Distributed energy storage system cooperative control method based on data fusion

Abstract

The invention discloses a distributed energy storage system cooperative control method based on data fusion, and belongs to the field of power system energy storage control. The edge layer performs dimensionality reduction on multi-source data through kernel principal component analysis, the cloud constructs a space-time correlation model by combining a graph neural network with a long-short-term memory network, and data privacy is guaranteed by means of federal learning distributed updating. An upper layer model prediction control framework is combined with an improved whale algorithm to solve multi-target optimization of system economy, energy storage life and power grid stability; the lower-layer power type energy storage adopts self-adaptive droop control, and the energy type energy storage model predicts, controls and quantifies the degradation cost of the battery. When communication is interrupted, the analytic hierarchy process and the consistency algorithm cooperate to disperse power distribution; and when the equipment fails, convolutional neural network diagnosis is combined with alliance chain redundancy switching. Multi-source deep correlation, characteristic differentiation regulation and control and high robustness are realized, the energy storage efficiency is improved, the service life of equipment is prolonged, and the method is suitable for micro-grids, smart grids and other scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy storage control of power systems, and specifically discloses a distributed energy storage system collaborative control method based on data fusion. BACKGROUND

[0002] With the rapid development of new energy generation technology, distributed energy storage systems, as a key link to stabilize renewable energy output fluctuations and ensure stable operation of power grids, have attracted widespread attention in collaborative control technology. Distributed energy storage systems are composed of multiple decentralized energy storage units, which work collaboratively to dynamically adjust power grid power, and have important application value in microgrids, smart grids and other scenarios.

[0003] In the existing collaborative control method of distributed energy storage systems, there are obvious technical limitations in data processing and control strategy design: First, the degree of multi-source data fusion is insufficient. Distributed energy storage system operation involves energy storage device state data (such as state of charge, operating temperature), power grid parameters (such as voltage, frequency), environmental data (such as light, wind speed), and load demand data, etc. The existing technology often only integrates or processes data in isolation, without establishing the internal relationship between different sources of data, resulting in one-sided control decisions that are difficult to adapt to dynamic adjustment needs under complex conditions.

[0004] Second, the flexibility of collaborative control strategy is lacking. Traditional control methods often use fixed thresholds or preset modes for charge and discharge management, without fully considering the differences in characteristics of different types of energy storage devices (such as power and energy storage), and are difficult to respond in real time to rapid changes in new energy output and load demand, which can lead to low utilization efficiency of energy storage resources, rapid wear and tear of equipment, and other problems.

[0005] Third, the system robustness needs to be improved. The existing control architecture relies on centralized decision-making mode, which is prone to global control failure when the communication link is interrupted or some energy storage units fail, making it difficult to ensure stable operation of the system under abnormal conditions, limiting the reliability and applicability of the distributed energy storage system.

[0006] Therefore, how to deeply integrate multi-source data, build a flexible and efficient collaborative control strategy, and improve the operation performance and robustness of the distributed energy storage system has become a technical problem that needs to be solved in this field. The present application is aimed at the shortcomings of the existing technology, and proposes a distributed energy storage system collaborative control method based on data fusion, which establishes a multi-source data correlation model, designs a layered collaborative strategy that adapts to the characteristics of the device, and a distributed autonomous mechanism, to overcome the shortcomings of insufficient data fusion, lack of control flexibility and poor robustness in the existing technology, and achieve efficient collaborative operation of the distributed energy storage system. SUMMARY

[0007] Therefore, the present application aims to overcome the shortcomings of the prior art, and provides a distributed energy storage system collaborative control method based on data fusion, to realize the deep fusion and internal correlation of multi-source data such as energy storage device state, power grid parameters, environmental data and load demand, and to provide comprehensive and accurate decision basis for collaborative control; design a hierarchical collaborative control strategy for different characteristics of power type and energy type energy storage devices, dynamically respond to changes in new energy output and load demand, improve the utilization efficiency of energy storage resources and reduce the life loss of the equipment; introduce a distributed autonomous mechanism to reduce the over-reliance on centralized decision-making, improve the operation stability and reliability of the system under abnormal working conditions such as communication interruption or equipment failure, and thus realize efficient, collaborative and stable operation of the distributed energy storage system.

[0008] In order to achieve the above-mentioned purpose, the present application provides a distributed energy storage system collaborative control method based on data fusion, comprising the following steps: (1) Multi-source data fusion and correlation modeling: the original data is processed by kernel principal component analysis for dimension reduction in the edge layer, and the key feature vector is extracted; a multi-source data spatio-temporal correlation model is constructed by using graph neural network combined with long short-term memory network in the cloud, and a multi-source data feature matrix with fused spatio-temporal features is output; the distributed update of model parameters is realized by using federated learning to protect data privacy; (2) Hierarchical collaborative control strategy: the upper layer global optimization layer is based on the model predictive control framework, and the system economy, energy storage life loss and power grid stability are used as multi-objective optimization functions, and the improved whale algorithm is used to solve the global optimal solution; in the lower layer local autonomous layer, the power type energy storage adopts adaptive droop control combined with reinforcement learning to dynamically adjust the droop coefficient, and the energy type energy storage optimizes the charging and discharging depth based on the model predictive control, and quantifies the battery degradation cost as the objective function; (3) Distributed autonomous mechanism: when the communication is interrupted, the priority queue based on the analytic hierarchy process and the consistency algorithm are used to realize the distributed power distribution; when the equipment fails, the health status is diagnosed by using convolutional neural network, and the redundant switching and state synchronization are realized by using distributed ledger technology.

[0009] Further, the multi-source data includes energy storage device state data, power grid parameters, environmental data and load demand data, the kernel principal component analysis dimension reduction process includes nonlinear mapping to high-dimensional feature space, covariance matrix calculation and feature vector orthogonalization screening; when the edge layer pre-processes the data, outlier rejection and normalization processing are used.

[0010] Further, the input of the graph neural network includes the geographic location of the energy storage unit, the spatial information of the power grid topology, and the feature vector and historical operation data output by the edge layer; the graph neural network adopts the GraphSAGE algorithm for neighbor node sampling, and the aggregation function combines the mean aggregation and the attention mechanism; meanwhile, an energy storage unit-power grid node-environment factor knowledge graph is constructed, and the entity relationship is extracted by the AutoKG model and stored in the Neo4j database.

[0011] Further, the parameter update of the federated learning is realized by local training of each edge node and cloud aggregation, the edge node training adopts the Adam optimizer, and the sensitive parameters are protected by the differential privacy technology during cloud aggregation; the edge layer deployment controller in the system architecture supports multi-protocol communication, and the cloud is deployed in a cluster to support high-concurrency processing.

[0012] Further, the constraint conditions of the upper model predictive control framework include the energy storage power constraint, the state of charge constraint, and the power grid voltage frequency deviation constraint; the improved whale algorithm introduces an adaptive weight factor and a flight strategy to optimize the solving process.

[0013] Further, the adaptive droop coefficient calculation formula of the power-type energy storage is: ; wherein, is the initial droop coefficient of the power-type energy storage unit, is the current state of charge of the power-type energy storage unit, is the rated state of charge of the power-type energy storage unit, is the sensitivity coefficient of the droop coefficient to the state of charge deviation, which is optimized online through deep reinforcement learning.

[0014] Further, the model predictive control optimization objective function of the energy-type energy storage is: ; wherein, is the battery degradation cost of the energy-type energy storage unit at the kth moment; is the weight coefficient of multi-objective optimization; is the actual output power of the energy-type energy storage unit at the kth moment; is the reference power instruction at the kth moment; is the state of charge of the energy-type energy storage unit at the kth moment; is the reference state of charge of the energy-type energy storage unit, is the prediction time domain of the model predictive control, which is dynamically adjusted through fuzzy logic.

[0015] Further, the priority queue indicators during communication interruption include the available power ratio, the state of charge, and the voltage deviation, and the priority is obtained through weighted calculation; the consistency algorithm realizes power distribution through neighbor node interaction and global target adjustment.

[0016] Further, the convolutional neural network fault diagnosis module inputs current, voltage and temperature time series data, and outputs fault probability distribution; the distributed ledger technology adopts a consortium chain architecture, and a consensus mechanism is practical Byzantine fault tolerance, so that device state information is synchronized.

[0017] Further, the dynamic partitioning and cooperation step comprises: based on a density clustering algorithm, a control area is divided according to energy storage unit density and communication time delay; when a load fluctuation threshold in the area is triggered, cross-area cooperation support is started.

[0018] The above technical scheme has at least the following beneficial effects: In the present application, the hierarchical cooperative control strategy is adapted to the characteristics of the device in the architecture of 'upper layer global optimization and lower layer local autonomy': the upper layer model predictive control framework combines the improved whale algorithm to comprehensively solve the multi-objective global optimal solution of system economy, energy storage life loss and power grid stability; the lower layer designs an exponential function type adaptive droop coefficient formula (combined with deep reinforcement learning to optimize the sensitive coefficient) for power type energy storage, so that power regulation is smoother; for energy type energy storage, the battery degradation cost is quantified into the model predictive control objective function, so that the depth of charge and discharge and the life loss are cooperatively optimized, and the utilization efficiency of energy storage resources is improved.

[0019] In the present application, the distributed autonomous mechanism enhances the robustness through 'communication interruption fault tolerance and device failure redundancy': when the communication is interrupted, the priority queue based on the analytic hierarchy process cooperates with the consistency algorithm to realize decentralized power distribution and get rid of the dependence on centralized decision-making; when the device fails, the convolutional neural network diagnoses the health state, the consortium chain synchronizes the state information and triggers the redundancy switching, so that the fault recovery time is shortened and the risk of abnormal operating conditions is reduced.

[0020] In the present application, the adaptive droop coefficient formula of the power type energy storage establishes a continuous mapping relationship between the state of charge deviation and the control parameter in the form of an exponential function, replaces the traditional segmented / linear adjustment, optimizes the sensitive coefficient combined with deep reinforcement learning, avoids the sudden impact of the control parameter, improves the smoothness of power regulation; the model predictive control objective function of the energy type energy storage first couples the battery degradation cost, the power tracking error and the state of charge deviation, breaks through the limitation of single objective optimization, guarantees the power tracking accuracy and significantly prolongs the cycle life of the energy storage.

[0021] In the present application, the dynamic partitioning and cooperation step adaptively divides the control area based on the density clustering algorithm (taking the energy storage unit density and the communication time delay as indexes), automatically starts cross-area support when the load fluctuation threshold in the area is triggered, breaks the rigidity of the preset partition, lets the system dynamically adjust the cooperation range according to the energy storage distribution and the operating conditions, improves the cross-area response capability in the case of local load mutation, and enhances the dynamic cooperation efficiency of the distributed energy storage system. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to make the technical solutions of the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0023] Figure 1 is a flow chart of the data fusion-based distributed energy storage system collaborative control method of the present application. DETAILED DESCRIPTION

[0024] The exemplary embodiments will be described in detail herein with reference to the accompanying drawings. Unless otherwise indicated, the same numbers on different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of apparatuses and methods consistent with some aspects of the present application as detailed in the appended claims.

[0025] Embodiment One

[0026] As shown in Figure 1 , the present embodiment provides a data fusion-based distributed energy storage system collaborative control method, comprising the following steps: (1) Multi-source data fusion and correlation modeling: dimensionality reduction processing of the original data is performed by kernel principal component analysis at the edge layer, and key feature vectors are extracted; a multi-source data spatio-temporal correlation model is constructed by combining a graph neural network with a long short-term memory network at the cloud end, and a multi-source data feature matrix with fused spatio-temporal features is output; distributed updating of model parameters is realized by federated learning, and data privacy is protected; (2) Hierarchical collaborative control strategy: the upper layer global optimization layer is based on a model predictive control framework, takes system economy, energy storage life loss, and power grid stability as multi-objective optimization functions, and uses an improved whale algorithm to solve the global optimal solution; in the lower layer local autonomous layer, power-type energy storage uses adaptive droop control combined with reinforcement learning to dynamically adjust the droop coefficient, and energy-type energy storage optimizes the charge and discharge depth based on model predictive control, and quantifies the battery degradation cost as the objective function; (3) Distributed autonomous mechanism: when communication is interrupted, a priority queue based on the analytic hierarchy process and a consistency algorithm are used to realize decentralized power distribution; when a device fails, a convolutional neural network is used to diagnose the health status, and a distributed ledger technology is used to realize redundancy switching and state synchronization.

[0027] Further, the multi-source data includes energy storage device state data, power grid parameters, environmental data, and load demand data, and the principal component analysis dimension reduction process includes nonlinear mapping to a high-dimensional feature space, covariance matrix calculation, and characteristic vector orthogonalization screening; the edge layer adopts outlier rejection and normalization processing when preprocessing data.

[0028] As an implementation, the input of the graph neural network in the embodiment includes the geographical position of the energy storage unit, the spatial information of the power grid topology, and the feature vectors and historical operation data output by the edge layer; the graph neural network adopts the GraphSAGE algorithm for neighbor node sampling, and the aggregation function combines the mean aggregation and the attention mechanism; meanwhile, an energy storage unit-power grid node-environment factor knowledge graph is constructed, entity relationships are extracted by an AutoKG model, and are stored in a Neo4j database.

[0029] As an implementation, the parameter update of the federated learning in the embodiment is achieved through local training of each edge node and cloud aggregation; the edge node training adopts the Adam optimizer, and the sensitive parameters are protected by the differential privacy technology during cloud aggregation; the edge layer deployment controller in the system architecture supports multi-protocol communication, and the cloud is deployed in a cluster to support high-concurrency processing.

[0030] As an implementation, the constraint conditions of the upper model predictive control framework in the embodiment include the energy storage power constraint, the state of charge constraint, and the power grid voltage frequency deviation constraint; the improved whale algorithm introduces an adaptive weight factor and a flight strategy to optimize the solving process.

[0031] As an implementation, the adaptive droop coefficient calculation formula of the power-type energy storage in the embodiment is: ; wherein, is the initial droop coefficient of the power-type energy storage unit, is the current state of charge of the power-type energy storage unit, is the rated state of charge of the power-type energy storage unit, is the sensitivity coefficient of the droop coefficient to the state of charge deviation, which is optimized online through deep reinforcement learning.

[0032] As an implementation, the model predictive control optimization objective function of the energy-type energy storage in the embodiment is: ; wherein, is the battery degradation cost of the energy-type energy storage unit at the kth moment; is the weight coefficient of multi-objective optimization; is the actual output power of the energy-type energy storage unit at the kth moment; is the reference power instruction at the kth moment; is the state of charge of the energy-type energy storage unit at the kth moment; Reference state of charge for energy type energy storage unit, Prediction horizon for model predictive control, dynamically adjusted by fuzzy logic.

[0033] As an implementation, the priority queue indicators in the communication interruption in the embodiment include the available power ratio, the state of charge, and the voltage deviation, and the priority is obtained by weighted calculation; the consistency algorithm realizes power distribution by neighbor node interaction and global target adjustment.

[0034] As an implementation, the convolutional neural network fault diagnosis module in the embodiment inputs current, voltage, and temperature time series data, and outputs fault probability distribution; the distributed ledger technology adopts a consortium chain architecture, and the consensus mechanism is practical Byzantine fault tolerance, realizing device state information synchronization.

[0035] As an implementation, the embodiment further includes a dynamic partitioning coordination step: based on the density clustering algorithm, the control area is divided based on the energy storage unit density and the communication time delay, and when the load fluctuation in the area triggers the threshold, cross-area coordination support is started.

[0036] To further detail the above (1) multi-source data fusion and correlation modeling, (2) hierarchical collaborative control strategy, and (3) distributed autonomous mechanism, the following is described in detail: The embodiment provides a distributed energy storage system collaborative control method based on data fusion, comprising the following steps: (1) Multi-source data fusion and correlation modeling In the edge layer, the original data is processed by kernel principal component analysis for dimension reduction, and the key feature vector is extracted; ① Feature extraction: for the distributed energy storage system, the original data focuses on the state of charge (state of charge SOC, operating temperature T) of the energy storage device, the real-time parameters (voltage deviation ΔU, frequency deviation Δf) of the power grid, and the load demand (real-time load power P_load) core dimension (screening basis: directly affecting the energy storage control strategy and power grid stability, covering the "source-grid-load-storage" correlation logic); ② Extraction method (kernel principal component analysis, KPCA): Nonlinear mapping: through the Gaussian kernel function:

[0037] =0.8, which is experimentally adapted to the energy storage data distribution characteristics), the original data is mapped to a high-dimensional feature space; Covariance calculation: construct the covariance matrix of the mapped data:

[0038] ​(n is the sample size to ensure statistical validity); feature selection: set the feature value ≥ 1.2, cumulative contribution rate ≥ 95% screening threshold, finally extract [SOC, ΔU, Δf, P_load] as the key feature vector (dimension compression to 4 dimensions, balance calculation efficiency and information retention).

[0039] A spatiotemporal correlation model of multi-source data is constructed in the cloud using a graph neural network combined with a long short-term memory network, and a multi-source data feature matrix that fuses spatiotemporal features is output. ① Graph neural network (GNN) module (spatial correlation modeling): Input dimension: energy storage unit geographic location coordinates (x, y), power grid topology adjacency matrix A (dimension = energy storage unit number × power grid node number), 4-dimensional feature vector output by the edge layer, and near-24-hour historical operation data (time step 5 minutes, a total of 288 time steps); Algorithm logic: GraphSAGE algorithm is used to sample neighbor nodes (8 neighbor nodes are selected for each node, including 4 photovoltaic inverter nodes and 4 load nodes, which conforms to the "energy storage-photovoltaic-load" spatial correlation scenario), and the aggregation function is mean aggregation (60% weight) + attention aggregation (40% weight); (attention weight is calculated by , The correlation score of the node and is calculated, and the feature contribution of the key connection is strengthened); Network structure: 3-layer GNN is set, input layer dimension = 10 (4-dimensional features + 6-dimensional spatial information), hidden layer neuron number = 16, and output layer dimension = 12 (parameters optimized by grid search, suitable for the data scale of distributed energy storage systems).

[0040] ② Long short-term memory network (LSTM) module (temporal correlation modeling): Input processing: 12-dimensional spatial features output by GNN and historical time series data are spliced to construct a spatiotemporal fusion input; Network structure: 2-layer LSTM network is set, each layer has 24 neurons, the activation function is ReLU, and the forgetting gate threshold is 0.5 (ReLU improves gradient transmission efficiency, and the forgetting gate threshold balances historical information retention and update); Training configuration: the data set is divided into training set, validation set and test set according to the ratio of 7:2:1, the loss function adopts mean square error (MSE), and the iteration termination condition is that the validation set loss does not decrease for 5 consecutive rounds (maximum iteration number = 100) (to avoid overfitting and ensure model generalization); Output dimension: the final output is a multi-source data feature matrix with a time step × energy storage unit number × 12, realizing deep fusion of spatiotemporal features.

[0041] Distributed updating of model parameters is achieved using federated learning, ensuring data privacy; ① Update parameter definition: includes GNN aggregation layer weight (W) ), LSTM input gate weight (W ), forget gate weight (W ), output gate weight (W ), and full connection layer bias (b) (covering the core trainable parameters of the model, ensuring the effectiveness of the update).

[0042] ② Distributed updating process: Edge node training: edge nodes such as photovoltaic and commercial areas train GNN-LSTM models based on local data, with optimizer = Adam, learning rate = 0.001 (photovoltaic area, data fluctuation) / 0.0005 (commercial area, data stable), batch_size = 32, training rounds = 10 (differential learning rate adaptation to scene characteristics, training rounds balance efficiency and convergence); Parameter upload: upload the model parameters trained locally every hour (only upload parameters, not raw data, to avoid privacy leakage risk); Cloud aggregation: use FedAvg algorithm to weight average the parameters (weight = data volume of each node to total data volume, to ensure that nodes with more data contribute more significantly), and add Gaussian noise (privacy budget ) through differential privacy technology to protect sensitive parameters from being reverse-engineered; Parameter distribution: the cloud distributes the aggregated global optimal parameters to all edge nodes, covering the local old parameters, completing one update cycle (ensuring the global consistency of model parameters).

[0043] (2) Hierarchical collaborative control strategy The upper global optimization layer is based on the model predictive control framework, with system economy, energy storage life loss, and grid stability as the multi-objective optimization function, and uses the improved whale algorithm to solve the global optimal solution; ① Multi-objective optimization function and variable definition: Optimization objectives: , Where: , is the power command sequence of the energy storage unit, Price(t) is the electricity price at time t, unit yuan / kWh;

[0044]

[0045] Where, battery life loss, Cost_batt is the unit price of the battery, 3000 is the cycle life, and ΔDOD_i is the charge and discharge depth of the i-th energy storage unit.

[0046] Grid fluctuation quantification

[0047] Optimization variable: power instruction sequence of each energy storage unit (t=1-24 hours, 1 hour step) and SOC control interval .

[0048] ②Model predictive control (MPC) and improved whale algorithm coupling solution process: MPC framework initialization: Prediction horizon (1 day), control horizon (every 6 hours rolling optimization); The constraint conditions are as follows:

[0049] Power constraint:

[0050] The rated power; SOC constraint: ; Grid constraint:

[0051] Improved whale algorithm solving steps: Population initialization: generate 50 candidate solutions, each solution is

[0052] Among them, the dimension = 24 x N + 2 x N, N is the number of energy storage units.

[0053] Fitness calculation: substitute the candidate solution into the MPC framework, simulate the economy, life loss and grid fluctuation within 24 hours, output the multi-objective function value F as the fitness (the smaller the value, the better); Iterative optimization: Adaptive weight: , where t is the iteration number, global exploration for the first 20 times, and local optimization for the last 30 times.

[0054] The search strategy is as follows: Global search : , where A and C are coefficients used to expand the search range.

[0055] Local development : , where X * is the current optimal solution, and b=1.5 strengthens the local fine search.

[0056] ​Convergence criterion: When the difference in optimal fitness over 5 consecutive generations is less than 1 / 3. Alternatively, iterate up to 50 times to output the globally optimal solution (power command sequence and SOC range).

[0057] MPC rolling optimization: Every 6 hours, the optimal solution output by the whale algorithm is used as the reference trajectory. Only the control instructions for the first 6 hours are executed, and the remaining time is iterated and solved again to adapt to real-time load fluctuations.

[0058] In the lower local autonomous layer, power-type energy storage adopts adaptive droop control combined with reinforcement learning to dynamically adjust the droop coefficient, while energy-type energy storage optimizes the charge and discharge depth based on model predictive control, with the quantification of battery degradation cost as the objective function. ① Power-type energy storage (adaptive droop control and reinforcement learning): The core formula for droop control: (Δf magnified 10 times to balance dimensions), where the adaptive droop coefficient is:

[0059] in, Based on the coefficient, The sensitivity coefficient is optimized through reinforcement learning.

[0060] Details of reinforcement learning (DDPG algorithm): State space:

[0061] Among them, the photovoltaic fluctuation amount = |current power - power 10 minutes ago|.

[0062] Action space: (Sensitivity coefficient range).

[0063] Reward function:

[0064] Where ΔP is the deviation between actual power and command, and a positive bonus is added for response time <0.5 seconds.

[0065] Training mechanism: 10,000 steps of offline training (experience replay pool capacity of 50,000), and fine-tuning every 5 minutes online to ensure real-time performance.

[0066] ② Energy storage (model predictive control and battery degradation costs): Optimize the objective function:

[0067] The prediction process involves 6 steps in the time domain, with each step lasting 10 minutes. Yuan, based on the fitting of lithium iron phosphate battery cycle experimental data, where ΔDOD(k) is the charge-discharge depth at the k-th step.

[0068] Constraints (limit SOC rate of change, reduce battery impact):

[0069] Solution method: use sequential quadratic programming (SQP) algorithm, output charge and discharge depth command every 10 minutes, track the power reference value of the upper global optimization.

[0070] (3) Distributed autonomous mechanism When communication is interrupted, use the priority queue and consistency algorithm based on analytic hierarchy process to realize distributed power distribution; ① Analytic hierarchy process (AHP) priority modeling: Evaluation index: available power ratio (weight 0.4), SOC (weight 0.3), voltage deviation (weight 0.3) (determined by AHP pairwise comparison matrix, consistency ratio (CR=0.08<0.1)); Priority calculation (normalized processing, the higher the value, the higher the priority):

[0071] Queue generation: arrange the energy storage units in descending order of Priority to form a "priority support group (top 30%), regular group (middle 40%), reserved group (bottom 30%) " three-level queue.

[0072] ② Consistency algorithm to realize distributed distribution: Algorithm type: first-order consistency algorithm (adapted to local communication of distributed system), update formula:

[0073] Where, is the neighbor node set, is the coupling gain, is the reference tracking gain.

[0074] Distribution steps: Each node broadcasts its local priority and available power; Iterative update power distribution based on consistency algorithm (iterate 5 times, convergence threshold 1%); The priority support group responds to load fluctuations first, the regular group supplements the adjustment, and the reserved group maintains the SOC standby.

[0075] When the equipment fails, diagnose the health status through convolutional neural network, and realize redundancy switching and state synchronization combined with distributed ledger technology; ① Convolutional neural network (CNN) fault diagnosis: Input features: current waveform (30 cycles), voltage ripple (standard deviation), temperature change rate (total of 32 features covering electrical and thermal characteristics); Network structure: 2 layers of convolutional layers (kernel size 3x3, channel number 16 / 32) + 2 layers of fully connected layers (neuron number 64 / 2), activation function is ReLU, output failure probability (0 to 1) (binary classification: normal / failure); Training data: 1000+ groups of energy storage device failure data (overcharge, short circuit, thermal runaway) are collected, data augmentation (translation, scaling) is used to improve generalization, and the test accuracy is ≥98%.

[0076] ② Distributed ledger and redundancy switching: Ledger architecture: consortium chain (only authorized energy storage nodes and grid dispatch participate), consensus mechanism is practical Byzantine fault tolerance (PBFT) (tolerance 1 / 3 node failure, consensus delay <200ms); State synchronization: upload device state (SOC, health, power) to the ledger every 1 minute, use Merkle tree (MerkleTree) to compress data, reduce storage overhead by 70%; Redundancy switching logic: Fault trigger: when the CNN diagnosis failure probability is >0.95, trigger redundancy switching; Switching process: the ledger retrieves backup energy storage units of the same capacity and type (based on SOC≥0.5, health≥0.8 screening), and automatically completes power takeover through smart contract (switching time <500ms); State synchronization: the backup unit synchronizes the historical SOC and control strategy parameters of the failed unit to ensure seamless connection.

[0077] In this embodiment, multi-source data fusion and association modeling are used to realize the deep association and dynamic adaptation of energy storage device state, grid parameter, environmental data and load demand, providing comprehensive and accurate decision basis for collaborative control; through hierarchical collaborative control strategy, differentiated control schemes are designed according to the characteristic differences of power and energy storage, taking into account system global optimization and local flexible response, effectively improving the utilization efficiency of energy storage resources and reducing the life loss of equipment; through distributed autonomous mechanism and dynamic partitioning collaboration, the robustness and collaboration efficiency of the system in the scenarios of communication interruption, equipment failure and load mutation are enhanced. The control method of this embodiment can be directly applied to microgrid, smart grid and other scenarios, providing a practical technical solution for efficient, collaborative and stable operation of distributed energy storage systems.

[0078] Embodiment two The embodiment provides a distributed energy storage system cooperative control method based on data fusion, the application scene of which is aimed at a commercial park microgrid (containing 2 power type energy storages (supercapacitors, model: Maxwell 36V / 100F), 3 energy type energy storages (lithium batteries, model: CATL 51.2V / 100Ah), 500kW photovoltaic array) with a photovoltaic penetration rate of 40%, and the effectiveness of the method in the "photovoltaic-air conditioning load time mismatch + high-frequency cloud layer blocking" scene is verified.

[0079] (1) Multi-source data fusion and association modeling Data acquisition and preprocessing Equipment and sampling logic: supercapacitor voltage is sampled by ADI AD7606 analog-to-digital converter (50Hz, 16-bit precision) to capture 5ms high-frequency fluctuations; lithium battery temperature is collected by DS18B20 sensor (10Hz, accuracy ±0.5℃) to monitor the risk of thermal runaway; Photovoltaic inverter power is collected through Modbus-RTU protocol (1Hz), synchronized with photovoltaic output and power grid timestamp; air conditioning load is connected to the building automation system (1Hz) through BACnet protocol, to obtain the start-stop state of each air conditioner; The edge layer is synchronized through NTP clock (accuracy ≤1ms) to ensure that the "photovoltaic output-air conditioner start-stop-energy storage state" data timestamps are aligned; for photovoltaic power drop caused by cloud layer blocking (30% drop within 1 second), cubic spline interpolation is used to repair missing data to avoid model misjudgment.

[0080] Graph neural network spatiotemporal modeling GraphSAGE configuration: each energy storage node samples 8 neighbor nodes (4 photovoltaic inverters + 4 air conditioning load nodes), the aggregation function uses mean aggregation (60% weight) and attention aggregation (40% weight) to strengthen the association learning of "photovoltaic-air conditioning adjacent nodes"; the training parameters are set as batch_size=32, epoch=100, learning rate=0.001 (Adam optimizer); Knowledge graph construction: extract the entity relationship (nodes containing 1 photovoltaic, 5 energy storages, and 10 air conditioning partitions, relationship types are "power supply" and "regulation") of "photovoltaic array-energy storage unit-air conditioning partition" through the AutoKG model and store it in the Neo4j database.

[0081] Federal learning parameter update Edge node division: photovoltaic area nodes train the "light mutation-supercapacitor response" model (Adam optimizer, learning rate 0.001), and commercial area nodes train the "peak-valley load-lithium battery scheduling" model (Adam optimizer, learning rate 0.0005); Model parameters are aggregated once per hour. During cloud aggregation, differential privacy technology (ε=1.0) is used to protect sensitive data such as photovoltaic output curves and air conditioning scheduling strategies. The edge controllers are RK3568 (photovoltaic area, supporting 5G communication) and Jetson Nano (commercial area, caching 3 days of historical data).

[0082] (2) Hierarchical collaborative control strategy upper-level global optimization The prediction time domain N of the model predictive control framework is dynamically adjusted (N=3 when photovoltaic fluctuations are large, N=10 when the load is stable, with a step size of 5 minutes). The constraints include supercapacitor power ±20kW (hard constraint) and lithium battery SOC [0.2, 0.9] (soft constraint, out-of-bounds penalty coefficient ×10). The improved whale algorithm iterates 50 times (each iteration ≤ 200ms), and the adaptive weight formula is as follows: ( For the current iteration, (For the total iteration), the initial focus is on global search, and the later focus is on local convergence. The multi-objective function takes photovoltaic absorption rate, peak-valley electricity price difference and lithium battery temperature as optimization objectives.

[0083] Lower-level local autonomy Power-type energy storage (supercapacitor): The adaptive droop coefficient formula is: ; in, (Compatible with 36V rated voltage), α is optimized using the DDPG algorithm (learning rate = 0.0001, reward function). When photovoltaic fluctuations exceed 15%, α is increased to 1.1; Energy storage (lithium battery): The model predictive control objective function is:

[0084] in, , For depth of charge / discharge, weighting coefficient (The sum is 1).

[0085] (3) Distributed autonomous mechanism Communication interruption fault tolerance When communication is interrupted, the priority queue index weights are set to "available power ratio (0.5) > SOC (0.3) > voltage deviation (0.2)", and the consensus algorithm is used. Power allocation, convergence after 3 iterations, with an error ≤5%; where,

[0086] Let i be the output power of the i-th energy storage unit in the k-th iteration; The neighbor node set (such as adjacent energy storage, load node) of the i-th energy storage unit in the communication topology; The current total power demand (or total power allocation instruction) of the system; The total number of energy storage units participating in this power allocation; The learning rate of the power difference of the neighbor node (strengthen local cooperation); The adjustment coefficient of the global power mean (guarantee overall power balance). The algorithm converges after 3 iterations, and the power allocation error is ≤5%.

[0087] Backup strategy for insufficient capacity: if the super capacitor capacity is insufficient (remaining <15%), the lithium battery will supplement in "short-time response mode" (gradually intervene after 30 seconds to avoid high-frequency impact and aggravate life loss).

[0088] Device failure redundancy Super capacitor failure: when the voltage drops by >10% (for 2 seconds), the CNN diagnoses the "electrolyte leakage" probability, >0.75, then triggers the alliance chain (PBFT consensus, 3 backup nodes) to synchronize the state, and switches to the standby capacitor (capacity redundancy 20%) within 5 seconds to restore 90% power output; Lithium battery thermal runaway prevention: when the temperature rise rate is >2°C / min or the temperature is >45°C, start the active balancing strategy (transfer 10% capacity to adjacent batteries within 3 minutes), and the alliance chain records the "temperature-charge / discharge" correlation data for life prediction.

[0089] (4) Dynamic partitioning cooperation Based on the DBSCAN algorithm to divide the area: neighborhood radius 200 meters (covering the physical association range of photovoltaic-energy storage-air conditioner), minimum number of points 5 (ensure containing 1 energy storage + 2 load nodes), automatically generate photovoltaic response area (high frequency) and load scheduling area (long period); Trigger cooperation when regional load fluctuation >20%: preferentially call the remaining energy storage in this area (allocate according to available capacity proportion), and if insufficient, support across regions (delay ≤100ms, communication guaranteed through 5G slicing).

[0090] For the problem of "photovoltaic peak (11:00-15:00) and air conditioner peak (14:00-18:00) time mismatch", through the cooperation of "super capacitor short-time storage and lithium battery long-time scheduling": the super capacitor stores 10% to 15% excess power during the photovoltaic peak, and the lithium battery releases the stored photovoltaic power during the air conditioner peak. The graph neural network learns the correlation law of the overlapping period to optimize the power transfer strategy, which reduces the electricity bill by 15% in this period.

[0091] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and it is within the scope of the present application to make changes, modifications, alternatives and variations thereto without departing from the spirit of the present application.

Claims

1. A data fusion based distributed energy storage system coordinated control method, characterized in that: The method comprises the following steps: (1) Multi-source data fusion and correlation modeling: In the edge layer, the original data is processed by kernel principal component analysis for dimension reduction, and the key feature vector is extracted; in the cloud, a multi-source data space-time correlation model is constructed by using a graph neural network combined with a long short-term memory network, and a multi-source data feature matrix with fused space-time features is output; the distributed update of model parameters is realized by using federated learning to protect data privacy; (2) Hierarchical collaborative control strategy: The upper layer global optimization layer is based on a model predictive control framework, takes system economy, energy storage life loss and power grid stability as multi-objective optimization functions, and solves the global optimal solution by using an improved whale algorithm; in the lower layer local autonomous layer, the power type energy storage dynamically adjusts the droop coefficient by using adaptive droop control combined with reinforcement learning, and the energy type energy storage optimizes the charging and discharging depth based on the model predictive control, and quantifies the battery degradation cost as the objective function; (3) Distributed autonomous mechanism: When the communication is interrupted, the priority queue based on the analytic hierarchy process and the consistency algorithm are used to realize the decentralized power distribution; when the equipment fails, the health state is diagnosed by using a convolutional neural network, and the redundant switching and state synchronization are realized by using a distributed ledger technology.

2. The method of claim 1, wherein: The multi-source data includes energy storage device state data, power grid parameters, environmental data and load demand data, the kernel principal component analysis dimension reduction process includes nonlinear mapping to a high-dimensional feature space, covariance matrix calculation and feature vector orthogonalization screening; when the edge layer pre-processes the data, outlier rejection and normalization processing are used.

3. The method of claim 1, wherein: The input of the graph neural network includes the geographical position of the energy storage unit, the spatial information of the power grid topology structure, and the feature vector and historical operation data output by the edge layer; the graph neural network uses the GraphSAGE algorithm for neighbor node sampling, and the aggregation function combines the mean aggregation and the attention mechanism; meanwhile, an energy storage unit-power grid node-environment factor knowledge graph is constructed, and the entity relationship is extracted by using an AutoKG model and stored in a Neo4j database.

4. The method of claim 1, wherein: The parameter update of the federated learning is realized by local training of each edge node and cloud aggregation, the edge node training uses the Adam optimizer, and the sensitive parameters are protected by using the differential privacy technology when the cloud aggregates; the edge layer deployment controller in the system architecture supports multi-protocol communication, and the cloud is deployed in a cluster to support high-concurrency processing.

5. The method of claim 1, wherein: The constraint conditions of the upper layer model predictive control framework include energy storage power constraint, state of charge constraint and power grid voltage frequency deviation constraint; the improved whale algorithm introduces an adaptive weight factor and a flight strategy to optimize the solving process.

6. The method according to any one of claims 1 to 5, characterized in that: The adaptive droop coefficient calculation formula of the power type energy storage is: ; wherein, is the initial droop coefficient of the power type energy storage unit, is the current state of charge of the power type energy storage unit, is the rated state of charge of the power type energy storage unit, is the sensitivity coefficient of the droop coefficient to the state of charge deviation, which is optimized online through deep reinforcement learning.

7. The method of claim 6, wherein: The model predictive control optimization objective function of the energy type energy storage is: ; wherein, is the battery degradation cost of the energy storage unit at the kth time instant; is the weight coefficient of the multi-objective optimization; is the actual output power of the energy storage unit at the kth time instant; is the reference power command at the kth time instant; is the state of charge of the energy storage unit at the kth time instant; is the reference state of charge of the energy storage unit, is the prediction horizon of the model predictive control, which is dynamically adjusted by fuzzy logic.

8. The method of claim 7, wherein: The priority queue indicators when the communication is interrupted include the available power ratio, the state of charge and the voltage deviation, and the priority is obtained by weighted calculation; the consistency algorithm realizes power distribution by neighbor node interaction and global target adjustment.

9. The method of claim 8, wherein: The input of the convolutional neural network fault diagnosis module is current, voltage and temperature time series data, and the output is a fault probability distribution; the distributed ledger technology uses a consortium chain architecture, the consensus mechanism is practical Byzantine fault tolerance, and the device state information synchronization is realized.

10. The method according to any one of claims 1 to 5, characterized in that: It also comprises a dynamic partitioning and synchronization step: based on a density clustering algorithm, the control area is divided according to the energy storage unit density and the communication time delay, when the load fluctuation in the area triggers the threshold, the cross-area collaborative support is started.

Images