Multi-type energy storage staged capacity optimization configuration method for new energy uncertainty

Abstract

The invention relates to the technical field of data processing, in particular to a new energy uncertainty multi-type energy storage staged capacity optimization configuration method, which comprises the following steps: acquiring energy storage parameters, new energy output and user load data through edge equipment, and verifying photovoltaic conversion efficiency and lithium battery electrochemical behaviors through a physical model to generate a high-confidence data set; the federal cooperative system fuses distributed node data to output a health degree distribution diagram, and the memory network analyzes the temperature and the charge-discharge coupling effect to generate a dynamic compensation coefficient alpha; an optimization engine receives the alpha value and wind and light prediction, generates a staged scheme through a parallel algorithm, and inputs the staged scheme into a simulation system for verification; and the edge node executes the scheme and collects response data, and twin compares a simulation value with an actual deviation to update a collection strategy. The method eliminates data distortion through physical rule verification, corrects aging unit state deviation through a dynamic compensation mechanism, optimizes decision binding real-time health degree, and effectively defends charging and discharging analysis errors of a virtual power plant dispatching desk.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for optimizing the phased capacity configuration of multi-type energy storage with uncertainties in new energy sources. Background Art

[0002] Uncertainty in new energy sources stems primarily from the random volatility and unpredictability of renewable energy output power, posing challenges to the stable operation of power grids and requiring effective management through multiple types of energy storage systems. Energy storage technologies include electrochemical, mechanical, and capacitive energy storage, each with complementary response characteristics adapted to different timescales. Leveraging data processing and control technologies, neural network-based predictive models and real-time scheduling algorithms can analyze demand and coordinate the charging and discharging logic of various energy storage systems, thereby optimizing energy allocation, reducing wind and solar curtailment, and enhancing system resilience. Ultimately, this will enable reliable scheduling and maximize the economic benefits of a high-penetration new energy grid.

[0003] The pain point in phased capacity planning for multi-type energy storage for renewable energy lies in the multi-objective optimization distortion caused by data barriers and the failure of dynamic aging modeling for distributed resources. This pain point stems from the loss of heterogeneous information caused by data normalization under privacy constraints. This failure also fails to integrate energy storage lifespan degradation factors in real time, resulting in the optimization model's inability to accurately reflect dynamic capacity changes. This ultimately amplifies performance deviations and economic losses in actual scheduling. For example, in a virtual power plant scenario where distributed photovoltaics and user-side lithium iron phosphate batteries are aggregated, the control system can only obtain normalized charging and discharging strategies but lacks specific aging characteristics. Private data restrictions also hinder the construction of a complete lifespan degradation trajectory. When the system uses a stochastic optimization algorithm to make phased capacity decisions, the model input is biased due to incomplete information, leading to storage coordination failures. For example, power plant projects experience decreased peak-shaving and valley-filling efficiency, and combined with uncompensated aging factors, this significantly reduces peak-valley price differential profits and increases grid assessment risks. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a multi-type energy storage staged capacity optimization configuration method with new energy uncertainty, which solves the problem of charging and discharging analysis errors of the virtual power plant dispatching station caused by real-time energy storage status errors.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0006] The method for optimizing the capacity of multi-type energy storage in stages with new energy uncertainty of the present invention includes:

[0007] Obtain energy storage device parameters, renewable energy output data, and user load data, verify physical characteristics, and generate a high-confidence data set;

[0008] Inputting the high-confidence data set into the collaborative system to perform energy storage health assessment and output a health distribution map to provide an input basis for dynamic capacity compensation;

[0009] Inputting the health distribution map and historical charge and discharge data into a memory network to generate a dynamic compensation coefficient α representing actual capacity attenuation;

[0010] Inputting the compensation coefficient α and wind / solar output forecast data into the optimization engine, triggering the parallel algorithm to generate a phased capacity optimization plan;

[0011] Executing the optimization scheme through edge nodes and collecting device response data;

[0012] Compare the deviation between the twin simulation value and the device response data, and generate feedback signals to update the data acquisition strategy.

[0013] Furthermore, the method for optimizing the capacity of multi-type energy storage in stages with uncertainty in new energy sources described in the present invention obtains energy storage device parameters, new energy output data, and user load data, verifies physical characteristics, and generates a high-confidence data set, including:

[0014] The edge device collects the charge and discharge rate matrix, temperature array data, and photovoltaic original characteristic points of the energy storage battery pack through the protocol interface;

[0015] Push the collected data to the transport layer through the encrypted channel, perform encryption operations and then map it to a unified address space;

[0016] The transmission data trigger processing layer uses physical models to verify photovoltaic conversion efficiency and electrochemical inversion, and marks abnormal points for data that fails verification;

[0017] Call the processing module by type for the marked results:

[0018] Execute interpolation algorithms on interrupted data, call neighboring stations to compensate for abnormal data points or switch to redundant channels;

[0019] Noise is added to the verified user load data and device parameters are integrated to form a high-confidence data set.

[0020] Furthermore, the method for optimizing the capacity of multi-type energy storage in stages with uncertainty in new energy sources described in the present invention uses the high-confidence data set to evaluate the health of energy storage in a collaborative system and output a health distribution map, including:

[0021] The main system sends the high-confidence data set and the initial neural network model to the distributed nodes;

[0022] Distributed nodes update model parameters based on local data and send encrypted gradients back to the main system;

[0023] The simulation system injects pathological scenarios to trigger error verification: if the error is ≥ 5% for three consecutive times, the model is rolled back; otherwise, a verification pass signal is sent to the backup system;

[0024] After receiving the verification pass signal, the standby system loads the scheduling instruction stream and executes the stress test response;

[0025] The model that passes the stress test is published to the main system, mapping the normalized SOC data into dynamic capacity intervals;

[0026] A battery health distribution map is generated based on the dynamic capacity interval.

[0027] Furthermore, the method for optimizing the capacity of multi-type energy storage in stages with uncertainty in new energy sources described in the present invention, wherein the health distribution map and historical charge and discharge data are input into a memory network to calculate the dynamic capacity compensation coefficient α, includes:

[0028] Real-time collection of ambient temperature and humidity data through the sensor interface, and inputting the data into the time series alignment module together with the health distribution map and historical cycle depth;

[0029] The sliding window extracts continuous working condition feature vectors from the aligned data according to the preset step size as the input of the memory network;

[0030] When the feature vector is updated or reaches the time threshold, the memory network loads the pre-trained model inference output compensation coefficient α;

[0031] The compensation coefficient α value is periodically refreshed every minute and pushed to the real-time interface of the optimization engine.

[0032] Furthermore, the method for optimizing the capacity of multi-type energy storage in stages with uncertainty in new energy sources described in the present invention, wherein the compensation coefficient α and the wind and solar power output forecast are input into the optimization engine, and an optimization solution is generated through a parallel algorithm, includes:

[0033] The first algorithm engine obtains the long-term probability distribution data set of wind and solar power, and generates a multi-stage capacity ratio solution set as a candidate solution;

[0034] The second algorithm engine receives the compensation coefficient α, the short-term wind and solar power output forecast, and the user load forecast in real time, generates hourly strategy instructions on a rolling basis, and archives them to the strategy library;

[0035] Input the candidate solutions and rolling strategy instructions into the simulation system, inject the grid frequency fluctuation fault scenario, and quantify the economic benefit deviation and the number of charge and discharge rate limit violations;

[0036] When the economic indicator difference rate is greater than 15% or the safety indicator deviation is greater than 10%, the manual intervention interface review plan is triggered;

[0037] After the optimization plan is approved, it will be switched to the production environment through blue-green deployment.

[0038] Furthermore, the method for optimizing the capacity of multi-type energy storage in stages under uncertainty of new energy sources described in the present invention is characterized in that the optimization scheme is executed while feeding back a deviation signal to improve data collection, including:

[0039] Subscribe to the scheduling instructions corresponding to the optimization plan from the production environment scheduling bus;

[0040] The edge computing node responds to the scheduling instruction and performs control logic calculation in seconds;

[0041] Collect actual voltage and current response data of energy storage equipment;

[0042] The twin is loaded with a typhoon path or extreme temperature scenario library, and combined with the actual response data to simulate the theoretical behavior of the equipment under extreme working conditions;

[0043] Calculating the root mean square deviation between the theoretical simulation value and the actual response data, and generating an optimization strategy package when the deviation is greater than 5%;

[0044] The policy package specifies the location of new temperature sensors, the sampling frequency increase multiple, or the expansion of voltage monitoring points, and is pushed to the edge nodes to update the collection policy.

[0045] Furthermore, in the method for optimizing the capacity configuration of multi-type energy storage in stages with uncertainty of new energy sources described in the present invention, the simulation system injects a capacity drop pathological scenario during the low-temperature verification phase and generates an alarm signal when the model predicts a capacity error of ≥5%.

[0046] The alarm signal triggers the reconfiguration protocol of the data acquisition agent, increasing the sampling frequency of the temperature sensor in the problem area to 10 seconds per time.

[0047] Furthermore, in the method for optimizing the capacity configuration of multi-type energy storage in stages with uncertainty of new energy sources described in the present invention, the memory network analyzes the temporal association between temperature data below -10°C and charge and discharge records above 2C, and outputs a coupling effect identifier;

[0048] When the identifier strength is greater than the threshold, the compensation coefficient α is automatically adjusted downward by 0.15 ± 0.03;

[0049] The lowered α value forces the optimization engine to update its charge and discharge constraints through a real-time interface.

[0050] Furthermore, in the method for optimizing the capacity configuration of multi-type energy storage in stages with new energy uncertainty described in the present invention, when the twin computing device instruction response delay is greater than 200ms, a response hysteresis deviation is marked;

[0051] For areas with deviation values ​​greater than 15%, the optimization strategy package generates sensor supplementary instructions:

[0052] The density of temperature monitoring points is increased to 300% of the original configuration;

[0053] Deploy voltage monitoring modules for battery clusters with capacity decay > 20%.

[0054] Furthermore, the multi-type energy storage capacity optimization configuration method for phased capacity optimization under the uncertainty of new energy sources described in the present invention defines α < 0.85 as a high-risk threshold, which is derived from the failure analysis of the 72-hour stress test of the backup system;

[0055] The second algorithm engine enforces the prohibition instruction: refusing to dispatch tasks with a rate greater than 1C to high-risk units;

[0056] The policy instructions dynamically match the health heat map status released by the main system and update the charge and discharge authorization parameters.

[0057] Beneficial effects of the present invention:

[0058] The present invention improves the accuracy of scheduling analysis through a multi-level technical linkage mechanism. The specific beneficial effects are as follows: the physical model verification layer filters the input noise caused by hardware distortion and data privacy, generates a high-confidence data set and eliminates initial state false alarms; the federal collaborative architecture dynamically integrates battery aging characteristics and low-temperature attenuation laws, and the compensation coefficient α refreshes the performance offset of the aging unit at the minute level to correct the scheduling system's deviation in the real-time capacity judgment of energy storage; the optimization engine binds the α value to execute high-risk task bans and dynamically matches the spatial coordinates of the health heat map, forcing the charging and discharging parameters to adapt to the health status of the equipment in real time, effectively resolving the scheduling analysis failure caused by the superposition of sensor drift, data barriers and capacity attenuation misjudgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0060] Figure 1 A flow chart of a method for optimizing the capacity of multiple types of energy storage in stages with uncertainties in new energy sources provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. In order to better understand the purpose of the present invention, the present invention is further described in detail below.

[0062] See also Figure 1 The method for optimizing the capacity of multi-type energy storage in stages under the uncertainty of new energy sources of the present invention comprises:

[0063] Step 1: Obtain energy storage device parameters, renewable energy output data, and user load data, verify physical characteristics, and generate a high-confidence data set;

[0064] Step 2: Input the high-confidence data set into the collaborative system to perform energy storage health assessment and output a health distribution map to provide an input basis for dynamic capacity compensation;

[0065] Step 3: Input the health distribution map and historical charge and discharge data into a memory network to generate a dynamic compensation coefficient α that represents actual capacity attenuation;

[0066] Step 4: Input the compensation coefficient α and wind / solar output forecast data into the optimization engine, triggering the parallel algorithm to generate a phased capacity optimization plan;

[0067] Step 5: Execute the optimization solution through the edge node and collect device response data;

[0068] Step 6: Compare the deviation between the twin simulation value and the device response data, and generate a feedback signal to update the data acquisition strategy.

[0069] Step 1: At the edge layer, the charge and discharge rate matrix, distributed temperature sensor array data, and raw IV characteristic points of the PV inverter are collected via the Modbus / TCP protocol from the battery pack's BMS. The communication layer establishes a TLS-encrypted channel to transmit this data to the processing center, implementing heterogeneous data mapping using the OPC UA address space. The processing layer verifies conversion efficiency compliance using the PV module IV characteristic model and inverts the electrochemical behavior logic using the lithium battery double-layer effect model. Anomalous data is processed in a hierarchical manner: time series interruptions are addressed using sliding window ARIMA interpolation, spatial anomalies are compensated using Kriging neighboring stations, and hardware failures are handled by switching to redundant channels. Verified user load data is supplemented with standard data, and device parameters are integrated to form an anonymized, high-confidence dataset. This step, through physical rule constraints, enhances the inherent credibility of the data and provides a traceable input foundation for health assessment.

[0070] In step 2, the main production system sends the high-confidence data set and the initial federated learning model to the edge node. After the node updates the neural network parameters in the local training environment, it transmits the gradient back to the main system through homomorphic encryption. The simulation test system simultaneously injects pathological scenarios such as -10°C low-temperature capacity drop and high-rate cycle attenuation. When the verification error is ≥5% for three consecutive times, the model rollback alarm is automatically triggered; the verified model is pushed to the backup system, and the simulated peak shaving and valley filling task flow is loaded to perform a 72-hour stress test. The test-compliant model is published to the main system for operation, and the normalized SOC data is mapped to the actual dynamic range of available capacity (such as 80% SOC corresponds to 65-82kWh), generating a visual battery health distribution heat map. The health distribution map integrates temperature attenuation characteristics to provide a dynamic input basis for capacity compensation.

[0071] Step 3 receives the health distribution map, historical charge and discharge cycle depth, and ambient temperature and humidity sensor data, and normalizes the multi-source input timestamps to a 1-minute granularity. A sliding window extracts a 72-hour continuous operating condition feature vector and inputs it into the LSTM network. A pre-trained transfer learning model is then loaded to analyze the nonlinear coupling relationship between temperature and cycle depth. The network outputs a compensation coefficient α in real time: when a temperature below -5°C is detected and a charge and discharge record above 2°C is detected, the α value is dynamically reduced by 0.1-0.2. The calculation results are published to the optimization engine interface via the OPC UA real-time data bus, with minute-by-minute updates to ensure timely capacity decay compensation. The compensation coefficient α quantifies the actual available capacity decay, eliminating any discrepancies between the health assessment and the real-time status.

[0072] In step 4, after receiving the compensation coefficient α and the 15-minute granularity wind and solar output forecast data, the optimization engine launches the genetic algorithm and the adaptive MPC engine in parallel. The genetic algorithm analyzes the annual wind and solar probability distribution dataset and generates a Pareto solution for the lithium battery / supercapacitor ratio using chromosome encoding. The MPC engine combines the α value with the short-term load forecast to generate hourly charge and discharge strategy instructions on a rolling basis and stores them in the strategy library. The simulation system injects a grid frequency fluctuation fault scenario of ±0.5Hz to quantify the economic benefit deviation rate between the solution set and the strategy instructions and the number of charge and discharge rate limit violations. If the discrepancy rate exceeds 15%, a WebSocket alert is triggered to the manual review interface, and the blue-green deployment mechanism switches the verified solution to the production environment. The compensation coefficient α forces an update of the optimization constraints to avoid incorrect scheduling of aging units.

[0073] In step 5, the production environment scheduling bus issues optimization solution instructions to the KubeEdge edge computing node. After subscribing to the instructions, the node executes millisecond-level charge and discharge control logic calculations. The actual voltage and current response data streams from the energy storage device's BMS are synchronously collected and time-aligned to Coordinated Universal Time. The twin is loaded with a typhoon path model or an ASME-certified extreme temperature scenario library, and simulates the theoretical response waveform based on the device's physical parameters. The edge execution output forms a closed-loop verification with the actual collected data, supporting subsequent quantitative deviation analysis.

[0074] In step 6, the root mean square deviation (RMS) between the theoretical simulation value and the actual response data is calculated. When the deviation exceeds 5%, an optimization strategy package in JSON format is generated. The strategy package specifies the geographic coordinates of the new temperature sensor, the sampling frequency increase (e.g., from 5 minutes to 10 seconds), or the deployment of a voltage monitoring module for battery clusters with capacity degradation exceeding 20%. Instructions are pushed to the edge layer collection agent via the MQTT protocol, dynamically updating the Modbus / TCP collection strategy parameters. The updated collection strategy improves the quality of the data source in step 1, completing the technical closed loop from execution feedback to source optimization.

[0075] Specifically, the method for optimizing the capacity of multi-type energy storage in stages with uncertainty in new energy sources described in the present invention obtains energy storage device parameters, new energy output data, and user load data, verifies physical characteristics, and generates a high-confidence data set, including:

[0076] The edge device collects the charge and discharge rate matrix, temperature array data, and photovoltaic original characteristic points of the energy storage battery pack through the protocol interface;

[0077] Push the collected data to the transport layer through the encrypted channel, perform encryption operations and then map it to a unified address space;

[0078] The transmission data trigger processing layer uses physical models to verify photovoltaic conversion efficiency and electrochemical inversion, and marks abnormal points for data that fails verification;

[0079] Call the processing module by type for the marked results:

[0080] Execute interpolation algorithms on interrupted data, call neighboring stations to compensate for abnormal data points or switch to redundant channels;

[0081] Noise is added to the verified user load data and device parameters are integrated to form a high-confidence data set.

[0082] A lightweight data collection agent is deployed at the device edge layer to acquire, in real time, the cell-level charge and discharge rate matrix, distributed temperature sensor array data, and raw IV characteristic sampling points of the photovoltaic inverter, output by the energy storage battery management system, via the Modbus / TCP industrial protocol. The collected data includes the voltage change time series of each cell within the battery pack and the balance status between groups. The photovoltaic data covers the DC output characteristics under different irradiance levels.

[0083] The communication transport layer establishes an MQTT over TLS 1.3 encrypted channel, mapping the raw data stream to the OPC UA unified address space for transmission to the central processing system. This mapping operation establishes a binding relationship between the device's physical address and the virtual data tag, achieving spatial alignment of heterogeneous data from PV power plants and energy storage sites at the logical layer. The encrypted channel ensures that the transmission process complies with the IEC 62351 security standard and prevents data tampering caused by man-in-the-middle attacks.

[0084] Data packets transmitted to the central processing layer trigger the PV panel physical verification module, which calls the IEC 61853 standard PV model library to verify the inverter's reported power. When output values ​​are detected exceeding the theoretical conversion efficiency limit, abnormal data points are automatically flagged and a repair process is activated. Double-layer effect inversion is simultaneously performed on lithium battery data, and the authenticity of charge and discharge rates is verified through electrochemical impedance spectroscopy. Rate mutation data caused by temperature sensor failure is listed as an anomaly.

[0085] A hierarchical processing protocol is initiated for marked anomalies: For data flow interruptions, a sliding window ARIMA interpolation algorithm is used to restore time series continuity. For spatial distribution anomalies, the Kriging spatial interpolation engine is triggered to retrieve data from similar operating conditions at neighboring power plants to reconstruct the data distribution. For hardware-level failures, redundant sensor channels are switched to maintain data integrity. The anomaly handling process inherits parameters from the equipment's historical operating signature database, ensuring that the repair results conform to the equipment's physical operating principles.

[0086] Physically verified user load data enters the privacy processing module, where layered noise is applied at the edge node: differentially private Laplace noise is added to the load time series, and Gaussian noise is added to the device operating parameters. The noise parameters dynamically adapt to the load curve's fluctuations, preserving key peaks and valleys while meeting privacy compliance requirements. The noisy data and device parameters are aligned with standardized timestamps and integrated into a high-confidence dataset for output to the downstream federated learning system.

[0087] Specifically, the method for optimizing the capacity of multi-type energy storage in stages with uncertainty in new energy sources described in the present invention uses the high-confidence data set to evaluate the health of energy storage in a collaborative system and output a health distribution map, including:

[0088] The main system sends the high-confidence data set and the initial neural network model to the distributed nodes;

[0089] Distributed nodes update model parameters based on local data and send encrypted gradients back to the main system;

[0090] The simulation system injects pathological scenarios to trigger error verification: if the error is ≥ 5% for three consecutive times, the model is rolled back; otherwise, a verification pass signal is sent to the backup system;

[0091] After receiving the verification pass signal, the standby system loads the scheduling instruction stream and executes the stress test response;

[0092] The model that passes the stress test is published to the main system, mapping the normalized SOC data into dynamic capacity intervals;

[0093] A battery health distribution map is generated based on the dynamic capacity interval.

[0094] The main production system schedules resources through a container orchestration platform and uses the gRPC microservice interface to deliver high-confidence datasets and pre-trained neural network initial models to distributed edge nodes. The initial model includes a convolution kernel weight matrix for lithium battery capacity decay, and the dataset covers timestamp-aligned versions of features related to photovoltaic output fluctuations and health. This data model collaborative architecture supports the operation of the federated learning mechanism.

[0095] Distributed nodes load data shards into a local encrypted training sandbox and apply a stochastic gradient descent optimizer to update the fully connected layer parameter gradients. The gradients are converted into ciphertext packets using the Paillier homomorphic encryption algorithm and transmitted back to the main system aggregation server via a dedicated channel. This encrypted transmission mitigates the risk of sensitive parameter leakage and ensures privacy-compliant computing.

[0096] The simulation test system uses a pathological scenario library to load a -10°C low-temperature capacity drop model and injects it into the federated learning model's output layer for error verification. The root mean square error (RMSE) between the predicted health distribution and the actual attenuation curve is calculated. If the error reaches or exceeds 5% three times in a row, an automatic rollback mechanism is activated, restoring the model to the previous stable version. If verification passes, a digital signature certificate is issued, transmitting an authorization signal to the backup system.

[0097] After receiving the digitally signed authorization, the backup system loads a 72-hour standard scheduling instruction stream and performs stress testing. Hardware metrics such as peak memory usage and inference latency are monitored. If response latency exceeds 200ms or the failure rate is ≥ 0.1%, the test process is interrupted. Models that pass the test are marked as reliable and released to the main production system environment in a phased manner using a canary release strategy.

[0098] The deployed model parses the normalized SOC data stream uploaded by the battery management system and maps percentage values ​​to dynamic capacity physical ranges based on the association rules between temperature compensation coefficients and cycle aging factors. For example, 80% SOC at 25°C is mapped to a practical capacity range of 65-82 kWh. The boundary values ​​are associated with the color coding rules of the health heat map.

[0099] Based on the upper and lower thresholds of the dynamic capacity range, a vectorized heat map of the spatial distribution of battery cluster health is generated. The heat map coordinates map the physical location of the energy storage devices. Red areas indicate high-risk units with capacity degradation greater than 15%, while blue areas indicate healthy units. This heat map is then output to the virtual power plant dispatcher visualization terminal.

[0100] Specifically, the method for optimizing the capacity of multi-type energy storage in stages with uncertainty in new energy sources of the present invention includes inputting the health distribution map and historical charge and discharge data into a memory network to calculate the dynamic capacity compensation coefficient α, which includes:

[0101] Real-time collection of ambient temperature and humidity data through the sensor interface, and inputting the data into the time series alignment module together with the health distribution map and historical cycle depth;

[0102] The sliding window extracts continuous working condition feature vectors from the aligned data according to the preset step size as the input of the memory network;

[0103] When the feature vector is updated or reaches the time threshold, the memory network loads the pre-trained model inference output compensation coefficient α;

[0104] The compensation coefficient α value is periodically refreshed every minute and pushed to the real-time interface of the optimization engine.

[0105] In a containerized cluster environment, the main production system delivers a high-confidence dataset and a pre-trained convolutional neural network initial model to distributed edge nodes via the gRPC microservice interface. The initial model includes a basic weight matrix for the temperature-dependent characteristics of lithium-ion battery capacity decay, and the dataset includes a timestamp-aligned version of the characteristics associated with photovoltaic output fluctuations and battery aging.

[0106] Distributed nodes load dedicated data shards into their local training sandboxes, use stochastic gradient descent optimizers to update fully connected layer parameters, and encrypt the gradient parameters using the Paillier homomorphic encryption algorithm to generate ciphertext data packets. The encrypted gradients are then transmitted back to the main system aggregation server via a dedicated data channel, completing the iterative update of the global model. This gradient encryption mechanism mitigates the risk of privacy leaks during transmission.

[0107] The simulation test system uses a library of pathological scenarios, including a sudden capacity drop at -10°C and a cyclic accelerated degradation scenario at a rate of 0.5°C, to inject into the model verification process. The system then compares the root mean square error (RMS) between the model's health probability distribution and the actual capacity curve. If the error reaches or exceeds 5% for three consecutive tests, an automatic rollback mechanism is activated, restoring the model to the previous stable version and generating an alarm log. Upon successful verification, a digital signature certificate is issued, signaling authorization to the backup system.

[0108] After receiving the digitally signed authorization, the backup system loads a 72-hour standard scheduling instruction stream, simulating peak load shaving and valley filling tasks for stress testing. Metrics such as peak memory usage and inference latency are monitored, and testing is interrupted if response latency consistently exceeds 200ms or the failure rate is ≥ 0.1%. Models that pass the test are marked as reliable and released to the main production system in a phased release using a canary release strategy.

[0109] The deployed model parses the normalized SOC data stream uploaded by the battery management system and maps it into dynamic capacity ranges based on the lithium battery state-of-health characteristic curve. This mapping associates a temperature compensation coefficient with a cycle aging factor. For example, 80% SOC in a 25°C environment is mapped to 75-85% of the nominal capacity. Based on the interval boundaries, a vectorized heat map of the battery health distribution is generated, noting the spatial locations of battery clusters with varying degrees of degradation.

[0110] Specifically, the method for optimizing the capacity of multi-type energy storage in stages with uncertainties in new energy sources described in the present invention, wherein the compensation coefficient α and the wind and solar power output forecast are input into the optimization engine, and an optimization scheme is generated through a parallel algorithm, includes:

[0111] The first algorithm engine obtains the long-term probability distribution data set of wind and solar power, and generates a multi-stage capacity ratio solution set as a candidate solution;

[0112] The second algorithm engine receives the compensation coefficient α, the short-term wind and solar power output forecast, and the user load forecast in real time, generates hourly strategy instructions on a rolling basis, and archives them to the strategy library;

[0113] Input the candidate solutions and rolling strategy instructions into the simulation system, inject the grid frequency fluctuation fault scenario, and quantify the economic benefit deviation and the number of charge and discharge rate limit violations;

[0114] When the economic indicator difference rate is greater than 15% or the safety indicator deviation is greater than 10%, the manual intervention interface review plan is triggered;

[0115] After the optimization plan is approved, it will be switched to the production environment through blue-green deployment.

[0116] The first algorithm engine accesses a long-term probability distribution dataset of wind and solar resources from a historical meteorological database, analyzing the joint distribution characteristics of photovoltaic irradiance and wind speed over a time scale covering the next 1-3 years. A genetic algorithm chromosome encoding is used to generate a Pareto solution set for the phased capacity allocation of various energy storage devices, including lithium battery packs and supercapacitors. This strategy space is then generated, encompassing quarterly capacity expansion plans and investment cost constraint boundaries. The candidate solution set is annotated with the expected wind curtailment rate and payback period parameters for each configuration strategy.

[0117] The second algorithm engine subscribes to compensation coefficient α update events via a real-time data bus, synchronously acquiring ultra-short-term wind and solar output forecast curves and user load patterns. Using an adaptive model predictive control framework, it optimizes hourly charge and discharge strategy instructions on a 15-minute rolling basis. These instructions include maximum charge and discharge rate constraints and SOC safety intervals. All strategy instructions are timestamped and stored in a distributed strategy library, supporting historical strategy backtesting.

[0118] The simulation system retrieves frequency fluctuation scenario data (±0.5Hz fluctuation lasting 10 minutes) from a grid fault case library and loads candidate capacity allocation solutions and real-time policy instructions as input parameters. The system simulates actual operating conditions to calculate economic indicators (the rate of change in peak-valley arbitrage returns and the cost of curtailed solar power) and safety indicators (the number of battery charge and discharge rate violations and the duration of voltage overshoots), and quantitatively compares the deviations between the dual-engine outputs.

[0119] When the economic benefit deviation rate exceeds 15% or the safety indicator deviation rate exceeds 10%, a WebSocket protocol alarm signal is automatically triggered and sent to the manual review interface of the dispatch console. The review panel simultaneously displays a comparison view of simulation data and supports online annotation of policy adjustment instructions. The manual review conclusion is recorded in the blockchain evidence node to ensure the traceability of the decision-making process.

[0120] Approved optimization solutions are packaged as container images and switched to the production environment using a blue-green deployment mechanism. Initially, 5% of traffic is diverted to the new version container group, and scheduling command response latency and memory leaks are monitored for 12 consecutive hours. Once the error rate falls below 0.1%, the diversion ratio is gradually increased to 100%. The old version container group is placed in hot standby mode, maintaining a 72-hour rollback capability.

[0121] Specifically, the method for optimizing the capacity of multi-type energy storage in stages under uncertainty of new energy sources according to the present invention is characterized in that the optimization scheme is executed while feeding back a deviation signal to improve data collection, including:

[0122] Subscribe to the scheduling instructions corresponding to the optimization plan from the production environment scheduling bus;

[0123] The edge computing node responds to the scheduling instruction and performs control logic calculation in seconds;

[0124] Collect actual voltage and current response data of energy storage equipment;

[0125] The twin is loaded with a typhoon path or extreme temperature scenario library, and combined with the actual response data to simulate the theoretical behavior of the equipment under extreme working conditions;

[0126] Calculating the root mean square deviation between the theoretical simulation value and the actual response data, and generating an optimization strategy package when the deviation is greater than 5%;

[0127] The policy package specifies the location of new temperature sensors, the sampling frequency increase multiple, or the expansion of voltage monitoring points, and is pushed to the edge nodes to update the collection policy.

[0128] Edge computing nodes capture the charge and discharge scheduling instruction stream corresponding to the optimization solution from the scheduling bus in real time via an OPC UA subscription interface. The instruction stream includes the target SOC change curve, power allocation weights, and time synchronization tags. Upon receiving the instruction stream, the node activates the microsecond-level control logic thread, parses the instruction, and generates PWM wave control signals to drive the IGBT power module for precise charging and discharging.

[0129] During execution, real-time data streams, such as voltage transient response waveforms and current ripple coefficients, are collected synchronously from the energy storage unit's BMS. IEEE 1588 precision clock protocol is used to align data timestamps and ancillary environmental parameters such as temperature and humidity, generating a dataset of actual device responses with temporal and spatial integrity.

[0130] The digital twin uses an ASME-certified typhoon path prediction model (WRF model output) or a library of extreme temperature scenarios (operating conditions ranging from -40°C to 70°C), loading device physical parameters (internal resistance, heat capacity, etc.) and collected actual response data. The fluid mechanics-electrochemical coupled simulation engine calculates the theoretical device behavior under extreme operating conditions, outputting the theoretical current and voltage waveform envelope characteristics.

[0131] Calculate the root mean square error (RMS) between the theoretical waveform envelope and the actual response data. If the error exceeds 5%, the analysis tool generates a JSON-formatted optimization strategy package: defining the GPS coordinates for the new temperature sensor deployment point; setting a sampling frequency increase multiplier (e.g., 2× / 5×) for the specified area; and expanding voltage monitoring points to the individual cell level within the battery cluster. The strategy includes a digital signature verification identifier.

[0132] An MQTT-based command channel pushes optimization policy packages to the edge-layer collection agent. The agent dynamically updates the collection point configuration table in the Modbus / TCP protocol, adjusting device polling intervals, adding monitoring point address codes, or switching redundant data source channels. The updated collection policy takes effect immediately, improving the integrity of subsequent data collection.

[0133] Specifically, the present invention provides a multi-type energy storage capacity optimization configuration method for phased capacity under uncertainty of new energy sources. The simulation system injects a capacity drop pathological scenario during the low-temperature verification phase and generates an alarm signal when the model predicts a capacity error of ≥5%.

[0134] The alarm signal triggers the reconfiguration protocol of the data acquisition agent, increasing the sampling frequency of the temperature sensor in the problem area to 10 seconds per time.

[0135] During the low-temperature verification phase, the simulation system loads a library of lithium battery pathological characteristics and injects a data model for a typical scenario involving a 30% capacity decay at -10°C. The system then uses a federated learning model to predict the battery capacity under the current operating conditions. This model then performs a time-series alignment calculation with the theoretical decay curve for the injected scenario, outputting a sequence of root mean square errors (RMS) in the capacity decay of each battery cluster. If the model's prediction error exceeds or equals 5% three times in a row, a digitally signed alarm signal containing the device's location code is generated.

[0136] The alarm signal is transmitted to the data collection agent's reconfiguration engine via the message queue middleware. The engine parses the location tag in the alarm signal and locates the physical area where the target battery cluster is located. The reconfiguration protocol is executed to update the temperature sensor parameters: the polling interval for data collection points is changed to 10 seconds, and synchronized collection mode for redundant temperature sensors is enabled. The protocol change instructions are then written to the collection configuration database after being identified by a digital certificate.

[0137] The data collection agent activates a dynamic monitoring thread to implement enhanced monitoring of temperature data in the target area. This monitoring data is uploaded in real time to the federated learning model training data pool, supplementing it with samples of low-temperature operating condition characteristics. The updated temperature data is transmitted back to the simulation system via an encrypted channel, establishing a closed-loop feedback loop for anomaly detection.

[0138] Specifically, in the method for optimizing the capacity of multi-type energy storage in stages with uncertainty in new energy sources described in the present invention, the memory network analyzes the temporal association between temperature data below -10°C and charge and discharge records above 2C, and outputs a coupling effect identifier;

[0139] When the identifier strength is greater than the threshold, the compensation coefficient α is automatically adjusted downward by 0.15 ± 0.03;

[0140] The lowered α value forces the optimization engine to update its charge and discharge constraints through a real-time interface.

[0141] The memory network is loaded with a transfer learning model for lithium-ion battery low-temperature attenuation. Using LSTM units, it analyzes ambient data below -10°C reported by the temperature sensor and the time series of charge and discharge events above 2C recorded by the battery management system. The model extracts the charge and discharge feature vectors three minutes before and after the temperature drop node. The covariance matrix of the two in the hidden layer is calculated to generate coupling effect identifiers. The identifier strength is normalized to a range of [0, 1].

[0142] The identifier strength value is compared against a preset dynamic threshold, derived from empirical parameters in a database of accelerated battery aging tests over the entire lifecycle. When the model outputs an identifier strength greater than 0.85, an automatic downward adjustment mechanism for the compensation factor α is triggered: the α value is reduced by 0.15±0.03 based on the temperature gradient correlation coefficient. This downward adjustment utilizes a sliding window smoothing algorithm to prevent control command oscillation caused by compensation jumps.

[0143] The reduced compensation coefficient α is published via the OPC UA real-time data bus, forcing an update to the charge and discharge constraint parameter table within the optimization engine. This update covers the maximum allowable rate limit and the SOC operating range boundary parameters, for example, adjusting the upper limit of the charge and discharge rate for cells with α less than 0.8 from 1.5C to 1.2C. Constraint update instructions are marked with a high priority and execute before regular policy calculation tasks.

[0144] Specifically, in the multi-type energy storage staged capacity optimization configuration method for new energy uncertainty described in the present invention, when the twin computing device instruction response delay is greater than 200ms, the response hysteresis deviation is marked;

[0145] For areas with deviation values ​​greater than 15%, the optimization strategy package generates sensor supplementary instructions:

[0146] The density of temperature monitoring points is increased to 300% of the original configuration;

[0147] Deploy voltage monitoring modules for battery clusters with capacity decay > 20%.

[0148] The twin uses a high-precision timing synchronization module to record the difference between the dispatch instruction issuance time and the energy storage device's BMS feedback action timestamp, calculating the device's command response delay time series. If the system detects five consecutive response delays exceeding 200ms, it flags the battery cluster area where the device is located as experiencing response lag deviation. This delay data is then linked to the device's location tags to generate a thermal distribution map.

[0149] For spatial grids in areas where the thermal map shows deviations exceeding 15%, the optimization engine uses a geographic information system (GIS) coordinate mapping algorithm to generate a sensor addition plan. The existing density of temperature monitoring points in the area is analyzed to calculate the number of new sensor coordinate points and the required topological configuration. The density increase strategy implements a three-fold expansion rule: increasing the number of monitoring points per cluster from the original three to nine, ensuring uniform coverage of the battery cluster surface area.

[0150] For battery cluster cells in the thermal map showing capacity degradation exceeding 20%, an enhanced voltage monitoring strategy is deployed simultaneously: high-precision voltage sampling circuits are installed at the positive and negative electrodes of each battery cell, increasing sampling accuracy to 0.1mV. The new monitoring module is connected to the edge data collection network via the Modbus RTU protocol, and its address is registered in the device management database. Configuration parameters are written to a JSON-structured instruction packet with version control.

[0151] After digital signature verification, the sensor supplementary instruction packet is published to the target region's edge gateway via the MQTT protocol. The gateway dynamically updates the collection task list: creating a new temperature sensor data collection thread and expanding the voltage monitoring point register address mapping table. Configuration changes take effect immediately, and device status scans are initiated with a microsecond-level polling cycle.

[0152] Specifically, the multi-type energy storage phased capacity optimization configuration method for new energy uncertainty described in the present invention defines α < 0.85 as a high-risk threshold, which is derived from the failure analysis of the 72-hour stress test of the backup system;

[0153] The second algorithm engine enforces the prohibition instruction: refusing to dispatch tasks with a rate greater than 1C to high-risk units;

[0154] The policy instructions dynamically match the health heat map status released by the main system and update the charge and discharge authorization parameters.

[0155] The high-risk threshold of α < 0.85 was established based on the failure analysis report from a 72-hour continuous stress test of the backup system. The test loaded a stepped charge and discharge task flow, recording the inflection point data for the failure probability surge of battery cells with α < 0.85. The risk boundary was then quantified using a lithium dendrite growth acceleration model. This threshold was fixed as a static parameter in the optimization engine configuration center and regularly updated based on new test data.

[0156] The second algorithm engine monitors updates to the compensation coefficient α via the real-time bus. If it detects that the α value falls below the high-risk threshold of 0.85, it immediately triggers a protection mechanism. The engine invokes the task allocation strategy modifier to generate a mandatory prohibition instruction: all charge and discharge task requests with a rate greater than 1C are removed from the scheduled task pool; the task dispatch queue priority is modified, marking the task authorization status of such units as "current limiting mode." This prohibition instruction is tagged with the highest security priority, bypassing the regular task scheduling queue and taking effect directly.

[0157] The policy instruction system synchronizes the health heat map status data published by the main system through an interface. The red and yellow area codes in the heat map map map to the physical unit locations. The authorization parameter updater matches the heat map spatial coordinates with the device identification code and dynamically adjusts three types of charge and discharge authorization parameters: the maximum allowable current drop rate coefficient, the temperature compensation slope correction value, and the SOC operating window compression ratio. The parameter update results are written to the real-time control data warehouse and the dual-machine hot standby is activated.

[0158] This invention solves the scheduling analysis errors caused by real-time energy storage state errors from three dimensions: data source management, dynamic state compensation, and optimized decision adaptation.

[0159] A multi-source physical model verification mechanism is established at the edge: The logical rationality of charge and discharge rates is inverted using a lithium battery double-layer effect model, and the physical boundaries of output data are constrained using photovoltaic IV characteristic curves. Kriging neighbor compensation is used to reconstruct data distribution for anomalies such as temperature sensor failure. User-side load data undergoes layered noise processing to preserve key peak and valley characteristics. High-confidence datasets filter out false status reports caused by hardware distortion, transmission interference, and privacy normalization at the source, providing verifiable input for health assessment and preventing downstream analysis from relying on inherent flaws.

[0160] The collaborative system integrates distributed node parameters based on a federated learning framework and outputs a health distribution map reflecting actual capacity. The memory network simultaneously incorporates the coupled characteristics of temperature drops and high-frequency charging and discharging, outputting a compensation coefficient α to quantify battery degradation in real time. This α value is updated minute-by-minute by the time series alignment module, dynamically compensating for performance drift caused by low-temperature capacity drops and cycle aging in energy storage units. The compensation mechanism converts normalized SOC data into available capacity intervals, correcting the dispatch system's misjudgment of the status of aging units.

[0161] The second algorithm engine establishes a high-risk unit identification rule based on the compensation coefficient α: when α is less than 0.85, a ban on dispatching tasks with a rate greater than 1C is enforced. Strategy instructions dynamically match the spatial coordinates of the health heat map, updating the maximum current drop rate and SOC window boundaries in the charge and discharge authorization parameters. The simulation system injects frequency fluctuation faults to verify the effectiveness of the constraint strategy. Deviations exceeding the threshold trigger reverse optimization of the acquisition strategy. This constraint mechanism ensures that dispatch instructions precisely adapt to the actual health status of the battery, avoiding incorrect load allocation to aging units.

Claims

1. A multi-type energy storage capacity optimization configuration method in stages with new energy uncertainty, characterized by: include: Obtain energy storage device parameters, renewable energy output data, and user load data, verify physical characteristics, and generate a high-confidence data set; Inputting the high-confidence data set into the collaborative system to perform energy storage health assessment and output a health distribution map to provide an input basis for dynamic capacity compensation; Inputting the health distribution map and historical charge and discharge data into a memory network to generate a dynamic compensation coefficient α representing actual capacity attenuation; Inputting the compensation coefficient α and wind / solar output forecast data into the optimization engine, triggering the parallel algorithm to generate a phased capacity optimization plan; Executing the optimization scheme through edge nodes and collecting device response data; Compare the deviation between the twin simulation value and the device response data, and generate feedback signals to update the data acquisition strategy.

2. The method for optimizing the capacity of multi-type energy storage in stages under the uncertainty of new energy according to claim 1 is characterized in that: Obtain energy storage device parameters, renewable energy output data, and user load data, verify physical characteristics, and generate a high-confidence data set, including: The edge device collects the charge and discharge rate matrix, temperature array data, and photovoltaic original characteristic points of the energy storage battery pack through the protocol interface; Push the collected data to the transport layer through the encrypted channel, perform encryption operations and then map it to a unified address space; The transmission data trigger processing layer uses physical models to verify photovoltaic conversion efficiency and electrochemical inversion, and marks abnormal points for data that fails verification; Call the processing module by type for the marked results: Execute interpolation algorithms on interrupted data, call neighboring stations to compensate for abnormal data points or switch to redundant channels; Noise is added to the verified user load data and device parameters are integrated to form a high-confidence data set.

3. The method for optimizing the capacity of multi-type energy storage in stages under the uncertainty of new energy according to claim 2 is characterized in that: Using the high-confidence dataset, the energy storage health is evaluated in the collaborative system, and a health distribution map is output, including: The main system sends the high-confidence data set and the initial neural network model to the distributed nodes; Distributed nodes update model parameters based on local data and send encrypted gradients back to the main system; The simulation system injects pathological scenarios to trigger error verification: if the error is ≥ 5% for three consecutive times, the model is rolled back; otherwise, a verification pass signal is sent to the backup system; After receiving the verification pass signal, the standby system loads the scheduling instruction stream and executes the stress test response; The model that passes the stress test is published to the main system, mapping the normalized SOC data into dynamic capacity intervals; A battery health distribution map is generated based on the dynamic capacity interval.

4. The method for optimizing the capacity of multi-type energy storage in stages under the uncertainty of new energy according to claim 3 is characterized in that: The health distribution map and historical charge and discharge data are input into the memory network to calculate the dynamic capacity compensation coefficient α, including: Real-time collection of ambient temperature and humidity data through the sensor interface, and inputting the data into the time series alignment module together with the health distribution map and historical cycle depth; The sliding window extracts continuous working condition feature vectors from the aligned data according to the preset step size as the input of the memory network; When the feature vector is updated or reaches the time threshold, the memory network loads the pre-trained model inference output compensation coefficient α; The compensation coefficient α value is periodically refreshed every minute and pushed to the real-time interface of the optimization engine.

5. The method for optimizing the capacity of multi-type energy storage in stages with uncertainty of new energy according to claim 4 is characterized in that: The compensation coefficient α and the wind and solar power output forecast are input into the optimization engine, and an optimization solution is generated through a parallel algorithm, including: The first algorithm engine obtains the long-term probability distribution data set of wind and solar power, and generates a multi-stage capacity ratio solution set as a candidate solution; The second algorithm engine receives the compensation coefficient α, the short-term wind and solar power output forecast, and the user load forecast in real time, generates hourly strategy instructions on a rolling basis, and archives them to the strategy library; Input the candidate solutions and rolling strategy instructions into the simulation system, inject the grid frequency fluctuation fault scenario, and quantify the economic benefit deviation and the number of charge and discharge rate limit violations; When the economic indicator difference rate is greater than 15% or the safety indicator deviation is greater than 10%, the manual intervention interface review plan is triggered; After the optimization plan is approved, it will be switched to the production environment through blue-green deployment.

6. The method for optimizing the capacity of multi-type energy storage in stages with uncertainty of new energy according to claim 5 is characterized in that: Executing the optimization scheme while feeding back the deviation signal to improve data acquisition includes: Subscribe to the scheduling instructions corresponding to the optimization plan from the production environment scheduling bus; The edge computing node responds to the scheduling instruction and performs control logic calculation in seconds; Collect actual voltage and current response data of energy storage equipment; The twin is loaded with a typhoon path or extreme temperature scenario library, and combined with the actual response data to simulate the theoretical behavior of the equipment under extreme working conditions; Calculating the root mean square deviation between the theoretical simulation value and the actual response data, and generating an optimization strategy package when the deviation is greater than 5%; The policy package specifies the location of new temperature sensors, the sampling frequency increase multiple, or the expansion of voltage monitoring points, and is pushed to the edge nodes to update the collection policy.

7. The method for optimizing the capacity of multi-type energy storage in stages with uncertainties in new energy sources according to claim 3 is characterized in that: The simulation system injects a capacity drop pathological scenario during the low temperature verification phase and generates an alarm signal when the model predicts a capacity error of ≥5%; The alarm signal triggers the reconfiguration protocol of the data acquisition agent, increasing the sampling frequency of the temperature sensor in the problem area to 10 seconds per time.

8. The method for optimizing the capacity of multi-type energy storage in stages with uncertainty of new energy according to claim 4 is characterized in that: The memory network parses the temporal association between temperature data below -10°C and charge and discharge records above 2C, and outputs a coupling effect identifier; When the identifier strength is greater than the threshold, the compensation coefficient α is automatically adjusted downward by 0.15 ± 0.03; The lowered α value forces the optimization engine to update its charge and discharge constraints through a real-time interface.

9. The method for optimizing the capacity of multi-type energy storage in stages with uncertainty of new energy according to claim 6 is characterized in that: When the twin computing device instruction response delay is greater than 200ms, the response hysteresis deviation is marked; For areas with deviation values ​​greater than 15%, the optimization strategy package generates sensor supplementary instructions: The density of temperature monitoring points is increased to 300% of the original configuration; Deploy voltage monitoring modules for battery clusters with capacity decay > 20%.

10. The method for optimizing capacity allocation of multiple types of energy storage in phases with uncertainty of new energy sources according to claim 5 is characterized in that: The definition of α < 0.85 as a high-risk threshold is derived from the failure analysis of the backup system's 72-hour stress test; The second algorithm engine enforces the prohibition instruction: refusing to dispatch tasks with a rate greater than 1C to high-risk units; The policy instructions dynamically match the health heat map status released by the main system and update the charge and discharge authorization parameters.

Images