A method for charging and discharging an energy storage cabinet

Abstract

The application provides a charging and discharging method of an energy storage cabinet, comprising the following steps: S1, establishing an association mapping of space-time big data and digital twin data; S2, processing the space-time big data and the digital twin data; S3, extracting a change gradient of derived features and battery parameters, and fusing to mine related composite features; S4, labeling the related composite feature data set; S5, constructing an electricity demand prediction model, a renewable energy output prediction model and a battery health management model; S6, training the three models; S7, taking the output result of the model as input, and outputting a comprehensive prediction data set through a multilayer perception machine model; S8, training a fusion decision module; S9, generating a charging and discharging strategy based on the fusion decision module and real-time data; S10, issuing a charging and discharging strategy execution instruction; and S11, starting the charging and discharging process. The application integrates space-time big data and digital twin data, significantly reduces prediction error, more comprehensively fuses data, and has higher prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of energy storage cabinet charging and discharging technology, specifically to a charging and discharging method for an energy storage cabinet. Background Technology

[0002] The charging and discharging methods of energy storage cabinets are closely related to the real-world needs of global energy transition and grid operation. New energy power generation, such as photovoltaic and wind power, is highly intermittent and volatile. For example, photovoltaic output surges during the day when sunlight is strong, but sharply decreases at night or during cloudy or rainy weather. This leads to a mismatch between power supply and load. Without efficient charging and discharging control, excess energy is wasted, and insufficient power supply affects grid stability. As an energy buffer device, optimizing the charging and discharging methods of energy storage cabinets is key to solving the problem of new energy consumption, enabling peak shaving and valley filling, and smoothing the new energy power generation curve. Furthermore, to balance grid load, various regions implement peak-valley electricity pricing policies, where peak-hour prices are often several times higher than off-peak prices. Enterprises, especially energy-intensive industries, have a strong demand for precise charging and discharging of energy storage cabinets to reduce electricity costs. Traditional fixed-time charging and discharging methods are difficult to adapt to dynamically adjusted electricity prices and fluctuating production electricity demands, forcing the upgrading of charging and discharging methods towards intelligence and dynamism.

[0003] The existing charging and discharging methods for energy storage cabinets have the following main drawbacks:

[0004] 1. Traditional charging and discharging methods often rely on fixed patterns or manual experience, making it difficult to match the randomness of power output from renewable energy sources such as photovoltaics and wind power with the peak-valley fluctuations in user electricity consumption; 2. Battery charging and discharging are prone to safety accidents such as lithium plating and thermal runaway due to overcharging, over-discharging, and abnormal temperatures, and traditional methods are unable to detect hidden equipment faults in a timely manner; 3. Inappropriate charging and discharging depths and frequencies will accelerate battery aging, while simply pursuing equipment lifespan may lead to idle assets; 4. In large-scale scenarios such as electric vehicle charging pile clusters and park microgrids, user charging and discharging behaviors are randomly distributed in time and space, and traditional scheduling is prone to local overload or resource waste.

[0005] The applicant intends to solve the above-mentioned technical problems by integrating spatiotemporal big data with digital twins, but after a detailed search, no relevant technical problems were found.

[0006] Therefore, a new technical solution is needed to solve the above-mentioned technical problems. Summary of the Invention

[0007] This application provides a method for charging and discharging an energy storage cabinet, comprising the following steps:

[0008] S1: Collect spatiotemporal big data and digital twin data, and establish a correlation mapping between spatiotemporal big data and digital twin data; the spatiotemporal big data includes meteorological data, macro power grid data, user behavior data, historical charging and discharging data of energy storage cabinets, and battery health change trend data; the digital twin data includes micro power grid data, internal physical parameters of energy storage cabinets, dynamic response data of temperature control system, and geometric data of energy storage cabinets;

[0009] S2: Normalize spatiotemporal big data and digital twin data to the [0, 1] interval;

[0010] S3: Extract derived features from spatiotemporal big data, extract the gradient of battery parameter changes from digital twin data, and mine relevant composite features after fusing the two types of data; relevant composite features include charge and discharge efficiency related composite features, battery health related composite features, and load prediction related composite features.

[0011] S4: Combining historical charging and discharging data of energy storage cabinets, macro power grid data, and battery health trend data, label the relevant composite feature dataset of S3 with three types of labels, including electricity demand level, renewable energy generation status, and battery health risk level;

[0012] S5: Construct electricity demand forecasting models, renewable energy output forecasting models, and battery health management models;

[0013] S6: Using the dataset labeled in S4, train the electricity demand forecasting model, the renewable energy output forecasting model, and the battery health management model respectively.

[0014] S7: The output results of the trained electricity demand prediction model, renewable energy output prediction model and battery health management model are used as input. The results are fused and processed by the multilayer perceptron model to output a comprehensive prediction dataset. The comprehensive prediction data includes the predicted values ​​of future electricity demand, renewable energy output characteristics and battery health status.

[0015] S8: Using the comprehensive prediction dataset of S7 as input, train the fusion decision module and output the optimal charging and discharging strategy parameters; during training, the state space of the fusion decision module is constructed simultaneously. The state space includes core variables of battery health, core variables of power supply guarantee, core variables of economic benefit and corresponding value constraints. At the same time, the benchmark values ​​of core parameters and weight adjustment constraints are fixed in advance.

[0016] S9: Based on the fusion decision module and real-time data, specific charging and discharging strategies are generated;

[0017] S10: Send the charging / discharging strategy execution command to the terminal device;

[0018] S11: The charging and discharging strategy results are converted into standardized control commands, which are sent to the battery management system and energy storage converter of the energy storage cabinet through the communication interface. After receiving the commands, the battery management system performs battery protection-related operations, and after receiving the commands, the energy storage converter officially starts the charging and discharging process.

[0019] As a preferred embodiment, the process of establishing the association mapping between spatiotemporal big data and digital twin data in step S1 includes the following steps:

[0020] SA1: Establishes a unique interface identifier for spatiotemporal big data and digital twin data to ensure that each piece of data can be accurately matched to the same spatiotemporal scenario;

[0021] SA2: Preprocessing spatiotemporal big data and digital twin data;

[0022] SA3: Establish explicit association rule tables for spatiotemporal big data and digital twin data to achieve direct mapping;

[0023] Embed explicit rule tables into the data processing system, and automatically trigger mapping through association keys to achieve real-time matching;

[0024] SA4: Mining implicit correlation features of spatiotemporal big data and digital twin data, and constructing implicit correlation models to achieve indirect mapping;

[0025] SA5: Verify the mapping.

[0026] As a preferred option, SA6 is also included: dynamically updating the explicit rule table and the implicit association model.

[0027] As a preferred embodiment, the derived features include those derived from meteorological data, macro-grid data, user behavior data, historical charge / discharge data of energy storage cabinets, and battery health trend data. Specifically, the meteorological data features include 24-hour average temperature, temperature change rate, and rainfall intensity level; the macro-grid data features include regional load peak-valley difference, renewable energy output fluctuation coefficient, voltage stability index, and load growth rate; the user behavior data features include peak electricity consumption duration, daily average electricity load growth rate, peak-valley electricity consumption ratio, electricity load fluctuation characteristics, and merchant promotional activity identifiers; the historical charge / discharge data features include charge / discharge cycle efficiency, monthly average depth of discharge, and idle time percentage; and the battery health trend data features include monthly average SOH decay rate, cumulative SOH decay, remaining cycle life percentage, decay acceleration coefficient, weekly SOH change rate, health status change warning value, predicted remaining SOH life, and correlation coefficient between charge / discharge cycle count and SOH decay.

[0028] As a preferred approach, the gradients for battery parameter changes include capacity gradients, voltage / current gradients, temperature gradients, and state of health gradients. Among them, capacity gradients include SOC change gradients and available capacity decay gradients; voltage / current gradients include individual cell voltage change gradients, connection point current change gradients, and charge / discharge power adjustment gradients; temperature gradients include temperature rise gradients and temperature control response gradients; and state of health gradients include SOH decay gradients, internal resistance growth gradients, and polarization impedance change gradients.

[0029] As a preferred embodiment, the charge / discharge efficiency-related composite features include temperature-voltage-charging efficiency composite features, renewable energy-charge / discharge adaptation efficiency composite features, peak-valley load-discharge efficiency composite features, weather-temperature-PCS efficiency composite features, and idle time-charging efficiency composite features; battery health-related composite features include voltage fluctuation-temperature-internal resistance damage composite features, high-frequency cycling-SOH attenuation warning composite features, depth of discharge-health mutation risk composite features, and rainfall-temperature-battery corrosion risk composite features; load forecasting-related composite features include peak electricity consumption-load growth forecasting composite features, merchant activity-load fluctuation forecasting composite features, renewable energy-load gap forecasting composite features, and weather-air conditioning load forecasting composite features.

[0030] As a preferred embodiment, step S4 includes the following steps:

[0031] S41: Align historical charging and discharging data of energy storage cabinets, macro power grid data, and battery health trend data, and classify thresholds for electricity demand level, renewable energy generation status, and battery health risk level;

[0032] S42: Label the electricity demand level using historical charging and discharging data of the energy storage cabinet;

[0033] S43: Label the status of renewable energy generation using macro-grid data;

[0034] S44: Mark the battery health risk level by using historical charging and discharging data of the energy storage cabinet and battery health change trend data;

[0035] S45: Tag verification and integration.

[0036] As a preferred option, it also includes:

[0037] S12: During the charging and discharging process, the sensor continuously collects the actual operating data of the energy storage cabinet and the actual operating data of the macro power grid. The data processing and deviation analysis module automatically compares the actual operating data of the energy storage cabinet and the actual operating data of the macro power grid with the output data of the comprehensive prediction dataset to calculate the core deviation.

[0038] S13: Set the deviation threshold. The data processing and deviation analysis module automatically determines whether the core deviation in S12 exceeds the set deviation threshold; otherwise, it maintains the current strategy and parameters and continues to monitor the data. If the deviation is found, it jumps to S14.

[0039] S14: Immediately trigger emergency response; the regional energy management system automatically calls the preset emergency strategies in the strategy library.

[0040] S15: The regional energy management system issues a re-forecasting instruction to the electricity demand forecasting model, the renewable energy output forecasting model, and the battery health management model. The electricity demand forecasting model, the renewable energy output forecasting model, and the battery health management model generate updated output results based on the latest real-time data, and then the new comprehensive forecast dataset is obtained by the fusion of the multilayer perceptron model.

[0041] S16: Input the updated comprehensive prediction results from S15 into the fusion decision module. The fusion decision module combines the core deviation type and deviation magnitude calculated in S12, adjusts the core parameters according to the weight benchmark and adjustment constraints preset in S8, and recalculates to generate optimization strategies and refined parameters that are suitable for the current actual scenario.

[0042] S17: The battery management system sends the optimized strategy and parameters generated in S16 to the battery management systems and energy storage converters of all relevant energy storage cabinets through a standardized interface, replacing the emergency response in S14; the sensors continuously collect the actual operating data after the strategy is executed and transmit it back to the battery management system in real time. The battery management system continues to compare the actual operating data with the updated comprehensive prediction results in S15, calculate the deviation, and if the deviation still exceeds the standard, repeat steps S14-S16; if the deviation does not exceed the standard, maintain the current strategy and continue monitoring.

[0043] This application has the following advantages:

[0044] 1. By integrating spatiotemporal big data and digital twin data, and after multi-step data processing, the prediction model can capture the linkage between macro trends and micro conditions, significantly reducing prediction errors, providing more comprehensive data fusion, and achieving higher prediction accuracy; 2. By constructing electricity demand prediction models, renewable energy output prediction models, and battery health management models, and achieving fusion output of prediction results through a multi-layer perceptron model, it can accurately predict load fluctuations, renewable energy output changes, and battery health risks within the next 24 hours; 3. Compared with traditional fixed-period charging and discharging strategies, this method can predict based on changes in renewable energy output, 4. Improves battery operation safety and extends equipment lifespan; 5. Compared with traditional strategies, the power supply guarantee rate during peak hours is increased to over 99.5%, and the power load deviation is controlled within ±15%, fully meeting the power supply reliability requirements of scenarios such as commercial complexes and industrial parks; 6. The renewable energy output prediction model can accurately predict the peak periods of renewable energy, improving the stability of energy utilization; 7. Accurate prediction of battery health can reduce operation and maintenance costs; 8. Strong engineering adaptability, compatible with energy storage systems in multiple scenarios. Detailed Implementation Example 1

[0045] This application provides a charging and discharging method for an energy storage cabinet, specifically a predictive charging and discharging scheme that integrates spatiotemporal big data and digital twins, including the following steps:

[0046] S1: Collect spatiotemporal big data and digital twin data, and establish a correlation mapping between spatiotemporal big data and digital twin data; the spatiotemporal big data includes meteorological data, macro power grid data, user behavior data, historical charging and discharging data of energy storage cabinets, and battery health change trend data; the digital twin data includes microgrid data, internal physical parameters of energy storage cabinets, dynamic response data of temperature control system, and geometric data of energy storage cabinets;

[0047] The meteorological data is obtained in real-time from hourly meteorological data from platforms such as the China Meteorological Administration and AccuWeather via API interfaces, including information such as light intensity, wind speed, temperature, and weather conditions, accurate to the latitude and longitude of the energy storage cabinet's location. Macro-grid data is integrated with the power company's system to obtain real-time data such as the total regional grid load, regional voltage, regional load growth rate, grid load curve, peak-valley electricity price periods, and the proportion of renewable energy generation, with a data update frequency of 15 minutes per update. User behavior data includes information collected from intelligent management systems in places such as shopping malls and office buildings, such as pedestrian traffic, merchant business hours, historical electricity demand, and promotional activity arrangements, synchronously recording historical consumption data for analyzing electricity demand characteristics. Historical charge and discharge data of the energy storage cabinet is collected through built-in sensors, including battery state of charge (SOC), monthly SOC, charge and discharge power, battery temperature, charge and discharge duration, and charge and discharge status, with a data collection frequency of 1 second per update to ensure accurate monitoring of the energy storage cabinet's status. Precise monitoring is achieved; battery health trend data includes SOH value reflecting the current health level, real-time capacity decay, health status level, historical battery health status, and trend statistics timestamps and decay-related operating condition labels supporting trend analysis. The battery health trend data is obtained by collecting battery operating parameters in real time through the Battery Management System (BMS) and combining them with background algorithms; the microgrid data includes the voltage at the energy storage cabinet access point, the input current of the energy storage converter (PCS), and the rate of change of the access point current; the internal physical parameters of the energy storage cabinet include the battery cell voltage, battery temperature, cell internal resistance, and other core physical quantitative indicators; the dynamic response data of the temperature control system includes quantitative parameters: response delay, heat dissipation power, coolant flow rate, and fan speed; status parameters: response indicators are normal, response indicators are delayed, etc.; the geometric data of the energy storage cabinet includes physical structure quantitative data such as cabinet dimensions and battery pack spacing; the above digital twin data relies on high-precision 3D modeling and direct equipment acquisition.

[0048] S2: Normalize spatiotemporal big data and digital twin data to the [0, 1] interval; the units and numerical ranges of different data vary greatly, and direct calculation will result in excessive weighting of large numerical fields and neglect of the weighting of small numerical fields. After normalization to the [0, 1] interval, all data are unified as dimensionless relative values, realizing direct comparison and calculation of macroscopic scene data of spatiotemporal big data and microscopic device data of digital twin data;

[0049] S3: Extract derived features from spatiotemporal big data and extract the gradient of battery parameter changes from digital twin data; after fusing the two types of data, mine composite features related to charging and discharging efficiency, composite features related to battery health, and composite features related to load prediction; in this step, the macro-scene data of spatiotemporal big data and the device's own data of digital twin data are combined to respectively address the three objectives of optimizing charging and discharging efficiency, protecting battery health, and accurately predicting electricity load.

[0050] Spatiotemporal big data extraction of derived features is based on raw data such as meteorological data, macro power grid data, user behavior data, historical charging and discharging data of energy storage cabinets, and battery health change trends. These features are extracted through statistical or time-series analysis. Derived features of meteorological data include 24-hour average temperature, temperature change rate, and rainfall intensity level. The 24-hour average temperature is calculated by taking the arithmetic mean of the raw temperature data within 24 hours. The temperature change rate is calculated as (current temperature - temperature N hours ago) / N hours. The rainfall intensity level is based on the cumulative rainfall in 24 hours. For example, if the cumulative rainfall is <10mm, it is classified as "none". If the cumulative rainfall is between 10mm and 25mm, it is classified as "small".

[0051] Derivative characteristics of macro-grid data include regional load peak-valley difference, renewable energy output fluctuation coefficient, voltage stability index, and load growth rate. Regional load peak-valley difference = maximum value of regional total load on the day - minimum value. Renewable energy output fluctuation coefficient = (maximum renewable energy output within N hours - minimum renewable energy output within N hours) / average output during the period. Voltage stability index is the standard deviation of regional voltage within 15 minutes. The smaller the standard deviation, the more stable the system. ,in, This represents the baseline load value, which is the total load of the power grid area during a certain reference period in the past; This represents the current load value, also known as the target load value, which is the total load of the power grid area during the current monitoring period or the future forecast period; if > A positive result indicates a positive growth rate, representing year-on-year / month-on-month growth in load; if < A negative result indicates a negative growth rate, representing a year-on-year / month-on-month decrease in load.

[0052] Derivative features of user behavior data include: peak electricity consumption duration, average daily electricity load growth rate, peak-valley electricity consumption ratio, electricity load fluctuation characteristics, and merchant promotion activity indicators. Peak duration is obtained by summing the continuous periods when the daily load is ≥80% of the rated load. Average daily electricity load growth rate = (average load of the day - average load of the previous day) / average load of the previous day × 100%. Peak-valley electricity consumption ratio = peak period electricity consumption / off-peak period electricity consumption. Electricity load fluctuation characteristics = load change amplitude / average load.

[0053] The derived characteristics of historical charging and discharging data of energy storage cabinets include: charge and discharge cycle efficiency, monthly average depth of discharge (DOD), and idle time percentage. Charge and discharge cycle efficiency = (total capacity of a single discharge / corresponding total charging capacity) × 100%. Monthly average depth of discharge is the average of the depth of discharge (1 - remaining SOC) for each discharge within the month. Idle time percentage is the duration within the month when SOC is between 30% and 70% and there is no charging or discharging / total duration.

[0054] Derivative features of battery health trend data include: monthly average SOH (State of Health) degradation rate, cumulative SOH degradation, remaining cycle life percentage, degradation acceleration factor, weekly SOH change rate, health status change warning value, predicted remaining SOH life, and correlation coefficient between charge / discharge cycle count and SOH degradation. Monthly average SOH degradation rate = (SOH value at the beginning of the month - SOH value at the end of the month) / SOH value at the beginning of the month × 100% ÷ number of days in the month × 30; Cumulative SOH degradation = initial SOH value - current SOH value, where the initial SOH value is the factory or initial calibration value, typically 100%; Remaining cycle life percentage = (rated cycle life - number of completed cycles) / rated cycle life × 100%; Degradation acceleration factor = actual monthly average degradation rate ÷ baseline degradation rate; Weekly SOH change rate = (SOH value on the last day of the week - SOH value on the first day of the week) / SOH value on the first day of the week × 100%.

[0055] The calculation method for the health status change warning value is as follows: set a short-term change threshold for SOH, such as a single-day decrease ≥0.1%. When the SOH decrease exceeds the threshold for two consecutive days, a warning value is generated, such as 1 indicating a warning and 0 indicating normal.

[0056] The calculation method for the predicted remaining SOH lifetime is as follows: Based on the SOH decay data of the past 6 months, the time when SOH drops to 80% is calculated using linear fitting y = kx + b, where x is time in months, y is the SOH value, k represents the monthly average decay rate, and b is the theoretical SOH value at the starting point x=0 of the fitted trend line. Specifically, the fitted decay trend line is calculated first, then the remaining decay is calculated as: remaining decay = current SOH - 80%, and finally the remaining SOH lifetime is calculated as: remaining SOH lifetime = remaining decay ÷ monthly average decay rate. The correlation coefficient between the number of charge / discharge cycles and SOH decay is analyzed using the Pearson correlation coefficient to quantify the linear correlation strength between the two. The formula is: r = Cov(X, Y) / (σ X × σ Y ), r∈[-1,1], the closer r is to 1, the stronger the correlation, Cov(X,Y) represents the covariance, σ X、 σ Y It represents the standard deviation.

[0057] The process of extracting battery parameter change gradients from digital twin data, using time-series differential calculations, primarily reflects the trend of parameter changes over time. These gradients include capacity gradients, voltage / current gradients, temperature gradients, and health status gradients. Capacity gradients include SOC change gradients and available capacity decay gradients. The SOC change gradient (% / min) = (current SOC - SOC N minutes ago) / Nmin. For example, if N is 15, a positive SOC change gradient indicates charging, and a negative SOC change gradient indicates discharging. The available capacity decay gradient (kWh / month) = (available capacity at the beginning of the month - available capacity at the end of the month) / (number of days in the month × 30).

[0058] Voltage / current gradients include individual cell voltage change gradients, connection point current change gradients, and charge / discharge power adjustment gradients; voltage change gradient (V / min) = (current cell voltage - voltage N minutes ago) / Nmin, if N is 5, the voltage change gradient reflects voltage stability; current change gradient (A / min) = (current connection point current - current N minutes ago) / Nmin, if N is 10, the current change gradient reflects the charge / discharge power adjustment rate. ,in, It represents the maximum charging and discharging power within a certain period of time, with charging being positive and discharging being negative, and the unit is kW / MW; This indicates the minimum charging and discharging power within a certain time period. T represents the duration of the time period, and the unit is s / min / h, such as a 15-minute time period.

[0059] Temperature gradients include temperature rise gradient and temperature control response gradient. Temperature rise gradient (°C / min) = (current battery pack temperature - temperature N minutes ago) / Nmin. If N is 1, the temperature rise gradient is used to warn of thermal runaway risk. Temperature control response gradient (°C / s) = (temperature after temperature control starts - temperature before start) / response delay time. The temperature control response gradient is used to evaluate the efficiency of the temperature control system.

[0060] Health status gradients include SOH decay gradient, internal resistance growth gradient (mΩ / cycle), and polarization impedance change gradient. SOH decay gradient (% / month) = (current SOH - SOH N months ago) / N. If N is 3, the SOH decay gradient is used to predict long-term health trends. Internal resistance growth gradient (mΩ / cycle) = (current cell internal resistance - internal resistance N cycles ago) / N. If N is 100, the internal resistance growth gradient reflects the cell aging rate. , This represents the polarization impedance (mΩ) after the nth cycle. The polarization impedance (mΩ) represents the initial state, i.e., the polarization impedance after 0 cycles, and n represents the cumulative number of charge-discharge cycles. When the polarization impedance change gradient is positive, the polarization impedance increases with the number of cycles, which is a typical manifestation of battery aging. The larger the value, the faster the polarization decays.

[0061] The two types of data are fused. First, spatiotemporal alignment is performed. Using time period ID and device ID as association keys, the spatiotemporal big data derived features of the same energy storage cabinet at the same time are bound to the battery parameter gradient of the digital twin data. Second, the fused data is normalized to [0,1] to eliminate numerical scale differences.

[0062] For example, the daily average electricity load growth rate is mapped from 0-50% to 0-1, and the SOH attenuation gradient is mapped from 0-0.01 / day to 0-1; finally, through correlation analysis such as Pearson coefficient, highly repetitive features are removed to reduce the amount of calculation.

[0063] By combining the impact of macroscopic scenarios with the microscopic state of equipment, we can extract relevant features that directly serve the target. Specific examples are as follows:

[0064] The goal of this study is to optimize charge and discharge efficiency by exploring the correlation between derived features of spatiotemporal big data and the gradient of battery parameter changes, in order to find the optimal charge and discharge conditions. Specifically, this includes:

[0065] Temperature-voltage-charging efficiency composite characteristic: This characteristic is formed by the fusion of 24-hour average temperature, voltage stability index, internal resistance growth gradient, and temperature rise gradient. The calculation formula is: Temperature-voltage-charging efficiency composite characteristic = (1 - normalized value of temperature rise gradient) × voltage stability index × (1 - normalized value of internal resistance growth gradient). It reflects the comprehensive impact of ambient temperature, macro grid voltage, and battery loss on charging efficiency. When the value is high, such as ≥0.7, high power charging can be maintained. When the value is low, power needs to be reduced to reduce losses.

[0066] Renewable energy-charge-discharge adaptation efficiency composite characteristic: This characteristic is composed of renewable energy output fluctuation coefficient, charge-discharge power adjustment gradient, and SOH monthly average decay rate. The calculation formula is: Renewable energy-charge-discharge adaptation efficiency composite characteristic = (1 - normalized value of charge-discharge power adjustment gradient) × (1 - renewable energy output fluctuation coefficient) × (1 - normalized value of SOH monthly average decay rate). It is used to quantify the efficiency of renewable energy consumption. When the value is high, such as ≥0.65, the energy storage cabinet can quickly track the fluctuation of renewable energy electricity with low loss. When the value is low, the adjustment speed needs to be reduced to avoid efficiency decline.

[0067] Peak-valley load-discharge efficiency composite characteristic: This characteristic is formed by the fusion of regional load peak-valley difference, peak-valley electricity consumption ratio, discharge cycle efficiency, and SOC change gradient. The calculation formula is: Peak-valley load-discharge efficiency composite characteristic = peak-valley electricity consumption ratio × discharge cycle efficiency × (normalized value of SOC change gradient + 0.5). It is used to match the discharge efficiency of the grid peak-valley load. When the characteristic value is high during the peak period (large peak-valley difference), priority discharge can take into account both efficiency and grid energy replenishment needs.

[0068] Meteorological-Temperature-PCS Efficiency Composite Characteristic: This characteristic is formed by the fusion of temperature change rate, temperature response gradient, and charge-discharge cycle efficiency. The formula is: Meteorological-Temperature-PCS Efficiency Synergistic Characteristic = (1 - Normalized value of temperature change rate) × Normalized value of temperature response gradient × Charge-discharge cycle efficiency. The higher the characteristic value, the better the synergy between the temperature control system and the PCS efficiency, and the energy storage cabinet can maintain rated power charging and discharging.

[0069] Idle Time-Charging Efficiency Composite Feature: This feature is composed of the idle time ratio, available capacity decay gradient, and voltage stability index. The calculation formula is: Idle Time-Charging Efficiency Composite Feature = (1 - Normalized value of available capacity decay gradient) × Voltage stability index × (1 - Normalized value of idle time ratio). It is used to optimize the charging strategy after idle time. When the idle time is short, the capacity decay is small, and the voltage is stable, the feature value is high, and the efficiency can be quickly restored by full power charging.

[0070] Battery health-related composite characteristics aim to protect battery health and prevent damage in advance; higher characteristic values ​​indicate greater health risks, requiring timely intervention, specifically including:

[0071] The voltage fluctuation-temperature-internal resistance damage composite characteristic is composed of voltage stability index, temperature rise gradient, and internal resistance growth gradient. The calculation formula is: Voltage fluctuation-temperature-internal resistance damage composite characteristic = voltage fluctuation coefficient × normalized value of temperature rise gradient × normalized value of internal resistance growth gradient. This characteristic is used to quantify the triple damage of voltage instability, high temperature, and internal resistance growth. When the characteristic value is ≥0.6, power should be reduced immediately and temperature control should be strengthened to avoid irreversible aging.

[0072] High-frequency cycle-SOH attenuation early warning composite feature: The feature is formed by the fusion of the daily average electricity load growth rate, the correlation coefficient between the number of charge-discharge cycles and SOH attenuation, and the SOH attenuation gradient. The calculation formula is: High-frequency cycle-SOH attenuation early warning composite feature = daily average electricity load growth rate × correlation coefficient × normalized value of SOH attenuation gradient. It is used to warn of the aging risk of high-frequency charging and discharging. When the user's electricity consumption is growing rapidly and the correlation between cycle and attenuation is strong, the feature value is high, and unnecessary charging and discharging should be reduced.

[0073] Depth of Discharge-Health Mutation Risk Composite Feature: This feature is a fusion of the monthly average depth of discharge, the health status mutation warning value, and the individual cell voltage change gradient. The calculation formula is: Depth of Discharge-Health Mutation Risk Composite Feature = Normalized value of monthly average depth of discharge × (1 - Normalized value of health status mutation warning value) × Normalized value of voltage change gradient. It is used to avoid health mutations caused by deep discharge. When DOD ≥ 80% and the feature value ≥ 0.5, the monthly average depth of discharge needs to be limited, such as reducing it to 60%.

[0074] Rainfall-Temperature-Battery Corrosion Risk Composite Feature: This feature is derived from the rainfall intensity level, 24-hour average temperature, and polarization impedance change gradient. The calculation formula is: Rainfall-Temperature-Battery Corrosion Risk Composite Feature = Normalized value of rainfall intensity level × (Normalized value of 24-hour average temperature + 0.3) × Normalized value of polarization impedance change gradient. It is applicable to outdoor energy storage scenarios. When rainfall is heavy and the temperature is high, the feature value is high, and the battery compartment moisture protection needs to be activated to avoid accelerated corrosion and aging.

[0075] Load forecasting-related composite features are used to accurately predict electricity load and improve forecast accuracy; higher feature values ​​indicate more significant load growth / fluctuations, specifically including:

[0076] Peak electricity consumption-load growth forecast composite feature: The forecast feature is formed by the fusion of peak electricity consumption duration, average daily electricity load growth rate, and SOC change gradient. The calculation formula is: Peak electricity consumption-load growth forecast composite feature = normalized value of peak electricity consumption duration × average daily electricity load growth rate × (normalized value of SOC change gradient + 0.5). It is used to predict peak load growth. When the feature value is ≥0.6, the load will continue to rise, and the energy storage cabinet needs to be charged in advance for backup energy.

[0077] Merchant Activity-Load Fluctuation Prediction Composite Feature: This feature is formed by fusing the electricity load fluctuation feature, the merchant promotion activity identifier, and the current change gradient at the access point. The calculation formula is: Merchant Activity-Load Fluctuation Prediction Composite Feature = Electricity Load Fluctuation Feature × Promotion Activity Identifier (Yes = 1 / No = 0) × Normalized Current Change Gradient Value; It predicts the load fluctuation caused by merchant promotions. When the feature value is high, the load will increase or decrease sharply, and the energy storage cabinet needs to respond and adjust quickly.

[0078] Renewable energy-load gap prediction composite feature: This feature is formed by the fusion of renewable energy output fluctuation coefficient, regional load peak-valley difference, and available capacity decay gradient. The calculation formula is: Renewable energy-load gap prediction composite feature = renewable energy output fluctuation coefficient × normalized value of regional load peak-valley difference × (1 - normalized value of available capacity decay gradient). This feature is used to predict the load gap caused by insufficient renewable energy. When the feature value is ≥0.55, the renewable energy power fluctuation is large, the grid gap is large, and energy storage needs to be discharged in advance to replenish energy.

[0079] Meteorological-Air Conditioning Load Forecast Composite Feature: This forecast feature is formed by the fusion of 24-hour average temperature, temperature change rate, and charging / discharging power adjustment gradient. The calculation formula is: Meteorological-Air Conditioning Load Forecast Composite Feature = (24-hour average temperature normalized value + 0.2) × temperature change rate × charging / discharging power adjustment gradient normalized value. It is used to predict the air conditioning load caused by extreme temperatures. The feature value is high at high / low temperatures, and energy storage needs to store electricity in advance to cope with peak air conditioning electricity consumption.

[0080] S4: Combining historical charging and discharging data of energy storage cabinets, macro power grid data, and battery health trend data, label the relevant composite feature dataset of S3 with three types of labels. The three types of labels include electricity demand level, renewable energy generation status, and battery health risk level. All labels are from historical measured data to avoid logical loops.

[0081] S5: Construct an electricity demand prediction model based on Transformer, the output of which is the future electricity demand; construct a renewable energy output prediction model based on LSTM-CNN fusion, the output of which is the future renewable energy output characteristics; construct a battery health management model based on CNN and physical mechanism model, the output of which is the future battery health status or aging trend.

[0082] S6: The dataset labeled in S4 is divided into a training set and a validation set. The training set is used to train the electricity demand forecasting model, the renewable energy output forecasting model, and the battery health management model, respectively. The validation set is used to optimize the parameters of the above three models to ensure that the prediction accuracy of the single model meets the standards: the electricity demand forecasting error is <5%, the average absolute percentage error of the renewable energy output forecasting model is <8%, and the average absolute percentage error of the battery health management model is <3%.

[0083] S7: The outputs of the trained electricity demand forecasting model, renewable energy output forecasting model, and battery health management model are used as inputs. These are then fused using a multilayer perceptron (MLP) model to output a comprehensive forecast dataset. This dataset includes future electricity demand forecasts, future renewable energy output forecasts, and future battery health status forecasts. The future electricity demand forecasts include time-period electricity load forecasts for the next 1-24 hours, electricity demand level forecasts, peak / valley load forecasts, and electricity load fluctuation coefficient forecasts, such as fluctuation amplitude within 15 minutes and peak duration forecasts. Renewable energy output forecasts include time-period renewable energy output forecasts for the next 1-24 hours, renewable energy generation status forecasts, peak / valley output forecasts, output fluctuation coefficients, output fluctuation amplitude forecasts (e.g., within 15 minutes), and renewable energy generation share forecasts (i.e., the proportion of renewable energy in the total regional load). Future battery health status forecasts include SOH forecasts for the next 1-3 months, SOC trends, SOH changes, weekly / monthly updates, capacity degradation forecasts for the next 1-6 months, internal resistance growth forecasts for the next 1-6 months, battery health risk level forecasts, aging acceleration warnings, and remaining charge / discharge cycle life forecasts.

[0084] S8: Using the comprehensive prediction dataset of S7 as input, train the fusion decision module and output the optimal charging and discharging strategy parameters, such as charging power, discharge depth, cutoff voltage, etc. During training, the state space of the fusion decision module is constructed simultaneously. The state space includes core variables of battery health, core variables of power supply guarantee, core variables of economic benefit and corresponding value constraints. At the same time, the benchmark values ​​of core parameters and weight adjustment constraints are fixed in advance.

[0085] The reward function weighting benchmark is as follows: Battery health protection has a weight of 0.5, corresponding to indicators such as SOC, temperature, and SOH, which is the highest priority to ensure equipment safety; User-side power supply protection has a weight of 0.3, which is the second priority for application load and power supply reliability to meet load demand; Economic benefits have a weight of 0.2, which corresponds to green electricity consumption and peak-valley electricity price arbitrage, which is the basic priority and takes into account operational efficiency.

[0086] Weight adjustment constraints: The magnitude of a single weight adjustment shall be ≤0.3 to avoid strategy imbalance; the sum of the three weights after adjustment shall always be 1, without adding any additional weight dimensions; the physical constraints of the equipment are hard boundaries, such as charging current not exceeding 0.8C, discharge depth not less than 20%, and charging and discharging power not exceeding the rated value of PCS. Parameter adjustments shall not exceed these boundaries.

[0087] The aforementioned battery health variables include SOC, SOH, battery temperature, cell internal resistance, remaining cycle life, battery temperature gradient, and SOC change rate; power supply security variables include real-time electricity load, electricity demand level, peak electricity consumption period identifier, power supply gap, and power supply redundancy; economic benefit variables include peak-valley-flat electricity price periods, photovoltaic output, green electricity consumption, and renewable energy output fluctuation coefficient.

[0088] Taking a typical summer day in a certain province as an example, the electricity price is: peak 1.2 yuan / kWh, flat 0.7 yuan / kWh, and valley 0.3 yuan / kWh. The average daily photovoltaic power generation is 325kWh. The optimal charging and discharging strategy parameters output by the integrated decision-making module are shown in Table 1 below. Using the charging and discharging strategy of this method, the daily peak-valley arbitrage income is increased by 30%-40% compared with the traditional strategy. At the same time, the green electricity consumption rate is increased to more than 90%, reducing the dependence on grid power purchase.

[0089] Table 1:

[0090]

[0091] S9: Based on the output of the fusion decision module, combined with real-time data such as the current real-time SOC and real-time grid load, a specific charging and discharging strategy is generated. The strategy generation must strictly follow the weight benchmark of S8. For example, if the power demand is high in the next 2 hours and the battery health risk is low, then under the premise of ensuring that the SOC is not lower than 20%, a 100kW discharge is executed with a discharge depth of 70%, which does not exceed the 80% safety boundary. If the battery health risk is medium / high, then the discharge depth must be reduced to within 50% to prioritize battery safety.

[0092] S10: Send the charging / discharging strategy execution command to the terminal device;

[0093] S11: The decision result is converted into standardized control commands and sent to the battery management system and energy storage converter of the energy storage cabinet through the communication interface. After receiving the command, the battery management system executes battery protection-related operations, and after receiving the command, the energy storage converter officially starts the charging and discharging process. Example 2

[0094] This embodiment provides a detailed explanation of establishing the association mapping between spatiotemporal big data and digital twin data in S1, specifically including the following steps:

[0095] SA1: Establish a unique interface identifier for spatiotemporal big data and digital twin data to ensure that each piece of data can be accurately matched to the same spatiotemporal scenario. Specifically, unify the time granularity of spatiotemporal big data and digital twin data, such as 15 minutes / piece, standardize the format, such as YYYY-MM-DD HH:MM:SS, and use methods such as linear interpolation to fill in missing timestamps; use hierarchical identification by region ID + energy storage cabinet number + equipment component ID to ensure that the physical location of digital twin data corresponds one-to-one with the regional scenario of spatiotemporal big data, with hierarchical identification, such as region A-energy storage cabinet 001-battery pack 01; use the above timestamp and hierarchical identifier as a unique association key to ensure that each piece of data can be accurately matched to the same spatiotemporal scenario, such as 2025-11-20 12:00:00-region A-energy storage cabinet 001;

[0096] SA2: Preprocesses spatiotemporal big data and digital twin data, addressing data redundancy and format differences to ensure direct correlation and calculation between the two types of data. Specifically, it identifies duplicate fields in macro-grid data such as the total load of regional power grids and micro-grid data such as the load at the energy storage cabinet access point. It prioritizes retaining micro-grid data directly collected by equipment in the digital twin data due to its higher accuracy, using macro-grid data from the spatiotemporal big data as a verification reference and deleting completely duplicated redundant fields. It unifies the formats of both spatiotemporal big data and digital twin data into the Parquet structured format, standardizing field types and physical units. For field type standardization, power is set to float and time to datetime. For physical unit standardization, temperature is set to °C, power to kW, and capacity to kWh. It removes outliers and fills in missing values ​​to ensure data quality, such as removing values ​​exceeding the rated voltage range of equipment in the micro-grid data and supplementing missing meteorological data using a moving average method.

[0097] SA3: Establish explicit association rule tables for spatiotemporal big data and digital twin data to achieve direct mapping; that is, establish a direct mapping relationship between macro-scenes and micro-devices for rapid response to simple scenarios; specifically:

[0098] Meteorological data such as temperature and weather conditions are correlated with battery temperature, heat dissipation power, and fan speed. The correlation logic is that when the ambient temperature rises, the battery temperature rises, which in turn requires increasing the heat dissipation power and / or increasing the fan speed. When it is raining, the fan speed is limited to prevent short circuits. This explicit correlation rule is used to support the calculation of the composite features of meteorology, temperature, and PCS efficiency, and to optimize the temperature control strategy.

[0099] Macro-level grid data such as regional voltage, the proportion of renewable energy generation, and peak-valley electricity price periods (0 represents off-peak, 1 represents peak) are correlated with access point voltage, PCS input current, and cell internal resistance. The correlation logic is as follows: When regional voltage fluctuates, the access point voltage fluctuates synchronously and sends adjustment commands to the power conversion system (PCS). To ensure the stability of the energy storage system and avoid overvoltage and voltage damage to equipment, the battery management system (BMS) monitors the access point voltage changes in real time and dynamically adjusts the magnitude and phase of the PCS input current to offset the impact of regional voltage fluctuations, ensuring that the charging and discharging voltage of the battery pack inside the energy storage cabinet is within a safe threshold range. As the proportion of renewable energy generation increases, the intermittency and randomness of its output will directly lead to instability in the power output on the grid side. This instability will be transmitted to the energy storage cabinet access point through the grid, manifesting as frequent fluctuations in the power at the access point. Rate fluctuations can cause rapid changes in the charging and discharging current of energy storage batteries. During the current change, the Joule heat generated by the battery's internal resistance will increase synchronously with the fluctuation amplitude, eventually leading to increased heating of the battery's internal resistance. Long-term accumulation may accelerate battery aging. When the macro power grid is in peak electricity consumption period, the total regional load demand increases significantly, and there is a power replenishment gap on the grid side. At this time, the energy storage system needs to respond to the grid load demand. The BMS triggers a discharge command, and the PCS will actively increase the output current according to the size of the grid load gap and the current SOC state of the energy storage cabinet, so as to efficiently feed the energy stored in the battery into the grid, achieve precise energy replenishment for peak loads, and at the same time ensure the stability of grid voltage and frequency. This explicit correlation rule is used to support the composite characteristics of voltage fluctuation-temperature-internal resistance damage and protect battery health.

[0100] The data on user behavior, such as foot traffic, is correlated with the rate of change of current at the access point. The correlation logic is as follows: When foot traffic in places such as shopping malls and office buildings increases significantly, the operational intensity of merchants will increase accordingly. The demand for starting or increasing the power of electrical loads such as lighting, air conditioning, POS equipment, and display devices will increase in a concentrated manner, leading to a rapid increase in the overall electrical load in the area. Since the access point of the energy storage cabinet is directly connected to the power distribution system of the place, the surge in electrical load will be transmitted to the access point through the power distribution network, which will manifest as an instantaneous change in the current at the access point. The greater the foot traffic, the higher the frequency of starting and stopping of electrical equipment or the higher the power adjustment range, the faster the rate at which the current at the access point rises from the baseline value to the new steady-state value, which will ultimately be reflected in a significant increase in the rate of change of current at the access point. The increase in the rate of change of current directly reflects the dynamic impact of the increase in foot traffic on the local power distribution system. This explicit correlation rule is used to support the composite feature of merchant activity-load fluctuation prediction and to predict the fluctuation of electrical load.

[0101] The current SOH value, real-time capacity decay, and battery health trend data such as decay-related operating condition labels (e.g., high-frequency cycling, high temperature) are correlated with response latency. The correlation logic is as follows: when the decay-related operating condition label is marked as high temperature, it indicates that the energy storage cabinet is currently operating in a high-temperature environment, such as outdoor exposure in summer or high-load charging and discharging leading to battery temperature rise. High-temperature environments will significantly affect the heat dissipation efficiency of the temperature control system. On the one hand, high temperature will reduce the temperature difference between the coolant and the battery body, weakening the heat exchange driving force. On the other hand, high temperature may cause the working efficiency of core components of the temperature control system, such as fans and pumps, to decrease, or even trigger overload protection to limit operating power. These two factors together cause the time from detecting that the battery temperature exceeds the standard to starting effective heat dissipation and bringing the temperature back to a safe range to be extended, that is, the response latency of the temperature control system increases significantly. The increase in response latency will further aggravate the battery's high-temperature exposure time, which may accelerate battery capacity decay and internal resistance growth. This explicit correlation rule is used to support the high-frequency cycling-SOH decay warning composite feature and predict the battery aging trend.

[0102] The regional load growth rate is correlated with the rate of change of the connection point current. The correlation logic is as follows: when the regional load growth rate increases, it means that the overall electricity demand in the region is rising rapidly. The rapid growth of the regional load will be directly transmitted to the energy storage cabinet connection point through the distribution network. In order to match the growth rate of regional electricity demand, the energy storage system needs to dynamically adjust its charging and discharging state, causing the connection point current to rapidly transition from the current steady-state value to the new demand value. The higher the regional load growth rate, the faster the electricity demand grows, and the greater the adjustment range and rate of the connection point current, which is ultimately reflected in the increase of the rate of change of the connection point current. The increase in the rate of change of current directly reflects the dynamic impact of regional load growth on the electrical state of the energy storage cabinet connection point. This explicit correlation rule is used to support the composite characteristics of peak-valley load and discharge efficiency, and to match the grid load demand.

[0103] By embedding explicit rule tables into the data processing system, mapping can be automatically triggered through association keys to achieve real-time matching.

[0104] SA4: Mining implicit correlations in spatiotemporal big data and digital twin data to construct implicit correlation models for indirect mapping; using algorithms to uncover field correlations without direct business logic to achieve deep linkage, specifically: selecting features requiring indirect correlation, such as the variation gradient of renewable energy output fluctuation coefficients and battery parameters, and the variation rate of regional load growth rate in macro-grid data and energy storage cabinet access point current in micro-grid data; algorithmically mining correlation patterns, quantifying field correlation strength through Pearson correlation coefficient analysis, and screening strong correlation features with correlation coefficients ≥ 0.7; training the implicit correlation model using decision tree / random forest algorithms to uncover implicit patterns, such as an increase of 0.08% / h in the variation gradient of battery parameters for every 15% increase in the regional renewable energy output fluctuation coefficient; solidifying the implicit correlation model by encapsulating the discovered implicit correlation patterns into a model as a supplement to explicit rules, improving the comprehensiveness of the mapping.

[0105] SA5: Verify the mapping to check its accuracy, adapt to changes in data and scenarios, and form a closed loop. Specifically, this involves sampling verification: randomly selecting 10% of the associated data to verify the mapping error and the accuracy of the associated feature matching. The mapping error must be ≤5%, and the accuracy of the associated feature matching must be ≥95%. Otherwise, the rule parameters should be adjusted or the associated model should be optimized. For example, if the power mapping error is 6%, the rule parameters need to be adjusted or the associated model optimized. If the temperature threshold mapping error is 4%, the rule parameters do not need to be adjusted or the associated model optimized.

[0106] As a preferred option, SA6 is also included: dynamically updating the explicit rule table and the implicit association model; specifically, binding the re-prediction instruction of the battery management system, when displaying changes in equipment status such as battery aging, increased line loss at the access point, or changes in macro-grid data display scenarios such as regional grid expansion, the explicit association rule table and the implicit association model are updated synchronously to ensure mapping adaptability. Example 3

[0107] This embodiment provides a detailed description of S4, specifically:

[0108] S41: Align historical charging and discharging data of energy storage cabinets, macro power grid data, and battery health trend data, and classify thresholds for electricity demand level, renewable energy generation status, and battery health risk level; use the timestamp (e.g., 5-minute / 15-minute slices) and device ID of the relevant composite feature dataset in S3 as association keys to bind historical charging and discharging data of energy storage cabinets, macro power grid data, and battery health trend data with relevant composite feature samples one by one, ensuring that one relevant composite feature sample corresponds to one set of labeled data;

[0109] The labeling thresholds are determined based on historical data statistics, as follows:

[0110] 1. Electricity Demand Level Threshold Classification: Electricity demand levels are classified into low, medium, high, and very high. Low level is defined as when the period load is <30% of the rated load; medium level is defined as when 30% of the rated load ≤ period load < 60% of the rated load; high level is defined as when 60% of the rated load ≤ period load < 85% of the rated load; and very high level is defined as when the period load ≥ 85%. Both period load and rated load refer to the user-side load. 2. Renewable Energy Generation Status Threshold Classification: Renewable energy generation status is classified into low generation, medium generation, high generation, and outage. Generation capacity <20% is defined as... The system is categorized as follows: low power generation when the installed capacity is less than 50% of the installed capacity, medium power generation when the installed capacity is less than 20% of the installed capacity, high power generation when the installed capacity is less than 80% of the installed capacity, and shutdown when the power is approximately 0. The system matches the measured output value of each sample with a timestamp and labels it with the corresponding status tag. The installed capacity refers to the nameplate installed capacity of the local renewable energy power generation unit connected to the energy storage cabinet, and the power generation refers to the actual power generation of the local renewable energy power generation unit connected to the energy storage cabinet. Local renewable energy power generation units include photovoltaic power plants, small wind farms, etc. 3. Battery Health Risk Level Threshold Classification: Battery health risk levels are classified as low risk, medium risk, and high risk. Low risk is defined as follows: SOH ≥ 90%, capacity decay ≤ 5%, monthly SOC below 20% 0 times, battery temperature ∈ [15℃, 35℃], monthly temperature ≥ 40℃ 0 times, and no severe operating condition label. Medium risk is defined as: 85% ≤ SOH < 90%, or 5% < capacity decay ≤ 10%, or monthly SOC below 20% ≤ 3 times, or battery temperature ≥ 40℃ ≤ 5 times, or one severe operating condition label. High risk is defined as: SOH < 85%, or capacity decay > 10%, or monthly SOC below 20% > 3 times, or battery temperature ≥ 40℃ > 5 times, or charging / discharging power exceeding rated values, or two or more severe operating condition labels. Among them, SOH is the current SOH value in the battery health trend data, capacity decay is the real-time capacity decay in the battery health trend data, the number of times the monthly SOC is below 20% is the monthly statistical value of the SOC time series in the historical charge and discharge data of the energy storage cabinet, the real-time battery temperature is the real-time collected operating data, the number of times the monthly temperature is ≥40℃ is the monthly statistical value of the battery temperature time series in the historical charge and discharge data of the energy storage cabinet, and the severe operating condition label is a label generated based on the abnormal operating condition records in the historical charge and discharge data of the energy storage cabinet.

[0111] S42: Label the electricity demand level using historical charging and discharging data of the energy storage cabinet;

[0112] Extract the battery state of charge (rated load) from the historical charging and discharging data of the energy storage cabinet, and label the power demand level according to the above-mentioned power demand level threshold classification standard; preferably, extract the daily traffic flow data, promotional activity arrangements, and merchant business hours from user behavior data, and modify the scenario according to the actual situation. If there is a promotional activity (identifier = 1) and the traffic flow is ≥80% of the historical average, the original power demand level is upgraded, such as upgrading from medium to high; if the merchant business hours are shortened by more than 2 hours, the original power demand level is downgraded, such as downgrading from medium to low; associate the modified level with the corresponding composite feature sample.

[0113] S43: Label the status of renewable energy generation through macro-grid data; combine macro-grid data to extract the installed capacity from the basic parameters of local renewable energy generation units, match it with the measured power generation, label the power generation status according to the threshold and associate it with the corresponding composite feature sample.

[0114] S44: Mark the battery health risk level using historical charge and discharge data of the energy storage cabinet and battery health trend data; extract SOC, number of times monthly SOC is below 20%, charge and discharge power, real-time battery temperature, number of times monthly temperature is ≥40℃, and harsh operating condition labels from the historical charge and discharge data of the energy storage cabinet; extract SOH value and real-time capacity decay from the battery health trend data, and mark the battery health risk level according to the above battery health risk level threshold classification standard; associate the marked battery health risk level with the corresponding composite feature sample.

[0115] S45: Tag verification and integration;

[0116] Check the logical matching degree between the labels and the S3 composite features. For example, the high demand label should correspond to the high value of the peak electricity consumption-load growth prediction composite feature, with a matching rate of ≥95%. Remove samples with missing labeled data or logical contradictions. For example, high risk, but SOH decay = 0. Generate a complete labeled dataset of composite features and three types of labels for S5 model training. For example, for the temperature-voltage-charging efficiency composite feature, the electricity demand level label is medium, the renewable energy generation status label is low, and the battery health risk level is low. Example 4

[0117] This embodiment provides a detailed description of S5, specifically:

[0118] I. Constructing an electricity demand forecasting model includes the following steps:

[0119] SA51: Preprocesses the input data, which includes raw data, derived features, battery parameter variation gradients, load forecast-related composite features, and tag data. The raw data includes 13 dimensions: light intensity, wind speed, temperature, weather conditions, total grid load, peak / valley electricity price periods, pedestrian traffic data, promotional activity schedules, SOC, charging / discharging power, battery temperature, charging / discharging duration, and charging / discharging status. Derived features include 24-hour average temperature, temperature change rate, rainfall intensity, regional load peak / valley difference, voltage stability index, peak electricity consumption duration, daily average electricity load growth rate, and peak / valley electricity consumption ratio. Electricity load fluctuation characteristics, merchant promotion activity identification, charge / discharge cycle efficiency, DOD, and idle time ratio; a total of 13 dimensions; battery parameter change gradients, including individual cell voltage change gradient, connection point current change gradient, charge / discharge power adjustment gradient, temperature rise gradient, and temperature control response gradient; a total of 5 dimensions; load forecast related composite features, including peak electricity consumption-load growth forecast composite features, merchant activity-load fluctuation forecast composite features, renewable energy-load gap forecast composite features, and meteorological-air conditioning load forecast composite features; tagged data includes electricity demand level and corresponding actual electricity load value for the time period;

[0120] All the above data are aligned in 5-minute time slices, and the mean value within each slice is calculated; the data acquisition frequency of S1 is adapted; the original data, derived features, composite features, and label data are concatenated into a 35-dimensional feature vector; it is ensured that all input data have been uniformly mapped to the [0, 1] interval, and the actual load value in the label is restored to kW unit through normalization and reverse mapping.

[0121] SA52: Design of the Transformer architecture;

[0122] The input layer receives the aforementioned 35-dimensional feature vectors, corresponding to the full-dimensional data of a single time slice, ensuring that the features cover the core factors affecting electricity demand; the temporal coding layer constructs a 7-day × 288-time-step input sequence from the 5-minute slice, capturing the historical 7-day load pattern; 288 = 24 hours × 60 minutes / 5 minutes; the embedding layer transforms the 35-dimensional features to a 64-dimensional space, enhancing feature expression capabilities and adapting to the temporal modeling requirements of the Transformer; the encoder module is stacked, with 6 layers in this embodiment, each layer having 64-dimensional input / output, and each layer performs the following sub-steps: a) multi-head attention mechanism, setting the number of heads to 3, focusing on key features and temporal correlations, such as the strong dependence of promotional arrangements on current change gradients; b) layer normalization, setting ε = 1e-6, stabilizing the training process and avoiding gradient vanishing; the feedforward network uses a 64-dimensional → 256-dimensional → 64-dimensional dimensional mapping, with GELU as the activation function to enhance nonlinear fitting capabilities; secondary layer normalization, setting ε = 1e- 6. Further optimize feature distribution; composite feature enhancement layer, assign initial weights to the four types of load forecast related composite features separately, with an initial weight of 0.4, and weighted fusion to the encoder output to highlight the role of core prediction features; global average pooling layer, compresses the temporal dimension and retains core feature information; fully connected layer, adopts 64→128→64→288 dimension mapping, output dimension 288 corresponding to the prediction points every 5 minutes of the next 24 hours; output layer, uses the Linear activation function to output 288 normalized load values, and then restores them to the actual electricity demand value in kW unit through the inverse mapping formula, finally outputting the electricity demand value (kW) every 5 minutes of the next 24 hours and the prediction confidence interval.

[0123] SA53: Training and optimization of electricity demand forecasting model; Model training: Data is divided into time series 7:2:1, with the training set being data from the past 7 months, the validation set being data from the past 2 months, and the test set being data from the past 1 month, without disrupting the time series, covering weekdays, weekends, all seasons, promotional and non-promotional scenarios;

[0124] The promotional activity logo is perturbed by ±1 day, such as adjusting the original August 1st promotion to July 31st and August 2nd; the temperature data is perturbed by ±5% to simulate random fluctuations in weather and user behavior; sensor fault data, such as sudden changes in SOC or constant temperature of 0℃, are removed, and missing values ​​are filled with the historical average of the same time period. For example, if a certain 5-minute light data is missing, it is filled with the average of the same period of the previous 3 days.

[0125] Training parameter configuration: Loss function: Mean Absolute Error (MAE) is used to adapt to load prediction and reduce the weight of peak load error; Optimizer: AdamW is used (learning rate = 0.001), weight decay = 1e-4 to prevent overfitting, such as avoiding excessive reliance on a single feature of promotional activities; Batch size: set to 32 to balance training efficiency and memory usage; ≥8GB of GPU memory is sufficient; Number of training epochs: maximum 50 epochs; early stopping mechanism is triggered if the MAE on the validation set does not decrease for 5 consecutive epochs to avoid overfitting; Learning rate scheduling: cosine annealing strategy is used, T... max =10, reduce the learning rate in the later stages of training to achieve accurate convergence.

[0126] Model Optimization: Dynamic Weight Adjustment: After each training round, the prediction error contribution of each composite feature is calculated. When the error is small, the corresponding feature weight is increased. For example, when the error of the composite feature of merchant activity-load fluctuation prediction is small, the weight is increased from 0.4 to 0.6; Temporal Attention Enhancement: Temporal position encoding is added to the encoder module to strengthen the focus on features during peak electricity consumption periods (18:00-22:00); Feature Selection: Highly redundant features are eliminated by using the Pearson coefficient to reduce the amount of computation; For example, if the correlation between the regional total load and the peak-valley difference is >0.9, the regional total load is retained.

[0127] SA54: Model Validation of the Electricity Demand Forecasting Model; The model validation metrics include: Test Set MAE (Target Value ≤ 4.5%), with an actual value of 4.2% in this embodiment, representing the average error across all scenarios, which is lower than the target value, indicating overall model accuracy; Test Set RMSE (Target Value ≤ 7.5%), with an actual value of 6.8% in this embodiment, showing low error dispersion and strong prediction stability; Peak Load Error (Target Value ≤ 9%), with an actual value of 7.8% in this embodiment, representing the average error during the evening peak hours of 18:00-22:00, meeting the requirements for energy storage dispatch; Promotional Scenario Error (Target Value ≤ 11%), with an actual value of 9.5% in this embodiment, indicating that even with large load fluctuations during promotional periods, the error remains lower than the target value, demonstrating strong model adaptability.

[0128] SA55: The electricity demand forecasting model outputs the electricity demand value every 5 minutes for the next 24 hours and the forecast confidence interval; for example, 2024-08-02 18:30-18:35: 178kW, the forecast confidence interval is 173kW-183kW, which is directly compatible with the energy storage dispatching system.

[0129] II. Constructing a renewable energy output prediction model includes the following steps:

[0130] SB51: Preprocess the input data, which includes raw data, derived features, battery parameter variation gradients, load forecasting related composite features, and tag data. The raw data includes 6 dimensions: irradiance, wind speed, temperature, weather conditions, renewable energy generation ratio, and local renewable energy generation unit installed capacity. The derived features include 24-hour average temperature, temperature change rate, rainfall intensity, and renewable energy output fluctuation coefficient, totaling 4 dimensions. The battery parameter variation gradients include the connection point current variation gradient, charge / discharge power adjustment gradient, and temperature rise gradient, totaling 3 dimensions. The load forecasting related composite features include a renewable energy-charge / discharge adaptation efficiency composite feature, totaling 1 dimension. The tag data includes the actual renewable energy output value and generation status tag for the corresponding time period, with generation status tags including low generation, medium generation, high generation, and outage status tags.

[0131] SB52: Design of an LSTM-CNN fusion model architecture;

[0132] The input layer receives the 14-dimensional feature vectors mentioned above; the time-series sequence is constructed by dividing 15-minute slices into a 4-day × 96-step input sequence to capture short-term weather dependencies; 96 = 24 hours × 60 minutes / 15 minutes; the CNN feature extraction layer has 3 convolutional kernels with 32, 64, and 128 channels respectively, extracting local features related to weather and power output, such as the nonlinear mapping between light intensity and photovoltaic power output; the bidirectional LSTM layer has 2 layers with 128 hidden units and dropout = 0.2, capturing the temporal fluctuations in power output, such as the lag response between wind speed changes and wind power output; the attention mechanism layer focuses on key features and weakens secondary features; and the composite feature fusion layer...

[0133] The system combines bidirectional LSTM output with renewable energy-charge-discharge adaptation efficiency composite features; the fully connected layer uses a 128→64→96 dimension mapping, with output dimension 96 corresponding to the prediction points every 15 minutes in the next 24 hours; the output layer outputs the relative power output value of the power generation state through the Sigmoid activation function, and converts it into the actual power output value (kW) by combining it with the local renewable energy installed capacity.

[0134] Advantages of LSTM and CNN fusion: CNN captures the spatial correlation between meteorological features and power output, such as the nonlinear mapping between light intensity and photovoltaic power output, while LSTM captures the temporal fluctuations in power output, such as wind speed changes and the lag response of wind power output, reducing the error by 8% compared to a single LSTM model; Attention mechanism focuses: highlighting core meteorological features, such as light intensity with a weight of 0.7 and wind speed with a weight of 0.6, while weakening secondary features, such as sunny and rainy conditions with a weight of 0.2, thus improving prediction accuracy.

[0135] SB53: Model training and optimization; Model training, data partitioning: splitting the data into 7:2:1 time series to cover extreme weather scenarios such as sunny, cloudy, rainy, snowy, strong wind, and calm weather, ensuring data diversity; Data augmentation: perturbing the light intensity and wind speed data by ±5% to simulate random weather fluctuations;

[0136] Training parameter configuration: Loss function, using mean squared error (MSE), penalizes output fluctuation error and improves the accuracy of output stability prediction; Optimizer, using Adam, learning rate = 0.002, momentum = 0.9, to accelerate training convergence; Batch size, set to 64, adapts to the length of the time series and improves training efficiency; Number of training epochs, maximum value of 40 epochs, early stopping mechanism is triggered at epochs 20-25 to avoid overfitting.

[0137] Model optimization: Multi-scale time series modeling: The LSTM layer simultaneously captures the output trends at three time scales: 1 hour, 6 hours, and 24 hours, improving the accuracy of short-term and long-term predictions; Composite feature weighting: The composite feature weight of renewable energy-charge and discharge adaptation efficiency is set to 0.5 to strengthen the correlation between absorption efficiency and output prediction; Extreme weather adaptation: Oversampling is performed on extreme weather samples (such as rainstorms and typhoons) to improve the model's prediction ability in special scenarios.

[0138] SB54: Model Validation:

[0139] The verification metrics include: MAE (Maximum Amount Error) of the test set, with a target value of ≤5.5% and an actual value of 5.2%, representing the average error across all scenarios and meeting engineering requirements; peak output error, with a target value of ≤11% and an actual value of 9.8%, indicating accurate peak output prediction and support for energy storage charging and discharging adaptation; and accuracy in identifying periods without power generation, with a target value of ≥95% and an actual value of 97.2%, demonstrating accurate identification of periods without power generation during cloudy or rainy weather and avoiding ineffective scheduling.

[0140] SB55: The renewable energy output prediction model outputs the renewable energy output value (kW) and output fluctuation coefficient every 15 minutes for the next 24 hours. For example, 2024-08-02 12:00-12:15: 1200kW, fluctuation coefficient 0.15, supporting the regulation of energy storage charging and discharging power.

[0141] III. Constructing a battery health management model includes the following steps:

[0142] SC51: Preprocesses the input data, which includes raw data, spatiotemporal big data derived features, battery parameter change gradients, battery health-related composite features, and tag data. The raw data includes 11 dimensions: SOC, charge / discharge power, battery temperature, charge / discharge duration, charge / discharge state, SOH value, real-time capacity decay, health status level, individual cell voltage, and cell internal resistance, providing basic operational and physical parameters for battery health. The spatiotemporal big data derived features include 9 dimensions: charge / discharge cycle efficiency, monthly average DOD, idle time percentage, monthly average SOH decay rate, cumulative SOH decay, remaining cycle life percentage, decay acceleration coefficient, health status change warning value, and correlation coefficient between charge / discharge cycle count and SOH decay, supplementing battery aging trend information. The battery parameter change gradients include 5 dimensions: SOC change gradient, available capacity decay gradient, SOH decay gradient, internal resistance growth gradient, and polarization impedance change gradient, reflecting the dynamic rate of change of battery parameters. The battery health-related composite features include voltage fluctuation-temperature-internal resistance damage composite features, high-frequency cycling- The system comprises four dimensions: SOH attenuation warning composite feature, discharge depth-health mutation risk composite feature, and rainfall-temperature-battery corrosion risk composite feature. These are health risk quantification features derived from multi-source data fusion. The labeled data includes battery health risk level, actual SOH value for the next 30 days, and remaining capacity value (Ah). All input features are uniformly mapped to the [0, 1] interval. For example, the SOH value of 80%-100% in the label is mapped to 0-1, and the remaining capacity value is mapped to 0-1. After prediction, the actual value is restored through reverse mapping.

[0143] SC52: A model architecture design that integrates CNN and physical mechanisms;

[0144] The input layer receives the aforementioned 29-dimensional feature vector; the reshape layer converts the time-series data into a 29×10×1 matrix to adapt to the CNN input format; the CNN feature extraction module has two convolutional kernels (3×3), with 32 and 64 channels respectively, and ReLU activation function, extracting the spatial correlation between battery parameters, such as the coupling relationship between voltage fluctuations, temperature increases, and internal resistance growth; the pooling layer uses 2×2 max pooling to compress dimensions; the LSTM time-series modeling layer has one layer with 128 hidden units and dropout=0.3, capturing long-term changes in battery health status, such as the SOH decay pattern within 6 months; the physical mechanism module constrains the model output based on battery decay theory, with the core formula: SOH(t) = SOH0 - k×t -α×DOD²-β×T, where SOH0 is the initial SOH value, k is the natural decay coefficient, α is the deep discharge decay coefficient, β is the temperature influence coefficient, t is the running time, and T is the average running temperature; Fusion layer, splicing CNN-LSTM output and physical mechanism output to ensure that the prediction results conform to the physical decay law of battery; Fully connected layer, using 128→64→32 dimensional mapping; Dual output layer: classification branch, 32→3 dimensional mapping, Softmax activation, outputs battery health risk level, regression branch, 32→2 dimensional mapping, Linear activation, outputs SOH value and remaining capacity (Ah) for the next 30 days.

[0145] SC53: Model Training and Optimization; Model Training: Data Partitioning: Data is split according to the number of battery cycles. The training set consists of cycles 1-800, the validation set consists of cycles 801-900, and the test set consists of cycles 901-1000, which conforms to the battery aging time series. Sample Balancing: The SMOTE algorithm is used to oversample low-risk and high-risk samples to ensure that the proportion of the three risk levels is 1:1:1, avoiding the model's bias towards medium-risk samples.

[0146] Training parameter configuration includes: loss function, using joint loss = cross-entropy (CE) classification loss × 0.5 + MAE regression loss × 0.5, to balance risk level classification and SOH / capacity tasks; optimizer, using AdamW, learning rate = 0.002, weight decay = 1e-4, to prevent overfitting; batch size, set to 32, to adapt to sample complexity and improve training stability; number of training epochs, maximum 60 epochs, with early stopping mechanism triggered at epochs 30-35 to avoid overfitting.

[0147] Model optimization: Phased training: The first phase trains the CNN-LSTM module and regression branch, and the second phase incorporates the physical mechanism module and classification branch to optimize the risk level; Risk weight adjustment: The classification loss weight of high-risk samples is increased to 1.2 to enhance the identification ability of high-risk scenarios, such as SOH mutation warning; Gradient constraint: Weight decay is set for gradient features of healthy states, such as internal resistance growth gradient, to avoid the model being oversensitive.

[0148] SC54: Model Validation; Core validation metrics include: Risk level accuracy (target value ≥ 93%, actual value 94.5%), accurate risk level identification, supporting maintenance decisions; State of Health (SOH) prediction error (target value ≤ 1.8%, actual value 1.5%), high SOH prediction accuracy, meeting health management needs; Remaining capacity error (target value ≤ 2.5%, actual value 2.2%), accurate remaining capacity prediction, avoiding damage from overcharging and discharging; Early warning lead time (target value ≥ 24 hours, actual value 36 hours), providing early warning for high risks and allowing maintenance time.

[0149] SC55: The battery health management model outputs the battery health risk level, SOH trend curve, remaining capacity (Ah), and maintenance recommendations for the next 30 days. If the risk is high, it is recommended to reduce the monthly average DOD to 60% and enhance the response of the temperature control system to directly support battery maintenance decisions. Example 5

[0150] This embodiment provides a preferred solution, including:

[0151] S12: During the charging and discharging process, the sensor continuously collects the actual operating data of the energy storage cabinet and the actual operating data of the macro power grid. The data processing and deviation analysis module automatically compares the actual operating data of the energy storage cabinet and the actual operating data of the macro power grid with the output data of the comprehensive prediction dataset to calculate the core deviation.

[0152] S13: Set deviation thresholds, such as SOC change deviation ±10%, electricity demand deviation ±15%, and temperature deviation ±2℃. The data processing and deviation analysis module automatically determines whether the core deviation in S12 exceeds the set deviation threshold; otherwise, it maintains the current strategy and parameters and continues to execute, continuously monitoring the data. If it does, it jumps to S14.

[0153] S14: Immediately trigger emergency response. The regional energy management system automatically calls the preset emergency strategies in the strategy library, such as the critical load priority power supply strategy and the battery emergency protection strategy. These strategies do not require complex calculations and are pre-fixed life-saving solutions. Temporary control commands are issued through the BMS and PCS of the energy storage cabinet to limit the discharge output of non-critical loads, forcibly reduce the charging and discharging power to a safe range, and activate the battery over-discharge / over-temperature protection mechanism, such as cutting off unnecessary discharge circuits and activating powerful heat dissipation. The emergency response in this step is an immediate action that does not rely on new forecast data. The core is speed and stability, which buys time for subsequent optimization.

[0154] S15: The regional energy management system issues a re-forecasting instruction to the electricity demand forecasting model, the renewable energy output forecasting model, and the battery health management model. The electricity demand forecasting model, the renewable energy output forecasting model, and the battery health management model generate updated output results based on the latest real-time data, and then the new comprehensive forecast dataset is obtained by the multilayer perceptron model. The re-forecasting in this step is to ensure that the subsequent strategy optimization has reliable data support and to avoid generating invalid strategies based on incorrect data.

[0155] S16: Input the updated comprehensive prediction results from S15 into the fusion decision module. The fusion decision module combines the core deviation types and magnitudes calculated in S12, adjusts the core parameters according to the weight benchmarks and adjustment constraints preset in S8, and recalculates to generate optimization strategies and refined parameters adapted to the current actual scenario. The specific adjustment logic is as follows:

[0156] If the deviation originates from battery health indicators, such as battery temperature deviation exceeding ±2℃ or SOC change deviation exceeding ±10%:

[0157] Core parameter adjustments: The weight of battery health assurance in the reward function is increased from 0.5 to 0.7, while the weight of economic priority is decreased from 0.2 to 0, and the weight of power supply priority remains unchanged at 0.3. At the same time, the feature weights of battery temperature gradient and SOC change rate in the state space are increased from 0.4 to 0.6, strengthening the decision-making influence of battery safety-related features. Refined parameters are generated: the upper limit of charging current is reduced from 0.5C to 0.35-0.4C, the depth of discharge is reduced by 5%-8%, and the temperature control system start-up threshold is reduced from 35℃ to 32℃, prioritizing the suppression of battery health deterioration while ensuring basic power supply for users.

[0158] If the deviation originates from electricity demand indicators, such as an electricity demand deviation exceeding ±15%, it may cause a power supply gap or redundancy. The core parameters are adjusted as follows: the power supply guarantee weight is increased from 0.3 to 0.5, the battery health weight remains unchanged at 0.5, and the economic benefit weight is reduced to 0. Simultaneously, the feature weights of real-time load fluctuations and peak electricity consumption period identifiers in the state space are increased from 0.3 to 0.6 to enhance load adaptability. Refined parameters are generated: the upper limit of discharge power is increased by 10%-15kW, the charging sequence avoids peak electricity consumption periods, and if a power supply gap exists, the grid is prioritized for supplementary power supply to avoid battery damage due to overload discharge of a single energy storage device.

[0159] If the deviation originates from renewable energy output indicators, such as a deviation exceeding 20%, indicating insufficient or excessive green electricity consumption, the core parameters will be adjusted as follows: the weight of economic benefits will be increased from 0.2 to 0.4, the weight of battery health will remain at 0.5, and the weight of power supply security will be reduced to 0.1. At the same time, the characteristic weight of the renewable energy output fluctuation coefficient in the state space will be increased from 0.3 to 0.6 to strengthen the orientation towards green electricity consumption. Refined parameters will be generated: the upper limit of charging power will be increased by 5-10kW, and the charging sequence will be precisely matched with the peak periods of renewable energy. If there is a surplus of green electricity, energy storage will be prioritized for consumption. If there is a shortage of green electricity, charging will be adjusted to the off-peak electricity price period of the grid. All parameter adjustments must meet the battery health and safety thresholds.

[0160] S17: The EMS sends the optimization strategy and parameters generated in S16 to the BMS and PCS of all relevant energy storage cabinets through a standardized interface, replacing the emergency response in S14. The system switches from emergency mode to optimized operation mode. Sensors continuously collect actual operating data after the strategy is executed, such as changes in battery SOC, actual charging and discharging power, and grid coordination effects, and transmit this data back to the EMS in real time. The EMS continues to compare the actual operating data with the updated comprehensive prediction results in S15 and calculate the deviation. If the deviation still exceeds the standard, steps S15-S17 are repeated. If the deviation does not exceed the standard, the current strategy is maintained and monitoring continues. This step ensures that the system can dynamically adapt to changes in the actual scenario and avoids the deviation from being out of control again after optimization.

[0161] Specifically, S12 includes the following steps:

[0162] S121: Real-time acquisition of equipment-side and macro-side data; more specifically, it can acquire battery SOC, charging and discharging current, charging and discharging voltage, battery temperature, heat flow distribution, PCS operating status, and BMS operating status at a high frequency of 1 second / time, and generate a data snapshot every 5 minutes; it can also synchronize real-time regional power load and real-time renewable energy output data from the power grid dispatching platform and user intelligent management system at a frequency of 5 minutes / time.

[0163] S122: Perform preprocessing and anomaly filtering on the collected data; remove sensor fault data, such as sudden SOC changes and constant voltage of 0, and fill missing values ​​with the average of the previous 5 minutes; perform validity verification on PCS and BMS status data, such as determining whether it is in normal operation, fault lockout, or standby state, and mark abnormal status data.

[0164] S123: Decompose the equipment-side data, macro-side data, and comprehensive prediction dataset into three matching subsets and clarify the mapping relationship;

[0165] Battery health status subset: battery SOC, battery temperature, heat flow distribution, and predicted battery health values ​​for the same period, such as SOC trend, temperature safety threshold, and health risk level; Macro-scenario load subset: regional real-time electricity load, real-time renewable energy output, predicted future electricity demand, and predicted renewable energy output characteristics; Equipment operation status subset: PCS real-time power, BMS protection status, charging and discharging current, voltage threshold matching degree, and preset equipment safe operation thresholds.

[0166] S124: For different subsets, data alignment is achieved in both time and space dimensions to ensure the timeliness and relevance of the comparison: Time alignment: For battery health status: The predicted battery health value for a future period, such as the predicted SOC=80% and temperature≤35℃ for period T+1, is bound to the actual battery health data collected in period T+1, and the actual SOC value and actual temperature value are bound one by one.

[0167] For macro-level load scenarios: The current macro-level scenario forecast value for time period T, such as the predicted electricity demand for time period T = 165kW, is linked to the actual macro-level data collected for time period T, with the actual electricity load for the same period = 180kW.

[0168] For equipment operating status: The actual operating parameters during time period T are compared with the predicted values ​​of equipment operating status in the comprehensive prediction dataset, such as predicted PCS power and predicted charging and discharging current.

[0169] Spatial alignment uses the energy storage cabinet device ID and the regional power grid node ID as the association key to ensure that the actual data and the predicted data correspond to the same energy storage device and the same power grid region, thus avoiding data confusion across regions.

[0170] In summary, due to the adoption of the above technical solution, this application has the following advantages:

[0171] 1. By integrating spatiotemporal big data and digital twin data, and after multi-step data processing, the prediction model can capture the linkage between macro trends and micro conditions, significantly reducing prediction errors, achieving more comprehensive data fusion, and higher prediction accuracy.

[0172] 2. Construct electricity demand forecasting models, renewable energy output forecasting models, and battery health management models, and achieve fusion output of forecasting results through multi-layer perceptron models; the three types of models focus on the core variables of different business dimensions, and the fusion of multi-layer perceptron models realizes the collaborative verification of multi-dimensional forecasting results. Compared with a single forecasting model, its comprehensive forecasting error is reduced by 40%-50%, and it can accurately predict load fluctuations, renewable energy output changes, and battery health risks in the next 24 hours;

[0173] 3. A complete technical chain has been formed, including data acquisition, feature mining, model prediction, decision output, and command execution. Through the linkage between the prediction model and the fusion decision module, the charging and discharging strategy is no longer a static preset scheme, but a dynamic optimization scheme based on real-time operating conditions and future predictions. Compared with the traditional fixed-period charging and discharging strategy, this method can adjust parameters in real time according to the changes in renewable energy output and user load fluctuations, realizing a deep adaptation between strategy and operating conditions.

[0174] 4. Improve battery operation safety and extend equipment lifespan. The battery health management model can predict battery SOH changes, capacity decay, and health risk levels in the next 1-3 months. Combined with the highest priority weighting of battery health in the S8 decision module, the charging and discharging strategy will prioritize avoiding conditions that damage the battery. After adopting this method, the prediction error of the remaining battery cycle life is ≤1.5%, the incidence of safety hazards such as battery overheating, overcharging / over-discharging is reduced by 75%, and the overall battery lifespan is extended by 20%-30%.

[0175] 5. Based on the accurate prediction of the electricity demand forecasting model, the integrated decision module can plan the charging and discharging sequence in advance, prioritize discharging to ensure the load during peak electricity consumption periods, and accurately replenish energy during off-peak periods, avoiding power outages or energy waste caused by power supply gaps or redundancy. Compared with traditional strategies, the power supply guarantee rate during peak periods is increased to over 99.5%, and the power load deviation is controlled within ±15%, fully meeting the power supply reliability requirements of scenarios such as commercial complexes and industrial parks.

[0176] 6. The renewable energy output prediction model can accurately predict the peak periods of renewable energy, maximize the consumption of green electricity, avoid the mismatch between energy storage charging and discharging and renewable energy output caused by output fluctuations, and improve the stability of energy utilization.

[0177] 7. Accurate prediction of battery health can reduce operation and maintenance costs. By identifying battery health risks in advance, preventive maintenance can replace emergency repairs, reducing the cost of a single maintenance by 50% and avoiding downtime losses caused by battery failure.

[0178] 8. Strong engineering adaptability, compatible with energy storage systems in multiple scenarios; the data acquisition interface of our application can be adapted to mainstream battery management systems (BMS), energy storage converters (PCS) and grid dispatching platforms. The feature mining and model training stages can be fine-tuned according to the operating conditions of different scenarios, which has broad engineering promotion value and can quickly adapt to the charging and discharging dispatching needs of various user-side energy storage projects.

[0179] However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims. All of these are within the protection scope of this application.

Claims

1. A charge and discharge method of an energy storage tank, characterized by, Comprise the following steps: S1: Collecting spatiotemporal big data and digital twin data, establishing the association mapping of spatiotemporal big data and digital twin data; The spatiotemporal big data includes meteorological data, macro power grid data, user behavior data, energy storage cabinet historical charging and discharging data, battery health degree trend data; The digital twin data includes micro power grid data, energy storage cabinet internal physical parameters, temperature control system dynamic response data, energy storage cabinet geometric data; S2: Normalizing the spatiotemporal big data and the digital twin data, and unifying them to the [0, 1] interval; S3: Extracting derived features from spatiotemporal big data, extracting the change gradient of battery parameters from digital twin data, and mining related composite features after fusing the two types of data; The related composite features include charging and discharging efficiency related composite features, battery health related composite features, and load prediction related composite features; S4: Combining the energy storage cabinet historical charging and discharging data, the macro power grid data, and the battery health degree trend data, labeling the related composite feature data set of S3 with three types of labels, including electricity demand level, renewable energy generation state, and battery health risk level; S5: Constructing electricity demand prediction model, renewable energy output prediction model, and battery health management model; S6: Using the data set labeled in S4, training the electricity demand prediction model, the renewable energy output prediction model, and the battery health management model respectively; S7: Taking the output results of the trained electricity demand prediction model, renewable energy output prediction model, and battery health management model as inputs, fusing and processing through a multilayer perception machine model, and outputting a comprehensive prediction data set, which includes future electricity demand, renewable energy output characteristics, and predicted values of battery health state; S8: Taking the comprehensive prediction data set of S7 as input, training the fusion decision module, and outputting the optimal charging and discharging strategy parameters; During training, a state space of the fusion decision module is constructed synchronously, which contains battery health class core variables, power supply guarantee class core variables, economic benefit class core variables, and corresponding value constraints, as well as fixed core parameter benchmark values and weight adjustment constraints in advance; S9: Based on the fusion decision module combined with real-time data, generating specific charging and discharging strategies; S10: Issuing charging and discharging strategy execution instructions to terminal devices; S11: Converting the charging and discharging strategy results into standardized control instructions, sending them to the battery management system and energy storage converter of the energy storage cabinet through the communication interface, executing battery protection related operations after the battery management system receives the instructions, and starting the charging and discharging process after the energy storage converter receives the instructions.

2. The method of claim 1, wherein, The association mapping of spatiotemporal big data and digital twin data in S1 comprises the following steps: SA1: Establishing a unique docking identifier for spatiotemporal big data and digital twin data to ensure that each data can be accurately matched to the same spatiotemporal scene; SA2: Preprocessing spatiotemporal big data and digital twin data; SA3: Establishing an explicit association rule table for spatiotemporal big data and digital twin data to realize direct mapping; Embedding the explicit rule table into the data processing system to automatically trigger mapping through the association key and realize real-time matching; SA4: Mine the implicit correlation characteristics of spatiotemporal big data and digital twin data, and construct an implicit correlation model to realize indirect mapping. SA5: Verify the mapping.

3. The method of claim 1, wherein the energy storage cabinet is charged and discharged. The derived features include derived features of weather data, derived features of macro power grid data, derived features of user behavior data, derived features of historical charge and discharge data of energy storage cabinets, and derived features of battery health degree change trend data, wherein the derived features of weather data include 24-hour average temperature, temperature change rate, rainfall intensity level, etc.; the derived features of macro power grid data include regional load peak-valley difference, renewable energy output fluctuation coefficient, voltage stability index, and load growth rate; the derived features of user behavior data include electricity consumption peak duration, daily electricity load growth rate, peak-valley electricity consumption ratio, electricity load fluctuation characteristics, and merchant promotion activity identifier; the derived features of historical charge and discharge data of energy storage cabinets include charge and discharge cycle efficiency, monthly average discharge depth, and idle time length proportion; and the derived features of battery health degree change trend data include SOH monthly decay rate, SOH cumulative decay amount, cycle life remaining proportion, decay acceleration coefficient, SOH weekly change rate, health state mutation early warning value, SOH remaining life prediction value, and charge and discharge cycle number and SOH decay correlation coefficient.

4. The method of claim 3, wherein the energy storage tank is charged and discharged. The change gradient of the battery parameter includes a capacity type gradient, a voltage / current type gradient, a temperature type gradient, and a health state type gradient, wherein the capacity type gradient includes an SOC change gradient and an available capacity decay gradient; the voltage / current type gradient includes a single battery voltage change gradient, an access point current change gradient, and a charge and discharge power adjustment gradient; the temperature type gradient includes a temperature rise gradient and a temperature control response gradient; and the health state type gradient includes an SOH decay gradient, an internal resistance growth gradient, and a polarization impedance change gradient.

5. The method of claim 4, wherein the energy storage tank is charged and discharged by, The charge-discharge efficiency related composite features include temperature-voltage-charge efficiency composite feature, renewable energy-charging and discharging adaptation efficiency composite feature, peak-valley load-discharge efficiency composite feature, weather-temperature-PCS efficiency composite feature, and idle duration-charge efficiency composite feature, wherein the temperature-voltage-charge efficiency composite feature is fused from 24-hour average temperature, voltage stability index, internal resistance growth gradient, and temperature rise gradient; the weather-temperature-PCS efficiency composite feature is fused from temperature change rate, temperature response gradient, and charge-discharge cycle efficiency; the battery health related composite features include voltage fluctuation-temperature-internal resistance damage composite feature, high-frequency cycle-SOH attenuation early warning composite feature, discharge depth-health mutation risk composite feature, and rainfall-temperature-battery corrosion risk composite feature, wherein the voltage fluctuation-temperature-internal resistance damage composite feature is fused from voltage stability index, temperature rise gradient, and internal resistance growth gradient, and the rainfall-temperature-battery corrosion risk composite feature is fused from rainfall intensity level, 24-hour average temperature, and polarization impedance change gradient; the load prediction related composite features include electricity consumption peak-load growth prediction composite feature, merchant activity-load fluctuation prediction composite feature, renewable energy-load gap prediction composite feature, and weather-air conditioning load prediction composite feature; and the above PCS represents an energy storage converter.

6. The method of claim 1, wherein, The S4 includes the following steps: S41: aligning historical charge-discharge data of the energy storage cabinet, macro-grid data, and battery health degree change trend data, and threshold dividing electricity demand level, renewable energy generation state, and battery health risk level; S42: labeling electricity demand level by the historical charge-discharge data of the energy storage cabinet; S43: labeling renewable energy generation state by the macro-grid data; S44: labeling battery health risk level by the historical charge-discharge data of the energy storage cabinet and the battery health degree change trend data; S45: label verification and integration.

7. The method of claim 1, wherein the energy storage cabinet is charged and discharged. Further comprising: S12: during the charge-discharge execution, the sensor continuously collects actual operation data of the energy storage cabinet and actual operation data of the macro-grid, and the data processing and deviation analysis module automatically compares the actual operation data of the energy storage cabinet and the actual operation data of the macro-grid with the output data of the comprehensive prediction data set, and calculates the core deviation; S13: setting a deviation threshold, and the data processing and deviation analysis module automatically judges whether the core deviation of S12 exceeds the set deviation threshold; otherwise, maintaining the current strategy and parameters to continue execution, continuously monitoring data, and yes, jumping to S14; S14: immediately triggering an emergency response, and the regional energy management system automatically calling the preset emergency strategy in the strategy library; S15: the regional energy management system issuing a re-prediction instruction to the electricity demand prediction model, the renewable energy output prediction model, and the battery health management model, and the electricity demand prediction model, the renewable energy output prediction model, and the battery health management model generating updated output results based on the latest real-time data, and then obtaining a new comprehensive prediction data set by the multi-layer perception machine model fusion. S16: input the updated comprehensive prediction result of S15 into the fusion decision module, combine the core bias type and bias amplitude calculated in S12, adjust the core parameters according to the weight benchmark and adjustment constraint preset in S8, and re-operate to generate an optimized strategy and refined parameters adapted to the current actual scene; S17: the battery management system sends the optimized strategy and parameters generated in S16 to all related battery management systems and energy storage converters through a standardized interface, replaces the emergency response in S14; the sensor continuously collects actual operation data after the strategy is executed and real-time feedback is transmitted to the battery management system; the battery management system continues to compare the actual operation data with the updated comprehensive prediction result of S15, calculates the deviation, and if the deviation is still out of standard, repeats steps S14-S16, and if the deviation is not exceeded, maintains the current strategy and continues to monitor.

8. The method of claim 7, wherein the energy storage tank is charged and discharged. The S12 includes the following steps: S121: real-time collection of device-side data and macro-side data; S122: pre-processing and abnormal filtering of the collected data; S123: disassembling the device-side data, macro-side data and comprehensive prediction data set into three matchable subsets to clearly map the relationship; the three subsets are: a battery health state subset, a macro scene load subset and a device operation state subset; S124: for different subsets, realize data alignment in time and space dual dimensions to ensure the timeliness and relevance of comparison.

Images