An intelligent energy storage system and control method based on AI multi-modal fusion

Abstract

The application relates to the technical field of energy storage, and particularly provides an intelligent energy storage system based on AI multi-modal fusion and a control method, the intelligent energy storage system comprising a household fixed energy storage host, a portable energy storage sub-host and an AI intelligent control unit; the fixed energy storage host and the portable energy storage sub-host are physically locked, energy is bidirectionally circulated, and multi-modal data is synchronously interacted through an intelligent docking station; the AI intelligent control unit realizes integrated intelligent regulation and control of energy scheduling, battery health management, dual-mode cooperation and user behavior adaptation; the control method comprises four stages of data acquisition, model analysis, execution regulation and dynamic iteration; the application has a hardware unit and an AI intelligent control unit, each unit module can cooperatively realize multi-dimensional data acquisition, analysis, decision and execution, is suitable for the cooperative linkage scene of the household fixed energy storage and the portable energy storage, and realizes AI-driven whole-life-cycle energy optimization, fault prediction and scene self-adaptive regulation and control.

Description

Technical Field

[0001] This invention relates to the field of energy storage power technology, and in particular to an intelligent energy storage system and control method based on AI multimodal fusion. Background Technology

[0002] Existing home energy storage systems and portable energy storage power supplies are largely fragmented in their application scenarios. The former focuses on peak-valley arbitrage and emergency power supply for the main power grid, while the latter focuses on outdoor mobile power consumption. Both lack a detachable and interconnected intelligent collaborative mechanism, and their connection structures only allow for basic plug-and-play functionality, lacking anti-detachment, high-protection, bidirectional energy and data interaction designs. This results in repeated investment for users and low energy utilization. Furthermore, existing AI energy storage technologies are mostly limited to single-function dimensions: for example, AI-based charge and discharge protection only achieves basic safety control, voice control only addresses ease of operation, and load forecasting only serves fixed-scenario energy dispatching. None of these technologies employ multi-modal dynamic weight allocation algorithms to achieve adaptive regulation across all scenarios, and they lack a collaborative architecture for rapid edge response and cloud iteration. Moreover, battery health management relies heavily on single-parameter monitoring, lacking multi-dimensional data correlation analysis.

[0003] In addition, existing technologies have three major defects: (1) Battery health management relies on monitoring a single parameter, and cannot achieve aging prediction and cell self-repair through the correlation of six dimensions of data such as voltage, current, temperature, cycle number, vibration and humidity. Moreover, they mostly adopt a single thermal management scheme, which is difficult to adapt to complex scenarios; (2) The emergency power supply mode lacks intelligent switching logic. After the fixed energy storage is exhausted, it is impossible to reverse activate the host and accurately allocate power to key loads through portable energy storage devices; (3) Energy dispatch only considers electricity price and load factors, without integrating the dynamic adaptation of user behavior, environmental parameters and equipment status, and there is no scenario-based differentiated weight control mechanism. The above defects result in insufficient intelligence level and poor scenario adaptability of existing energy storage systems, making it difficult to meet users' energy needs in all scenarios. Summary of the Invention

[0004] To overcome the shortcomings of existing residential fixed energy storage and portable energy storage systems, which are limited by structure and function as described in the background art, this invention provides a dual-mode architecture of "residential fixed energy storage host + portable energy storage slave unit". The two are physically connected, with bidirectional energy flow and data interaction through a dedicated intelligent interface. It has hardware units and AI intelligent control units. Each unit can work together to achieve multi-dimensional data acquisition, analysis, decision-making and execution. It is suitable for the collaborative linkage scenario of residential fixed energy storage and portable energy storage. It realizes an AI-driven full life cycle energy optimization, fault prediction and scenario adaptive regulation. It is an intelligent energy storage system and control method based on AI multimodal fusion.

[0005] The technical solution adopted by this invention to solve its technical problem is: A smart energy storage system based on AI multimodal fusion includes a home fixed energy storage host, a portable energy storage slave unit, and an AI intelligent control unit. The fixed energy storage host and the portable energy storage slave unit are physically locked, with bidirectional energy flow and synchronous multimodal data interaction via an intelligent expansion dock. The AI ​​intelligent control unit adopts an edge-end middle-layer feature alignment + dynamic weight allocation algorithm, integrating multi-dimensional data from voice, sensors, and networks to construct a closed loop of "data acquisition - analysis and decision-making - execution feedback - model iteration," achieving integrated intelligent control of energy dispatch, battery health management, dual-mode collaboration, and user behavior adaptation. The fixed energy storage host and the portable energy storage slave unit form a detachable and interconnectable dual-mode energy storage architecture. The portable energy storage slave unit can independently achieve photovoltaic self-charging and mobile power supply without the fixed energy storage host, making it convenient for users to carry and use when traveling.

[0006] Furthermore, the AI ​​intelligent control unit includes the following five core modules, which form a collaborative closed loop: the multimodal data fusion module cleans and extracts features from the collected heterogeneous data, and generates a standardized data matrix through dynamic weight allocation; the all-scenario adaptive energy scheduling module uses the matrix as input and combines electricity price, light intensity, and user habits to construct a multi-objective optimization model; the battery health full life cycle management module and the dual-mode collaborative emergency activation module output results to the user behavior learning module to iteratively optimize the control strategy. The AI ​​architecture adopts an edge lightweight model + cloud iteration mode, and does not rely on large language models for operation and maintenance dialogue.

[0007] Furthermore, the battery health lifecycle management module constructs an AI health prediction model through correlation analysis of six dimensions of data: voltage, current, temperature, cycle count, vibration, and humidity. This model predicts capacity degradation trends in advance and achieves self-repair by dynamically adjusting the charge and discharge current of individual cells for minor faults such as uneven cell voltage. Combined with the adaptive adjustment strategy of the cell liquid cooling temperature control module, it extends the battery cycle life.

[0008] Furthermore, the home-use fixed energy storage host adopts a high-capacity lithium-ion battery or sodium-ion battery and a high-conversion-efficiency bidirectional inverter, while the portable slave unit is equipped with a foldable photovoltaic charging panel and a Bluetooth / WiFi dual-mode communication module. The two achieve bidirectional energy transfer through an expansion dock, and the slave unit can serve as a backup power source for the host or an independent outdoor power supply unit.

[0009] Furthermore, the intelligent expansion dock has a built-in mechanical locking mechanism with pre-tightening force, IP65-rated waterproof conductive contacts, and a high-speed data transmission interface, supporting plug-and-play fixing and anti-drop-off of the slave unit; when the main unit's power is less than the threshold and the power grid is interrupted, the slave unit activates the main unit's inverter by inputting reverse power through the expansion dock, charging the main unit at a rate of 0.3-0.5C, while automatically identifying and prioritizing power to critical household loads, and temporarily shutting off non-critical loads to extend battery life.

[0010] A smart energy storage control method based on AI multimodal fusion includes the following four-stage closed-loop process: (1) Data acquisition stage: The multimodal data fusion module synchronously collects four-dimensional multimodal data, and generates a standardized matrix after multiple cleaning and alignment with intermediate layer features; (2) Model analysis stage: The multi-objective optimization model is input through the full-scenario adaptive energy scheduling module, and the charging and discharging and linkage strategies are output in combination with electricity price prediction, light intensity, and user behavior habits; (3) Execution and regulation stage: The energy flow and temperature control mode are adjusted by the hardware unit according to the instructions, and the power supply of key loads is prioritized; (4) Dynamic iteration stage: The model parameters are updated in the cloud based on the running data and user behavior feedback by the user behavior learning module, without involving fault prediction logic of factor decomposition or thermal flow field topology analysis.

[0011] Furthermore, the dynamic weight allocation algorithm of the multimodal data fusion module sets differentiated weight thresholds for different scenarios: sensor data weight ratio ≥40% in emergency power supply scenario, user behavior related data weight ratio ≥30% in daily home scenario, and photovoltaic illumination and voice command data weight ratio ≥50% in outdoor mobile scenario. The weights can be dynamically adjusted according to the scenario, and the feature alignment accuracy is ≥95%.

[0012] Furthermore, the user behavior learning module constructs a user habit feature library by collecting user electricity consumption data over several consecutive days. This library covers four dimensions: peak electricity consumption periods, power range of commonly used equipment, frequency and cycle of outdoor electricity consumption, and priority settings for key loads. The feature library is updated every 1-7 days and supports users to manually correct habit prediction results, with a correction command response delay of ≤500ms.

[0013] Furthermore, during the execution and control phase, the battery health management module can execute self-repair adjustment logic. For cell voltage imbalance faults, when the voltage difference of a single cell is ≥50mV, it automatically starts adjustment. Through differential charging and discharging with a micro-current of 0.05-0.1C, the voltage difference is controlled within 20mV, the adjustment time is ≤2 hours, and the adjustment process does not affect the normal power supply of the system. During the execution and control phase, the mechanical locking mechanism of the smart expansion dock can adopt an electromagnetic assisted locking method and is equipped with an unlocking button and an emergency manual unlocking mechanism.

[0014] Furthermore, in the model analysis stage, for the multi-objective optimization model, the weighted summation method is used to solve for the optimal solution. In the emergency scenario, the battery loss weight coefficient is ≤0.2 to prioritize ensuring power supply continuity; in the daily scenario, the energy utilization rate weight coefficient is ≥0.4 to prioritize improving economic efficiency.

[0015] Compared with existing technologies, the beneficial effects of this invention are: (1) Scene integration innovation: It breaks through the problem of scene separation between home and portable energy storage, and achieves full scene energy coverage through dual-mode collaborative architecture, avoiding repeated investment by users and improving energy utilization. Users can purchase only the host and rent one or more sub-units to meet actual needs while reducing one-time investment. (2) AI function upgrade: It adopts multi-modal data fusion technology to break through the limitations of existing single AI functions, realizes integrated intelligent control of energy scheduling, health management and behavior adaptation, and significantly improves the level of intelligence. (3) Battery life extension: Through AI health prediction and self-repair adjustment, it solves the problems of fast aging and difficult fault prediction of existing energy storage batteries, and extends the battery life cycle. (4) Emergency capability enhancement: It innovates the emergency reverse activation mechanism. When the host runs out of power, the sub-unit can quickly activate the system and ensure the power supply of key loads, improving the reliability of emergency power supply. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of an intelligent energy storage system based on AI multimodal fusion according to the present invention. Detailed Implementation

[0017] Figure 1 As shown, an intelligent energy storage system based on AI multimodal fusion includes the following modules: 1. Hardware Unit The hardware unit adopts a dual-mode architecture of "home fixed energy storage main unit + portable energy storage sub-unit". The two are physically connected, enabling bidirectional energy flow and data interaction through a dedicated smart interface, specifically including: (1) Home fixed energy storage host: capacity 5-20kWh, using high specific capacity lithium-ion battery or sodium-ion battery, equipped with bidirectional inverter, liquid cooling temperature control module (adapted to working environment from -20℃ to 60℃), multi-channel sensor group (voltage, current, temperature, humidity, vibration sensor), voice interaction module (noise-canceling microphone and speaker) and smart expansion dock; the expansion dock has built-in electromagnetic auxiliary locking mechanism, IP65 protection level gold-plated waterproof conductive contacts and high-speed data transmission interface, supports portable slave unit locking within 0.5 seconds, plug-in fixed anti-fall, bidirectional energy transmission, multi-modal data synchronization, and is also equipped with unlock button and emergency manual unlocking mechanism.

[0018] (2) Portable energy storage sub-unit: capacity 1-3kWh, weight ≤20kg, equipped with foldable photovoltaic charging panel, Type-C / USB A / AC multi-output interface, miniaturized sensor group, Bluetooth / Wi-Fi dual-mode communication module and emergency lighting module; after the sub-unit is connected to the host expansion dock, it can be used as the host's energy expansion unit or backup power supply. After being separated from the host, it can independently realize outdoor mobile power supply and photovoltaic self-charging.

[0019] (3) Auxiliary hardware: including smart home linkage module (supports communication with home appliances, photovoltaic power generation panels and other devices), mobile terminal APP and cloud server, which can realize remote monitoring, data storage and AI model iteration.

[0020] 2. AI Intelligent Control Unit The AI ​​intelligent control unit is based on a large deep learning model and integrates multimodal data for analysis and decision-making. It includes five core modules, the names and functions of which are as follows: (1) Multimodal data fusion module: synchronously collects voice commands (user operation requirements), visual data (user behavior, equipment status, environmental scene), sensor data (battery parameters, temperature and humidity, vibration, power grid status), and network data (electricity price curve, weather forecast, equipment linkage information). It generates a unified standardized data matrix through data cleaning, feature extraction and dynamic weight allocation algorithms, rather than factor decomposition. Among them, dynamic weight allocation sets differentiated thresholds for different scenarios (sensor data weight ≥ 40% in emergency scenarios, photovoltaic and voice data weight combined ≥ 50% in outdoor scenarios), and feature alignment accuracy ≥ 95%, solving the limitations of single data dimension. The voice module supports dialect recognition and accurate matching of commands in noisy environments. It relies on the edge terminal to achieve fast processing throughout the process, with a response time ≤ 100ms.

[0021] (2) Full-scenario adaptive energy dispatch module: Based on fused data, a multi-objective optimization model is constructed to dynamically adjust the charging and discharging strategy and energy flow direction. Specifically, it includes the following: Peak-valley electricity price optimization: Combined with day-ahead / real-time electricity price forecasts (accuracy ≥ 90%), it automatically charges from the grid during low-price periods and supplies power to household loads or feeds back to the grid during peak periods to maximize user benefits; Photovoltaic collaborative optimization: Based on weather forecasts and real-time light intensity, it adjusts the charging power of photovoltaic panels to prioritize meeting the charging needs of the slave unit, and stores the remaining electricity in the host or directly supplies it to household loads; Accurate load allocation: It identifies critical household loads (emergency lighting, medical equipment, routers) and non-critical loads (fans, entertainment equipment, etc.), prioritizes critical loads during emergency power supply, and dynamically adjusts the power supply power according to the remaining electricity of the slave unit.

[0022] (3) Battery health life cycle management module: Through the correlation analysis of six dimensions of data, including voltage, current, temperature, cycle number, vibration and humidity, an AI health prediction model is constructed to achieve the following three major functions: Aging prediction: Predict the capacity decay trend 3-6 months in advance. When the decay rate exceeds the threshold, the charging and discharging depth and liquid cooling temperature control strategy are automatically adjusted to extend the battery life, which is different from the single thermal flow field coupling early warning scheme; Fault early warning: Identify potential faults such as internal short circuits, poor interface contact, and poor cell consistency. Push early warning information through the APP and start the protection mechanism; Self-repair adjustment: In coordination with the aging prediction function, for minor faults such as cell voltage imbalance (single cell voltage difference ≥ 50mV), self-repair is achieved through differential charging and discharging with a micro current of 0.05-0.1C. The voltage difference is controlled within 20mV within 2 hours. The adjustment process does not affect the normal power supply of the system and avoids the expansion of the fault.

[0023] (4) Dual-mode collaboration and emergency activation module: realizes intelligent linkage between the host and the slave unit, breaks through the limitation of the existing parent-child architecture that can only supply power in one direction. Normal mode: the host prioritizes charging from the grid / photovoltaic, and automatically charges the slave unit after it is fully charged. The slave unit can be plugged in and unplugged at any time for outdoor power use. Emergency mode: when the grid is cut off and the host power is ≤10%, plugging in the fully charged slave unit can reverse activate the host inverter. The slave unit's power is prioritized to supply the critical loads of the household, while the host is slowly charged at a rate of 0.3-0.5C to achieve emergency power supply. Off-grid mode: in the absence of a grid, the slave unit can supply power to the host through the expansion dock after photovoltaic charging to meet the short-term power needs of the household.

[0024] (5) User behavior learning module: Through continuous collection of user electricity consumption data for 7-30 days, a habit feature library covering four dimensions is constructed, including peak electricity consumption periods, power range of commonly used equipment, frequency and cycle of outdoor electricity consumption, and priority settings of key loads. The feature library is updated every 1-7 days and supports users to manually correct habit prediction results (correction command response delay ≤ 500ms). Based on the feature library, the energy scheduling strategy is adaptively optimized. For example, the user's outdoor camping habit on weekends is identified, and the sub-unit is fully charged on a certain day in advance. The user's nighttime air conditioning usage habit is identified, and the nighttime energy storage discharge power is optimized to ensure the continuous operation of the air conditioner.

[0025] Figure 1As shown, an intelligent energy storage control method based on AI multimodal fusion includes the following four-stage closed-loop process: (1) Data acquisition stage: After the system starts, the multimodal data fusion module synchronously collects data from four dimensions: voice, vision, sensing, and network. After cleaning, feature extraction and alignment with the edge-end intermediate layer features (alignment accuracy ≥ 95%), a standardized data matrix is ​​generated through a dynamic weight allocation algorithm. The feature library update cycle is 1-7 days. (2) Model analysis stage: The standardized data matrix is ​​input into the multi-objective optimization model through the full-scenario adaptive energy scheduling module. With "maximizing energy utilization, minimizing battery loss, and maximizing user demand adaptation" as the core objective, the optimal solution is solved by weighted summation method (battery loss weight coefficient ≤ 0.2 in emergency scenarios and energy utilization weight coefficient ≥ 0.4 in daily scenarios). Combined with electricity price prediction (accuracy ≥ 90%), light intensity, and user behavior habit feature library, the charging and discharging strategy, battery status assessment results and user habit adaptation suggestions are output respectively. (3) Execution and control phase: The hardware unit adjusts the charging and discharging status, energy flow direction, temperature control strategy and equipment linkage mode of the host and slave according to the AI ​​decision instructions, prioritizes the power supply of key loads, and provides real-time feedback on the execution results; (4) Dynamic iteration phase: The AI ​​model parameters are updated regularly in the cloud based on the operation data and user behavior feedback through the user behavior learning module, optimizes the analysis accuracy and control strategy, realizes the intelligent self-evolution of the system, and does not involve fault prediction logic of factor decomposition or thermal flow field topology analysis.

[0026] The dynamic weight allocation algorithm of the multimodal data fusion module sets differentiated weight thresholds for different scenarios: sensor data weight ratio ≥40% in emergency power supply scenarios, user behavior-related data weight ratio ≥30% in daily household scenarios, and photovoltaic illumination and voice command data weight ratio ≥50% in outdoor mobile scenarios. The weights can be dynamically adjusted according to the scenario, and the feature alignment accuracy is ≥95%. The user behavior learning module collects user electricity consumption data for 7-30 consecutive days to build a user habit feature library, covering four dimensions: peak electricity consumption periods, power range of commonly used equipment, outdoor electricity consumption frequency and cycle, and priority settings of key loads. The feature library is updated every 1-7 days, and users can manually correct the habit prediction results with a correction command response latency of ≤500ms. During the regulation and control phase, the battery health management module executes self-repair regulation logic. For cell voltage imbalance faults, it automatically initiates regulation when the voltage difference between individual cells is ≥50mV. Through differential charging and discharging with a micro-current of 0.05-0.1C, the voltage difference is controlled within 20mV, with a regulation time ≤2 hours, and the regulation process does not affect the normal power supply of the system. During the regulation and control phase, the intelligent docking station's mechanical locking mechanism uses an electromagnetically assisted locking method, completing locking within 0.5 seconds after the slave unit is inserted, with a locking force ≥50N. It is also equipped with an unlocking button and an emergency manual unlocking mechanism. After unlocking, the slave unit insertion / removal force is ≤15N. The waterproof conductive contacts are gold-plated, with a contact resistance ≤10mΩ, and a data transmission rate ≥1Gbps. During the model analysis phase, for the multi-objective optimization model, the three core objectives are "maximizing energy utilization, minimizing battery loss, and maximizing user demand adaptability". The weighted summation method is used to solve for the optimal solution. In emergency scenarios, the weight coefficient of battery loss is ≤0.2 to prioritize ensuring power supply continuity; in daily scenarios, the weight coefficient of energy utilization is ≥0.4 to prioritize improving economic efficiency.

[0027] The matrix factorization-free multimodal data fusion calculation process during the data acquisition phase is as follows: 1. Definition and Parameter Configuration 1.1 Definition of Three Types of Multimodal Raw Input Channels Total raw dimensions: 6D sensing + 3D voice + 4D network = 13D single-frame raw vector

[0028] Channel 1: Battery Six-Dimensional Sensing Data S1=[U,I,T,Cyc,Hum,Vib] (U: cell voltage, I: charge / discharge current, T: cell temperature, Cyc: cumulative cycle count, Hum: internal humidity of the equipment, Vib: vibration amplitude of the enclosure) Channel 2: Data collection via voice interaction S2=[VoiceSNR,CmdType,Urgency] (VoiceSNR: Voice Signal-to-Noise Ratio, CmdType: Command Encoding, Urgency: Power Demand Urgency) Channel 3: Network Weather and Electricity Price Data S3=[ElecPrice,Irradiance,Forecast,LoadPlan] (ElecPrice: real-time electricity price, Irradiance: photovoltaic irradiance, Forecast: irradiance forecast factor, LoadPlan: household preset load capacity) 1.2 Baseline Weight Pool for Three Types of Operating Scenarios The weight vector Wbase = [w1, w2, w3] corresponds to the sensing, voice, and network channels respectively, and satisfies the quantization constraints: Scenario A (Daily Home Use Mode): Wbase, A = [0.35, 0.15, 0.50] Scenario B (Power outage emergency mode, sensor weight ≥ 40%): Wbase, B = [0.50, 0.10, 0.40] Scenario C (Outdoor sub-unit operates independently, total weight of voice + network ≥ 50%): Wbase, C = [0.25, 0.30, 0.45] Phase 2, Stage 1: Step-by-step process for multimodal data cleaning The five-step fixed process is executed sequentially: boundary threshold truncation → linear imputation of missing values ​​→ impulse transition filtering → three-point moving average noise reduction → channel confidence marking. 2.1 Step 1: Truncation of physical threshold boundaries for each channel 2.1.1 Calculation Formula

[0029] 2.1.2 Threshold Comparison Table for Each Channel Dimension Table 1: All-Dimensional Hardware Physical Upper and Lower Limits

[0030] 2.1.3 Execution Logic The original sampled values ​​are traversed channel by channel and dimension by dimension, and values ​​exceeding the limit are directly truncated to the boundary; no sampled frames are discarded throughout the process to ensure temporal continuity; a truncation vector is output. .

[0031] 2.2 Step 2: Linear imputation of missing values ​​within the temporal neighborhood 2.2.1 Triggering conditions: Sensor disconnection, radio failure, cloud network delay generates null / NaN flags.

[0032] 2.2.2 Fallback Rule (1) When the system just started, there were not enough frames in the cache, so the hardware default value was used to fill the cache. (2) If all 5 consecutive frames in the same channel are missing, it is determined to be a hardware failure, and the subsequent confidence coefficient is forcibly assigned a value of 0.8; the output vector after padding is then used. .

[0033] 2.3 Step 3: Electromagnetic Pulse Abnormality Filtering 2.3.1 Sources of Interference The switching on and off of inverters and relays generates instantaneous electromagnetic spikes, which have no practical significance in terms of physical operating conditions.

[0034] Set a global threshold Δth; for slow changes in cell temperature T and cycle count Cyc, set a dedicated threshold Δth.

[0035] 2.3.2 Processing Logic Δ > Δth: Identify impulse interference and replace the current value with the stable value of the previous frame; Δ ≤ Δth: Normal operating condition fluctuations, values ​​are retained; Output jump filter vector .

[0036] 2.4 Step 4: Three-point moving average time-domain noise reduction Execution rules: Fixed-window 3-frame elimination of AD and high-frequency audio noise; when the buffer has less than 3 frames, average the existing valid frames without waiting for the buffer to fill up; output a clean basic vector. .

[0037] 2.5 Step 5: Channel Reliability Coefficient Marking Count the number of anomaly handling operations (truncation, padding, and jump corrections) across all dimensions of a single channel, and generate a confidence correction coefficient. : (1) Level 1: No abnormal corrections, data is completely reliable, rk=1.2 (2) Level 2: Only a small number of values ​​are truncated, with no missing values ​​or consecutive jumps, rk=1.0 (3) Level 3: Missing padding / continuous pulse jumps exist, data distortion exists, rk=0.8 Output the confidence vector r = [r1, r2, r3], and pass it directly to the dynamic weight allocation stage.

[0038] Phase 3, Stage 2: Sub-channel feature extraction and cross-channel feature alignment (without matrix factorization) Objective: To map three types of heterogeneous data to the value range [0, 1], with a mandatory feature alignment accuracy of ≥95%, and to perform independent channel operations throughout the process without cross-channel matrix decomposition or dimensionality reduction.

[0039] 3.1 Step 1: Channel Splitting After splitting and cleaning, the vector consists of three independent sub-vectors:

[0040] 3.2 Step 2: Independent Min-Max Normalization within Channels

[0041] k=1,2,3 Channel number, i is the dimension number within channel i.

[0042] 3.3 Step 3: Channel-level feature aggregation The arithmetic mean of all normalized features within a channel is taken and compressed into a single-channel global representation value:

[0043] d1=6, d2=3, d3=4; Output the three-dimensional basic feature vector. .

[0044] 3.4 Step 4: Feature Alignment Accuracy Verification Value range deviation calculation:

[0045] Alignment precision:

[0046] Correction rules: When the accuracy is less than 95%, fine-tune the normalization upper and lower limits of linear correction for each channel, and prohibit matrix decomposition and dimensionality reduction throughout the process until the accuracy meets the standard.

[0047] Phase 4, Stage 3: Dynamic weight allocation for different scenarios, generating a unified and standardized data matrix. 4.1 Step 1: Automatic determination of system operating scenario (1) Power grid disconnection + host SOC ≤ 10% → Scenario B (emergency), retrieve

[0048] (2) Portable sub-unit disconnected from the expansion dock and connected to the photovoltaic interface → Scenario C (outdoor), retrieve

[0049] (3) Other operating conditions → Scene A (daily household use), retrieve

[0050] 4.2 Step 2: Dynamic Adjustment of Benchmark Weight Credibility

[0051] 4.3 Step 3: Weight normalization (sum of weights for all three channels = 1)

[0052] Output the final dynamic weight vector .

[0053] 4.4 Step 4: Multi-channel weighted fusion to generate a single-frame global feature scalar

[0054] Standardized fusion feature value at time t.

[0055] 4.5 Step 5: Construct a standardized data matrix by temporal stacking. A fixed-length temporal buffer at the edge continuously stores the fused features of M frames, which are then vertically stacked to generate the input matrix for the downstream AI model.

[0056] To unify and standardize the data matrix.

[0057] Specifically, the innovations of the matrix factorization-free multimodal data fusion technology of this invention and its differences from existing technologies are as follows: (1) Hardware simplification innovation: removing visual acquisition hardware and relying only on three low-cost data sources, namely sensing, voice and network, to achieve multimodal AI fusion, which is different from the energy storage intelligent system with built-in cameras in the industry; (2) Algorithm path innovation: avoiding all matrix factorization in the entire link, using only lightweight single-channel linear operation, which can be deployed on low-cost 8 / 32-bit MCUs without cloud computing power support; (3) Link linkage innovation: outputting the credibility coefficient in the data cleaning process and automatically and dynamically weakening the weight of the distortion channel, while the existing technology has independent cleaning and weighting processes with no linkage correction mechanism; (4) Quantitative constraint innovation: solidifying the hard boundaries of the weights of the three types of scenarios, with the sensor weight ≥40% in emergency mode and the voice + network weight ≥50% in outdoor mode, and the algorithm logic is quantifiable and reproducible; (5) Precision quantification innovation: setting a hard index of feature alignment accuracy ≥95%, solving the defects of chaotic value range and large model prediction deviation in traditional multimodal fusion.

[0058] The calculation process of the multi-objective optimization model for the full-scenario adaptive energy dispatch module in the model analysis phase is as follows: Prerequisites 1. Input: The standardized data matrix Mstd output from the previous multimodal fusion (generated by weighted fusion of sensor, speech, and network modalities without decomposition). 2. Core Constraints: The optimization aims to achieve three main objectives: ① maximize the economic benefits of energy utilization, ② minimize battery aging and loss, and ③ maximize the satisfaction of users' electricity needs. Differentiated weighting coefficients for different scenarios: Battery loss weight for emergency scenarios ≤ 0.2, and daily household energy utilization rate weight ≥ 0.4; Output control quantities: main unit charging and discharging power, portable slave unit charging and discharging power, critical load on / off control signal, photovoltaic charging power adjustment value; 3. Operating hardware: Energy storage edge MCU, linear lightweight solution, no complex iterative optimization, single frame solution time <2ms, meeting the requirement of total system response ≤100ms; 4. Current technology status: Conventional multi-objective optimization in the industry uses iterative optimization methods such as particle swarm optimization and genetic algorithms. This technology uses weighted linear summation analytical solutions, which do not require iteration and are suitable for low-cost edge hardware.

[0059] I. Definition of Basic Variables 1.1 System Control Decision Variables (Data to be solved is shown in the table below) Table 2

[0060] 1.2 Multimodal fusion input features (from (Analysis and extraction) Real-time electricity price normalization characteristics (network channel) Normalized characteristics of photovoltaic irradiance (network channel) : Remaining SOC (Sensor Channel) of the host device : Remaining SOC (State of Charge) of the portable device Cell temperature normalization characteristics (sensor channel); Fuser: User power demand urgency (voice channel) 1.3 Target Weight Allocation by Scenario Let the three-dimensional target weight vector be...

[0061] Economic benefit weighting Battery loss weighting User needs satisfaction weight satisfy

[0062] Scenario A: Daily home use mode

[0063] Scenario B: Power outage emergency mode Scenario C: Outdoor portable sub-unit operates independently.

[0064] 1.4 Three types of single-objective quantization functions (all mapped to the [0,1] interval, with unified dimensions) Objective 1: Economic benefit function J1 (the larger the value, the higher the benefit).

[0065] Physical logic: Discharge more when electricity prices are high and charge more when electricity prices are low; prioritize self-consumption of photovoltaic power when there is sufficient sunlight to reduce electricity purchase costs.

[0066] Objective 2: Battery aging loss function J2 (the smaller the value, the lower the battery loss).

[0067] Physical logic: High temperature, high-power charging and discharging, and SOC deviating from the 50% range will all accelerate cell degradation.

[0068] Objective 3: User needs satisfaction function J3 (the larger the value, the higher the degree of user needs matching).

[0069] Physical logic: The higher the urgency of the user's voice command, the more loads will maintain power supply; in emergency scenarios, priority will be given to critical loads such as medical devices and refrigerators.

[0070] 1.5 Comprehensive Multi-Objective Evaluation Function (Weighted Linear Summation, the Core Non-Iterative Solution Scheme in this Embodiment) The loss objective is mapped inversely to unify the maximization optimization direction:

[0071] The core objective of the scheduling model is to maximize the overall objective function under all hardware, power, and load constraints.

[0072] II. System Operation Constraints (Solve Boundary Conditions, which cannot be exceeded) Constraint 1: Boundary constraint of energy storage capacity SOC

[0073] Δt: Single-frame scheduling control period (100ms). Host capacity, Sub-machine capacity, typically ranging from 20% to 90% of SOC.

[0074] Constraint 2: Upper limit constraint on power hardware

[0075] Constraint 3: Load priority constraint (mandatory in emergency scenarios)

[0076] Critical loads in medical / refrigeration / router applications. Lighting, entertainment, and other general loads are forcibly disconnected in emergency mode. .

[0077] Constraint 4: Two-way energy balance constraint Photovoltaic output + Energy storage discharge = Energy storage charging + Total load consumption

[0078] III. Complete Serial Calculation Process of Multi-Objective Optimization Model Step 1: Scene recognition, loading the corresponding target weight vector λ (1) Read the system operating condition markers and determine the current operating scenario (daily / emergency / outdoor); (2) Retrieve the fixed weights for this scene ,satisfy ; (3) Verification constraints: Automatic locking after emergency scenario determination. everyday scenarios locked It cannot be modified by human intervention.

[0079] Step 2: From the standardized fusion matrix Extract normalized feature parameters, read the latest frame fusion features from the temporal fusion matrix, and decompose them to obtain 6 types of core input features: Read the latest frame fused features from the temporal fusion matrix and decompose them to obtain 6 core input feature classes:

[0080] All eigenvalues ​​are unified to [0, 1], eliminating the differences in the dimensions of electricity price, temperature, and power.

[0081] Step 3: Hierarchical constraint preprocessing to narrow the feasible region of solution variables. (1) SOC boundary filtering: Eliminate power combinations that will lead to overcharging or over-discharging of the battery; (2) Emergency load classification processing: In scenario B, all non-critical loads are directly set to 0 to reduce the number of variables to be solved; (3) Photovoltaic power limiting: based on real-time sunlight. Locking in the maximum output limit of photovoltaics; Function: Significantly reduces the computational load on the MCU, enabling millisecond-level solutions.

[0082] Step 4: Construct single-objective component functions J1, J2, J3 Substitute the current feature parameters and calculate the three indicators of economic benefits, battery loss, and user demand satisfaction respectively, and output them all to the interval [0, 1].

[0083] Step 5: Obtain the comprehensive evaluation function by weighted fusion.

[0084] Perform linear weighted calculation: This step is the core difference of this invention: it only involves linear arithmetic operations, without PCA / factor decomposition, and without particle swarm optimization / genetic iterative optimization, which is different from the multi-objective optimization algorithms used in the energy storage industry.

[0085] Step 6: Solve the optimal combination of control variables using linear analytical methods Based on the energy balance equation constraint, a piecewise traversal analytical method (lightweight and non-iterative) is adopted: (1) Traverse the feasible discrete power levels of the host's charging and discharging capabilities; (2) Matching portable energy storage charging and discharging, and photovoltaic power regulation; (3) Traverse the load on / off combinations; (4) Calculate the corresponding values ​​for each group of variables. Record the power and load switch combination corresponding to the maximum value; (5) Output the globally optimal set of control variables as the scheduling output instruction.

[0086] Step 7: Control command amplitude limiting and secondary safety verification (1) Verify whether the optimal power exceeds the maximum hardware power and whether it triggers SOC out-of-bounds operation; (2) The charging and discharging power of the battery cell is automatically reduced under high temperature conditions to reduce battery loss J2; (3) Verify the power supply continuity of critical loads; medical and refrigeration loads cannot be disconnected in emergency mode. If the verification fails, the suboptimal solution will be automatically replaced.

[0087] Step 8: Output scheduling and control commands and send feedback back to the next higher-level AI unit. Output complete control instruction set: (1) The setpoint for the charging / discharging power of the main unit; (2) Portable handset charging / discharging power setpoint; (3) Photovoltaic DC-DC power regulation; (4) On / off signals of each load switch; Simultaneously, the output of this scheduling and the current comprehensive target value will be included. The data is packaged and sent back to the multimodal fusion module for weight correction of the next frame's data fusion, forming a global closed-loop control.

[0088] The innovation of the multi-objective optimization model of the full-scenario adaptive energy dispatch module and its differences from existing technologies: (1) Multi-objective linear weighted non-iterative solution: Existing multi-objective optimization of energy storage generally adopts a large number of iterations such as genetic algorithm, particle swarm optimization, and simulated annealing, which consumes a lot of computing power and can only run in the cloud; This solution is purely linear arithmetic operation, without matrix decomposition and iteration, and can run in real time on the local edge MCU with a response speed of ≤100ms. (2) Hard weight boundary constraints for different scenarios: Fixed upper and lower limits of weights for emergency, daily and outdoor scenarios, forming a quantifiable and reproducible dispatch logic. Existing technologies have no fixed weight constraints, the adaptive logic is fuzzy, and they do not have patent verifiability. (3) Strong closed-loop linkage with front-end multimodal fusion: The dispatch result is fed back to the data fusion module to correct the dynamic weight allocation of the next frame, forming a complete closed loop of "collection-fusion-dispatch-feedback"; Traditional energy dispatch modules are independent units without data layer linkage correction mechanisms. (4) Load grading and differentiated constraint mechanism: In emergency scenarios, non-critical loads are automatically and forcibly cut off, and priority is given to ensuring the power supply of medical and refrigeration core loads. This is adapted to the emergency core needs of household energy storage. The existing multi-objective model does not distinguish the load priority hierarchical constraint. (5) Feature unified normalization: All input features are uniformly output to [0, 1] by the non-decomposed multimodal fusion matrix. No separate unit conversion is required, which reduces the model coupling complexity and is different from the scheduling scheme that processes multi-source data separately.

[0089] The complete calculation process of the six-dimensional AI health prediction model for the battery health lifecycle management module during the regulation and control phase is as follows: Prerequisites 1. Module positioning Input: Cleaned and feature-extracted six-dimensional battery sensing standardized features (cell voltage U, charge / discharge current I, cell temperature T, cumulative cycle count Cyc, enclosure vibration Vib, internal humidity Hum); Output: (1) Current SOH health status value of the battery; (2) Predicted capacity decay trend value in the next 3 to 6 months; (3) Cell imbalance fault judgment signal; (4) Adaptive charge and discharge regulation control command.

[0090] 2. Core Innovation (1) A six-dimensional coupled linear AI fitting model is adopted, without using PCA / SVD matrix factorization for dimensionality reduction, matching the multimodal fusion unified non-decomposition algorithm link mentioned above; (2) Six-parameter cross-coupled damage factor calculation, which is different from the industry's three-dimensional independent evaluation scheme of voltage / current / temperature; (3) Layered two-level prediction: real-time current SOH assessment + medium- and long-term capacity decay timeline prediction, which can provide early warning of aging 3 to 6 months in advance; (4) Built-in cell voltage difference self-repair judgment logic, linked to the energy storage main control to dynamically fine-tune the micro-current balance of the individual cells; (5) All operations are lightweight arithmetic operations, calculated locally by the edge MCU, with a single frame calculation time of <1ms and a total response time of ≤100ms.

[0091] 3. Input variable definition (all are normalized features after the aforementioned data cleaning output, with a value range of [0,1]) Normalized single-cell voltage characteristics Normalized charge and discharge current characteristics Normalized cell temperature characteristics Normalized loop count feature Normalized box vibration amplitude characteristics Normalized internal humidity characteristics I. Preset of Basic Damage Weights and Coupling Coefficients (Reproducible) 1.1 Six-Dimensional Basic Damage Weights As shown in Table 3 below Table 3

[0092] constraint:

[0093] 1.2 Six-dimensional cross-coupling damage coefficient matrix (no matrix factorization, only pairwise coupling operations) Vibration-humidity coupling Temperature-current coupling Voltage-cyclic coupling

[0094] High value: High temperature + high current combined, damage amplified. High value: Frequent high and low voltage charge and discharge cycles combined with multiple cycles accelerate the attenuation. High values: High humidity combined with continuous vibration exacerbates corrosion of metal components. For the remaining dimensions, the pairwise coupling coefficient is 1.0 (no superposition of damage amplification). II. Calculation Process Phase 1: Calculation of basic damage in a single dimension For the six normalized features, the single-dimensional independent damage value Dx is calculated respectively. The damage value range is [0, 1]. The larger the value, the more severe the aging damage caused by that dimension.

[0095]

[0096] Phase 2: Calculation of the total real-time damage value by superimposing six-dimensional cross-coupling. Three sets of key coupling amplification coefficients are introduced to superimpose the additional damage caused by the synchronous deterioration of multiple factors, thus obtaining the real-time comprehensive damage value.

[0097] Phase 3: Real-time battery current health status (SOH) calculation The overall damage is mapped to the standard SOH range (energy storage battery benchmark: new battery SOH=100%, end-of-life threshold SOH=80%). Constraint verification: (1) If Forced bandwidth limitation Brand new battery, undamaged (2) If If the battery is determined to have reached the scrap warning threshold, a first-level fault alarm will be pushed to the host computer.

[0098] Phase 4: Building a time-series cache and constructing an AI decay trend fitting dataset. A fixed-time buffer is allocated at the edge to continuously store the time-series data of the most recent N frames: Cache array

[0099] The corresponding sampling time interval Δt (single frame sampling period 100ms) is stored synchronously.

[0100] Unlike existing technologies, it does not require massive cloud-based historical datasets; lightweight linear trend fitting can be achieved with local short-term time-series caching, without the need for training complex deep learning networks.

[0101] Phase 5: AI linear trend fitting, medium- to long-term capacity decay prediction (3-6 months) 5.1 Calculate the slope of damage change (attenuation rate) ) We employ a lightweight least-squares linear fitting method, solving only the first-order slope, without matrix inversion or factorization.

[0102] Unit: Damage value / frame This indicates that the battery is aging at an accelerated rate.

[0103] 5.2 Projecting SOH for the next 3 and 6 months Based on the monthly equivalent total number of sampling frames decay rate Real-time comprehensive damage value Estimated comprehensive damage value after 3 months of parameter calculation Estimated comprehensive damage value after 6 months The long-term health status is obtained by mapping: , .

[0104] 5.3 Rules for Determining Early Warning of Aging Classification (1) Level II warning (warning 3 months in advance): The app sends an aging and degradation warning and automatically reduces the depth of charge and discharge. (2) Level 1 scrapping warning: If the battery is determined to reach the end-of-life standard within 3 months, high-power discharge will be restricted, and a replacement reminder will be sent.

[0105] Phase 6: Cell Voltage Imbalance Fault Identification and Self-Repair Trigger Calculation (1) Read the normalized voltage difference of each individual cell. ; (2) Threshold determination: when (For actual voltage difference ≥ 50mV), initiate the cell micro-current self-repair balancing program; when (For actual pressure difference ≤ 20mV), balancing is complete, and the repair mode is exited.

[0106] (3) Calculation of equalization microcurrent adjustment: Output differentiated micro-charge and discharge current commands of 0.05~0.1C and send them to the battery management unit (BMS).

[0107] Phase 7: Output multiple types of control signals and provide closed-loop feedback to the energy storage scheduling module. The module uniformly outputs four types of results and synchronously uploads them to the multi-modal fusion and energy scheduling module. (1) Real-time indicators: Comprehensive damage value ; (2) Medium- to long-term forecasts: Aging warning level; (3) BMS control command: cell equalization microcurrent Charging and discharging power limit thresholds; (4) Fault reporting signals: high temperature / high humidity / vibration superimposed damage alarm, scrap warning, differential pressure fault signal.

[0108] Note: The innovation of the six-dimensional AI health prediction model of the battery health full life cycle management module and its difference from existing technologies: (1) Six-dimensional fully coupled damage modeling, breaking through the limitations of the three-dimensional traditional assessment. Existing battery health assessment only uses three-dimensional parameters of voltage, current and temperature, ignoring the coupled damage of vibration, humidity and cycle number; this solution synchronously correlates six parameters, accurately captures the degradation characteristics of energy storage batteries in outdoor, humid and vibration scenarios, and is suitable for the use environment of home + portable dual-mode energy storage. (2) Avoid matrix factorization throughout the process and unify the architecture with the multimodal algorithm mentioned above. The mainstream SOH model in the industry uses PCA dimensionality reduction and SVD feature decomposition to process multidimensional data. This model only uses linear weighting and pairwise coupling superposition, without any matrix decomposition operation. The entire AI system algorithm link maintains a unified innovative feature. (3) Local short-term time series linear fitting, no need for cloud big data training. Conventional AI battery health models rely on massive historical samples in the cloud to train neural networks. This module only needs a local short-term cache of 90 frames to complete the fitting of the degradation slope. The edge MCU runs independently and can accurately predict the capacity degradation in 3 to 6 months even when the network is disconnected or offline. (4) Damage coupling layering mechanism to quantify the accelerated aging effect of multiple harsh working conditions. Innovative setting of three sets of coupling amplification coefficients of temperature-current, voltage-cycle, vibration-humidity, etc., accurately reproduces the accelerated degradation law of battery under composite harsh working conditions such as high temperature and high current, high humidity and vibration. Existing technology does not have a multi-factor superposition damage quantification mechanism. (5) Health prediction directly links BMS self-repair and upper-level energy scheduling closed-loop model output. Not only does it provide fault warning, but it also outputs cell equalization micro-current and power limit instructions in real time, and synchronously feeds them back to the full-scenario energy scheduling module. According to SOH, the charging and discharging strategy is dynamically adjusted to form a complete closed loop of "sensor acquisition-health assessment-scheduling optimization equalization repair". Existing battery management modules run independently and do not participate in upper-level energy scheduling linkage. Specific Implementation Example 1: Building a Home Intelligent Energy Management System This embodiment establishes a smart energy storage system consisting of a 10kWh home-use fixed energy storage main unit and two 2kWh portable slave units, with the following specific configuration: Hardware configuration: The main unit uses lithium iron phosphate battery cells with a cycle life of 10,000 cycles (70% DOD), a bidirectional inverter conversion efficiency of 96.5%, and a liquid-cooled temperature control module that supports operating environments from -20℃ to 60℃; the slave unit weighs 15kg, is equipped with a 200W portable folding photovoltaic panel, and supports AC / DC multi-interface output; the expansion dock has a built-in electromagnetic auxiliary locking mechanism (locking force ≥50N, contact resistance ≤10mΩ) and IP65 waterproof contacts, a data transmission rate ≥1Gbps, and completes locking within 0.5 seconds after the slave unit is inserted, with an unlocking insertion / removal force ≤15N.

[0110] AI module configuration: It adopts an edge computing + cloud collaborative architecture, deploying lightweight deep learning models at the edge to achieve real-time data processing and decision-making (response time ≤100ms); the cloud connects to big data analysis models and regularly updates electricity price forecasts, weather forecasts and user behavior model parameters.

[0111] Operational Process: After system startup, the multimodal module collects data such as grid electricity price, photovoltaic irradiance, indoor temperature and humidity, and the status of user electrical equipment. The AI ​​energy scheduling module generates optimization strategies. During off-peak hours (0-7 AM), the main unit charges from the grid to 80% capacity. From 8 AM to 5 PM, the photovoltaic panels charge the slave unit, and the remaining power is used for household appliances. During peak electricity consumption from 6 PM to 10 PM, the main unit supplies power to household loads, while the AI ​​behavior learning module reserves power for nighttime lighting equipment based on user habits.

[0112] Example 2: Emergency Reverse Activation Function Test Simulating a power outage scenario, the main unit starts with 10% battery. After the main unit is fully charged (100%), the system activates emergency mode: the slave unit supplies power to the main unit's inverter via the docking station, activating the main unit's output function and prioritizing power to emergency lighting, refrigerators, routers, and other appliances, maintaining a stable power output of around 200W; simultaneously, the main unit is charged at a 0.3C rate. The AI ​​module monitors the battery level in real time, and when the slave unit's battery level drops to 20%, it automatically cuts off the main unit's charging circuit, maintaining power only for critical loads, with a continuous power supply time of up to 8 hours to meet household emergency needs.

[0113] The following are application examples of this invention: (1) Home intelligent energy management scenario: The user's home is equipped with this system and distributed photovoltaic panels. The AI ​​module automatically formulates optimization strategies by analyzing weather forecasts, electricity price curves and user electricity consumption habits. During the off-peak hours of weekday electricity prices, the main unit charges from the grid; during the day when there is sufficient sunlight, the photovoltaic panels prioritize charging the sub-unit, and the remaining electricity is used for the household load. Excess electricity is fed back to the grid to obtain revenue; during the peak hours of evening electricity consumption, the main unit supplies electricity to the household load to avoid peak electricity price expenditures. At the same time, the AI ​​vision module automatically shuts down non-critical loads after the user leaves home to reduce energy consumption. (2) Outdoor camping multi-device power supply scenario: The user takes the portable sub-unit to outdoor camping on the weekend. The sub-unit charges itself by folding the photovoltaic panel. The AI ​​module adjusts the charging power according to the light intensity and identifies the connected camping equipment (drone, projector, electric oven) and dynamically allocates the output power to avoid overload. During the camping process, the user can control the sub-unit to turn on and off and adjust the output voltage through voice commands. The AI ​​module supports dialect recognition and can still respond accurately in noisy environments. After camping, the slave unit is plugged into the host expansion dock, and the host automatically recharges the slave unit for the next use. (3) Emergency power supply scenario for sudden power outage: A typhoon causes the regional power grid to be interrupted, and the host's power gradually depletes to 8% (emergency threshold). The system automatically issues an early warning. The user uses a fully charged slave unit delivered by the rescue team. After plugging it into the host expansion dock, the slave unit immediately activates the host inverter in reverse, prioritizing power supply to critical loads such as refrigerators, medical ventilators, and routers, while charging the host at a 0.5C rate. The AI ​​module monitors the power of the slave unit and the host in real time. When the slave unit's power is ≤20%, it automatically switches to power only critical loads to ensure the core needs continue until the power grid is restored. (4) Continuous power supply scenario for home medical equipment: There are patients with chronic diseases at home who need to rely on medical equipment such as oxygen concentrators and blood glucose meters for continuous power supply. The AI ​​module of this system identifies medical equipment as critical loads and monitors the equipment's operating status and battery health in real time. When the power grid is interrupted, the emergency power supply mode is immediately activated, and the host and slave unit work together to ensure that the medical equipment does not stop operating. Meanwhile, the AI ​​health module predicts the battery status in real time. If the battery capacity is predicted to be insufficient, it will push a warning message through the APP in advance to remind users to prepare a backup device to avoid the risk of medical equipment shutdown. (5) RV mobile energy supply scenario: The RV is equipped with this system. The main unit is fixedly installed at the bottom of the RV, and the slave unit can be flexibly plugged in and out. During the journey, the RV generator charges the main unit. The AI ​​module adjusts the charging current according to the generator speed and battery status to avoid overcharging the battery. After parking at the campsite, the slave unit can be taken out for outdoor barbecue and lighting power supply, while the main unit supplies power to the RV air conditioner and refrigerator. The AI ​​module dynamically allocates energy by analyzing the RV's power load and battery power. At the same time, it optimizes the charging and discharging timing by combining the campsite electricity price (if applicable) to reduce the energy cost of RV travel. (6) Remote area household off-grid power supply scenario: Remote mountain households have no stable power grid coverage and are equipped with this system and high-power photovoltaic panels.During the day, the photovoltaic panels charge the main unit and the slave unit, and the AI ​​module adjusts the charging strategy according to changes in sunlight to maximize the use of solar energy. At night, the main unit supplies power to household loads, while the slave unit serves as a backup power source, automatically replenishing power when the main unit's battery is low. The AI ​​health module monitors the battery status continuously and automatically adjusts the liquid cooling temperature control strategy for environments with large temperature differences in mountainous areas, preventing the battery from aging prematurely due to temperature fluctuations and ensuring long-term stable operation of the system.

[0114] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of the invention and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of the invention should be included within the protection scope of the invention. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. An intelligent energy storage system based on AI multimodal fusion, comprising a residential fixed energy storage host, a portable energy storage slave unit, and an AI intelligent control unit; Its features are, The fixed energy storage host and the portable energy storage slave unit are physically locked, have bidirectional energy flow, and can synchronously interact with multimodal data through an intelligent expansion dock; The AI ​​intelligent control unit adopts an edge-end intermediate layer feature alignment + dynamic weight allocation algorithm, integrates multi-dimensional data from voice, sensors, and networks, and constructs a closed loop of "data acquisition - analysis and decision-making - execution feedback - model iteration" to achieve integrated intelligent control of energy scheduling, battery health management, dual-mode collaboration, and user behavior adaptation. The fixed energy storage host and the portable energy storage sub-unit form a detachable and interconnectable dual-mode energy storage architecture. The portable energy storage sub-unit can be independently charged by photovoltaics and powered by mobile devices when separated from the fixed energy storage host, making it convenient for users to carry and use when they go out.

2. The intelligent energy storage system based on AI multimodal fusion according to claim 1, characterized in that, The AI ​​intelligent control unit comprises the following five core modules, which form a collaborative closed loop: The multimodal data fusion module cleans and extracts features from the collected heterogeneous data, and then generates a standardized data matrix through dynamic weight allocation; The full-scenario adaptive energy dispatch module uses a matrix as input and combines electricity price, light intensity, and user habits to construct a multi-objective optimization model. The battery health lifecycle management module and the dual-mode collaborative emergency activation module output results to the user behavior learning module, iteratively optimizing the control strategy. The AI ​​architecture adopts an edge lightweight model + cloud iteration mode, and does not rely on a large language model for operation and maintenance dialogue.

3. The intelligent energy storage system based on AI multimodal fusion according to claim 1, characterized in that, The battery health lifecycle management module uses six dimensions of data correlation analysis—voltage, current, temperature, cycle count, vibration, and humidity—to build an AI health prediction model that predicts capacity degradation trends in advance. For minor faults such as uneven cell voltage, it achieves self-repair by dynamically adjusting the charge and discharge current of individual cells. Combined with the adaptive adjustment strategy of the cell liquid cooling temperature control module, it extends the battery cycle life.

4. The intelligent energy storage system based on AI multimodal fusion according to claim 1, characterized in that, The home-use fixed energy storage main unit uses high-capacity lithium-ion or sodium-ion batteries and a high-conversion-efficiency bidirectional inverter. The portable slave unit is equipped with a foldable photovoltaic charging panel and a Bluetooth / WiFi dual-mode communication module. The two units achieve bidirectional energy transfer through an expansion dock. The slave unit can serve as a backup power source for the main unit or an independent outdoor power supply unit.

5. The intelligent energy storage system based on AI multimodal fusion according to claim 1, characterized in that, The smart docking station features a built-in mechanical locking mechanism with pre-tightening force, IP65-rated waterproof conductive contacts, and a high-speed data transmission interface, supporting plug-and-play fixing and anti-drop-off of the slave unit. When the main unit's battery level is below the threshold and the power grid is interrupted, the slave unit activates the main unit's inverter by inputting reverse power through the docking station, charging the main unit at a rate of 0.3-0.5C. At the same time, it automatically identifies and prioritizes powering critical loads in the home, while temporarily shutting off non-critical loads to extend battery life.

6. A smart energy storage system based on AI multimodal fusion according to any one of claims 1 to 5, characterized in that... The control method includes the following four-stage closed-loop process. (1) Data acquisition stage: The multimodal data fusion module synchronously acquires four-dimensional multimodal data, and generates a standardized matrix after multiple cleaning and alignment with intermediate layer features; (2) Model analysis stage: The multi-objective optimization model is input through the full-scenario adaptive energy dispatch module, and the charging and discharging and linkage strategies are output in combination with electricity price prediction, light intensity and user behavior habits; (3) Execution and control phase: The hardware unit adjusts the energy flow and temperature control mode according to the instructions, and prioritizes the power supply to critical loads. (4) Dynamic iteration stage: The user behavior learning module updates the model parameters in the cloud based on the running data and user behavior feedback, without involving the fault prediction logic of factor decomposition or thermal flow field topology analysis.

7. The intelligent energy storage control method based on AI multimodal fusion according to claim 6, characterized in that, The dynamic weight allocation algorithm of the multimodal data fusion module sets differentiated weight thresholds for different scenarios: sensor data weight ratio ≥40% in emergency power supply scenario, user behavior related data weight ratio ≥30% in daily household scenario, and photovoltaic illumination and voice command data weight ratio ≥50% in outdoor mobile scenario. The weights can be dynamically adjusted according to the scenario, and the feature alignment accuracy is ≥95%.

8. The intelligent energy storage control method based on AI multimodal fusion according to claim 6, characterized in that, The user behavior learning module collects user electricity consumption data over several consecutive days to build a user habit feature library, covering four dimensions: peak electricity consumption periods, power range of commonly used equipment, frequency and cycle of outdoor electricity consumption, and priority settings of key loads. The feature library is updated every 1-7 days and supports users to manually correct habit prediction results, with a correction command response latency of ≤500ms.

9. The intelligent energy storage control method based on AI multimodal fusion according to claim 6, characterized in that, During the control phase, the battery health management module can execute self-repair adjustment logic. For cell voltage imbalance faults, when the voltage difference of a single cell is ≥50mV, it will automatically start adjustment. Through differential charging and discharging with microcurrent of 0.05-0.1C, the voltage difference will be controlled within 20mV, the adjustment time will be ≤2 hours, and the normal power supply of the system will not be affected during the adjustment process. During the control phase, the mechanical locking mechanism of the smart expansion dock can adopt electromagnetic assisted locking method and is equipped with an unlocking button and an emergency manual unlocking mechanism.

10. The intelligent energy storage control method based on AI multimodal fusion according to claim 6, characterized in that, During the model analysis phase, for the multi-objective optimization model, the weighted summation method is used to solve for the optimal solution. In the emergency scenario, the battery loss weight coefficient is ≤0.2, prioritizing the protection of power supply continuity. In everyday scenarios, the energy utilization rate weighting coefficient should be ≥0.4, prioritizing the improvement of economic efficiency.

Images