A high-efficiency self-adaptive source-network-load storage optimization scheduling method for uncertain source-load
By improving the DDPG algorithm through deep embedding clustering and meta-learning, a fast-adaptive power grid dispatching strategy is constructed, which solves the problems of poor real-time performance and weak robustness of the power grid dispatching system when facing the uncertainty of renewable energy, and improves the flexibility and economy of power grid operation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- MAINTENANCE BRANCH OF STATE GRID FUJIAN ELECTRIC POWER
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-02
AI Technical Summary
The existing power grid dispatching system suffers from poor real-time performance and weak robustness in the face of the uncertainties of renewable energy and the coordination and interaction of multiple energy sources. It is difficult to respond quickly to changes in supply and demand, resulting in a decline in power grid operation stability and a reduction in resource utilization.
We employ the DDPG algorithm, which is improved by deep embedding clustering and meta-learning, to construct a fast adaptation strategy. By learning the latent space representation and uncertainty measure through a deep autoencoder and combining it with the MAML meta-learning framework, we achieve source-network-load-storage collaborative optimization and generate an efficient scheduling strategy.
It has enabled the power grid dispatching system to quickly adapt, improving the flexibility, economy, and reliability of power grid operation, and enabling the efficient absorption of renewable energy.
Smart Images
Figure CN122136918A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power system technology, and specifically relates to an efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty. Background Technology
[0002] As the penetration rate of renewable energy sources such as wind and solar power in the power system continues to rise, grid dispatch faces unprecedented challenges. Renewable energy output exhibits significant intermittency, volatility, and randomness, and its power generation capacity is highly dependent on weather conditions, a stark contrast to the stable and controllable characteristics of traditional thermal power. This uncertainty places enormous pressure on the real-time power balance of the grid, especially in scenarios with a high proportion of renewable energy integration, where the complexity of grid dispatch increases exponentially. Simultaneously, the rapid development of distributed energy resources has led to a dramatic increase in the number of energy entities in the grid, making the traditional centralized dispatch model of "source follows load" insufficient to meet the operational needs of the new power system.
[0003] With increasing load fluctuations, existing dispatching systems lack sufficient flexibility to respond quickly to changes in supply and demand, leading to decreased grid stability and potentially causing safety issues such as frequency overruns and voltage fluctuations. Existing dispatching mechanisms are significantly inadequate in coordinating the interaction of multiple energy forms. Wind power, solar power, and other energy sources differ significantly in output characteristics, regulation capabilities, and economics, and traditional dispatching methods fail to adequately consider these differences. Regarding uncertainty modeling, existing methods often rely on precise probability distribution assumptions or overly conservative robust optimization strategies. The former struggles to accurately describe the complex uncertainties in real-world systems, while the latter can lead to overly conservative dispatching schemes, increasing operating costs and reducing resource utilization. At the optimization algorithm level, while traditional deep deterministic policy gradient and other reinforcement learning methods can handle continuous action spaces, they suffer from slow training convergence, high computational resource consumption, and poor real-time performance. Especially when facing new uncertainties, policy adjustments are slow and lack rapid adaptability.
[0004] These technological limitations severely restrict the improvement of grid operation efficiency and hinder the large-scale consumption of renewable energy. For example, Chinese patent CN110880789B discloses an economic dispatch method for a wind and solar power combined generation system. This invention cannot dynamically and adaptively adjust in real time based on the real-time output of wind and solar power and the grid status, resulting in insufficient multi-energy coordination and interaction. Therefore, there is an urgent need to develop a data-driven dispatch method that can effectively cope with source-load uncertainty and has rapid adaptability to improve the flexibility, economy, and reliability of grid operation and provide technical support for building a new power system. Summary of the Invention
[0005] The purpose of this invention is to provide an efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty. By dividing uncertain tasks through deep embedding clustering and combining meta-learning to improve the DDPG algorithm, a fast adaptation strategy is constructed to solve the problems of poor real-time performance and weak robustness of traditional methods, and to achieve economic efficiency, stability and efficient consumption of new energy under source-grid-load-storage collaborative optimization.
[0006] The technical solution of the present invention is as follows: On the one hand, this invention provides an efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty, comprising the following steps: Obtain historical source load prediction and actual data of the power grid to calculate the source load prediction error sequence; Based on the source load prediction error sequence, a deep autoencoder is used to learn the latent space representation, and combined with uncertainty measure, an improved deep embedding clustering algorithm is used to generate a set of clusters in the latent space as a set of highly cohesive scheduling tasks. A basic DDPG model is constructed based on a set of scheduled tasks. The training method of the basic DDPG model is improved by applying the MAML meta-learning framework to train meta-parameters. For new uncertain scenarios, the trained meta-parameters are used to update the inner loop to quickly generate optimized scheduling instructions. At the same time, the system operation data is monitored and the actual adjustment capability of the system is fed back to correct the model parameters. After correction, the optimal scheduling strategy is obtained. Output the optimal scheduling strategy and deploy it to the system. Collect system operation data in real time and trigger retraining periodically. Update the model parameters in each step through incremental learning.
[0007] Preferably, the step of learning the latent space representation based on the historical prediction error sequence using a deep autoencoder specifically involves: Samples were segmented according to the operating day, with each sample... The source load prediction error sequence representing one running day is a sequence of dimensions. The matrix, where In terms of time dimension, The error dimension includes wind power, photovoltaic, and load error variables; a deep autoencoder is used to learn the latent space representation as follows:
[0008]
[0009] In the formula, It is a sample The latent space vector; It is its reconstructed output; It is the set of learnable parameters of the encoder network; It is the set of learnable parameters of the decoder network; It is an encoder; It is a decoder; the depth autoencoder minimizes the reconstruction loss. Conduct pre-training.
[0010] Preferably, the joint information entropy of wind power, photovoltaic, and load forecasting errors is calculated for each sample as an uncertainty measure, expressed as:
[0011] In the formula, As a measure of uncertainty; For the sample , representing the source load prediction error sequence for one running day, is a sequence with dimension . The matrix, where In terms of time dimension, The error dimensions include wind power, photovoltaic, and load error variables; It is a sample Information entropy; The probability distribution is estimated from the sample using kernel density estimation. It was obtained by fitting the error data corresponding to the operating day.
[0012] Preferably, the loss function used in the improved deep embedding clustering algorithm is expressed as:
[0013] In the formula, The reconstruction loss is used to ensure the representational capability of the latent space; This represents the KL divergence loss, used to compute the latent space vector. With cluster center similarity And optimize its distribution with the target distribution. The KL divergence is used to force data points closer to a cluster center; This represents the consistency loss, achieved by minimizing the uncertainty metric of samples within the cluster. Compared with the average uncertainty measure of the cluster The differences are used to achieve this.
[0014] Preferably, the construction of the DDPG basic model based on the set of scheduled tasks specifically involves: Design the state space of the DDPG basic model:
[0015] In the formula, For state space; Indicates the current time The source load prediction error vector is a direct input for dealing with uncertainty; This indicates the real-time output of wind power and solar power, as well as the real-time load demand. This indicates the real-time state of charge of the energy storage system; This indicates the system's real-time marginal electricity price or time-of-use electricity price; This represents the sine and cosine encoding of a 24-hour day; Design the motion space of the DDPG basic model:
[0016] In the formula, For action space; This indicates the output adjustment of the conventional unit relative to the baseline plan; This indicates the charging and discharging power of the energy storage system, where positive represents discharging and negative represents charging. This represents the adjustable load change achieved through demand response; Design the reward function for the DDPG basic model:
[0017] In the formula, For the reward function; Indicates operating costs, This is the unit price of electricity purchased on the main network. It is the power purchased from the mainnet at time t. This is the unit power generation cost of traditional generating units. This is the benchmark output of a traditional generator set at time t. It is the unit charge-discharge degradation cost of energy storage, including the cost of purchasing electricity from the main grid, the cost of generating electricity from traditional units, and the cost of energy storage degradation; Indicates a penalty for power imbalance. It is the total power provided by the main network. This is the total power demand of the distribution network; This indicates punishment for exceeding the limits; This indicates a carbon emission penalty; This represents a weighting coefficient used to balance the proportions of economic efficiency, safety, and low carbon emissions.
[0018] Preferably, the method for improving the training of the DDPG base model using the MAML meta-learning framework, specifically for meta-parameter training, involves: Each cluster generated in the latent space by the improved deep embedding clustering algorithm This corresponds to a meta-learning task. The task environment is defined by the statistical characteristics of the samples in the cluster, the task data consists of the DDPG base model, and the task objective is to learn to effectively cope with clusters. Scheduling strategies for the specific uncertainty patterns they represent; The objective function of the MAML meta-learning framework is expressed as:
[0019] In the formula, It is a meta-parameter; The model in the task New parameters after adaptation; It is the loss function for this task; The MAML meta-learning framework trains the meta-parameters of the scheduling strategy through inner-loop task-specific adaptation and outer-loop meta-knowledge aggregation.
[0020] Preferably, the inner loop update process specifically includes: From the generated set of scheduling tasks A batch of tasks is randomly sampled from the data. For each batch of tasks , will the current meta-parameters Make a copy to obtain the task-specific parameters. ; Use task Experience replay buffer A batch of transferred samples from the sampling process, for Perform several steps of standard DDPG gradient updates to obtain the adapted parameters. ; The outer loop update process is specifically as follows: For each task Use the adapted parameters Calculate the loss for a new batch of data for this task that is different from the data used for updating the inner loop. ; Based on the performance of all sampling tasks after adaptation, the meta-parameters are updated following the direction of "how to initialize to achieve the best fast adaptation effect". ; The final output of the optimal meta-parameters after training convergence serves as the core for rapid adaptation.
[0021] Preferably, the actual adjustment capability of the feedback system includes: the maximum adjustable power of wind power, photovoltaic power and load, and the safety boundary of energy storage SOC; The modified model parameters include: adjusting the weights of the loss function in the improved deep embedding clustering algorithm and updating the outer loop learning rate of the MAML meta-learning framework. The weights of the loss function are adjusted using adaptive coefficients that are dynamically adjusted based on the feedback error, and the outer loop learning rate is adjusted using a decay strategy to improve stability.
[0022] On the other hand, the present invention provides an efficient adaptive source-grid-load-storage optimization scheduling system for source-load uncertainty, including a data input layer, a task partitioning module, an optimization scheduling module, an online application module, and a system deployment layer; The data input layer is used to acquire historical source-load prediction and actual data of the power grid to calculate the source-load prediction error sequence. The task partitioning module is used to learn the latent space representation based on the source load prediction error sequence using a deep autoencoder and combined with uncertainty measure. It generates a set of clusters in the latent space through an improved deep embedding clustering algorithm, which serves as a set of highly cohesive scheduling tasks. The scheduling module is optimized to build a basic DDPG model based on a set of scheduled tasks. The training method of the basic DDPG model is improved by applying the MAML meta-learning framework to train meta-parameters. The online application module is used to update the trained meta-parameters in an inner loop to quickly generate optimized scheduling instructions for new uncertain scenarios. At the same time, it monitors the system operation data and provides feedback on the actual adjustment capability of the system to correct the model parameters. After correction, the optimal scheduling strategy is obtained. The system deployment layer is used to output the optimal scheduling strategy and deploy it to the system, collect system operation data in real time and trigger retraining periodically, and update the model parameters in each step through incremental learning.
[0023] In another aspect, the present invention also provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any embodiment of the present invention.
[0024] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.
[0025] Compared with the prior art, the present invention has the following technical effects: 1. Rapid Adaptive Capability: This invention utilizes a "deep embedding clustering + meta-learning" framework, enabling scheduling strategies to adapt to new uncertain scenarios with only a small amount of data (a few updates), greatly improving the real-time performance of the scheduling system.
[0026] 2. High cohesive task partitioning: This invention utilizes joint information entropy and improved deep embedding clustering to partition historical uncertainty patterns into a qualitatively consistent set of tasks, providing a high-quality training foundation for meta-learning.
[0027] 3. Multi-objective collaborative optimization: The DDPG reward function designed in this invention integrates economic cost, power balance, limit violation penalty and carbon emissions, and achieves a balance between economy, safety and low carbon emissions.
[0028] 4. Closed-loop online optimization: This invention corrects model parameters through online monitoring and feedback, and combines incremental learning with periodic retraining, enabling the system to continuously adapt to long-term changes in the power grid and extreme operating conditions. Attached Figure Description
[0029] Figure 1 This is the overall flowchart of an efficient adaptive source-grid-load-storage optimization scheduling method for uncertain source-load conditions; Figure 2 It is an improvement and optimization of the scheduling flowchart; Figure 3 This is the overall architecture diagram of a highly efficient adaptive source-grid-load-storage optimization scheduling system for uncertain source-load conditions. Detailed Implementation
[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.
[0031] Example 1 This embodiment provides an efficient source-grid-load-storage optimization scheduling method that adapts to source-load uncertainty. (See reference...) Figure 1 As shown, it includes the following steps: Historical source-load prediction and actual data of the power grid are obtained to calculate the source-load prediction error sequence.
[0032] Specifically, historical power source (wind power, solar power) and load forecast and actual data are obtained from the power grid, and the forecast error sequence is calculated. Let the scheduling period be T, and the source-load forecast error sequence is expressed as:
[0033] In the formula, Let be the error vector at time t, including wind power error. Photovoltaic error and load error .
[0034] In addition, parameter initialization is performed, including parameters for the deep embedding clustering algorithm such as the number of autoencoder layers and the number of clusters, parameters for the DDPG basic model such as the learning rate and discount factor, and meta-learning parameters such as the inner loop learning rate and the outer loop learning rate.
[0035] Based on the source-load prediction error sequence, a deep autoencoder is used to learn the latent space representation. Combined with an uncertainty metric, an improved deep embedding clustering algorithm is used to generate a set of clusters in the latent space, serving as a set of highly cohesive scheduling tasks. This step intelligently groups runtime days that have historically faced different uncertainty patterns, providing structured training tasks for subsequent meta-learning.
[0036] In a preferred embodiment of this practice, the step of learning the latent space representation based on the historical prediction error sequence using a deep autoencoder specifically involves: The samples were segmented according to the operating day (24 hours), and each sample This represents the source load prediction error sequence for one run day. For a dataset containing N run days, the input is... , of which a single sample It is a dimension of The matrix, where In terms of time dimension, The error dimensions include wind power, photovoltaic, and load error variables, with T=24 hours and D=3 error variables; a deep autoencoder is used to learn its low-dimensional, dense latent space representation. Encoder Mapping the high-dimensional input to the latent space, the decoder... Used to reconstruct the input to ensure that the latent space retains key information from the original data, for:
[0037]
[0038] In the formula, It is a sample The latent space vector; It is its reconstructed output; It is the set of learnable parameters of the encoder network; It is the set of learnable parameters of the decoder network; It is an encoder; It is a decoder; the depth autoencoder minimizes the reconstruction loss. Conduct pre-training.
[0039] In this preferred embodiment, the core function of uncertainty measurement is to quantitatively characterize the qualitative differences in uncertainty faced on different operating days, rather than merely the differences in the shape of the error sequence. This embodiment uses each sample... Calculate a scalar uncertainty measure The joint information entropy of wind power, solar power, and load forecasting errors is calculated. This entropy reflects the magnitude of the overall uncertainty for that day and is expressed as:
[0040] In the formula, As a measure of uncertainty; For the sample , representing the source load prediction error sequence for one running day, is a sequence with dimension . The matrix, where In terms of time dimension, The error dimensions include wind power, photovoltaic, and load error variables; It is a sample Information entropy; The probability distribution is estimated from the sample using kernel density estimation. It was obtained by fitting the error data corresponding to the operating day.
[0041] As a preferred embodiment of this practice, an improved deep embedding clustering algorithm is employed to enhance the feature space. Clustering is performed within the cluster. The clustering loss function consists of three parts that jointly guide the formation of clusters, and is expressed as:
[0042] In the formula, The reconstruction loss is used to ensure the representational capability of the latent space; Denotes the KL divergence loss, where It is the target distribution matrix. It is a soft-assignment probability matrix. Used to calculate latent space vectors With cluster center similarity And optimize its distribution with the target distribution. The KL divergence is used to force data points closer to a cluster center. ; , ,in This is the preset number of clusters. It belongs to the first The sample set of each cluster This represents the consistency loss, achieved by minimizing the uncertainty metric of samples within the cluster. Compared with the average uncertainty measure of the cluster The difference is achieved by utilizing the previously calculated uncertainty measure, which is the core of generating a qualitatively consistent set of scheduled tasks.
[0043] After the above clustering process, all historical running days were divided into K clusters. Each cluster That is, it corresponds to a meta-learning task. The final output is a set of scheduled tasks. This set covers the main types of uncertainty the system has faced in its history ("days with high wind power fluctuations", "days with strong correlation between photovoltaic power and load", "days with overall stability", etc.), providing a solid foundation for the next step of training a meta-scheduling strategy with rapid adaptability.
[0044] A basic DDPG model is constructed based on a set of scheduled tasks. The training method of the basic DDPG model is improved by applying the Model Independent Meta-Learning (MAML) framework to train meta-parameters. In this step, to address the problems of slow training and poor generalization ability of traditional DDPG when facing source-load uncertainty, this embodiment first constructs a refined basic DDPG model oriented towards a source-grid-load-storage system. Then, the Model Independent Meta-Learning (MAML) framework is introduced to fundamentally enhance it, training a set of "meta-parameters" with rapid adaptability.
[0045] The core of the DDPG model designed in this embodiment is to train an agent that, through interaction with the environment, learns to make optimal scheduling decisions while satisfying various operational constraints. As a preferred implementation of this embodiment, the construction of the DDPG basic model based on the set of scheduling tasks specifically involves: Design the state space of the DDPG basic model, and the state... Comprehensive capture of time periods is required The system's dynamic information provides the agent with a basis for decision-making. Therefore, the designed state vector is:
[0046] In the formula, For state space; Indicates the current time The source load prediction error vector is a direct input for dealing with uncertainty; This indicates the real-time output of wind power and solar power, as well as the real-time load demand. It indicates the real-time state of charge of the energy storage system, which is related to its regulation potential; It indicates the system's real-time marginal electricity price or time-of-use electricity price, providing economic incentive signals; This means that the 24 hours of a day are encoded using sine and cosine, enabling the DDPG basic model to perceive the periodicity of time.
[0047] Design the motion space and motion of the DDPG basic model. These are continuous control commands issued by the dispatch center to the controllable unit, defined as:
[0048] In the formula, For action space; This indicates the output adjustment of the conventional unit relative to the baseline plan; This indicates the charging and discharging power of the energy storage system, where positive represents discharging and negative represents charging. This represents the adjustable load change achieved through demand response.
[0049] Design the reward function for the DDPG basic model. The reward function is key to guiding the agent's learning. In this embodiment, it is designed as a multi-objective comprehensive reward function:
[0050] In the formula, For the reward function; Indicates operating costs, This is the unit price of electricity purchased on the main network. It is the power purchased from the mainnet at time t. This is the unit power generation cost of traditional generating units. This is the benchmark output of a traditional generator set at time t. It is the unit charge-discharge degradation cost of energy storage, including the cost of purchasing electricity from the main grid, the cost of generating electricity from traditional units, and the cost of energy storage degradation; The penalty for power imbalance is a core element of system stability. It is the total power provided by the main network. This is the total power demand of the distribution network; This indicates punishment for exceeding the limits; This signifies carbon emission penalties and encourages the adoption of clean energy. This represents a weighting coefficient used to balance the proportions of economic efficiency, safety, and low carbon emissions.
[0051] Traditional DDPG, when faced with a new and uncertain scenario (i.e., a new task), needs to start from scratch or spend a significant amount of time retraining, failing to meet the real-time scheduling requirements. The core innovation of this embodiment lies in the deep integration of the MAML framework with DDPG to train a set of "meta-parameters". This set of parameters is not a final strategy for a specific task, but rather an initialization point for "high-level prior knowledge" or a "fast learner." As a preferred implementation in this embodiment, such as... Figure 2 As shown, the method for improving the training of the DDPG base model using the MAML meta-learning framework, specifically for meta-parameter training, is as follows: Each cluster generated in the latent space by the improved deep embedding clustering algorithm This corresponds to a meta-learning task ,Task The specific definition is: Task environment: A simulation environment in which the distribution characteristics of source load prediction errors are determined by clusters. It is defined by the statistical characteristics of all samples (mean of error, covariance matrix, time series pattern).
[0052] Task data: cluster The DDPG base model (states, actions, rewards) for all historical run days contained herein constitutes the experience replay buffer for this task. The main body is used for the internal loop adaptation of meta-learning.
[0053] Task objective: To learn a method that can effectively deal with clusters The scheduling strategy for the specific uncertainty pattern it represents.
[0054] The goal of MAML is to find a set of initial parameters. This makes it possible for tasks to be distributed... Any new task in the middle sampling The model can maximize its performance after only a small amount of data for the task (such as one or a few gradient updates). Therefore, the objective function of the MAML meta-learning framework is expressed as:
[0055] In the formula, It is a meta-parameter; The model in the task New parameters after adaptation; It is the loss function for this task.
[0056] The MAML meta-learning framework trains the meta-parameters of the scheduling policy (the mapping function π:s→a learned by the agent, i.e., the complete policy from state to action) through inner-loop task-specific adaptation and outer-loop meta-knowledge aggregation.
[0057] In a preferred embodiment of this invention, the inner loop update process is specifically as follows: Sampling tasks, from the generated set of scheduled tasks Randomly sample a batch of tasks, for example .
[0058] Cloning the model for each batch of tasks , will the current meta-parameters Make a copy to obtain the task-specific parameters. .
[0059] Quick updates, use tasks Dedicated experience replay buffer A small amount of data from (which stores all transition samples generated during the historical execution days of this task) (i.e., sampling a batch of transition samples, where each transition sample is a quadruple (s, a, r, s', i.e., state, action, reward, and next state) is used to... Perform several steps of standard DDPG gradient updates, simultaneously updating the Actor and Critic networks, to obtain the adapted parameters. :
[0060] In the formula, It is the inner loop learning rate; This is the total loss of DDPG on this task.
[0061] The inner loop step simulates the process by which an agent quickly adapts to a new scenario.
[0062] The outer loop update process is specifically as follows: Assess the effectiveness of adaptation for each task. Use the adapted parameters Calculate the loss on a new batch of data for this task (different from the data used for the inner loop update). This step is crucial because it evaluates the generalization performance of the "rapid adaptation" strategy on new data, rather than its ability to memorize information.
[0063] Update the meta-parameters based on the performance after adaptation across all sampling tasks, following the direction of "how to initialize to achieve the best rapid adaptation effect". :
[0064] In the formula, It is the outer loop learning rate.
[0065] The gradient calculation here involves the second derivative; to improve efficiency, a first-order approximation is used, namely:
[0066] This step is the "Learn to Learn" process, which continuously refines and encodes general knowledge for dealing with different uncertainty patterns into meta-parameters. middle.
[0067] The final output is the optimal meta-parameters after training convergence. This set of parameters is the core element needed for rapid adaptation.
[0068] Use trained meta-parameters to address new uncertainties A small number of DDPG updates are performed, which is the inner loop update in the MAML framework (the update uses a small amount of data from the new scenario to perform several steps (usually 1-5 steps) of gradient updates on the meta-parameters, allowing the model to quickly adapt to the new task; one inner loop update refers to performing one gradient descent update), to quickly generate optimized scheduling instructions (the specific control actions calculated by the strategy for the current state at a specific moment), controlling the source-grid-load-storage equipment (adjusting unit output, energy storage charging and discharging power, and load regulation). At the same time, the system operation data is monitored (here, the actual system operation data is monitored as a short-term, online feedback loop, used to instantly evaluate the execution effect of the scheduling instructions (such as whether limits are exceeded, whether the regulation capacity is sufficient), and the model parameters are corrected online accordingly (such as adjusting the clustering loss weights), which is part of the closed-loop optimization) and the actual regulation capacity of the system is fed back to correct the model parameters, forming a closed-loop optimization, and the optimal scheduling strategy is obtained after correction.
[0069] In a preferred embodiment of this invention, the actual adjustment capability of the feedback system includes: the maximum adjustable power of wind power, photovoltaic power, and load, as well as the SOC safety boundary of energy storage. The maximum adjustable power is expressed as:
[0070] In the formula, Maximum adjustable power; These represent the maximum adjustable power of wind power, the maximum adjustable power of photovoltaic power, and the maximum adjustable power of the load at the current moment, respectively.
[0071] The energy storage SOC safety boundary is represented as follows:
[0072] In the formula, This serves as the lower boundary of the safety boundary for energy storage SOCs. This is the upper boundary of the safety boundary of the energy storage SOC.
[0073] The modified model parameters include: adjusting the weights of the loss function in the improved deep embedding clustering algorithm and updating the outer loop learning rate of the MAML meta-learning framework, wherein the adjustment of the loss function weights uses adaptive coefficients to dynamically adjust according to the feedback error, such as... , These are adaptive coefficients (adjusted based on feedback error). The outer loop learning rate employs a decay strategy to improve stability, such as... .
[0074] The system outputs the optimal scheduling strategy and deploys it to the system. It collects real-time system operation data (such as actual source load values, scheduling costs, carbon emissions, etc.). This data accumulation is a long-term, continuous process used to periodically (e.g., weekly / monthly) trigger model retraining and incremental learning to update the deep autoencoder, DDPG, and meta-learning models, adapting to long-term system changes. Regular retraining is triggered, and incremental learning updates model parameters at each step, including the deep autoencoder, DDPG base model parameters, and meta-learning model. Uncertainty metric weights and reward functions are recalibrated to improve adaptability to extreme conditions. For example, the model is retrained with new data periodically to ensure continuous strategy optimization and maintain efficient, low-carbon system operation.
[0075] To verify the effectiveness and superiority of the method provided in this embodiment, some specific examples are provided below: Experimental environment: Based on the IEEE 30-node system modification, wind power, photovoltaic, energy storage and adjustable loads are connected.
[0076] Key parameters: Number of clusters K=5, hidden layer dimension of autoencoder=16. DDPG: Both Actor / Critic networks are 3 fully connected layers with a learning rate of 0.001. MAML: Inner loop learning rate α=0.01, outer loop learning rate β=0.001.
[0077] Experimental results comparison: Under the new uncertainty scenario, the method described in this embodiment achieves the effects of reducing the operating cost throughout the day, increasing the renewable energy consumption rate, and reducing the power imbalance penalty. Stable scheduling can be achieved with only a few updates, while the standard DDPG requires more than 2,000 steps.
[0078] Example 2 Accordingly, this embodiment provides a highly efficient source-grid-load-storage optimization scheduling system that adapts to source-load uncertainty, such as... Figure 3 As shown, the system for implementing the method as described in any embodiment of the present invention includes a data input layer, a task partitioning module, an optimization scheduling module, an online application module, and a system deployment layer. The data input layer is used to acquire historical source-load prediction and actual data of the power grid to calculate the source-load prediction error sequence. The task partitioning module is used to learn the latent space representation based on the source load prediction error sequence using a deep autoencoder and combined with uncertainty measure. It generates a set of clusters in the latent space through an improved deep embedding clustering algorithm, which serves as a set of highly cohesive scheduling tasks. The scheduling module is optimized to build a basic DDPG model based on a set of scheduled tasks. The training method of the basic DDPG model is improved by applying the MAML meta-learning framework to train meta-parameters. The online application module is used to update the trained meta-parameters in an inner loop to quickly generate optimized scheduling instructions for new uncertain scenarios. At the same time, it monitors the system operation data and provides feedback on the actual adjustment capability of the system to correct the model parameters. After correction, the optimal scheduling strategy is obtained. The system deployment layer is used to output the optimal scheduling strategy and deploy it to the system, collect system operation data in real time and trigger retraining periodically, and update the model parameters in each step through incremental learning.
[0079] Example 3 This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.
[0080] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.
[0081] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0082] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0083] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0084] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A highly efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty, characterized in that, Includes the following steps: Obtain historical source load prediction and actual data of the power grid to calculate the source load prediction error sequence; Based on the source load prediction error sequence, a deep autoencoder is used to learn the latent space representation, and combined with uncertainty measure, an improved deep embedding clustering algorithm is used to generate a set of clusters in the latent space as a set of highly cohesive scheduling tasks. A basic DDPG model is constructed based on a set of scheduled tasks. The training method of the basic DDPG model is improved by applying the MAML meta-learning framework to train meta-parameters. For new uncertain scenarios, the trained meta-parameters are used to update the inner loop to quickly generate optimized scheduling instructions. At the same time, the system operation data is monitored and the actual adjustment capability of the system is fed back to correct the model parameters. After correction, the optimal scheduling strategy is obtained. Output the optimal scheduling strategy and deploy it to the system. Collect system operation data in real time and trigger retraining periodically. Update the model parameters in each step through incremental learning.
2. The efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty according to claim 1, characterized in that, The method of learning the latent space representation based on historical prediction error sequences using a deep autoencoder specifically involves: Samples were segmented according to the operating day, with each sample... The source load prediction error sequence representing one running day is a sequence of dimensions. The matrix, where In terms of time dimension, The error dimension includes wind power, photovoltaic, and load error variables; a deep autoencoder is used to learn the latent space representation as follows: In the formula, It is a sample The latent space vector; It is its reconstructed output; It is the set of learnable parameters of the encoder network; It is the set of learnable parameters of the decoder network; It is an encoder; It is a decoder; the depth autoencoder minimizes the reconstruction loss. Conduct pre-training.
3. The efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty according to claim 1, characterized in that, The joint information entropy of wind power, solar power, and load forecasting errors is calculated for each sample as a measure of uncertainty, expressed as: In the formula, As a measure of uncertainty; For the sample , representing the source load prediction error sequence for one running day, is a sequence with dimension . The matrix, where In terms of time dimension, The error dimensions include wind power, photovoltaic, and load error variables; It is a sample Information entropy; The probability distribution is estimated from the sample using kernel density estimation. It was obtained by fitting the error data corresponding to the operating day.
4. The efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty according to claim 2, characterized in that, The loss function used in the improved deep embedding clustering algorithm is expressed as: In the formula, The reconstruction loss is used to ensure the representational capability of the latent space; This represents the KL divergence loss, used to compute the latent space vector. With cluster center similarity And optimize its distribution with the target distribution. The KL divergence is used to force data points closer to a cluster center; This represents the consistency loss, achieved by minimizing the uncertainty metric of samples within the cluster. Compared with the average uncertainty measure of the cluster The differences are used to achieve this.
5. The efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty according to claim 1, characterized in that, The specific steps for constructing the DDPG basic model based on the set of scheduled tasks are as follows: Design the state space of the DDPG basic model: In the formula, For state space; Indicates the current time The source load prediction error vector is a direct input for dealing with uncertainty; This indicates the real-time output of wind power and solar power, as well as the real-time load demand. This indicates the real-time state of charge of the energy storage system; This indicates the system's real-time marginal electricity price or time-of-use electricity price; This represents the sine and cosine encoding of a 24-hour day; Design the motion space of the DDPG basic model: In the formula, For action space; This indicates the output adjustment of the conventional unit relative to the baseline plan; This indicates the charging and discharging power of the energy storage system, where positive represents discharging and negative represents charging. This represents the adjustable load change achieved through demand response; Design the reward function for the DDPG basic model: In the formula, For the reward function; Indicates operating costs, This is the unit price of electricity purchased on the main network. It is the power purchased from the mainnet at time t. This is the unit power generation cost of traditional generating units. This is the benchmark output of a traditional generator set at time t. It is the unit charge-discharge degradation cost of energy storage, including the cost of purchasing electricity from the main grid, the cost of generating electricity from traditional units, and the cost of energy storage degradation; Indicates a penalty for power imbalance. It is the total power provided by the main network. This is the total power demand of the distribution network; This indicates punishment for exceeding the limits; This indicates a carbon emission penalty; This represents a weighting coefficient used to balance the proportions of economic efficiency, safety, and low carbon emissions.
6. The efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty according to claim 1, characterized in that, The method for improving the training of the DDPG base model using the MAML meta-learning framework, specifically for meta-parameter training, is as follows: Each cluster generated in the latent space by the improved deep embedding clustering algorithm This corresponds to a meta-learning task. The task environment is defined by the statistical characteristics of the samples in the cluster, the task data consists of the DDPG base model, and the task objective is to learn to effectively cope with clusters. Scheduling strategies for the specific uncertainty patterns they represent; The objective function of the MAML meta-learning framework is expressed as: In the formula, It is a meta-parameter; The model in the task New parameters after adaptation; It is the loss function for this task; The MAML meta-learning framework trains the meta-parameters of the scheduling strategy through inner-loop task-specific adaptation and outer-loop meta-knowledge aggregation.
7. The efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty according to claim 6, characterized in that, The inner loop update process is specifically as follows: From the generated set of scheduling tasks A batch of tasks is randomly sampled from the data. For each batch of tasks , will the current meta-parameters Make a copy to obtain the task-specific parameters. ; Use task Experience replay buffer A batch of transferred samples from the sampling process, for Perform several steps of standard DDPG gradient updates to obtain the adapted parameters. ; The outer loop update process is specifically as follows: For each task Use the adapted parameters Calculate the loss for a new batch of data for this task that is different from the data used for updating the inner loop. ; Based on the performance of all sampling tasks after adaptation, the meta-parameters are updated following the direction of "how to initialize to achieve the best fast adaptation effect". ; The final output of the optimal meta-parameters after training convergence serves as the core for rapid adaptation.
8. The efficient adaptive source-grid-load-storage optimization scheduling method for source-load uncertainty according to claim 1, characterized in that, The actual adjustment capability of the feedback system includes: the maximum adjustable power of wind power, photovoltaic power and load, and the safety boundary of energy storage SOC; The modified model parameters include: adjusting the weights of the loss function in the improved deep embedding clustering algorithm and updating the outer loop learning rate of the MAML meta-learning framework. The weights of the loss function are adjusted using adaptive coefficients that are dynamically adjusted based on the feedback error, and the outer loop learning rate is adjusted using a decay strategy to improve stability.
9. A highly efficient adaptive source-grid-load-storage optimization scheduling system for source-load uncertainty, characterized in that, The system is used to implement the method as described in any one of claims 1 to 8, and includes a data input layer, a task partitioning module, an optimization scheduling module, an online application module, and a system deployment layer; The data input layer is used to acquire historical source-load prediction and actual data of the power grid to calculate the source-load prediction error sequence. The task partitioning module is used to learn the latent space representation based on the source load prediction error sequence using a deep autoencoder and combined with uncertainty measure. It generates a set of clusters in the latent space through an improved deep embedding clustering algorithm, which serves as a set of highly cohesive scheduling tasks. The scheduling module is optimized to build a basic DDPG model based on a set of scheduled tasks. The training method of the basic DDPG model is improved by applying the MAML meta-learning framework to train meta-parameters. The online application module is used to update the trained meta-parameters in an inner loop to quickly generate optimized scheduling instructions for new uncertain scenarios. At the same time, it monitors the system operation data and provides feedback on the actual adjustment capability of the system to correct the model parameters. After correction, the optimal scheduling strategy is obtained. The system deployment layer is used to output the optimal scheduling strategy and deploy it to the system, collect system operation data in real time and trigger retraining periodically, and update the model parameters in each step through incremental learning.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.