechanism, it is possible to more accurately capture the long-term dependence and key characteristics of the campus energy consumption pattern, significantly improving the accuracy of energy demand forecasting.

[0021] Realize personalized control: The load control strategy model trained by the reinforcement learning algorithm can generate personalized control instructions according to the characteristics of different energy-consuming units, which not only meets the needs of a single energy-consuming unit, but also achieves the overall control goals of the park.

[0022] Protecting data privacy: The federated learning method is used to update the model, so that each energy-consuming unit can train the model locally and only need to upload the encrypted model parameters, effectively protecting the privacy and security of energy consumption data. Improve system adaptability: Through real-time monitoring and iterative optimization, the system can quickly adapt to changes in the park's energy consumption patterns and fluctuations in the external environment, maintaining the stability of the regulation effect.

[0023] Optimize computing efficiency: The distributed model training and update mechanism greatly reduces the computing burden of the central server and improves the system's response speed and real-time performance.

[0024] Enhanced scalability: This method is applicable to parks of different sizes and types, has good scalability, and can easily cope with the expansion of park scale and the increase of energy consumption units.

[0025] Improve energy efficiency: Through accurate demand forecasting and flexible load regulation, it is possible to better balance energy demand and distributed energy supply within the park and improve overall energy efficiency.

[0026] Reduce operating costs: Optimized energy scheduling solutions and load control strategies can effectively reduce the energy costs of the park and the peak load regulation pressure of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings constituting a part of the present application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 A flow chart of a resource control method for a park-level adjustable load according to an embodiment of the present invention; Figure 2 The present invention is a structural block diagram of a resource control device for campus-level adjustable loads according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0029] The following detailed description is an exemplary description, which is intended to provide further detailed description of the present invention. Unless otherwise specified, all technical terms used in the present invention have the same meaning as those generally understood by those skilled in the art to which the present application belongs. The terms used in the present invention are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention.

[0030] Example 1 like Figure 1As shown, a resource control method for campus-level adjustable loads includes: S1. Obtain real-time energy consumption data of multiple energy consumption units in the park.

[0031] Specifically, the energy-consuming units include buildings, industrial equipment and new energy facilities.

[0032] S2. Input the real-time energy consumption data into a pre-trained energy consumption prediction model to obtain the energy consumption demand of the park in the future period; wherein the energy consumption prediction model is trained based on a deep learning algorithm.

[0033] Specifically, the deep learning algorithm includes a long short-term memory network and an attention mechanism.

[0034] Specifically, the energy consumption prediction model is obtained based on deep learning algorithm training, including: The energy consumption data of the park is preprocessed, the original timestamp is converted into periodic features and one-hot encoded, the continuous features are standardized, and sliding window samples are constructed; the preprocessed energy consumption data is extracted with multi-scale time features using multiple one-dimensional convolutional layers with different convolution kernel sizes, the extracted features are concatenated in the channel dimension, and the dimension is reduced through one-dimensional convolution; the reduced-dimensional features are input into the bidirectional long short-term memory network encoder to generate the encoder output; The encoder output is decoded using a unidirectional long short-term memory network decoder. During the decoding process, the prediction result of the previous step is used as the input of the current step. At each time step of the decoder, the attention weight between the current hidden state of the decoder and the encoder output is calculated. Feed the decoder output into a multi-head attention layer, which contains multiple attention heads. Each attention head independently calculates the query matrix, key matrix, and value matrix, and uses the scaled dot product attention mechanism to calculate the attention score. Concatenate the outputs of multiple attention heads and map them back to the original dimension through a linear layer. Add residual connections after the convolutional layer, the LSTM layer, and the attention layer, and apply layer normalization after each residual connection; use the fully connected layer and activation function to process the output of the attention layer to generate the final energy consumption prediction result; The model was trained using a loss function containing root mean square error and L1 regularization term, combined with the Adam optimizer and cosine annealing learning rate strategy. The early stopping strategy was applied during the training process, and the training process was dynamically adjusted according to the performance of the validation set. Multiple models with different random initializations were trained, and the prediction results of these models were integrated through the averaging method to obtain the final park energy consumption prediction results.

[0035] S3. Based on the predicted energy demand, combined with the output characteristics of distributed energy in the park and the grid load conditions, formulate an initial energy scheduling plan.

[0036] Specifically, according to the predicted energy demand, combined with the output characteristics of distributed energy in the park and the grid load, an initial energy dispatch plan is formulated, including: Construct a multi-objective optimization model, taking minimization of operating costs, maximization of renewable energy utilization and minimization of carbon emissions as optimization goals, while setting power balance constraints, thermal balance constraints, energy storage system constraints and distributed energy output constraints; An improved multi-objective particle swarm optimization algorithm is designed, which includes adaptive inertia weight, crowding distance sorting mechanism and ε-constraint method. The adaptive inertia weight is dynamically adjusted with the number of iterations, the crowding distance sorting mechanism is used to maintain the diversity of non-dominated solution sets, and the ε-constraint method is used to deal with complex constraints. Integrate deep learning energy demand prediction model and renewable energy output prediction model to generate multiple energy demand scenarios and renewable energy output probability distribution; Using the improved multi-objective particle swarm optimization algorithm, a Pareto optimal solution set is generated based on the multi-objective optimization model, energy demand scenarios and renewable energy output probability distribution; Performing cluster analysis on the Pareto optimal solution set, and selecting the central solution of each cluster as a representative scheduling solution; Monte Carlo simulation is performed on each representative scheduling scheme to calculate the expected performance and risk metrics of the scheme under different scenarios and evaluate the robustness of the scheme; Based on the robustness evaluation results, multiple alternative scheduling schemes are screened out, and a 24-hour scheduling schedule is generated that includes the output of each power generation unit, energy storage charging and discharging plan, and controllable load scheduling; Conduct sensitivity analysis on the selected alternative scheduling schemes to evaluate the impact of key parameters on the schemes; Based on the results of sensitivity analysis, the optimal initial energy scheduling scheme is selected, which can maintain good performance under different conditions. The selected initial energy scheduling scheme is output as the basis for the subsequent load control strategy generation.

[0037] S4. Input the initial energy scheduling plan into a pre-trained load control strategy model to generate personalized load control instructions for each energy-consuming unit; wherein the load control strategy model is trained using a reinforcement learning algorithm.

[0038] Specifically, the reinforcement learning algorithm includes a deep Q network and a policy gradient method.

[0039] Specifically, the load control strategy model is trained by using a reinforcement learning algorithm, including: Construct a park load control environment model, including state space, action space and reward function; the state space includes time characteristics, load status, energy price, weather conditions and energy supply information, the action space is the discretized load adjustment amount, and the reward function comprehensively considers energy cost savings, user comfort and grid stability; Design a multi-agent collaborative learning framework, assign an independent deep Q-network agent to each adjustable energy unit, and set up a central coordinator to be responsible for information aggregation and global reward distribution; Construct a dual deep Q network structure, including a main network for selecting actions and updating, and a target network for calculating the target Q value. Both networks have the same multi-layer neural network structure. Integrated Actor-Critic architecture, where the Actor network outputs the action probability distribution, the Critic network estimates the state value, and calculates the advantage function for policy gradient update; Design a priority experience replay mechanism, including priority calculation based on temporal difference error, priority-based sampling probability calculation, and importance sampling weight calculation; Implement a communication mechanism among multiple agents, use the attention mechanism to aggregate key information of other agents, and regularly broadcast global status information through a central coordinator; Design a hybrid reward allocation scheme that combines local rewards for individual energy-consuming units with global rewards for the entire park; The course learning strategy is adopted to divide the training process into three stages: load control of a single energy-consuming unit, coordinated control of multiple similar energy-consuming units, and comprehensive control of multiple types of energy-consuming units in the entire park. During the training process, the ε-greedy strategy is used for action selection; Regularly evaluate model performance, including cumulative rewards, energy cost savings, user comfort satisfaction, and load volatility, and conduct load forecast error sensitivity analysis, energy price fluctuation adaptability testing, and performance evaluation under extreme weather conditions; Perform knowledge distillation, quantization, and ONNX conversion on the trained model to obtain a compressed model suitable for deployment; Using the compressed model, a load control strategy is generated according to the real-time park status, and load adjustment instructions for each energy-consuming unit are output.

[0040] In a preferred embodiment, the method further comprises: The personalized load regulation instruction is sent to the intelligent control terminal of each energy-consuming unit; the energy consumption status of each energy-consuming unit after executing the load regulation instruction is monitored in real time; based on the energy consumption status, the deviation between the actual load regulation effect and the expected target is calculated; according to the deviation, the energy consumption prediction model and the load regulation strategy model are updated using a federated learning method; wherein each energy-consuming unit trains the model locally, and encrypts the model parameters and uploads them to a central server for aggregation; the energy scheduling plan and the load regulation instruction are iteratively optimized until the predetermined regulation target is met or the maximum number of iterations is reached.

[0041] Specifically, according to the deviation, the energy consumption prediction model and the load regulation strategy model are updated by a federated learning method, wherein each energy consumption unit trains the model locally and encrypts the model parameters and uploads them to a central server for aggregation, including: Initializing global model parameters and encryption keys, wherein the global model parameters include energy consumption prediction model parameters and load regulation strategy model parameters, and the encryption keys include a public key and a private key; Distributing the global model parameters and the public key to multiple energy consuming units; Each of the energy-consuming units collects local data, preprocesses the local data, and calculates a data quality score; each of the energy-consuming units uses the local data to update local model parameters and calculates a prediction accuracy score; each of the energy-consuming units adds differential privacy noise to the updated local model parameters; each of the energy-consuming units uses the public key to encrypt the local model parameters after adding noise; each of the energy-consuming units uploads the encrypted local model parameters to a central server; The central server calculates the contribution weight of each of the energy-consuming units, where the contribution weight is based on the data quality score and the prediction accuracy score; the central server performs weighted aggregation on the encrypted local model parameters; the central server uses the private key to decrypt the aggregation result to obtain updated global model parameters.

[0042] In a preferred embodiment, the method further comprises: Constructing a knowledge distillation model, and distributing the updated global model parameters and the knowledge distillation model to each of the energy consuming units; Each of the energy-consuming units calculates the similarity between the local data distribution and the global distribution; each of the energy-consuming units dynamically adjusts the knowledge distillation weight according to the similarity; each of the energy-consuming units defines a new local loss function based on the original task loss and the knowledge distillation loss; each of the energy-consuming units uses the new local loss function to perform local model fine-tuning; The central server evaluates the performance of the updated global model parameters on a validation dataset; When the performance degradation exceeds a preset threshold, a rollback mechanism is triggered to roll back the global model parameters to the last stable version and temporarily freeze the energy-consuming units that cause the performance degradation; Repeat the above steps until the preset termination condition is met.

[0043] Figure 2 FIG. 1 is a schematic diagram of the structure of a resource mining and control system for a park-level adjustable load according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to obtain historical energy consumption data and real-time energy consumption data of multiple energy consumption units in the park, wherein the energy consumption units include buildings, industrial equipment and new energy facilities; based on the historical energy consumption data, an energy consumption prediction model is constructed using a deep learning algorithm, wherein the deep learning algorithm includes a long short-term memory network and an attention mechanism; and the energy consumption demand of the park in the future period is predicted using the energy consumption prediction model and the real-time energy consumption data; The second unit is used to formulate an initial energy dispatching plan according to the predicted energy demand, combined with the output characteristics of distributed energy in the park and the load of the power grid; based on the initial energy dispatching plan, a reinforcement learning algorithm is used to train a load control strategy model, and the reinforcement learning algorithm includes a deep Q network and a policy gradient method; using the load control strategy model, a personalized load control instruction for each energy unit is generated; and the personalized load control instruction is sent to the intelligent control terminal of each energy unit; The third unit is used to monitor the energy consumption status of each energy-consuming unit after executing the load control instruction in real time; based on the energy consumption status, calculate the deviation between the actual load control effect and the expected target; according to the deviation, use the federated learning method to update the energy consumption prediction model and the load control strategy model, wherein each energy-consuming unit trains the model locally and encrypts the model parameters and uploads them to the central server for aggregation; iteratively optimize the energy scheduling plan and load control instructions until the predetermined control target is met or the maximum number of iterations is reached.

[0044] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned resource control method.

[0045] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the resource control method described above is implemented.

[0046] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A resource control method for adjustable loads at the park level, characterized in that: include: Obtain historical and real-time energy consumption data for multiple energy consumption units in the park, including buildings, industrial equipment, and new energy facilities; Based on the historical energy consumption data, an energy consumption prediction model is constructed using a deep learning algorithm, wherein the deep learning algorithm includes a long short-term memory network and an attention mechanism; and the energy consumption prediction model and the real-time energy consumption data are used to predict the energy consumption demand of the park in the future period; Formulate an initial energy dispatch plan based on the predicted energy demand, combined with the output characteristics of distributed energy in the park and the grid load; Based on the initial energy scheduling scheme, a reinforcement learning algorithm is used to train a load control strategy model, wherein the reinforcement learning algorithm includes a deep Q network and a policy gradient method; using the load control strategy model, a personalized load control instruction for each energy consuming unit is generated; and the personalized load control instruction is sent to an intelligent control terminal of each energy consuming unit; Real-time monitoring of the energy consumption status of each energy-consuming unit after executing the load control instruction; based on the energy consumption status, calculating the deviation between the actual load control effect and the expected target; according to the deviation, using a federated learning method to update the energy consumption prediction model and the load control strategy model, wherein each energy-consuming unit trains the model locally and encrypts the model parameters and uploads them to the central server for aggregation; The energy scheduling scheme and load control instructions are iteratively optimized until a predetermined control target is met or a maximum number of iterations is reached.

2. The resource control method according to claim 1, characterized in that: Based on the historical energy consumption data, an energy consumption prediction model is constructed using a deep learning algorithm, wherein the deep learning algorithm includes a long short-term memory network and an attention mechanism including: The energy consumption data of the park is preprocessed, the original timestamp is converted into periodic features and one-hot encoded, the continuous features are standardized, and sliding window samples are constructed; the preprocessed energy consumption data is extracted with multi-scale time features using multiple one-dimensional convolutional layers with different convolution kernel sizes, the extracted features are concatenated in the channel dimension, and the dimension is reduced through one-dimensional convolution; the reduced-dimensional features are input into the bidirectional long short-term memory network encoder to generate the encoder output; The encoder output is decoded using a unidirectional long short-term memory network decoder. During the decoding process, the prediction result of the previous step is used as the input of the current step. At each time step of the decoder, the attention weight between the current hidden state of the decoder and the encoder output is calculated. Feed the decoder output into a multi-head attention layer, which contains multiple attention heads. Each attention head independently calculates the query matrix, key matrix, and value matrix, and uses the scaled dot product attention mechanism to calculate the attention score. Concatenate the outputs of multiple attention heads and map them back to the original dimension through a linear layer. Add residual connections after the convolutional layer, the LSTM layer, and the attention layer, and apply layer normalization after each residual connection; use the fully connected layer and activation function to process the output of the attention layer to generate the final energy consumption prediction result; The model was trained using a loss function containing root mean square error and L1 regularization term, combined with the Adam optimizer and cosine annealing learning rate strategy. The early stopping strategy was applied during the training process, and the training process was dynamically adjusted according to the performance of the validation set. Multiple models with different random initializations were trained, and the prediction results of these models were integrated through the averaging method to obtain the final park energy consumption prediction results.

3. The resource control method according to claim 1, characterized in that: According to the predicted energy demand, combined with the output characteristics of distributed energy in the park and the load of the power grid, the initial energy dispatch plan is formulated, including: Construct a multi-objective optimization model, taking minimization of operating costs, maximization of renewable energy utilization and minimization of carbon emissions as optimization goals, while setting power balance constraints, thermal balance constraints, energy storage system constraints and distributed energy output constraints; An improved multi-objective particle swarm optimization algorithm is designed. The algorithm includes adaptive inertia weight, crowding distance sorting mechanism and ε-constraint method. The adaptive inertia weight is dynamically adjusted with the number of iterations. The crowding distance sorting mechanism is used to maintain the diversity of non-dominated solution sets. The ε-constraint method is used to deal with complex constraints. Integrate deep learning energy demand prediction model and renewable energy output prediction model to generate multiple energy demand scenarios and renewable energy output probability distribution; Using the improved multi-objective particle swarm optimization algorithm, a Pareto optimal solution set is generated based on the multi-objective optimization model, energy demand scenarios and renewable energy output probability distribution; Performing cluster analysis on the Pareto optimal solution set, and selecting the central solution of each cluster as a representative scheduling solution; Monte Carlo simulation is performed on each representative scheduling scheme to calculate the expected performance and risk metrics of the scheme under different scenarios and evaluate the robustness of the scheme; Based on the robustness evaluation results, multiple alternative scheduling schemes are screened out, and a 24-hour scheduling schedule is generated that includes the output of each power generation unit, energy storage charging and discharging plan, and controllable load scheduling; Conduct sensitivity analysis on the selected alternative scheduling schemes to evaluate the impact of key parameters on the schemes; Based on the sensitivity analysis results, the optimal initial energy scheduling scheme is selected, which can maintain good performance under different conditions; The selected initial energy scheduling scheme is output as the basis for generating subsequent load control strategies.

4. The resource control method according to claim 1, characterized in that: Based on the initial energy scheduling scheme, a load control strategy model is trained using a reinforcement learning algorithm, wherein the reinforcement learning algorithm includes a deep Q network and a policy gradient method including: Construct a campus load control environment model, which includes state space, action space and reward function. The state space includes time characteristics, load status, energy price, weather conditions and energy supply information. The action space is the discretized load adjustment amount. The reward function comprehensively considers energy cost savings, user comfort and grid stability. Design a multi-agent collaborative learning framework, assign an independent deep Q-network agent to each adjustable energy unit, and set up a central coordinator to be responsible for information aggregation and global reward distribution; Construct a dual deep Q network structure, including a main network for selecting actions and updating, and a target network for calculating the target Q value. Both networks have the same multi-layer neural network structure. Integrated Actor-Critic architecture, where the Actor network outputs the action probability distribution, the Critic network estimates the state value, and calculates the advantage function for policy gradient update; Design a priority experience replay mechanism, including priority calculation based on temporal difference error, priority-based sampling probability calculation, and importance sampling weight calculation; Implement a communication mechanism among multiple agents, use the attention mechanism to aggregate key information of other agents, and regularly broadcast global status information through a central coordinator; Design a hybrid reward allocation scheme that combines local rewards for individual energy-consuming units with global rewards for the entire park; The course learning strategy is adopted to divide the training process into three stages: load control of a single energy-consuming unit, coordinated control of multiple similar energy-consuming units, and comprehensive control of multiple types of energy-consuming units in the entire park. During the training process, the ε-greedy strategy is used for action selection; Regularly evaluate model performance, including cumulative rewards, energy cost savings, user comfort satisfaction, and load volatility, and conduct load forecast error sensitivity analysis, energy price fluctuation adaptability testing, and performance evaluation under extreme weather conditions; The trained model is subjected to knowledge distillation, quantization, and ONNX conversion to obtain a compressed model suitable for deployment.

5. The resource control method according to claim 4, characterized in that: Using the compressed model, a load control strategy is generated according to the real-time park status, and load adjustment instructions for each energy-consuming unit are output.

6. The resource control method according to claim 1, characterized in that: According to the deviation, the energy consumption prediction model and the load regulation strategy model are updated by a federated learning method, wherein each energy consumption unit trains the model locally and encrypts the model parameters and uploads them to a central server for aggregation, including: Initializing global model parameters and encryption keys, wherein the global model parameters include energy consumption prediction model parameters and load regulation strategy model parameters, and the encryption keys include a public key and a private key; Distributing the global model parameters and the public key to multiple energy consuming units; Each of the energy-consuming units collects local data, preprocesses the local data, and calculates a data quality score; each of the energy-consuming units uses the local data to update local model parameters and calculates a prediction accuracy score; each of the energy-consuming units adds differential privacy noise to the updated local model parameters; each of the energy-consuming units uses the public key to encrypt the local model parameters after adding noise; each of the energy-consuming units uploads the encrypted local model parameters to a central server; The central server calculates the contribution weight of each of the energy-consuming units, where the contribution weight is based on the data quality score and the prediction accuracy score; the central server performs weighted aggregation on the encrypted local model parameters; the central server uses the private key to decrypt the aggregation result to obtain updated global model parameters.

7. The resource control method according to claim 6, characterized in that: The method further comprises: Constructing a knowledge distillation model, and distributing the updated global model parameters and the knowledge distillation model to each of the energy consuming units; Each of the energy-consuming units calculates the similarity between the local data distribution and the global distribution; each of the energy-consuming units dynamically adjusts the knowledge distillation weight according to the similarity; each of the energy-consuming units defines a new local loss function based on the original task loss and the knowledge distillation loss; each of the energy-consuming units uses the new local loss function to perform local model fine-tuning; The central server evaluates the performance of the updated global model parameters on a validation dataset; When the performance degradation exceeds a preset threshold, a rollback mechanism is triggered to roll back the global model parameters to the last stable version and temporarily freeze the energy-consuming units that cause the performance degradation; Repeat the above steps until the preset termination condition is met.

8. A resource control device for adjustable loads at the park level, used to implement the resource control method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain historical energy consumption data and real-time energy consumption data of multiple energy consumption units in the park, wherein the energy consumption units include buildings, industrial equipment and new energy facilities; Based on the historical energy consumption data, an energy consumption prediction model is constructed using a deep learning algorithm, wherein the deep learning algorithm includes a long short-term memory network and an attention mechanism; and the energy consumption prediction model and the real-time energy consumption data are used to predict the energy consumption demand of the park in the future period; The second unit is used to formulate an initial energy dispatch plan based on the predicted energy demand, combined with the output characteristics of distributed energy in the park and the grid load; Based on the initial energy scheduling scheme, a reinforcement learning algorithm is used to train a load control strategy model, wherein the reinforcement learning algorithm includes a deep Q network and a policy gradient method; using the load control strategy model, a personalized load control instruction for each energy consuming unit is generated; and the personalized load control instruction is sent to an intelligent control terminal of each energy consuming unit; The third unit is used to monitor the energy consumption status of each energy-consuming unit after executing the load control instruction in real time; based on the energy consumption status, calculate the deviation between the actual load control effect and the expected target; according to the deviation, use the federated learning method to update the energy consumption prediction model and the load control strategy model, wherein each energy-consuming unit trains the model locally and encrypts the model parameters and uploads them to the central server for aggregation; The energy scheduling scheme and load control instructions are iteratively optimized until a predetermined control target is met or a maximum number of iterations is reached.

9. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the resource control method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by a processor, the resource control method according to any one of claims 1 to 7 is implemented.

Images