Distributed energy storage regulation method and system based on edge computing and artificial intelligence

Abstract

The application discloses a distributed energy storage regulation method and system based on edge computing and artificial intelligence, relates to the technical field of power system energy storage regulation and intelligent microgrid, and constructs a three-layer architecture of a sensing layer, a decision layer and a control layer: the sensing layer collects operation parameters of energy storage units, power grid state information and a clearness index; the decision layer is composed of an artificial intelligence prediction module and a global scheduling optimization engine, a random forest is used to select features and construct an RF-LSTM hybrid model to predict load power, a clearness index method is used to predict photovoltaic output power, a comprehensive objective function is constructed, and a global scheduling instruction is generated by solving through a particle swarm optimization algorithm; the control layer adopts an edge computing architecture, and energy storage unit collaborative regulation and an event triggering mechanism are realized through a local controller. The application realizes hierarchical scheduling of edge autonomy and global optimization, effectively improves prediction accuracy, voltage stability and economy, and reduces communication burden and network loss.

Description

Technical Field

[0001] This invention relates to the field of smart grid and energy management technology, specifically to a distributed energy storage control method and system based on edge computing and artificial intelligence. Background Technology

[0002] In recent years, the installed capacity of new energy power generation has continued to grow rapidly, with photovoltaic and wind power dominating. However, the inherent volatility, intermittency, and unpredictability of renewable energy bring instability to the power system, causing problems such as grid power fluctuations, reverse power flow, voltage deviation exceeding limits, and wind and solar curtailment. Traditional centralized energy storage systems have limitations in terms of site selection flexibility and construction scale. When massive distributed energy resources are connected, their centralized control architecture faces bottlenecks in information transmission delays and processing capacity, making it difficult to meet real-time control requirements.

[0003] To address these challenges, distributed energy storage systems have attracted widespread attention due to their advantages such as flexible installation, strong environmental adaptability, and the ability to achieve on-site energy consumption. However, existing research on the optimal configuration of distributed energy storage suffers from a single-objective problem, lacking comprehensive consideration of multiple factors such as grid security, equipment lifespan, and economic benefits, resulting in economic drawbacks such as high investment costs and long payback periods. Furthermore, artificial intelligence algorithms face engineering challenges in practical applications, such as long verification cycles and a lack of real-world testing scenarios, hindering the industrialization of research results. Simultaneously, the internal physical and electrochemical mechanisms of energy storage devices are complex, making it difficult for traditional mathematical models to fully describe the dynamic characteristics under varying operating conditions, and they also place stringent demands on computing resources.

[0004] Therefore, there is an urgent need to propose a technical solution for distributed energy storage systems to improve the grid's capacity to support distributed new energy sources and enhance its economic operation. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to address the difficulty of multi-objective coordination in existing research on the optimal configuration of distributed energy storage.

[0006] This invention solves the above-mentioned technical problems through the following technical means: a distributed energy storage and control system based on edge computing and artificial intelligence, comprising:

[0007] The perception layer, deployed at the energy storage units and photovoltaic nodes in the distribution area, is used to collect the operating parameters of the energy storage units, grid status information and sunshine index in real time, and transmit them to the decision layer and control layer through a low-latency communication protocol. The decision-making layer integrates an artificial intelligence prediction module and a global scheduling optimization engine. The artificial intelligence prediction module includes: a random forest feature selection unit, used to select features from multi-dimensional time features and environmental features; an RF-LSTM hybrid prediction unit, used to predict user load power based on the selected features and historical load data; and a photovoltaic output prediction unit, used to predict photovoltaic power using the sunny index method combined with historical data. The global scheduling optimization engine is used to construct a comprehensive objective function, solve the objective function through a particle swarm optimization algorithm, and generate global scheduling instructions to be sent to the control layer. The control layer is deployed locally in each energy storage unit and adopts an edge computing architecture. Each energy storage unit is configured with a local controller, which is used to realize the coordinated control between energy storage units and integrate the event triggering mechanism.

[0008] This invention achieves decoupling of data acquisition, optimization decision-making, and execution control through a three-layer distributed architecture consisting of a perception layer, a decision layer, and a control layer: the perception layer uses low-latency communication to ensure real-time reporting of grid status; the decision layer integrates an artificial intelligence prediction module and a global scheduling optimization engine, significantly improving scheduling accuracy through feature filtering, load and photovoltaic output prediction, and particle swarm optimization; the control layer adopts an edge computing architecture to enable local controllers of each energy storage unit to have autonomous and collaborative capabilities, enhancing system fault tolerance and reliability; the local controllers integrate an event triggering mechanism to dynamically respond to grid anomalies and reduce ineffective computational overhead, thus achieving efficient and economical operation and voltage stability control of the distributed energy storage system in complex distribution network environments.

[0009] Furthermore, the random forest feature selection unit performs the following operations: Timestamp data and corresponding meteorological data for different time periods are extracted from historical power system operation data. Multidimensional time and environmental features are then parsed to obtain these features. After preprocessing, an original candidate feature set is constructed by concatenating dimensions. Let the original candidate feature set be:

[0010] in, For the first feature vector, This represents the total number of candidate feature variables; The feature set X is input into the random forest model, and the importance of each feature is evaluated by the change in out-of-bag sample error. The importance score for each feature variable is defined as follows:

[0011] in, The number of decision trees in the random forest. Indicates the first The OOB error corresponding to each decision tree Indicates the random perturbation of the first The corresponding OOB error after each feature variable This represents the contribution of each feature vector to the classification process; Scoring based on feature importance All candidate features are ranked, and the top K features are selected as important feature variables to form a feature subset:

[0012] in, This represents the Kth important eigenvector.

[0013] Furthermore, the RF-LSTM hybrid prediction unit performs the following operations: After feature selection, the selected K important feature variables and the corresponding historical load data are used as the input sequence to construct an LSTM model. Let the input vector of the LSTM model at time t be:

[0014] in, The historical load power at time t; After mapping using the LSTM model, the predicted user load power at the predicted time is obtained:

[0015] in, This represents the nonlinear mapping function of the LSTM obtained after training.

[0016] Furthermore, the photovoltaic output prediction unit performs the following operations: The solar power output is predicted using a solar irradiance-based sunshine index. Based on historical solar power output data, the solar power output data is statistically analyzed and the average sunshine index at each time point is taken into account to obtain the solar power output prediction result for the current time. The formula for predicting photovoltaic power output is:

[0017]

[0018]

[0019]

[0020]

[0021]

[0022] in, For the predicted photovoltaic power, take its average value as , For the current time, Let k be the time step for the prediction, and k be the time interval for the prediction to extend into the future. The sunshine index is calculated by taking its average value. , This represents the start time of the selected historical data time window. For the nearby The maximum historical value of photovoltaic power at the same time within a day. , .

[0023] Furthermore, the comprehensive objective function is constructed as follows: The formula for calculating the voltage fluctuation assessment of the Ath distributed energy storage access node is as follows:

[0024] in, This represents the voltage amplitude of the A-th distributed energy storage access node at time t. This indicates the voltage amplitude of the node when there is no energy storage output. This represents the control voltage reference value for the A-th distributed energy storage access node, and p represents the filter coefficient. Indicates the voltage acquisition time interval; An economic model is established by combining the investment costs, operation and maintenance costs, network loss costs, and electricity trading revenue during power grid planning:

[0025] in, Indicates investment cost, Indicates operation and maintenance costs. R represents the network loss cost, and R represents the electricity trading revenue; The comprehensive objective function obtained by combining the node voltage fluctuation and economic models is expressed as:

[0026] in, Represents the overall objective function. This indicates the voltage fluctuation threshold.

[0027] Furthermore, the decision-making layer also enforces the following constraints: Power balance constraints:

[0028] in, This represents the traditional power generation capacity of node i at time t. This represents the predicted photovoltaic output power of node i at time t. This represents the energy storage discharge power of node i at time t. This represents the predicted power of user load at node i at time t. Let represent the energy storage charging power of node i at time t, and M represent the set of all nodes; Voltage amplitude constraint:

[0029] in, This represents the voltage magnitude at node i. and These represent the lower and upper limits of the voltage amplitude, respectively. Distributed power generation output constraints:

[0030] in, This represents the actual photovoltaic output of node i. , Let represent the lower and upper limits of the photovoltaic output of node i, respectively; Energy storage system charge and discharge constraints:

[0031] in, and Let represent the maximum values ​​of the energy storage charging power and energy storage discharging power at node i, respectively. This represents the energy state of the energy storage system at node i. and This indicates the lower and upper limits of its energy state. This represents the energy state of the energy storage system at node i at time t. and These represent charging efficiency and discharging efficiency, respectively.

[0032] Furthermore, the global scheduling optimization engine of the decision-making layer is also used for: The predicted net load is the difference between the predicted load power and the predicted photovoltaic output power at time t, and the calculation formula is as follows:

[0033] The formula for calculating the residual is:

[0034] The formula for calculating the residual penalty term is: .

[0035] Furthermore, the objective function is solved using the particle swarm optimization algorithm, specifically as follows: Initialize the particle swarm, and calculate the fitness value of each particle based on the objective function and the residual penalty term. The formula for calculating fitness is:

[0036] in, Represents the nth particle. Indicates the weighting coefficient. This represents the residual penalty coefficient; Each particle adjusts its velocity and position by referencing its own historical best position and the global best position of the entire particle swarm. The particle update formula is as follows:

[0037] in, and Let these represent the velocities of particle n at times t and t+1, respectively. This represents the inertial weights used to control the decay of particle velocity. and This indicates the step size for adjusting the particle's movement towards its own optimal point and the global optimal point. and A random number representing an increase in the randomness of the search. This represents the optimal solution found by particle n during the search process. This represents the globally optimal solution. This represents the nth particle.

[0038] Furthermore, the local controller of the control layer integrates a SOC partition management module and performs the following operations: Real-time state of charge of energy storage unit The calculation formula is:

[0039] in, For rated capacity, The remaining energy storage capacity at the current moment; Real-time state of charge of the energy storage unit Divided into three regions: those exceeding 90% of the full SOC. Regions below 10% of full SOC Regions, and those in between The region is defined, and corresponding charge and discharge constraint strategies are implemented based on the state of charge region.

[0040] This invention also provides a distributed energy storage regulation method based on edge computing and artificial intelligence, including: S1. Collect the operating parameters, grid status, and weather index of the energy storage units and distributed power nodes in the distribution area, and transmit them through a low-latency communication protocol. S2. Based on the artificial intelligence model, feature filtering and load power prediction are performed on the data, and the solar power output prediction value is generated by using the clear sky index method. S3. Construct a comprehensive objective function and apply voltage, power balance, and energy storage charging and discharging constraints; S4. Calculate the net load forecast and residual based on the load power forecast and photovoltaic output power forecast, construct the residual penalty term and weighted comprehensive objective function, and generate global scheduling instructions by solving the particle swarm optimization algorithm. S5. The global scheduling command is sent to each energy storage unit, and the SOC partition management strategy is executed by the local controller under the edge computing architecture to adjust the charging and discharging power. S6. In response to the event triggering condition, activate the fast response mode to adjust the energy storage operation strategy; S7. Periodically record operating data, evaluate control performance, and provide feedback to optimize the prediction model and trigger thresholds.

[0041] The advantages of this invention are: 1. This invention proposes a three-layer system architecture design for multiple energy storage units, combining an event-triggered mechanism to achieve intelligent adaptive regulation of energy storage units under global target constraints. This design balances the independence and collaboration of distributed systems, overcoming the limitations of traditional centralized control. The system architecture consists of a perception layer, a decision layer, and a control layer. The perception layer ensures real-time data acquisition through a low-latency communication protocol. The decision layer integrates an artificial intelligence prediction module and a global scheduling optimization engine to achieve global optimization scheduling. The control layer adopts an edge computing architecture to configure local controllers to perform collaborative regulation, ensuring efficient collaboration among the three layers.

[0042] 2. The decision layer of this invention integrates an artificial intelligence prediction module and a global scheduling optimization engine. Through feature screening, load and photovoltaic output prediction, and comprehensive consideration of constraints such as power balance, voltage amplitude, and distributed power output, it uses a particle swarm optimization algorithm to solve the problem, thereby improving the regulation efficiency and response speed of the energy storage system.

[0043] 3. This invention utilizes an edge computing architecture to deploy computing tasks such as SOC partition management, charging and discharging power adjustment, and event triggering response in a local controller close to the energy storage unit, thereby achieving localized rapid response.

[0044] 4. This invention monitors the state of charge (SOC) of the energy storage unit in real time through a local controller in the control layer. The real-time SOC is divided into three regions, and corresponding charge / discharge constraint strategies are executed according to these regions to avoid overcharging and discharging, thereby extending battery life. Through full-process data recording and performance evaluation, the prediction model and trigger thresholds are dynamically adjusted to promptly identify potential faults and performance degradation issues, ensuring the system's stability during long-term operation. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the system in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the energy scheduling optimization process in Embodiment 1 of the present invention; Figure 3 This is a flowchart of the distributed energy storage control process in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of data interaction and consistency control of the energy storage unit in Embodiment 1 of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1 like Figure 1 As shown, the distributed energy storage control system based on edge computing and artificial intelligence includes: The perception layer, deployed at the energy storage units and photovoltaic nodes in the distribution area, is used to collect the operating parameters of the energy storage units, grid status information and sunshine index in real time, and transmit them to the decision layer and control layer through a low-latency communication protocol.

[0048] Specifically, high-precision sensors and monitoring devices are deployed at the energy storage units and photovoltaic nodes in the distribution area, enabling real-time acquisition of operating parameters such as voltage, current, and state of charge of each energy storage unit with millisecond-level sampling frequency and data accuracy.

[0049] It achieves real-time communication with the power grid system, collecting power grid status information such as load fluctuations, frequency, voltage, and harmonic deviation values ​​to support dispatching decisions. It employs the high-bandwidth, low-latency MQTT communication protocol to enhance system reliability and fault tolerance.

[0050] The decision-making layer integrates an artificial intelligence prediction module with a global scheduling optimization engine.

[0051] Specifically, the artificial intelligence prediction module includes: a random forest feature selection unit, an RF-LSTM hybrid prediction unit, and a photovoltaic power output prediction unit.

[0052] The random forest feature filtering unit is used to filter features from multidimensional temporal and environmental features.

[0053] For the Random Forest algorithm, the key is to analyze the error of out-of-bag (OOB) data to obtain important feature vectors. The process is as follows: Timestamp data was extracted from the historical load database of the power system, and multi-dimensional time features were parsed, including month (range 1-12), day (range 1-31), weekday (range 1-7, Monday corresponds to 1, Sunday corresponds to 7), weekday, holiday, and hour of the day (range 1-24). Weekdays and holidays were coded using 0 / 1 labels; weekdays were marked as 1 and non-weekdays as 0, and holidays were marked as 1 and non-holidays as 0. The discrete features of month, day, and hour of the day were converted into binary vectors using one-hot encoding, with month features encoded as a 12-dimensional binary vector, day features as a 31-dimensional binary vector, and hour features as a 24-dimensional binary vector. Simultaneously, environmental features corresponding to the above timestamps were obtained from the meteorological data platform, including weather type and sunshine duration (range 1-24 hours). Among them, the weather type is classified and coded according to the "Ground Meteorological Observation Specifications". The coding rules are as follows: sunny / cloudy is coded as 1, light rain / light snow is coded as 2, and heavy rain, rainstorm, snowstorm, freezing rain and other heavy precipitation weather are coded as 3; the sunshine duration is normalized to the 0-1 range through Min-Max.

[0054] The multidimensional time features and environmental features that have undergone encoding and normalization are concatenated to construct the original candidate feature set; let the original candidate feature set be:

[0055] in, This represents the total number of candidate feature variables.

[0056] The feature set X is input into the random forest model, and the importance of each feature is evaluated by the change in out-of-bag sample error. The importance score for each feature variable is defined as follows:

[0057] in, The number of decision trees in the random forest. Indicates the first The OOB error corresponding to each decision tree Indicates the random perturbation of the first The corresponding OOB error after each feature variable It represents the contribution of each feature vector to the classification process and can measure the independent classification ability of each feature.

[0058] All candidate features are ranked according to their importance scores, and the top K features are selected as the most important features, forming a feature subset.

[0059] The value of K is determined based on the trade-off between prediction accuracy and computational complexity.

[0060] The RF-LSTM hybrid prediction unit is used to predict user load power based on filtered features and historical load data.

[0061] For an LSTM model, an LSTM unit consists of an input gate, a forget gate, an output gate, and a memory unit. Its core memory unit describes the current state of the unit. The three control gates in the model are connected to the multiplication unit, controlling the input, memory unit, and output of that unit. The forgetting part of the memory unit is handled by the memory unit itself. Input of the Forgot Gate and intermediate output The retention vector of the memory unit is jointly determined by the input gate. The results after transformation by the tanh function and the sigmoid function, respectively, jointly determine the updated... and output Together determine the intermediate output The calculation formula is as follows:

[0062] in, and These represent the sigmoid function and the tanh function, respectively. , , , , and These represent the input gate, output gate, forget gate, intermediate output node, intermediate input node, and state unit, respectively. as well as They represent the inputs respectively. and intermediate output The matrix weights when multiplying with the corresponding gate; ⊙ represents the element-wise multiplication of the two vectors. and These represent the correction vectors for each corresponding gate.

[0063] After feature selection, the selected K key features and the corresponding historical load data are used as the input sequence to construct an LSTM model. Let the input vector of the LSTM model at time t be:

[0064] in, Let t be the historical load power at time t.

[0065] After mapping via the LSTM network, the predicted user load values ​​for the predicted time are obtained:

[0066] in, This represents the nonlinear mapping function of the LSTM obtained after training.

[0067] The photovoltaic power output prediction unit is used to predict photovoltaic power output by combining the sunshine index method with historical data.

[0068] The formula for predicting photovoltaic power output is as follows:

[0069]

[0070]

[0071]

[0072]

[0073]

[0074] in, For the predicted photovoltaic power, take its average value as t is the current time. To determine the predicted time step, take... =0.25, h=15min, k is the time interval to be added to the future time in the prediction, and k takes the range of 1 to 64 in the natural integer field. The sunshine index is calculated by taking its average value. , This represents the start time of the selected historical data time window. This represents the maximum historical photovoltaic power output at the same time within the next j days. , .about The Sunny Day Index is calculated by using the ratio of the measured photovoltaic power 3 hours before the forecast time to the maximum photovoltaic power at the same time in the next 10 days.

[0075] The global scheduling optimization engine is used to construct a comprehensive objective function, solve the objective function using the particle swarm optimization algorithm, and generate global scheduling instructions to be sent to the control layer.

[0076] A distributed energy storage system has multiple energy storage access nodes. The formula for calculating the voltage fluctuation assessment of the Ath distributed energy storage access node is as follows:

[0077] in, This represents the voltage amplitude of the A-th distributed energy storage access node at time t. This indicates the voltage amplitude of the node when there is no energy storage output. Let represent the control voltage reference value of the A-th distributed energy storage access node, and p be the filter coefficient that can be obtained through low-pass filtering. This indicates the voltage sampling time interval.

[0078] Considering the investment costs, operation and maintenance costs, network loss costs, and electricity trading revenue required for power grid planning, an economic model is established:

[0079] in, Indicates investment cost, Indicates operation and maintenance costs. R represents the cost of grid loss, and R represents the revenue from electricity trading.

[0080] The comprehensive objective function obtained by combining the node voltage fluctuation and economic models is expressed as:

[0081] in, This represents the overall objective function, which aims to minimize the total cost, including investment, operation and maintenance, and network loss costs, while ensuring that voltage stability is met. This is the voltage fluctuation threshold.

[0082] Set constraints To ensure the safety and stability of the active distribution network, and to ensure the coordinated operation of distributed power sources and energy storage systems, the following constraints are set: (1) Power balance constraint:

[0083] in, This represents the traditional power generation capacity of node i at time t. This represents the predicted photovoltaic output power of node i at time t. This represents the energy storage discharge power of node i at time t. This represents the predicted power of user load at node i at time t. Let represent the energy storage charging power of node i at time t, and M represent the set of all nodes.

[0084] (2) Voltage amplitude constraint:

[0085] in, This represents the voltage magnitude at node i. and These represent the lower and upper limits of the voltage amplitude, respectively.

[0086] (3) Output constraints of distributed power sources:

[0087] in, This represents the actual photovoltaic output of node i. , These represent the lower and upper limits of the photovoltaic output of node i, respectively.

[0088] The photovoltaic output forecast results are used to characterize the available output level of distributed photovoltaic systems within the dispatch cycle. The actual grid-connected power of distributed power sources must simultaneously meet both the forecast available output and the physical constraints of the equipment. When the forecast photovoltaic output exceeds the rated capacity of the equipment, the output is limited according to the rated capacity of the equipment.

[0089] (4) Energy storage system charge and discharge constraints:

[0090] in, and Let represent the maximum values ​​of the energy storage charging power and energy storage discharging power at node i, respectively. This represents the energy state of energy storage system i. and This indicates the lower and upper limits of its energy state. This represents the energy state of the energy storage system at node i at time t. and These represent the charge and discharge efficiencies, respectively.

[0091] The energy storage system dynamically adjusts the difference between the predicted load power and the predicted photovoltaic output power through charging and discharging power to meet the power balance requirements within the dispatch cycle. The predicted net load is:

[0092] The residuals and residual penalty terms are as follows:

[0093]

[0094] Solving the model achieves energy scheduling optimization While solving the objective function, it is also necessary to ensure that the above constraints are met. A particle swarm optimization algorithm and a local search strategy are used for scheduling optimization.

[0095] Energy scheduling optimization flowchart as follows Figure 2First, initialize the particle swarm. Then, calculate the fitness value of each particle based on the objective function and the residual penalty term. The formula for calculating fitness is:

[0096] in, This represents a weighting coefficient used to balance the importance of economy and voltage stability. This represents the residual penalty coefficient.

[0097] Each particle adjusts its velocity and position by referencing its own historical best position and the global best position of the entire particle swarm. The particle update formula is as follows:

[0098] in, and Let these represent the velocities of particle n at times t and t+1, respectively. This represents the inertial weights used to control the decay of particle velocity. and This indicates the step size for adjusting the particle's movement towards its own optimal point and the global optimal point. and A random number representing an increase in the randomness of the search. This represents the optimal solution found by particle n during the search process. This represents the globally optimal solution. Represents the nth particle. It is a vector.

[0099] If the particle swarm fails to find the optimal solution in multiple iterations, a local search strategy is triggered, performing a more refined search within a local region near the particle swarm's current position. Through continued iterations of the particle swarm, the optimal solution is eventually output.

[0100] The control layer is deployed locally in each energy storage unit and adopts an edge computing architecture. Each energy storage unit is configured with a local controller, which is used to realize the coordinated control between energy storage units and integrate the event triggering mechanism.

[0101] Specifically, the local controller processes and executes the data collected by the perception layer and the prediction results and scheduling instructions issued by the decision layer in real time. For example... Figure 3 and Figure 4 As shown.

[0102] Based on the predicted net load calculated by the decision-making level, initial charging and discharging power reference values ​​are allocated to each energy storage unit as initial scheduling quantities for subsequent control and coordination.

[0103] Rated capacity of distributed energy storage The real-time state of charge of the energy storage unit is determined during equipment manufacturing. The calculation formula is:

[0104] in, The remaining energy storage capacity at the current moment.

[0105] To optimize energy storage management, the State of Charge (SOC) is divided into three zones: when the SOC is above 90%, it is... Region; when it is below 10% of the full SOC, it is Area; the remaining part is Region. Combining the collected data and predicted values, the SOC health constraint is set as follows: when... In In the designated area, only discharge is permitted; when In In regional scenarios, energy storage units can freely respond to dispatch demands based on predicted net load; when In In this area, charging is only permitted.

[0106] Consistency control is performed at edge nodes, and power output is dynamically adjusted using neighborhood information.

[0107] Each energy storage unit is equipped with a local controller, enabling preliminary scheduling based on local information. Through neighborhood data exchange, the energy storage unit obtains the power status and operating information of neighboring energy storage units. Under the premise of meeting SOC health constraints, negative feedback is formed based on local node power deviation and neighborhood consistency error to correct the charging and discharging power reference values ​​of the energy storage unit in real time. This achieves coordinated power allocation among energy storage units, ensuring the consistency and global coordination of the overall system operating status.

[0108] Perform event trigger judgment. If the event trigger condition is met, enter the event strategy and execute instructions such as fast charging / fast discharging / limited reverse transmission.

[0109] There are various event triggering conditions, including electrical isolation anomalies, leakage current anomalies, SOC exceeding limits, temperature anomalies, voltage / current exceeding limits, returned power exceeding limits, electricity price anomalies, and excessive multi-unit coordination deviations. The priority of importance is: personal safety > equipment safety > voltage stabilization > economic objectives > efficiency objectives. When different events are triggered, corresponding instructions will be executed to ensure system robustness.

[0110] In case of abnormal electrical isolation or leakage current, immediately disconnect the relevant circuits, activate the alarm, and report to the decision-making level; in case of SOC exceeding the limit, abnormal temperature, or current / voltage exceeding the limit, restrict or suspend charging and discharging, and report the fault to the decision-making level; in case of excessive reverse power, allow the energy storage unit to perform power regulation (fast charging / fast discharging / limited reverse power transmission) and adjust the output of the photovoltaic inverter; in case of abnormal electricity price, adjust the power trading strategy, such as suspending the sale of electricity to the grid when the selling price is too low; when there is a large deviation in the coordination of multiple units, re-optimize the consistency control parameters and adjust the power allocation strategy.

[0111] In summary, this system can achieve efficient collaborative control and operation management of multi-area and multi-level distributed energy storage units, solve the power balance problem in high-penetration distributed photovoltaic areas, improve grid stability, and optimize system economy.

[0112] Example 2 Based on Embodiment 1, Embodiment 2 also provides a distributed energy storage regulation method based on edge computing and artificial intelligence, including: S1 collects the operating parameters, grid status, and weather index of the energy storage units and distributed power nodes in the distribution area, and transmits them through a low-latency communication protocol.

[0113] S2. Based on the artificial intelligence model, the data is used for feature screening and load forecasting, and the solar power output forecast value is generated by using the sunny index method.

[0114] Specifically, timestamp data and corresponding meteorological data for the same period are extracted from historical power system operation data. Multidimensional time and environmental features are then analyzed. After preprocessing these features, an original candidate feature set is constructed by concatenating dimensions. Let the original candidate feature set be:

[0115] in, For the first feature vector, This represents the total number of candidate feature variables.

[0116] The feature set X is input into the random forest model, and the importance of each feature is evaluated by the change in out-of-bag sample error. The importance score for each feature variable is defined as follows:

[0117] in, The number of decision trees in the random forest. Indicates the first The OOB error corresponding to each decision tree Indicates the random perturbation of the first The corresponding OOB error after each feature variable This represents the contribution of each feature vector to the classification process.

[0118] All candidate features are ranked according to their importance scores, and the top K features are selected as important feature variables to form a feature subset.

[0119] in, This represents the Kth important eigenvector.

[0120] After feature selection, the selected K important feature variables and the corresponding historical load data are used as the input sequence to construct an LSTM model. Let the input vector of the LSTM model at time t be:

[0121] in, The historical load power at time t; After mapping using the LSTM model, the predicted user load power at the predicted time is obtained:

[0122] in, This represents the nonlinear mapping function of the LSTM obtained after training.

[0123] The solar power output is predicted using a solar irradiance-based sunshine index. Based on historical solar power output data, the solar power output data is statistically analyzed and the average sunshine index at each time point is taken into account to obtain the solar power output prediction result for the current time. The formula for predicting photovoltaic power output is:

[0124]

[0125]

[0126]

[0127]

[0128]

[0129] in, For the predicted photovoltaic power, take its average value as , For the current time, Let k be the time step for the prediction, and k be the time interval for the prediction to extend into the future. The sunshine index is calculated by taking its average value. , This represents the start time of the selected historical data time window. For the nearby The maximum historical value of photovoltaic power at the same time within a day. , .

[0130] S3. Construct a comprehensive objective function and apply voltage, power balance, and energy storage charging and discharging constraints.

[0131] Specifically, the calculation formula for the voltage fluctuation assessment of the Ath distributed energy storage access node is as follows:

[0132] in, This represents the voltage amplitude of the A-th distributed energy storage access node at time t. This indicates the voltage amplitude of the node when there is no energy storage output. This represents the control voltage reference value for the A-th distributed energy storage access node, and p represents the filter coefficient. Indicates the voltage acquisition time interval; Considering the investment costs, operation and maintenance costs, network loss costs, and electricity trading revenue required for power grid planning, an economic model is established:

[0133] in, Indicates investment cost, Indicates operation and maintenance costs. R represents the cost of grid loss, and R represents the revenue from electricity trading.

[0134] The comprehensive objective function obtained by combining the node voltage fluctuation and economic models is expressed as:

[0135] in, Represents the overall objective function. This indicates the voltage fluctuation threshold.

[0136] To ensure the safety and stability of the active distribution network, and to ensure the coordinated operation of distributed power sources and energy storage systems, the following constraints are set: Power balance constraints:

[0137] in, This represents the traditional power generation capacity of node i at time t. This represents the predicted photovoltaic output power of node i at time t. This represents the energy storage discharge power of node i at time t. This represents the predicted power of user load at node i at time t. Let represent the energy storage charging power of node i at time t, and M represent the set of all nodes.

[0138] Voltage amplitude constraint:

[0139] in, This represents the voltage magnitude at node i. and These represent the lower and upper limits of the voltage amplitude, respectively.

[0140] Distributed power generation output constraints:

[0141] in, This represents the actual photovoltaic output of node i. , These represent the lower and upper limits of the photovoltaic output of node i, respectively.

[0142] Energy storage system charge and discharge constraints:

[0143] in, and Let represent the maximum values ​​of the energy storage charging power and energy storage discharging power at node i, respectively. This represents the energy state of energy storage system i. and This indicates the lower and upper limits of its energy state. This represents the energy state of the energy storage system at node i at time t. and These represent the charge and discharge efficiencies, respectively.

[0144] S4. Calculate the net load forecast and residual based on the load power forecast and photovoltaic output power forecast, construct the residual penalty term and weighted comprehensive objective function, and generate global scheduling instructions by solving the particle swarm optimization algorithm.

[0145] Specifically, the predicted net load is the difference between the predicted load power and the predicted photovoltaic output power at time t, and the calculation formula is as follows:

[0146] The formula for calculating the residual is:

[0147] The formula for calculating the residual penalty term is: .

[0148] The objective function is solved using the particle swarm optimization algorithm, specifically as follows: Initialize the particle swarm, and calculate the fitness value of each particle based on the objective function and the residual penalty term. The formula for calculating fitness is:

[0149] in, Indicates the weighting coefficient. This represents the residual penalty coefficient; Each particle adjusts its velocity and position by referencing its own historical best position and the global best position of the entire particle swarm. The particle update formula is as follows:

[0150] in, and Let these represent the velocities of particle n at times t and t+1, respectively. This represents the inertial weights used to control the decay of particle velocity. and This indicates the step size for adjusting the particle's movement towards its own optimal point and the global optimal point. and A random number representing an increase in the randomness of the search. This represents the optimal solution found by particle n during the search process. This represents the globally optimal solution.

[0151] S5. The global scheduling command is sent to each energy storage unit, and the SOC partition management strategy is executed through the local controller under the edge computing architecture to adjust the charging and discharging power.

[0152] Specifically, the real-time state of charge of the energy storage unit The calculation formula is:

[0153] in, For rated capacity, The remaining energy storage capacity at the current moment; Real-time state of charge of the energy storage unit Divided into three regions: those exceeding 90% of the full SOC. Regions below 10% of full SOC Regions, and those in between The region is defined, and corresponding charge and discharge constraint strategies are implemented based on the state of charge region.

[0154] S6. In response to the event triggering condition, activate the fast response mode to adjust the energy storage operation strategy.

[0155] S7. Periodically record operating data, evaluate control performance, and provide feedback to optimize the prediction model and trigger thresholds.

[0156] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distributed energy storage and control system based on edge computing and artificial intelligence, characterized in that, include: The perception layer, deployed at the energy storage units and photovoltaic nodes in the distribution area, is used to collect the operating parameters of the energy storage units, grid status information and sunshine index in real time, and transmit them to the decision layer and control layer through a low-latency communication protocol. The decision-making layer integrates an artificial intelligence prediction module and a global scheduling optimization engine. The artificial intelligence prediction module includes a random forest feature selection unit, which is used to select features from multidimensional time features and environmental features. The RF-LSTM hybrid prediction unit is used to predict user load power based on filtered features and historical load data; the photovoltaic output prediction unit is used to predict photovoltaic power using the sunny index method combined with historical data; the global scheduling optimization engine is used to construct a comprehensive objective function, solve the objective function through the particle swarm optimization algorithm, and generate global scheduling instructions to be sent to the control layer. The control layer is deployed locally in each energy storage unit and adopts an edge computing architecture. Each energy storage unit is configured with a local controller, which is used to realize the coordinated control between energy storage units and integrate the event triggering mechanism.

2. The distributed energy storage and control system based on edge computing and artificial intelligence according to claim 1, characterized in that, The random forest feature selection unit performs the following operations: Timestamp data and corresponding meteorological data for different time periods are extracted from historical power system operation data. Multidimensional time and environmental features are then parsed to obtain these features. After preprocessing, an original candidate feature set is constructed by concatenating dimensions. Let the original candidate feature set be: in, For the first feature vector, This represents the total number of candidate feature variables; The feature set X is input into the random forest model, and the importance of each feature is evaluated by the change in out-of-bag sample error. The importance score of each feature variable is defined as follows: in, The number of decision trees in the random forest. Indicates the first The OOB error corresponding to each decision tree Indicates the random perturbation of the first The corresponding OOB error after each feature variable This represents the contribution of each feature vector to the classification process; Scoring based on feature importance All candidate features are ranked, and the top K features are selected as important feature variables to form a feature subset: in, This represents the Kth important eigenvector.

3. The distributed energy storage and control system based on edge computing and artificial intelligence according to claim 1, characterized in that, The RF-LSTM hybrid prediction unit performs the following operations: After feature selection, the selected K important feature variables and the corresponding historical load data are used as the input sequence to construct an LSTM model. Let the input vector of the LSTM model at time t be: in, The historical load power at time t; After mapping using the LSTM model, the predicted user load power at the predicted time is obtained: in, This represents the nonlinear mapping function of the LSTM obtained after training.

4. The distributed energy storage and control system based on edge computing and artificial intelligence according to claim 1, characterized in that, The photovoltaic power output prediction unit performs the following operations: The solar power output is predicted using a solar irradiance-based sunshine index. Based on historical solar power output data, the solar power output data is statistically analyzed and the average sunshine index at each time point is taken into account to obtain the solar power output prediction result for the current time. The formula for predicting photovoltaic power output is: in, For the predicted photovoltaic power, take its average value as , For the current time, Let k be the time step for the prediction, and k be the time interval for the prediction to extend into the future. The sunshine index is calculated by taking its average value. , This represents the start time of the selected historical data time window. For the nearby The maximum historical value of photovoltaic power at the same time within a day. , .

5. The distributed energy storage and control system based on edge computing and artificial intelligence according to claim 1, characterized in that, The comprehensive objective function is constructed as follows: The formula for calculating the voltage fluctuation assessment of the Ath distributed energy storage access node is as follows: in, This represents the voltage amplitude of the A-th distributed energy storage access node at time t. This indicates the voltage amplitude of the node when there is no energy storage output. This represents the control voltage reference value for the A-th distributed energy storage access node, and p represents the filter coefficient. Indicates the voltage sampling time interval; An economic model is established by combining the investment costs, operation and maintenance costs, network loss costs, and electricity trading revenue during power grid planning: in, Indicates investment cost, Indicates operation and maintenance costs. R represents the network loss cost, and R represents the electricity trading revenue; The comprehensive objective function obtained by combining the node voltage fluctuation and economic models is expressed as: in, Represents the overall objective function. This indicates the voltage fluctuation threshold.

6. The distributed energy storage and control system based on edge computing and artificial intelligence according to claim 1, characterized in that, The decision-making level also enforces the following constraints: Power balance constraints: in, This represents the traditional power generation capacity of node i at time t. This represents the predicted photovoltaic output power of node i at time t. This represents the energy storage discharge power of node i at time t. This represents the predicted power of user load at node i at time t. Let represent the energy storage charging power of node i at time t, and M represent the set of all nodes; Voltage amplitude constraint: in, This represents the voltage magnitude at node i. and These represent the lower and upper limits of the voltage amplitude, respectively. Distributed power generation output constraints: in, This represents the actual photovoltaic output of node i. , Let represent the lower and upper limits of the photovoltaic output of node i, respectively; Energy storage system charge and discharge constraints: in, and Let represent the maximum values ​​of the energy storage charging power and energy storage discharging power at node i, respectively. This represents the energy state of the energy storage system at node i. and This indicates the lower and upper limits of its energy state. This represents the energy state of the energy storage system at node i at time t. and These represent charging efficiency and discharging efficiency, respectively.

7. The distributed energy storage and control system based on edge computing and artificial intelligence according to claim 1, characterized in that, The global scheduling optimization engine of the decision-making layer is also used for: The predicted net load is the difference between the predicted load power and the predicted photovoltaic output power at time t, and the calculation formula is as follows: The formula for calculating the residual is: The formula for calculating the residual penalty term is: 。 8. The distributed energy storage and control system based on edge computing and artificial intelligence according to claim 1, characterized in that, The objective function is solved using the particle swarm optimization algorithm as follows: Initialize the particle swarm, and calculate the fitness value of each particle based on the objective function and the residual penalty term. The formula for calculating fitness is: in, Represents the nth particle. Indicates the weighting coefficient. This represents the residual penalty coefficient; Each particle adjusts its velocity and position by referencing its own historical best position and the global best position of the entire particle swarm. The particle update formula is as follows: in, and Let these represent the velocities of particle n at times t and t+1, respectively. This represents the inertial weights used to control the decay of particle velocity. and This indicates the step size for adjusting the particle's movement towards its own optimal point and the global optimal point. and A random number representing an increase in the randomness of the search. This represents the optimal solution found by particle n during the search process. This represents the globally optimal solution. This represents the nth particle.

9. The distributed energy storage and control system based on edge computing and artificial intelligence according to claim 1, characterized in that, The local controller of the control layer integrates the SOC partition management module and performs the following operations: Real-time state of charge of energy storage unit The calculation formula is: in, For rated capacity, The remaining energy storage capacity at the current moment; Real-time state of charge of the energy storage unit Divided into three regions: those exceeding 90% of the full SOC. Regions below 10% of full SOC Regions, and those in between The region is defined, and corresponding charge and discharge constraint strategies are implemented based on the state of charge region.

10. A distributed energy storage regulation method based on edge computing and artificial intelligence, characterized in that, include: S1. Collect the operating parameters, grid status, and weather index of the energy storage units and distributed power nodes in the distribution area, and transmit them through a low-latency communication protocol. S2. Based on the artificial intelligence model, feature filtering and load power prediction are performed on the data, and the solar power output prediction value is generated by using the clear sky index method. S3. Construct a comprehensive objective function and apply voltage, power balance, and energy storage charging and discharging constraints; S4. Calculate the net load forecast and residual based on the load power forecast and photovoltaic output forecast, construct the residual penalty term and weighted comprehensive objective function, and generate global scheduling instructions by solving the particle swarm optimization algorithm. S5. The global scheduling command is sent to each energy storage unit, and the SOC partition management strategy is executed by the local controller under the edge computing architecture to adjust the charging and discharging power. S6. In response to the event triggering condition, activate the fast response mode to adjust the energy storage operation strategy; S7. Periodically record operating data, evaluate control performance, and provide feedback to optimize the prediction model and trigger thresholds.

Images