Distributed energy storage system based on artificial intelligence real-time scheduling

Abstract

The application discloses a kind of distributed energy storage systems based on artificial intelligence real-time scheduling, belong to distributed energy storage technical field, system includes the demand scheduling end according to the demand data determined predicted scheduling instruction obtained;When receiving scheduling instruction, model analysis end continuously predicts load, new energy and demand event, generates global optimal scheduling strategy;Digital twin synchronization end is simulated to global optimal scheduling strategy and is checked, is sent to federal edge main station after passing through;Federal edge main station is according to the global optimal scheduling strategy after checking and updates slave station state regularly, maintains low-power consumption communication.By this way, the present application can solve the problem that the existing scheduling mode is to carry out energy scheduling after the demand event occurs, passive lag, slow response, poor scheduling rationality, to provide power for slave station in advance, ensure sufficient power.

Description

Technical Field

[0001] This invention relates to the field of distributed energy storage technology, and in particular to a distributed energy storage system based on real-time scheduling using artificial intelligence. Background Technology

[0002] In recent years, with the advancement of the construction of new power systems, a large number of distributed photovoltaic, small wind power, building waste energy, and elevator energy feedback have been connected to the grid, resulting in increased load fluctuations and a widening peak-valley difference in the distribution network. Remote demand dispatch has become a key means to ensure grid stability.

[0003] Among them, the electricity generated by distributed small-scale power generation devices such as wind power and photovoltaic power generation is intermittent and unstable, which can easily lead to energy waste. Distributed energy storage systems, as an effective means of energy storage and management, can store electricity during periods of low electricity demand and release it during peak periods, which helps to balance electricity supply and demand and improve energy utilization efficiency.

[0004] However, existing distributed energy storage systems suffer from the following problems: First, current scheduling methods only initiate energy dispatch after demand events occur, resulting in passive, delayed, slow, and poorly rational scheduling. Second, energy utilization is inefficient; existing distributed energy storage systems mostly use AC grid-connected charging and discharging, leading to high inverter losses, insufficient utilization of surplus and off-peak electricity, and significant energy waste. Third, existing distributed energy storage systems suffer from redundant and unreliable communication, relying on timed broadcasts and periodic reporting, resulting in high communication power consumption and susceptibility to congestion. In the event of public network delays or packet loss, this can directly lead to scheduling failure. Finally, the decision-making algorithms of existing distributed energy storage systems employ greedy rules or simple local optimizations, making them prone to getting trapped in local optima and unable to achieve global coordination and lifetime balance. Summary of the Invention

[0005] This invention relates to the field of distributed energy storage technology. The purpose of this invention is to solve the problems of existing scheduling methods, which only perform energy scheduling after demand events occur, resulting in passive and delayed scheduling, slow response, and poor scheduling rationality. The invention proposes a distributed energy storage system based on real-time scheduling using artificial intelligence.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A distributed energy storage system based on real-time scheduling using artificial intelligence includes:

[0008] The demand scheduling end determines the predicted scheduling instructions based on the acquired demand data, according to the constraints.

[0009] On the model analysis end, when a scheduling instruction is received, predictive analysis and multi-objective balance analysis are performed on the acquired multi-dimensional data to output an initial global optimal scheduling strategy; the initial global optimal scheduling strategy is adjusted according to the acquired pre-analysis report to obtain the global optimal scheduling strategy.

[0010] The digital twin synchronization terminal is used to perform scheduling rehearsals based on the initial global optimal scheduling strategy, analyze the rehearsal results, and output a rehearsal analysis report;

[0011] The federated edge master station matches suitable slave stations according to the global optimal scheduling strategy and generates slave station scheduling instructions;

[0012] An energy-saving intelligent slave station is used to execute the issued slave station scheduling commands and to collect the status data after execution and transmit it to the model analysis terminal.

[0013] The multi-energy complementary acquisition unit charges the slave station according to the preset complementary rules.

[0014] Preferably, the step of determining the predicted scheduling instruction based on the constraint conditions and demand data includes: calculating the real-time demand using a sliding window method of a first preset duration; determining the real-time demand overrun rate and event level; predicting the future demand value using a weighted moving average method; triggering a pre-scheduling instruction when the future demand value is greater than the upper limit of the demand; determining the scheduling target value for the target power reduction based on the event level; and generating a scheduling instruction package based on the pre-scheduling instruction and the scheduling target value.

[0015] Preferably, the step of performing predictive analysis and multi-objective equilibrium analysis on the acquired multi-dimensional data to output an initial globally optimal scheduling strategy includes: using a Transformer-based time-series prediction method to perform predictive analysis on the sequence data in the multi-dimensional data, and outputting load prediction results and new energy prediction results; and using a deep reinforcement learning-based global optimal scheduling decision method to perform multi-objective equilibrium analysis on the load prediction results, new energy prediction results, and related data in the multi-dimensional data, and output an initial globally optimal scheduling strategy.

[0016] Preferably, the Transformer-based time-series forecasting method performs predictive analysis on sequence data in multi-dimensional data and outputs load forecasting results and new energy forecasting results, including: performing feature concatenation on the preprocessed sequence data to obtain fused features; performing position encoding on the fused feature vector to obtain coded feature data; and using a multi-layer encoder to perform multi-dimensional analysis on the coded feature data to output load forecasting results and new energy forecasting results.

[0017] Preferably, the method of using a multi-layer encoder to analyze the encoded feature data in multiple dimensions and output load forecasting results and renewable energy forecasting results includes: the first layer using a multi-head self-attention mechanism to capture the short-term dependence of the encoded feature data between adjacent time steps, obtaining noise-filtered features containing short-term fluctuations; the second layer using multi-head self-attention to extract intraday trend features from the noise-filtered features; the third layer using multi-head self-attention to analyze the long-term regularity features formed by the intraday trend features; the fourth layer using multi-head self-attention to analyze the long-term regularity features, determine the mutual attention of all time steps, and output the globally optimal time series features containing short-term fluctuations, intraday trend cross-day cycles, meteorological coupling, and time features; and determining the load forecasting results and renewable energy forecasting results based on the globally optimal time series features.

[0018] Preferably, the global optimal scheduling decision-making method based on deep reinforcement learning performs multi-objective equilibrium analysis on load forecasting results, new energy forecasting results, and related data in multi-dimensional data to output an initial global optimal scheduling strategy. This includes: constructing an initial state vector from the load forecasting results, new energy forecasting results, and related data in multi-dimensional data; analyzing the initial state vector using an Actor network to output scheduling instructions; pruning the scheduling instructions based on preset hard constraints to obtain target scheduling instructions; acquiring the state data of the federated edge master station after executing the target scheduling instructions; processing the state data based on a preset multi-objective weighted reward function to determine the reward value; evaluating the value of this action using a Critic network; and updating the target scheduling instructions based on the calculated Critic loss and Actor loss to obtain the initial global optimal scheduling strategy.

[0019] Preferably, the method of predicting future demand using a weighted moving average includes: predicting the load for the next 1 to 5 minutes using a third-order weighted moving average model, the calculation formula of which is:

[0020] ,

[0021] in, , , These are the weighting coefficients. This represents the total active power of the transformer substation at the current moment. This represents the total active power of the transformer area one minute ago. This represents the total active power of the transformer area 2 minutes ago. This represents the predicted future demand value.

[0022] Preferably, the process of performing scheduling pre-simulation based on the initial global optimal scheduling strategy, analyzing the pre-simulation results, and outputting a pre-simulation analysis report includes: constructing a mirror model, executing the initial global optimal scheduling strategy using the mirror model to perform instruction pre-simulation simulation, obtaining pre-simulation results; analyzing the relationship between the pre-simulation results and the satisfaction of hard constraints, and outputting a pre-simulation analysis report.

[0023] Preferably, the step of matching suitable slave stations according to the globally optimal scheduling strategy and generating slave station scheduling instructions includes: acquiring slave station data that meets the scheduling conditions; and using a greedy algorithm to match the globally optimal scheduling strategy with the slave station data based on the second constraint condition, thereby generating slave station scheduling instructions.

[0024] Preferably, the charging of the slave station according to the preset complementary rules includes: first charging the slave station in the order of photovoltaic, elevator feedback, surplus energy, and off-peak electricity, and then using grid power supply.

[0025] Compared with existing technologies, the present invention provides a distributed energy storage system based on real-time scheduling of artificial intelligence, which has the following beneficial effects.

[0026] 1. After the system of this invention is powered on, the demand dispatching terminal determines the predicted dispatching instructions based on the acquired demand data; the model analysis terminal continuously predicts load, new energy sources, and demand events, generating a globally optimal dispatching strategy; the digital twin synchronization terminal simulates and verifies the globally optimal dispatching strategy, and after passing the verification, sends it to the federated edge master station; the federated edge master station selects the optimal combination of slave stations and executes demand response according to DC priority routing; when supplying power to slave stations, the multi-energy complementary acquisition unit prioritizes charging the slave stations, and the insufficient part is supplemented by off-peak electricity from the power grid. In this way, this invention can solve the problems of existing dispatching methods that only perform energy dispatching after demand events occur, resulting in passive and delayed dispatching, slow response, and poor dispatching rationality, by supplying power to slave stations in advance to ensure sufficient power.

[0027] 2. The present invention adopts a multi-energy complementary acquisition unit to prioritize photovoltaic, elevator feedback and building surplus energy for slave station charging, and supplements the insufficient part with off-peak electricity from the power grid, which can reduce power loss.

[0028] Other advantages, objectives and features of the invention will be set forth in part in the description which follows; and in part will be apparent to those skilled in the art upon examination of the following description; or may be learned from practice of the invention. Attached Figure Description

[0029] Figure 1 This is a schematic block diagram of a distributed energy storage system based on real-time scheduling using artificial intelligence, according to the present invention.

[0030] Figure 2 This is a schematic diagram of the structure of an electronic device according to the present invention. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0032] In some embodiments of the present invention, reference is made to... Figure 1-2 A distributed energy storage system based on real-time scheduling using artificial intelligence, comprising:

[0033] The demand dispatching unit determines the predicted dispatch instructions based on the acquired demand data, according to the constraints. The constraints include the allowable response time, recovery time, demand response assessment threshold, and allowable adjustment range for the distribution area.

[0034] Demand data includes real-time total power of the transformer area, maximum demand setting of the transformer area, three-phase current of the transformer area, and three-phase voltage of the transformer area. Time information: time, day type (weekday / weekend / holiday), historical demand event records, and historical maximum demand.

[0035] On the model analysis side, predictive analysis and multi-objective balance analysis are performed on the acquired multi-dimensional data to output the initial global optimal scheduling strategy; and the initial global optimal scheduling strategy is adjusted according to the acquired pre-analysis report to obtain the global optimal scheduling strategy.

[0036] Multi-dimensional data includes: power grid dispatching end: demand dispatching target value, event level, and constraint range; federated edge master station: available capacity, SOC, SOH, location, and losses; smart slave station: real-time power, temperature, fault status, and historical operating data; meteorology: sunshine, wind speed, temperature, and precipitation; electricity price: peak / flat / valley electricity price, demand price, and ancillary service price; historical load, historical dispatching strategies, and reward / penalty records.

[0037] The digital twin synchronization terminal is used to perform scheduling rehearsals based on the initial global optimal scheduling strategy, analyze the rehearsal results, and output a rehearsal analysis report.

[0038] The federated edge master station matches suitable slave stations according to the globally optimal scheduling strategy and generates slave station scheduling instructions.

[0039] Energy-saving intelligent slave stations are used to execute issued slave station scheduling commands and collect status data after execution and transmit it to the model analysis terminal.

[0040] The multi-energy complementary acquisition unit charges the slave station according to the preset complementary rules.

[0041] Specifically, after the system powers on, the demand scheduling terminal determines the predicted scheduling instructions based on the acquired demand data. When a scheduling instruction is received, the model analysis terminal continuously predicts load, new energy sources, and demand events to generate a globally optimal scheduling strategy. The digital twin synchronization terminal simulates and verifies the globally optimal scheduling strategy, and after passing the verification, sends it to the federated edge master station. The federated edge master station periodically updates the slave station status, maintains low-power communication, and quickly reassembles in case of anomalies.

[0042] When the power grid issues a demand command, the federated edge master station selects the optimal combination of slave stations and executes the demand response according to the DC priority route. After the response is completed, the slave stations smoothly switch back to standby mode, and the status data is transmitted back to the cloud to complete the review, settlement, and model evolution.

[0043] The energy-saving intelligent slave station collects battery status data in real time, estimates the state of charge and health of the energy storage unit, and reports this data to the federated edge master station, demand scheduling terminal, and model analysis terminal. Furthermore, when supplying power to the slave station, the multi-energy complementary acquisition unit prioritizes charging the slave station, with any shortfall supplemented by off-peak electricity from the grid.

[0044] In this way, the present invention can solve the problems of existing scheduling methods, which only perform energy scheduling after demand events occur, resulting in passive, delayed, slow, and poor scheduling rationality. It provides power to slave stations in advance, ensuring sufficient power. At the same time, by using a multi-energy complementary data acquisition unit, it prioritizes charging slave stations with photovoltaic power, elevator feedback, and building surplus energy, and supplements the shortfall with off-peak electricity from the grid, which can reduce power loss.

[0045] In one embodiment of the present invention, the step of determining the predicted scheduling instruction based on demand data according to constraints includes: calculating real-time demand using a sliding window method of a first preset duration; determining the real-time demand overrun rate and event level; predicting future demand values ​​using a weighted moving average method; triggering a pre-scheduling instruction when the future demand value is greater than the upper limit of demand; determining the scheduling target value for target power reduction based on the event level; and generating a scheduling instruction package based on the pre-scheduling instruction and the scheduling target value.

[0046] Preferably, the first preset duration is 15 minutes.

[0047] The sliding window method is used to obtain 15 sample data points of the total active power of the transformer area for the first preset time period, and the maximum value among them is taken as the current real-time demand.

[0048] The real-time demand exceedance rate and event level are determined. The real-time demand exceedance rate is the ratio of real-time demand to the upper limit of demand. When the real-time demand exceedance rate is less than 1, there is no exceedance and scheduling is not triggered. If the real-time demand exceedance rate is greater than or equal to 1.00 and less than 1.05, it is judged as a level 1 demand event, indicating a normal exceedance. If the real-time demand exceedance rate is greater than or equal to 1.05, it is judged as a level 2 demand event, indicating an emergency exceedance.

[0049] In one embodiment of the present invention, a weighted moving average method is used to predict future demand; to achieve advance prediction, a third-order weighted moving average model is used to predict the load for the next 1 to 5 minutes; the calculation formula is as follows:

[0050] ,

[0051] in, , , These are the weighting coefficients. This represents the total active power of the transformer substation at the current moment. This represents the total active power of the transformer area one minute ago. This represents the total active power of the transformer area 2 minutes ago. This represents the predicted future demand value.

[0052] Preferred, 0.6 0.3 It is 0.1.

[0053] When the future demand value exceeds the upper limit of the demand, a prediction over-limit flag is output, triggering a pre-scheduling instruction to start pre-scheduling in advance.

[0054] The target power reduction scheduling value is determined based on the event level. The target scheduling value is calculated as the sum of the difference between real-time demand and the upper limit of demand, plus a safety margin. The safety margin is a matching value between 5% and 10% of the upper limit of demand, which can be obtained through a lookup table. Then, a scheduling instruction package is generated based on the target power reduction scheduling value, response time limit, event level, instruction type, and allowable output range.

[0055] The target power is reduced according to the scheduling instruction packet, and the actual reduction value is obtained; it is determined whether the absolute value of the difference between the actual reduction value and the scheduling target value is less than or equal to the assessment threshold; if so, it is qualified; if not, it is unqualified, and the pass rate of the response is calculated.

[0056] A stop command will be issued when any of the following conditions are met: ① The real-time demand is less than 0.85 times the upper limit of demand; ② The cumulative triggering time reaches the preset scheduling time; ③ A stop command is issued manually; ④ A fault or maintenance signal is triggered in the transformer area.

[0057] The power grid demand dispatching terminal collects three-phase voltage, current, and real-time active power data of the distribution area, calculates real-time demand using a 15-minute sliding window algorithm, and predicts future load trends using a weighted moving average. It then determines the demand exceedance level and generates dispatch target values. According to preset rules, it generates encrypted dispatching instructions and sends them to the system. During the dispatching process, it monitors the energy storage system's response power in real time, judges the response qualification based on assessment thresholds, and issues a stop instruction when the demand falls back to the safe range or reaches the set time, thus completing the control, monitoring, and settlement of the entire demand response process.

[0058] Through the configuration method of this embodiment, the present invention can predict and pre-schedule preparations in advance, avoiding power shortages and allowing for a more composed backend response. Simultaneously, the tiered triggering method ensures smooth handling of normal events and rapid response to emergency events.

[0059] In one embodiment of the present invention, predictive analysis and multi-objective equilibrium analysis are performed on the acquired multi-dimensional data to output an initial global optimal scheduling strategy, including: a Transformer-based time-series prediction method to predictive analyze the sequence data in the multi-dimensional data and output load prediction results and new energy prediction results; and a deep reinforcement learning-based global optimal scheduling decision method to perform multi-objective equilibrium analysis on the load prediction results, new energy prediction results, and related data in the multi-dimensional data and output an initial global optimal scheduling strategy.

[0060] Furthermore, in one embodiment of the present invention, the Transformer-based time-series forecasting method performs predictive analysis on sequence data in multi-dimensional data and outputs load forecasting results and new energy forecasting results, including: performing feature concatenation on the preprocessed sequence data to obtain fused features; performing position encoding on the fused feature vector to obtain encoded feature data; and using a multi-layer encoder to perform multi-dimensional analysis on the encoded feature data to output load forecasting results and new energy forecasting results.

[0061] The sequence data includes historical load sequences, meteorological sequences, and historical renewable energy output sequences, while also acquiring corresponding hourly and daily calendar features; the daily types include weekdays, weekends, and holidays.

[0062] Specifically, the preprocessing process mainly includes: handling missing values, removing outliers, and normalizing the sequence data.

[0063] The four types of information—load, new energy, meteorology, and time—are concatenated into a fusion feature, with each row representing a complete feature vector for a time step. Sine and cosine position coding is used to encode the fusion feature, adding position information to the time series data to facilitate understanding of the time sequence.

[0064] Furthermore, in one embodiment of the present invention, a multi-layer encoder is used to analyze the encoded feature data in multiple dimensions to output load forecasting results and new energy forecasting results. This includes: a first layer using a multi-head self-attention mechanism to capture the short-term dependencies between adjacent time steps of the encoded feature data, obtaining noise-filtered features containing short-term fluctuations; a second layer using multi-head self-attention to extract intraday trend features from the noise-filtered features; a third layer using multi-head self-attention to analyze the long-term regularity features formed by the intraday trend features; a fourth layer using multi-head self-attention to analyze the long-term regularity features, determining the mutual attention of all time steps, and outputting globally optimal time-series features containing short-term fluctuations, intraday trend cross-day cycles, meteorological coupling, and time features; and determining the load forecasting results and new energy forecasting results based on the globally optimal time-series features.

[0065] Specifically, the first layer uses a multi-head self-attention mechanism to capture the short-term dependencies between adjacent time steps of the encoded feature data and outputs attention weights; after processing by residual connection, normalization and nonlinear transformation of the feedforward network, noise-filtered features with short-term fluctuations are obtained.

[0066] The second layer uses multi-head self-attention analysis to analyze the correlation between 4- to 24-hour periodic data formed by noise filtering features, focusing on the analysis of morning peak, midday flat, evening peak, and nighttime trough. After residual, normalization, and feedforward network (FFN) processing, intraday trend features containing intraday bimodal peaks and load curve shapes are obtained.

[0067] The third layer employs multi-head self-attention analysis to analyze the long-term regularity features formed by intraday trend characteristics. This involves integrating at least one month, one year, or all intraday trend characteristics; analyzing the correlation of intraday trend characteristics across days; learning the differences between weekdays, weekends, and holidays; and extracting long-term regularity features containing weekly cycles, daily type patterns, and historical similar daily patterns after residual analysis, normalization, and feedforward network (FFN) processing.

[0068] The fourth layer employs multi-head self-attention analysis to identify long-term regularity characteristics, determine the mutual attention of all time steps, and integrate the feature data corresponding to load, photovoltaic, wind speed, temperature, hourly and daily types. After processing by residuals, normalization and feedforward network FFN, the output is a globally optimal time series feature that includes short-term fluctuations, intraday trend cross-day cycle, meteorological coupling and time features.

[0069] Finally, the fully connected layer maps the globally optimal time-series features to predicted values ​​for the next 0-4 hours. After inverse normalization and constraint correction, it determines the load forecast results including the load forecast curve, and the new energy forecast results including the photovoltaic forecast curve and the wind power forecast curve.

[0070] This invention employs the Transformer time-series forecasting algorithm to perform high-precision forward forecasting of load, photovoltaic output, and wind power output in a distribution area for the next 0-4 hours, possessing proactive pre-scheduling capabilities to avoid passive impacts. Furthermore, it simultaneously extracts short-term fluctuations, intraday trends, and weekly patterns, resulting in strong forecast stability.

[0071] In one embodiment of the present invention, a global optimal scheduling decision method based on deep reinforcement learning performs multi-objective equilibrium analysis on load forecasting results, new energy forecasting results, and related data in multi-dimensional data to output an initial global optimal scheduling strategy. This includes: constructing an initial state vector from the load forecasting results, new energy forecasting results, and related data in multi-dimensional data; analyzing the initial state vector using an Actor network to output scheduling instructions; pruning the scheduling instructions based on preset hard constraints to obtain target scheduling instructions; acquiring the state data of the federated edge master station after executing the target scheduling instructions; processing the state data based on a preset multi-objective weighted reward function to determine the reward value; evaluating the value of this action using a Critic network; and updating the target scheduling instructions based on the calculated Critic loss and Actor loss to obtain the initial global optimal scheduling strategy.

[0072] The associated data includes scheduling target value, response time window, allowable power deviation, equipment status of each slave station (state of charge of energy storage unit, health status of energy storage unit, dispatchable capacity, maximum charge and discharge power, battery temperature, current voltage, etc.), and constraints (safe range of state of charge of energy storage unit, maximum output limit of single station, upper limit of three-phase imbalance, upper and lower voltage limits).

[0073] The state space is used to construct an initial state vector from all observable information, serving as input data for reinforcement learning. The action space contains the charging and discharging power commands for each energy storage slave station; these are continuous values ​​that can be directly issued and executed. The reward function is a multi-objective weighted reward function, which must simultaneously satisfy the objectives of sufficient demand, maximum revenue, minimum loss, and minimum lifetime decay.

[0074] The formula for the multi-objective weighted reward function is:

[0075]

[0076] in, As a reward value, Rewards for qualified demand response For electricity price arbitrage profits, Costs associated with inverter and line losses, For the cost of battery life degradation, , , , These are the weighting coefficients corresponding to these objectives; , The sum of is 1. , The sum of is 1.

[0077] The initial state vector is analyzed using an Actor network, and the scheduling instructions are output, which is a set of optimal charging and discharging power instructions.

[0078] Based on preset hard constraints, the dispatch instructions are trimmed and those that are not met are adjusted to obtain the target dispatch instructions. The preset hard constraints include: the total output meets the demand target, the remaining power of the energy storage unit is greater than or equal to 20% and less than or equal to 80%, the power is less than or equal to the maximum power, and the output is automatically allocated so that the three-phase imbalance is less than 5%.

[0079] Obtain status data after the federated edge master station executes the target scheduling command; this status data includes the actual output power, actual remaining power, actual health, actual temperature, and actual voltage of the energy storage unit.

[0080] Based on the pre-set multi-objective weighted reward function, the state data is processed, and the overall reward value of demand response qualification reward, electricity price arbitrage income, inverter and line loss cost, and battery life decay cost is analyzed to determine the quality of this scheduling. At the same time, a Critic network is used to evaluate the value of this action and evaluate its quality. Then, the target scheduling instruction is updated according to the overall calculated Critic loss and Actor loss to obtain an initial global optimal scheduling strategy that includes the optimal output instruction of each slave station, the optimal start time, the optimal stop time, the smooth power curve, and the life balance constraint.

[0081] This invention employs a deep deterministic policy gradient algorithm based on deep reinforcement learning to construct a globally optimal scheduling decision model. Using load and renewable energy forecasts as prior information, it constructs a state space encompassing system operating status, equipment health status, and grid constraints. The charging and discharging power of each energy storage unit serves as the continuous action space. A multi-objective reward function is designed, integrating demand response revenue, peak-valley arbitrage revenue, conversion loss cost, and battery lifespan degradation cost. The optimal scheduling strategy is output through an Actor network, executed after constraint pruning, and evaluated by a Critic network to update the network. This achieves globally self-optimizing scheduling that ensures qualified demand response, maximizes revenue, minimizes losses, and balances battery lifespan.

[0082] In one embodiment of the present invention, scheduling pre-simulation is performed according to an initial global optimal scheduling strategy, and the pre-simulation results are analyzed and a pre-simulation analysis report is output. This includes: constructing a mirror model, executing the initial global optimal scheduling strategy using the mirror model to perform instruction pre-simulation simulation, and obtaining pre-simulation results; analyzing the relationship between the pre-simulation results and the satisfaction of hard constraints, and outputting a pre-simulation analysis report.

[0083] Specifically, based on the pre-built digital twin model of the entire system, the key elements of the system are mirrored in real time to build a mirror model. The mirror content mainly includes the federated edge master station: topology, slave station list, real-time status; intelligent slave station: voltage, current, power, SOC, temperature, location; power grid: transformer impedance, voltage limit, power flow constraints.

[0084] The construction sequence is: master station—slave station—line—load—power grid twin, completing the entire mirroring process from skeleton to whole, forming a mirror model.

[0085] The initial global optimal scheduling strategy and power target are input into the mirror model; power flow calculation is performed, followed by calculations of voltage, current, and three-phase imbalance to obtain prediction results; the relationship between the prediction results and the satisfaction of hard constraints is analyzed, and a prediction analysis report is output.

[0086] Hard constraints include threshold conditions such as whether there is overvoltage, overcurrent, overheating, and three-phase balance, which can be adjusted according to the actual situation.

[0087] Then, on the model analysis side, the initial global optimal scheduling strategy is further adjusted based on the pre-analysis report, and unreasonable parameters are appropriately adjusted to obtain the global optimal scheduling strategy.

[0088] In one embodiment of the present invention, the step of matching a suitable slave station according to the globally optimal scheduling strategy and generating a slave station scheduling instruction includes: obtaining slave station data that meets the scheduling conditions; and using a greedy algorithm to match the globally optimal scheduling strategy with the slave station data based on a second constraint condition, thereby generating a slave station scheduling instruction.

[0089] Slave data includes dischargeable power, inverter loss factor, electrical distance / line impedance, phase, battery status, etc.

[0090] The system constraints corresponding to the second constraint condition include: minimum state of charge of 20%, maximum single-station output power, upper limit of three-phase imbalance ≤5%, and total power allowable error ≤3% of demand.

[0091] The hard constraints corresponding to the second constraint include power constraints, energy storage unit health constraints, remaining power constraints, power upper limit constraints, and temperature constraints.

[0092] A greedy algorithm is used to match the globally optimal scheduling strategy with slave station data to generate slave station scheduling instructions. This includes: analyzing loss, distance, and three-phase balance using a multi-objective scoring method to calculate the comprehensive priority score for each slave station; sorting the slave stations according to their comprehensive priority scores from highest to lowest; adding the highest-ranked slave stations to the optimal combination; stopping the selection process when the total power of the added slave stations reaches the preset target power, thus obtaining the initial optimal combination; and fine-tuning the initial optimal combination using constraint programming to address issues such as slightly over / under power, three-phase imbalance, and overload of individual stations. When the power meets the demand conditions, the state of charge is in a safe state, the three phases are balanced, and there is no overload or over-temperature, the optimal slave station combination is output.

[0093] Then, based on the optimal slave station combination, slave station scheduling instructions are generated.

[0094] The federated edge master station employs an optimal slave station combination algorithm that combines greedy algorithms with constraint programming. First, it performs constraint filtering on its assigned intelligent slave stations to obtain an available set. Then, it constructs a multi-objective priority scoring function based on inverter losses, electrical distance, and phase affiliation to rank the available slave stations. Through a greedy strategy, it progressively selects high-priority slave stations, bringing the total output close to the demand target. Finally, constraint programming is used for power calibration and fine-tuning of three-phase imbalance. Under the premise of satisfying power constraints, battery constraints, temperature constraints, and power quality constraints, it outputs the optimal slave station combination that achieves the target total output, minimizes losses, minimizes distance, and maximizes three-phase balance. This ensures the rationality and stability of slave station scheduling and avoids underpowerment issues.

[0095] In one embodiment of the present invention, charging the slave station according to a preset complementary rule includes: first charging the slave station in the order of photovoltaic, elevator feedback, surplus energy, and off-peak electricity, and then using grid power supply.

[0096] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any machine-readable storage medium for use by, or in conjunction with, instruction execution systems, apparatuses or devices (such as computer-based systems, systems including processor 510 or other systems that can fetch and execute instructions from, or in connection with, such instruction execution systems, apparatuses or devices).

[0097] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory 520 and executed by a suitable instruction execution system.

[0098] Typically, a machine-executable program 41 may include routines, programs, object programs, components, logic, data structures, etc., that perform specific tasks or implement specific abstract data types. In a distributed cloud computing environment, the machine-executable program 41 may reside on local or remote computing system storage media, including storage devices.

[0099] Electronic device 50 may include a processor 510 adapted to execute stored instructions and a memory 520 that provides temporary storage space for the operation of instructions during operation. The processor 510 may be a single-core processor 510, a multi-core processor 510, a computing cluster, or any other configuration. The memory 520 may include random access memory (RAM), read-only memory, flash memory, or any other suitable storage system.

[0100] The processor 510 can be connected via a system interconnect (e.g., PCI, PCI-Express, etc.) to an I / O interface (input / output interface) suitable for connecting the electronic device 50 to one or more I / O devices (input / output devices). I / O devices may include, for example, a keyboard and indicating devices, where indicating devices may include a touchpad or touchscreen, etc. I / O devices may be built into the electronic device 50 or may be external devices connected to a computing device.

[0101] The processor 510 can also be linked via a system interconnect to a display interface suitable for connecting the electronic device 50 to a display device. The display device may include a display screen that is a built-in component of the electronic device 50. The display device may also include an external computer monitor, television, or projector connected to the electronic device 50. Furthermore, a network interface controller (NIC) may be adapted to connect the electronic device 50 to a network via a system interconnect. In some embodiments, the NIC may use any suitable interface or protocol (such as an Internet Minicomputer System Interface) to transmit data. The network may be a cellular network, a radio network, a wide area network (WAN), a local area network (LAN), or the Internet, etc. Remote devices can connect to the computing device via the network.

[0102] The flowchart provided in this embodiment is not intended to indicate that the operations of the method will be performed in any particular order, or that all operations of the method are included in every case. Furthermore, the method may include additional operations. Within the scope of the technical concept provided by the method in this embodiment, additional variations can be made to the above method.

Claims

1. A distributed energy storage system based on real-time scheduling using artificial intelligence, characterized in that, include: The demand scheduling end determines the predicted scheduling instructions based on the acquired demand data, according to the constraints. On the model analysis end, when a scheduling instruction is received, predictive analysis and multi-objective balance analysis are performed on the acquired multi-dimensional data to output an initial global optimal scheduling strategy; the initial global optimal scheduling strategy is adjusted according to the acquired pre-analysis report to obtain the global optimal scheduling strategy. The digital twin synchronization terminal is used to perform scheduling rehearsals based on the initial global optimal scheduling strategy, analyze the rehearsal results, and output a rehearsal analysis report; The federated edge master station matches suitable slave stations according to the global optimal scheduling strategy and generates slave station scheduling instructions; Energy-saving intelligent slave station, used to execute the issued slave station scheduling instructions; And collect the status data after execution and transmit it to the model analysis terminal; The multi-energy complementary acquisition unit charges the slave station according to the preset complementary rules.

2. The distributed energy storage system according to claim 1, characterized in that, The method of determining the predicted scheduling instructions based on demand data under constraints includes: calculating real-time demand using a sliding window method of a first preset duration; determining the real-time demand overrun rate and event level; predicting future demand values ​​using a weighted moving average method; triggering a pre-scheduling instruction when the future demand value exceeds the upper limit of demand; determining the scheduling target value for target power reduction based on the event level; and generating a scheduling instruction package based on the pre-scheduling instruction and the scheduling target value.

3. The distributed energy storage system according to claim 1, characterized in that, The aforementioned predictive analysis and multi-objective equilibrium analysis of the acquired multi-dimensional data to output an initial globally optimal scheduling strategy includes: a Transformer-based time-series prediction method to predict and analyze the sequence data in the multi-dimensional data, and output load prediction results and new energy prediction results; and a deep reinforcement learning-based global optimal scheduling decision method to perform multi-objective equilibrium analysis on the load prediction results, new energy prediction results, and related data in the multi-dimensional data, and output an initial globally optimal scheduling strategy.

4. The distributed energy storage system according to claim 3, characterized in that, The Transformer-based time-series forecasting method described above performs predictive analysis on sequence data in multi-dimensional data and outputs load forecasting results and new energy forecasting results. It includes: performing feature concatenation on the preprocessed sequence data to obtain fused features; performing position encoding on the fused feature vector to obtain encoded feature data; and using a multi-layer encoder to perform multi-dimensional analysis on the encoded feature data to output load forecasting results and new energy forecasting results.

5. The distributed energy storage system according to claim 4, characterized in that, The method employing a multi-layer encoder for multi-dimensional analysis of encoded feature data to output load forecasting and renewable energy forecasting results includes: The first layer uses a multi-head self-attention mechanism to capture short-term dependencies between adjacent time steps of the encoded feature data, obtaining noise-filtered features containing short-term fluctuations; the second layer uses multi-head self-attention to extract intraday trend features from the noise-filtered features; the third layer uses multi-head self-attention to analyze the long-term regularity features formed by the intraday trend features; the fourth layer uses multi-head self-attention to analyze the long-term regularity features, determine the mutual attention of all time steps, and output a globally optimal time-series feature containing short-term fluctuations, intraday trend cross-day cycles, meteorological coupling, and time features; based on the globally optimal time-series feature, the load forecasting and renewable energy forecasting results are determined.

6. The distributed energy storage system according to claim 3, characterized in that, The aforementioned global optimal scheduling decision-making method based on deep reinforcement learning performs multi-objective equilibrium analysis on load forecasting results, new energy forecasting results, and related data from multi-dimensional data to output an initial global optimal scheduling strategy. This includes: constructing an initial state vector from the load forecasting results, new energy forecasting results, and related data from multi-dimensional data; analyzing the initial state vector using an Actor network to output scheduling instructions; pruning the scheduling instructions based on preset hard constraints to obtain target scheduling instructions; acquiring the state data of the federated edge master station after executing the target scheduling instructions; processing the state data based on a preset multi-objective weighted reward function to determine the reward value; evaluating the value of this action using a Critic network; and updating the target scheduling instructions based on the calculated Critic loss and Actor loss to obtain the initial global optimal scheduling strategy.

7. The distributed energy storage system according to claim 2, characterized in that, The method of predicting future demand using a weighted moving average includes: predicting the load for the next 1 to 5 minutes using a third-order weighted moving average model, the calculation formula of which is: ， in, , , These are the weighting coefficients. This represents the total active power of the transformer substation at the current moment. This represents the total active power of the transformer area one minute ago. This represents the total active power of the transformer area 2 minutes ago. This represents the predicted future demand value.

8. The distributed energy storage system according to claim 1, characterized in that, The process of performing scheduling pre-simulation based on the initial global optimal scheduling strategy, analyzing the pre-simulation results, and outputting a pre-simulation analysis report includes: constructing a mirror model, executing the initial global optimal scheduling strategy using the mirror model to perform instruction pre-simulation simulation, obtaining pre-simulation results; analyzing the relationship between the pre-simulation results and the satisfaction of hard constraints, and outputting a pre-simulation analysis report.

9. The distributed energy storage system according to claim 1, characterized in that, The process of matching suitable slave stations according to the globally optimal scheduling strategy and generating slave station scheduling instructions includes: acquiring slave station data that meets the scheduling conditions; and using a greedy algorithm to match the globally optimal scheduling strategy with the slave station data based on the second constraint to generate slave station scheduling instructions.

10. The distributed energy storage system according to claim 1, characterized in that, The method of charging the slave station according to the preset complementary rules includes: first charging the slave station in the order of photovoltaic, elevator feedback, surplus energy, and off-peak electricity, and then using grid power.

Images