Power distribution network wind and light storage capacity optimization method considering multi-microgrid energy storage cooperation

Abstract

The invention belongs to the field of microgrid resource capacity optimization. The invention provides a power distribution network wind and light storage capacity optimization method considering multi-microgrid energy storage cooperation. The method comprises the following steps: step 1, establishing a wind and light combined operation power information data set considering spatial correlation during multi-microgrid source load fluctuation; and step 2, multi-microgrid wind and light storage capacity configuration optimization is realized by using a reinforcement learning algorithm. And step 3, complementing the wind-light fluctuation scene with few samples by using a transfer learning algorithm. Based on deep fusion space-time correlation modeling, multi-agent reinforcement learning and cross-domain transfer learning, a multi-microgrid energy storage collaborative optimization framework with dynamic adaptive capacity is provided. According to the method, the deep association rule of the multi-dimensional operation data of the micro-grid group can be analyzed, global optimal capacity configuration is realized through a coevolution mechanism of an intelligent algorithm, and a brand new solution is provided for solving the problem of power distribution network optimization under high-proportion new energy access.

Description

Technical Field

[0001] The present invention relates to the field of microgrid resource capacity optimization, and in particular to a method for optimizing wind and solar energy storage capacity in a distribution network by considering multi-microgrid energy storage collaboration. Background Art

[0002] As the penetration of high-proportion renewable energy in distribution networks continues to increase, the high volatility and spatiotemporal heterogeneity of distributed wind and solar power sources have led to prominent issues such as insufficient grid regulation margin and node voltage overshooting. Especially in scenarios where multiple microgrids are interconnected, the dynamic interactions between source, storage, and load subsystems exhibit high-dimensional nonlinear coupling characteristics. Traditional capacity allocation paradigms based on independent optimization of single microgrids struggle to meet global economic and robustness requirements. Research on the coordinated optimization of multi-microgrid energy storage is not only a key path to achieving flexible regulation and efficient resource utilization in new power systems, but also a significant technological breakthrough in promoting deep collaboration across the "source, grid, load, and storage" network of the Energy Internet. This research holds significant engineering value for reducing the cost of renewable energy integration and enhancing grid resilience.

[0003] Current research focuses on single-microgrid optimization or simple cluster scheduling, which has the following problems: independent optimization models ignore the temporal and spatial complementarity between microgrids, resulting in repeated configuration of energy storage capacity and mismatched response actions; traditional mathematical programming methods are difficult to accurately characterize the temporal and spatial correlation characteristics of wind and solar power output and load fluctuations, causing the coordinated scheduling strategy to deviate from the actual operating scenario; existing methods are limited by the sparsity and regional differences of wind and solar power fluctuation scenario samples, and cannot construct an optimization model with cross-regional generalization capabilities. Summary of the Invention

[0004] 1. Technical problems to be solved:

[0005] How to solve the problem of coordinated optimization of multiple microgrids under the condition of high proportion of renewable energy access.

[0006] 2. Technical solution:

[0007] In order to solve the above problems, the present invention provides a method for optimizing the wind and solar storage capacity of a distribution network considering the coordination of multi-microgrid energy storage, including the following steps:

[0008] Step 1: Establish a wind-solar joint operation power information dataset that considers the spatial correlation of source and load fluctuations of multiple microgrids.

[0009] Step 2: Use reinforcement learning algorithm to optimize the configuration of wind, solar and storage capacity in multiple microgrids.

[0010] Step 3: Use transfer learning algorithm to complete the wind and solar fluctuation scenario with few samples.

[0011] Furthermore, the step 1 is specifically as follows:

[0012] S101: Multi-dimensional operation data collection and preprocessing.

[0013] S102: Spatiotemporal correlation feature analysis and quantification of feature data.

[0014] S103: Joint operation scenario database construction.

[0015] Furthermore, the step 2 is specifically as follows:

[0016] S201: Construction of multi-agent collaborative optimization framework.

[0017] S202: Multidimensional state-action space mapping design.

[0018] S203: Development of compound reward function mechanism.

[0019] S204: Distributed strategy training and experience sharing.

[0020] Furthermore, the step 3 is specifically as follows:

[0021] S301: Cross-region scene feature matching analysis.

[0022] S302: Dynamic adaptation training of the transfer learning model.

[0023] S303: Virtual scene generation and data enhancement.

[0024] S304: Closed-loop verification and iteration of migration effects.

[0025] 3.Beneficial effects:

[0026] This paper proposes a dynamically adaptable multi-microgrid energy storage collaborative optimization framework based on the deep integration of spatiotemporal correlation modeling, multi-agent reinforcement learning, and cross-domain transfer learning. This approach not only analyzes the underlying correlation patterns in multi-dimensional operational data across microgrids but also achieves globally optimal capacity allocation through the co-evolution of intelligent algorithms, providing a novel solution to the challenge of optimizing distribution networks under high-proportion renewable energy access. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of the present invention.

[0028] Figure 2 It is a flow chart of step 1 in the present invention.

[0029] Figure 3 It is a flow chart of step 2 in the present invention.

[0030] Figure 4 It is a flow chart of step 3 in the present invention.

[0031] Figure 5This is a diagram showing the effect of optimized configuration of energy storage capacity.

[0032] Figure 6 It is a rendering of the effect of generating a few-sample scene. DETAILED DESCRIPTION

[0033] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0034] like Figure 1 As shown, a method for optimizing wind and solar storage capacity in a distribution network considering multi-microgrid energy storage coordination includes the following steps:

[0035] Step 1: Establish a wind-solar joint operation power information dataset that considers the spatial correlation of source and load fluctuations of multiple microgrids.

[0036] Step 2: Use reinforcement learning algorithm to optimize the configuration of wind, solar and storage capacity in multiple microgrids.

[0037] Step 3: Use transfer learning algorithm to complete the wind and solar fluctuation scenario with few samples.

[0038] like Figure 2 As shown, the step 1 is specifically as follows:

[0039] S101: Multi-dimensional operation data collection and preprocessing, specifically:

[0040] Using synchronized phasor measurement units (PMUs) and microgrid energy management systems deployed at distribution network nodes, operational data such as wind and solar power output, load curves, and energy storage system status from multiple microgrids is collected with a temporal resolution of 15 minutes. Outlier detection and interpolation are performed on the raw data, and a sliding window method is used to eliminate measurement noise. Z-score standardization is used to standardize the data dimensions across different microgrids. Geographic coordinates, meteorological station data, and grid topology information are also collected for each microgrid to construct a benchmark parameter set for spatiotemporal correlation analysis between wind and solar power.

[0041] S102: Spatiotemporal correlation feature analysis and quantification of feature data, specifically:

[0042] Analyze the spatial correlation of wind and solar power output fluctuations between different microgrids, including the correlation of wind power output in the same wind speed corridor and the correlation of photovoltaic output in similar areas, and obtain quantitative characteristics of multi-dimensional operation data. In one embodiment, the specific method is:

[0043] S10201: Calculate the wind power output covariance matrix in the same wind speed corridor and the photovoltaic output covariance matrix in the similar area; specifically:

[0044]

[0045] Among them, C w，ij: the output covariance of wind farms i and j in the same wind speed corridor; P w，i (t): normalized output of wind farm i at time t; μ w，i : Time series mean of wind farm i output; μ w，j : the time series mean of wind farm j’s output; T: the total number of points in the time window.

[0046]

[0047] Among them, C pv，ld : output covariance of photovoltaic power stations k and l; P pv，k (t): normalized output of photovoltaic power station k at time t; μ pv，k : wind farm k output time series mean; μ pv，l : Time series mean of wind farm output.

[0048] S10202: The dynamic time warping (DTW) algorithm is used to quantify the temporal misalignment correlation of load fluctuations and reveal the alternating pattern of load peaks and valleys between regions.

[0049] S10203: Introduce the spatial weight model of the Geographic Information System (GIS) to establish quantitative results of spatial data association laws that take into account spatial differences.

[0050] S103: Joint operation scenario database construction, specifically:

[0051] The pre-processed multi-source data is multi-dimensionally fused with spatiotemporal correlation features to generate a joint operation scenario library with spatiotemporal labels. Specifically, this includes: constructing a cross-microgrid wind / solar output-load fluctuation joint distribution histogram, annotating the spatiotemporal correlation intensity level of each scenario; establishing a storage response delay feature mapping table to record the transmission delay of charging and discharging instructions at different spatial distances; and developing a typical scenario extraction module based on time series clustering, using the DBSCAN density clustering algorithm to summarize the massive data into a representative set of operation modes.

[0052] like Figure 3 As shown, the step 2 is specifically as follows:

[0053] S201: Construction of a multi-agent collaborative optimization framework, specifically:

[0054] A multi-agent reinforcement learning architecture is established for heterogeneous microgrids, defining each microgrid's energy storage system as an independent agent. A two-layer communication mechanism is designed, consisting of a grid control center and local microgrid controllers. The central coordinator aggregates global state information, while the local controllers perform distributed decision-making.

[0055] S202: Multi-dimensional state-action space mapping design, specifically:

[0056] The agent's observation state space is defined to include the microgrid's own energy storage SOC, the predicted wind and solar output probability for the next two hours, the load demand trend of adjacent microgrids, and the priority of grid dispatch instructions. The action space is divided into a discrete-continuous hybrid dimension, including three controllable variables: energy storage charging and discharging power decision gear, inter-microgrid power support request intensity level, and reserve capacity reservation ratio. Key state indicators are selected through feature importance analysis to eliminate observation data redundancy.

[0057] S203: Development of compound reward function mechanism,

[0058] The design incorporates a triple-incentive evaluation system based on economics, safety, and timeliness. The economic metric calculates the difference between energy storage loss costs and the grid's electricity purchase fee; the safety metric assesses the node voltage deviation rate and the duration of power exceeding limits; and the timeliness metric quantifies the impact of command response delay on the grid's compliance with regulation requirements. A dynamic weight adjustment module is introduced to automatically adjust the weight coefficients of each metric based on the grid's operating status, ensuring that the training process converges to a Pareto optimal strategy.

[0059] In one embodiment, the compound reward function is specifically:

[0060] R total (t) = w econ (t)·R econ (t)+w safe (t)·R safe (t))+w time (t)·R time (t)

[0061] Among them, R total (t): the compound reward value at time t, reflecting the comprehensive optimization effect; w econ , w safe , w time : Economic weight coefficient, safety weight coefficient, timeliness weight coefficient.

[0062] R econ (t): Economic reward function, specifically:

[0063]

[0064] Among them, C base is the benchmark electricity purchase cost (yuan), taking the 95% percentile value of historical data; C purchase (t) is the electricity purchase cost of the power grid during period t; C deg (t) is the energy storage loss cost during period t.

[0065] R safe (t): Safety reward function, specifically:

[0066]

[0067] Among them, V i (t) is the per-unit voltage value of node i; V ref is the voltage reference value; T violate (t) is the cumulative time of exceeding the limit in period t; ΔV max is the maximum allowable voltage deviation; T total is the total duration of the time period; n is the number of nodes.

[0068] R time (t): Time-sensitive reward function, specifically:

[0069] R time (t) = e -λτ(t)

[0070] Where τ(t) is the command response delay and λ is the attenuation coefficient.

[0071] S204: Distributed strategy training and experience sharing, specifically:

[0072] Parallel training is performed using a GPU cluster, with each agent deploying its own independent policy and value networks. A cross-agent experience replay pool mechanism is designed to store state-action-reward tuples with temporal and spatial correlation. Policy distillation technology is developed to refine the global optimization knowledge of the central coordinator into a lightweight policy model, which is regularly updated in synchronization with the local controller.

[0073] like Figure 4 As shown, the step 3 is specifically as follows:

[0074] S301: Cross-region scene feature matching analysis, specifically:

[0075] Based on a comprehensive historical scenario library, a source domain feature space is constructed. Wind and solar power output fluctuation patterns, load timing characteristics, and energy storage response characteristics are extracted to form a multidimensional feature vector. The morphological similarity between the target domain's few-sample scenarios and the source domain's scenarios is calculated, and a correlation mapping matrix for cross-regional scenarios is established. Feature importance ranking is used to select core feature dimensions with strong transferability, eliminating regional noise interference.

[0076] S302: Dynamic adaptation training of the transfer learning model, specifically:

[0077] A dual-channel deep transfer network architecture is designed. The source domain channel loads the parameters of a pre-trained reinforcement learning optimization model, while the target domain channel receives a small number of local samples. A domain adapter module is developed to embed an adversarial domain discriminator in the feature extraction layer. This mechanism uses a gradient reversal mechanism to force the network to learn domain-invariant features. A progressive fine-tuning strategy is employed to gradually unfreeze high-level network parameters to adapt to the unique fluctuation patterns of the target domain, preserving the integrity of the transferred general optimization knowledge from the source domain.

[0078] S303: Virtual scene generation and data enhancement, specifically:

[0079] A typical generative adversarial network was constructed, using the scenario feature vectors output by the transfer model as the generation condition input. A spatiotemporal consistency constraint module was designed to ensure that the wind-solar-load fluctuations in the generated scenarios conformed to the geographical and meteorological patterns of the target region. Latent space interpolation techniques were used to synthesize transition scenarios between the source and target domains, creating an enhanced scenario library covering operating conditions such as extreme fluctuations, typical days, and seasonal transitions.

[0080] S304: Closed-loop verification and iteration of migration effects

[0081] A two-way evaluation mechanism was established: Forward validation involves inputting virtual scenarios generated in the target domain into the reinforcement learning model to test the feasibility of the capacity allocation strategy; backward validation involves inverting the optimization results back into the source domain model to verify the physical consistency of the knowledge transfer. A scenario fidelity index was developed to quantify the deviation between the generated scenario's core metrics, such as temporal volatility and spatial correlation, and the real data. When the deviation exceeds 5%, an adaptive correction module is triggered to dynamically adjust the generator's attention weight distribution until the requirement for seamless cross-domain scenario switching is met.

[0082] Example:

[0083] Background parameters include:

[0084] Geographical location: Regions A (microgrid 1), B (microgrid 2), and C (microgrid 3) in a certain province are selected to form a triangular power supply area, with Region B located at the hub of the wind speed corridor. Wind and solar power installed capacity: Each microgrid is configured with wind power (30MW, 20MW, 25MW) + photovoltaic power (25MW, 15MW, 20MW). Load characteristics: Industrial load accounts for 58%, peak-to-valley difference rate is 32%, and daily maximum load is (85MW, 62MW, 73MW). Energy storage parameters: Lithium battery, cycle efficiency 92%, life cycle 6000 times, SOC limit [0.2, 0.95]; Simulation period: Full-year data for 2023, time resolution 15 minutes.

[0085] Figure 5 It shows that the present invention reduces the total configuration capacity by 26.1% through inter-microgrid energy storage coordination, among which region B (MG2) has the most significant capacity optimization due to its location at the hub of the wind power corridor.

[0086] Figure 6 The ability of transfer learning to complete small-sample scenarios was verified, and the correlation coefficient between the generated scenarios and real data reached 0.87 (p<0.01), meeting the requirements of engineering applications.

[0087] This paper systematically proposes a method for collaborative optimization of multi-microgrid energy storage under high-proportion renewable energy access by integrating spatiotemporal correlation modeling, multi-agent collaborative optimization, and cross-domain transfer learning. First, based on spatial correlation modeling of wind speed corridors and irradiation similarity zones (step 1-S10201), it effectively captures the spatiotemporal complementarity of distributed power sources, reducing energy storage capacity requirements. Second, a multi-agent deep reinforcement learning architecture (step 2-S201) achieves Nash equilibrium optimization of inter-microgrid energy storage strategies by designing action spaces and dynamic weighted reward functions, improving grid regulation efficiency while reducing capacity allocation. Finally, combined with virtual scenario generation technology, it improves the accuracy of optimization decisions in low-sample scenarios.

Claims

1. A method for optimizing wind and solar storage capacity in a distribution network considering multi-microgrid energy storage coordination, comprising the following steps: Step 1: Establish a wind-solar joint operation power information dataset that considers the spatial correlation of multi-microgrid source and load fluctuations; Step 2: Optimize the configuration of wind, solar and storage capacity in multiple microgrids using reinforcement learning algorithms; Step 3: Use transfer learning algorithm to complete the wind and solar fluctuation scenario with few samples.

2. The method for optimizing wind and solar storage capacity in a distribution network considering multi-microgrid energy storage coordination according to claim 1, characterized in that: The step 1 is specifically as follows: S101: Multi-dimensional operational data acquisition and preprocessing: Synchronized phasor measurement devices and microgrid energy management systems deployed at distribution network nodes are used to collect wind and solar power output, load curves, and energy storage system status operational data from multiple microgrids. The time resolution is set to 10-20 minutes. Outlier detection and interpolation repair are performed on the raw data. A sliding window method is used to eliminate measurement noise. The data dimensions of different microgrids are standardized using Z-score. The geographic coordinates, meteorological station data, and grid topology information of each microgrid are also collected to construct a benchmark parameter set for spatiotemporal correlation analysis between wind and solar power. S102: Spatiotemporal Correlation Feature Analysis and Quantitative Feature Data: Analyze the spatial correlation of wind and solar power output fluctuations between different microgrids, including the correlation of wind power output in the same wind speed corridor and the correlation of photovoltaic output in similar areas, to obtain quantitative features of multi-dimensional operating data; S103: Joint operation scenario database construction: The pre-processed multi-source data is multi-dimensionally integrated with the spatiotemporal correlation features to generate a joint operation scenario library with spatiotemporal labels.

3. The method for optimizing wind and solar storage capacity in a distribution network considering multi-microgrid energy storage coordination as claimed in claim 2, characterized in that: In step S102, the specific method is: S10201: Calculate the wind power output covariance matrix in the same wind speed corridor and the photovoltaic output covariance matrix in the similar area; specifically: Among them, C w，ij : the output covariance of wind farms i and j in the same wind speed corridor; P w，i (t): normalized output of wind farm i at time t; μ w，i : Time series mean of wind farm i output; μ w，j : the time series mean of wind farm j output; T: the total number of points in the time window; Among them, C pv，ld : output covariance of photovoltaic power stations k and l; P pv，k (t): normalized output of photovoltaic power station k at time t; μ pv，k : the time series mean of wind farm k output; μ pv，l :Wind farm l output time series mean S10203:Introduce the spatial weight model of geographic information system and establish the quantitative results of spatial data association law considering spatial differences.

4. The method for optimizing wind and solar storage capacity in a distribution network considering multi-microgrid energy storage coordination according to claim 2, characterized in that: In step 103, the joint operation scenario library with spatiotemporal labels specifically includes: constructing a joint distribution histogram of wind and solar power output and load fluctuations across microgrids, marking the spatiotemporal correlation intensity level of each scenario; establishing a storage response delay feature mapping table to record the transmission delay of charging and discharging instructions at different spatial distances; developing a typical scenario extraction module based on time series clustering, and summarizing massive data into a representative set of operation modes through the DBSCAN density clustering algorithm.

5. The method for optimizing wind and solar storage capacity in a distribution network considering multi-microgrid energy storage coordination according to claim 1, characterized in that: The step 2 is specifically as follows: S201: Multi-agent collaborative optimization framework construction: Build a multi-agent reinforcement learning architecture based on heterogeneous microgrid clusters, define the energy storage system of each microgrid as an independent agent, and include a two-layer communication mechanism between the grid control center and the microgrid local controllers. The central coordinator is responsible for aggregating global state information, and the local controllers perform distributed decision-making. S202: Multidimensional state-action space mapping design: Define the agent's observation state space to include the microgrid's energy storage SOC, the wind and solar output probability forecast for the next 1.5-2.5 hours, the load demand trend of adjacent microgrids, and the grid dispatch instruction priority. The action space is divided into a discrete-continuous hybrid dimension, including three types of controllable variables: energy storage charging and discharging power decision gear, cross-microgrid power support request intensity level, and reserve capacity reservation ratio. Key state indicators are selected through feature importance analysis to eliminate observation data redundancy. S203: Development of a composite reward function mechanism: Design a triple reward evaluation system encompassing economy, safety, and timeliness: The economy indicator calculates the difference between energy storage loss cost and grid power purchase costs; the safety indicator evaluates node voltage deviation rate and power limit duration; and the timeliness indicator quantifies the impact of command response delay on grid regulation demand satisfaction. A dynamic weight adjustment module is introduced to automatically adjust the weight coefficients of each indicator based on the grid operating status to ensure that the training process converges to the Pareto optimal strategy. S204: Distributed strategy training and experience sharing: Use GPU clusters for parallel training, deploy independent strategy networks and value networks for each agent, design a cross-agent experience replay pool mechanism, store state-action-reward tuples with spatiotemporal correlation, develop strategy distillation technology, refine the global optimization knowledge of the central coordinator into a lightweight strategy model, and regularly synchronize updates to the local controller.

6. The method for optimizing wind and solar storage capacity in a distribution network considering multi-microgrid energy storage coordination according to claim 5, characterized in that: In step S203, the composite reward function is specifically: R total (t)=w econ (t)·R econ (t)+w safe (t)·R safe (t)+w time (t)·R time (t) Among them, R total (t): the compound reward value at time t, reflecting the comprehensive optimization effect; w econ , w safe , w time : Economic weight coefficient, safety weight coefficient, timeliness weight coefficient; R econ (t): Economic reward function R econ (t) Specifically: Among them, C base is the benchmark electricity purchase cost (yuan), taking the 95% percentile value of historical data; C purchase (t) is the electricity purchase cost of the power grid during period t; C deg (t) is the energy storage loss cost during period t; Safety reward function R safe (t) Specifically: Among them, V i (t) is the per-unit voltage value of node i; V ref is the voltage reference value; T violate (t) is the cumulative duration of the limit violation during period t; ΔV max is the maximum allowable voltage deviation; T total is the total duration of the period; n is the number of nodes; Time-sensitive reward function R time (t) Specifically: R time (t)=e -λ·τ (t) Where τ(t) is the command response delay and λ is the attenuation coefficient.

7. The method for optimizing wind and solar storage capacity in a distribution network considering multi-microgrid energy storage coordination according to claim 1, characterized in that: The step 3 is specifically as follows: S301: Cross-regional scenario feature matching analysis: Based on a complete historical scenario library, a source domain feature space is constructed. Wind and solar power output fluctuation patterns, load timing characteristics, and energy storage response characteristics are extracted to form a multi-dimensional feature vector. The morphological similarity between the target domain few-sample scenario and the source domain scenario is calculated. An association mapping matrix for cross-regional scenarios is established. Core feature dimensions with strong transfer capabilities are selected through feature importance ranking to eliminate regional noise interference. S302: Dynamic Adaptive Training of Transfer Learning Models: Design a dual-channel deep transfer network architecture. The source domain channel loads pre-trained reinforcement learning optimization model parameters, while the target domain channel accesses a small number of local data samples. Develop a domain adapter module, embed an adversarial domain discriminator in the feature extraction layer, and use a gradient reversal mechanism to force the network to learn domain-invariant features. Adopt a progressive fine-tuning strategy to gradually unfreeze high-level network parameters to adapt to the unique fluctuation patterns of the target domain, preserving the transfer integrity of the general optimization knowledge in the source domain. S303: Virtual Scenario Generation and Data Augmentation: A typical generative adversarial network is constructed, using the scenario feature vectors output by the migration model as generation condition inputs. A spatiotemporal consistency constraint module is designed to ensure that the wind-solar-load fluctuations in the generated scenarios conform to the geographical and meteorological patterns of the target region. Latent space interpolation techniques are used to synthesize transition scenarios between the source and target domains, forming an enhanced scenario library covering conditions such as extreme fluctuations, typical days, and seasonal transitions. S304: Closed-loop verification and iteration of migration effect: Build a two-way evaluation mechanism: Forward verification inputs the virtual scene generated by the target domain into the reinforcement learning model to test the feasibility of the capacity configuration strategy; reverse verification inverts the optimization results to the source domain model to verify the physical consistency of knowledge migration, develop a scene fidelity index, and quantify the deviation rate between the core indicators such as the temporal volatility and spatial correlation of the generated scene and the real data. When the deviation exceeds 5%, the adaptive correction module is triggered to dynamically adjust the generator's attention weight distribution until the seamless switching requirements of cross-domain scenarios are met.

Images