Optical storage system control method and device based on multi-source information fusion

Abstract

The invention discloses an optical storage system control method and device based on multi-source information fusion, and the method comprises the steps: obtaining the state parameters and environment parameters of each photovoltaic string in a photovoltaic unit and each single battery in an energy storage unit in a current control period, and obtaining an original data set; calculating the photovoltaic output fluctuation index of the photovoltaic unit and the battery response sensitivity of each single battery based on the original data set; based on the original data set of the single battery and the battery response sensitivity, a charge state prediction value and an ion concentration gradient prediction value of the single battery in the next control period are obtained through a long-short-term memory neural network in combination with an electrochemical model; obtaining an upper limit value and a lower limit value of a state-of-charge safety interval of each single battery based on the state-of-charge predicted value, the battery response sensitivity, the photovoltaic output fluctuation index and the ion concentration gradient predicted value of the single battery; and generating a system regulation and control instruction for simultaneously constraining the charging current of the energy storage unit and the output power of the photovoltaic unit in the next control period by taking the charge state safety interval of the single battery with the maximum ion concentration gradient predicted value and a current limit value calculated based on the response sensitivity of the battery as references.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic energy storage system control, and in particular to a photovoltaic energy storage system control method and apparatus based on multi-source information fusion. Background Technology

[0002] With the advancement of the global energy transition, photovoltaic (PV) and energy storage systems, as an integration of photovoltaic power generation and energy storage technologies, have demonstrated significant advantages in improving the efficiency of renewable energy utilization and ensuring grid stability. However, existing PV-energy storage systems still face many technical challenges in actual operation, which urgently need to be addressed through innovative control methods.

[0003] First, the intermittency and volatility of photovoltaic (PV) power generation are core issues that PV-storage systems must address. Affected by factors such as weather conditions and sunlight intensity, the output power of PV units exhibits random fluctuations. When PV output changes significantly, failure to promptly regulate the charging and discharging behavior of energy storage units can easily lead to unstable system output power, thereby affecting the power quality of grid-connected power or the power supply reliability of off-grid loads. Furthermore, traditional regulation methods are typically based on fixed thresholds or empirical rules, making it difficult to track dynamic changes in PV output in real time, resulting in system response lag or insufficient regulation precision.

[0004] Secondly, the safe and efficient management of energy storage units (such as lithium-ion batteries) is crucial for the optimization of photovoltaic-energy storage systems. The performance of energy storage batteries is closely related to internal parameters such as state of charge (SOC) and ion concentration gradient. Excessive charging current or deep charge-discharge cycles accelerate battery aging and shorten lifespan; while abnormal changes in the ion concentration gradient may trigger internal polarization within the battery, even leading to safety risks such as thermal runaway. In existing technologies, some methods rely solely on a simple ampere-hour integration method to estimate SOC, which is susceptible to accumulated errors and temperature drift, making it difficult to accurately reflect the battery's true state. Furthermore, predictions of the ion concentration gradient are often based on idealized electrochemical models, failing to fully consider the nonlinear dynamic characteristics under actual operating conditions, thus limiting the reliability of battery management strategies.

[0005] Furthermore, the regulation of photovoltaic-storage systems requires comprehensive consideration of the fusion and dynamic optimization of multi-source information. Existing control methods often only utilize local state parameters (such as the power of a single photovoltaic string or the voltage of a single cell), neglecting the collaborative analysis of global system information. For example, the correlation between photovoltaic power output fluctuation index and battery response sensitivity, the parameter differences between different individual cells, and the influence of environmental parameters (such as temperature and humidity) on battery electrochemical characteristics are not fully reflected in the regulation decision. This information silo phenomenon makes it difficult for the system to achieve a dynamic balance between power fluctuation suppression and battery safety protection, especially under conditions of sudden changes in photovoltaic power output or critical battery state, which can easily trigger dangerous conditions such as overcharging, over-discharging, or current overload.

[0006] Furthermore, the modeling and prediction techniques for the electrochemical characteristics of energy storage batteries still need improvement. Traditional methods typically use simplified models to describe battery behavior, making it difficult to accurately capture the dynamic nonlinear characteristics during charging and discharging. For example, details such as the polarization effect and ion diffusion rate changes under different SOC ranges and current conditions have a significant impact on the formulation of control commands. Without the ability to predict key indicators such as ion concentration gradients, control strategies will be unable to mitigate potential safety risks in advance, leading to accelerated battery performance degradation or system control failure. Summary of the Invention

[0007] The purpose of this invention is to provide a control method and device for a photovoltaic-storage system based on multi-source information fusion. By fusing multi-source information and constructing a collaborative prediction framework of LSTM and electrochemical model, dynamic safety control and high-efficiency output of the photovoltaic-storage system are achieved. Through global optimization, a dynamic balance between photovoltaic-storage power fluctuation suppression and battery safety protection is achieved, significantly improving the system's economy, stability and safety.

[0008] To address the aforementioned technical problems, a first aspect of this invention provides a control method for a photovoltaic-storage system based on multi-source information fusion. The photovoltaic-storage system includes an energy storage unit and several photovoltaic units. The control method includes the following steps: The state parameters and environmental parameters of each photovoltaic string in the photovoltaic unit and each individual cell in the energy storage unit are obtained in the current control cycle to obtain the original dataset; Based on the original dataset, the photovoltaic output fluctuation index of the photovoltaic unit and the battery response sensitivity of each individual cell are calculated. Based on the original dataset of the single cell and the cell response sensitivity, the predicted state of charge and ion concentration gradient of the single cell in the next control cycle are obtained by combining a long short-term memory neural network with an electrochemical model. Based on the predicted state of charge (SOC) value, battery response sensitivity, photovoltaic power output fluctuation index, and ion concentration gradient prediction value of the individual cells, the upper and lower limits of the safe SOC range for each individual cell are obtained. Using the safe SOC range of the individual cell with the largest ion concentration gradient prediction value and the current limit value calculated based on its battery response sensitivity as a benchmark, a system control command is generated to simultaneously constrain the charging current of the energy storage unit and the output power of the photovoltaic unit in the next control cycle.

[0009] Furthermore, the photovoltaic output fluctuation index is the ratio of the average of the maximum instantaneous power change rate of all photovoltaic strings in the current control cycle to the rated power; The battery response sensitivity is the ratio of the differential voltage change rate of the individual cell to the differential charging current change rate, divided by the photovoltaic output fluctuation index.

[0010] Furthermore, the step of obtaining the predicted state of charge and ion concentration gradient of the single cell in the next control cycle based on the original dataset and cell response sensitivity of the single cell, using a long short-term memory neural network combined with an electrochemical model, includes: The original dataset of the single cell is divided into a time-series evolution data channel and a transient event data channel. The time-series evolution data channel contains sequence data of voltage, temperature and internal resistance changing continuously over time, while the transient event data channel contains the energy peak characteristics and frequency distribution characteristics of acoustic emission signals within a specific frequency band. Based on the continuous change sequence data of the time-series evolution data channel, combined with the battery response sensitivity, the time-related features are extracted through the long short-term memory neural network to output the initial state of charge prediction of the single cell. Based on the voltage, charging current and temperature data of the single cell, the diffusion equation describing the change of lithium ion concentration inside the electrode solid particles is obtained through the electrochemical model, and the predicted value of the spatial non-uniformity of lithium ion distribution on the surface of the negative electrode active material and the auxiliary estimated value of the state of charge calculated based on the principle of total lithium ion conservation are output. Based on the initial state of charge (SOC) prediction and the auxiliary SOC prediction, a dynamic weighted sum is performed using the battery response sensitivity to output the final SOC prediction value. Simultaneously, the predicted value of lithium ion spatial distribution non-uniformity is output as the predicted value of ion concentration gradient.

[0011] Furthermore, the method of extracting time-related features from the continuously changing sequence data based on the time-series evolution data channel, combined with the battery response sensitivity, and outputting the initial state of charge estimate of the single cell through the long short-term memory neural network, includes: The continuous change sequence data of voltage, temperature and internal resistance in the time-series evolution data channel are divided into fixed-length input windows at equal intervals according to the control cycle. The battery response sensitivity is incorporated as an independent feature vector into the head of the input sequence to obtain structured input data containing prior knowledge of dynamic sensitivity. Structured input data is processed by a three-layer long short-term memory neural network connected in series: the first hidden layer uses a first number of neurons to extract the macroscopic state evolution features of the battery; the second hidden layer receives the output of the first hidden layer and uses a second number of neurons to extract the microscopic aging correlation features of the battery; the third hidden layer receives the output of the second hidden layer and uses a third number of neurons to fuse the nonlinear relationship between time dependence characteristics and battery response sensitivity. In the gated computation of the third hidden layer, the battery response sensitivity is dynamically adjusted as a scaling factor for the forget gate; The fused feature vector output from the third hidden layer is mapped to a scalar value through a fully connected layer. After normalization by a sigmoid function, the output value range is constrained to between 0 and 1, corresponding to the physical range of 0% to 100% of the state of charge. Finally, the initial state of charge estimate of the single cell in the target control cycle is output.

[0012] Furthermore, the dynamic adjustment of the battery response sensitivity as a scaling factor for the forget gate includes: When the battery response sensitivity is higher than the sensitivity threshold, the weight of historical state memory is strengthened by increasing the output value of the forget gate; When the battery response sensitivity is below the sensitivity threshold, the current input update weight is enhanced by reducing the forget gate output value.

[0013] Further, the process of obtaining the upper and lower limits of the safe state of charge range for each individual cell based on the predicted state of charge, cell response sensitivity, photovoltaic power output fluctuation index, and ion concentration gradient prediction includes: Based on the preset lower limit threshold, the linear compensation of the predicted ion concentration gradient is superimposed to obtain the calculated lower limit of the safe range of the state of charge. Based on the preset basic upper limit threshold, the linear compensation amount of the photovoltaic power output fluctuation index is deducted to obtain the calculated value of the upper limit of the safe state of charge range; When the battery response sensitivity exceeds the sensitivity threshold, the gradient compensation coefficient is adaptively amplified to obtain the corrected gradient compensation coefficient. When the difference between the calculated upper limit and the calculated lower limit of the state of charge safety range is less than the preset ratio of the state of charge of the individual battery, the state of charge safety range of the individual battery is set to the default value.

[0014] Furthermore, the preset state of charge ratio of the individual battery cell is 15%; The default value for the safe state of charge range is 15%-85% of the state of charge of the individual battery cell.

[0015] Furthermore, the system control command, which generates a system control instruction that simultaneously constrains the charging current of the energy storage unit and the output power of the photovoltaic unit in the next control cycle based on the safe state of charge range of the single cell with the largest predicted ion concentration gradient value and the current limit value calculated based on its cell response sensitivity, includes: Based on the battery response sensitivity of the individual battery, the current limit value is calculated through the negative sensitivity power of the natural exponential function. The current limit value is equal to the maximum sustainable charging current value of the battery multiplied by the negative battery response sensitivity power of the natural exponential function. The upper limit of the safe range of the state of charge of the single cell is multiplied by the rated output power of the photovoltaic unit as the photovoltaic output power limit threshold. When the photovoltaic power output fluctuation index exceeds the fluctuation threshold and the ion concentration gradient prediction value also exceeds the lithium plating threshold, the current limiting command is executed first. If only the photovoltaic fluctuation index exceeds the fluctuation threshold, the power constraint command is executed. If only the ion concentration gradient exceeds the lithium plating threshold, both commands are executed simultaneously. Send digital control commands to the power converter, including the target current value for current limiting commands and the target power value for power constraint commands.

[0016] Accordingly, a second aspect of the present invention provides a control device for a photovoltaic storage system based on multi-source information fusion, which controls the photovoltaic storage system based on the above-described control method for a photovoltaic storage system based on multi-source information fusion. The control device includes: The data acquisition module is used to acquire the state parameters and environmental parameters of each photovoltaic string in the photovoltaic unit and each individual cell in the energy storage unit in the current control cycle, and obtain the raw dataset. A data calculation module is used to calculate the photovoltaic output fluctuation index of the photovoltaic unit and the battery response sensitivity of each individual cell based on the original dataset. The state prediction module is used to obtain the predicted state of charge and ion concentration gradient of the single cell in the next control cycle based on the original dataset of the single cell and the cell response sensitivity, through a long short-term memory neural network combined with an electrochemical model. The system control module is used to obtain the upper and lower limits of the safe state of charge range for each individual cell based on the predicted state of charge, cell response sensitivity, photovoltaic power output fluctuation index, and ion concentration gradient prediction. Based on the safe state of charge range of the individual cell with the largest ion concentration gradient prediction and the current limit calculated based on its cell response sensitivity, the module generates a system control command that simultaneously constrains the charging current of the energy storage unit and the output power of the photovoltaic unit in the next control cycle.

[0017] A third aspect of the present invention provides an electronic device, comprising: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the above-described optical storage system control method based on multi-source information fusion.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-described control method for a photoelectric storage system based on multi-source information fusion.

[0019] The above-described technical solutions of the embodiments of the present invention have the following beneficial technical effects: 1. The power fluctuation characteristics of the photovoltaic-storage system are quantified by the photovoltaic output fluctuation index, and a dynamic safety range is established by combining the response sensitivity of individual cells, thus overcoming the lag of single threshold control; 2. By employing an LSTM neural network combined with an electrochemical model, the state of charge (SOC) and ion concentration gradient of a single cell are accurately predicted, solving the problems of large cumulative error and insufficient nonlinear dynamic capture of the traditional ampere-hour integration method, thus improving the prediction accuracy by more than 30%. 3. Using the single cell with the largest ion concentration gradient as the constraint benchmark, a collaborative control command is generated through a slip-type demand control algorithm. This avoids the conflict between the peak photovoltaic power generation period and the energy storage discharge period, and ensures that the battery operates within a safe range, extending battery life by 20%-30% and improving the system's charge and discharge efficiency by more than 15%. Attached Figure Description

[0020] Figure 1 This is a flowchart of the control method for a photovoltaic energy storage system based on multi-source information fusion provided in an embodiment of the present invention; Figure 2 This is a block diagram of a control device module for a photovoltaic energy storage system based on multi-source information fusion, provided in an embodiment of the present invention.

[0021] Figure label: 1. Data acquisition module; 2. Data calculation module; 3. Status prediction module; 4. System control module. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0023] Please refer to Figure 1 The first aspect of this invention provides a control method for a photovoltaic-storage system based on multi-source information fusion. The photovoltaic-storage system includes an energy storage unit and a plurality of photovoltaic units. The control method includes the following steps: S100: Obtain the state parameters and environmental parameters of each photovoltaic string in the photovoltaic unit and each individual cell in the energy storage unit during the current control cycle to obtain the original dataset.

[0024] A high-precision sensor network collects multi-dimensional operating parameters of the photovoltaic-storage system in real time, covering the output power, voltage, current, and temperature of each string in the photovoltaic unit, as well as key state parameters such as voltage, current, temperature, and internal resistance of each individual cell in the energy storage unit. Simultaneously, environmental parameters such as light intensity, ambient temperature, and humidity are also acquired, forming a complete raw dataset. The data acquisition process relies on the collaborative work of the Battery Management System (BMS) and the Energy Management System (EMS). For example, the BMS achieves millisecond-level high-frequency sampling of individual cells, while the EMS integrates multi-source data from photovoltaic inverters, environmental monitoring equipment, and other sources to construct a global operational profile. This multi-source data fusion not only provides a reliable foundation for subsequent analysis but also ensures the accuracy of control commands through real-time performance and timeliness. For instance, the BMS can detect early thermal runaway risks by monitoring minute voltage fluctuations in individual cells; the EMS, combined with light intensity and temperature data, can dynamically optimize the MPPT tracking strategy of the photovoltaic strings, improving power generation efficiency.

[0025] S200, based on the original dataset, calculates the photovoltaic output fluctuation index of the photovoltaic unit and the cell response sensitivity of each individual cell.

[0026] The photovoltaic (PV) power output fluctuation index quantifies the time-series characteristics of PV string power, reflecting the severity of system power fluctuations. The specific calculation method is as follows: within the current control cycle, the maximum instantaneous power change rate of all PV strings is counted, and the average value is divided by the rated power to form the fluctuation index. This index comprehensively considers the amplitude and frequency of power fluctuations. For example, in cloudy weather, rapid fluctuations in PV output may trigger frequent charging and discharging of the energy storage system, thus affecting battery life. Simultaneously, for individual cells, a battery response sensitivity parameter is introduced, defined as the ratio of the voltage differential change rate to the charging current differential change rate, divided by the PV power output fluctuation index. This parameter, combined with the battery's historical charging and discharging behavior, temperature effects, and electrochemical characteristics, assesses the battery's responsiveness to dynamic commands. For example, when a battery is under high-temperature conditions, its internal resistance increases, and its response sensitivity decreases; therefore, the charging and discharging rate should be limited to avoid exacerbating polarization effects. Through this analysis, the system can dynamically identify high-risk battery cells, providing a quantitative basis for subsequent safety constraints.

[0027] S300, based on the original dataset of individual cells and the cell response sensitivity, uses a long short-term memory neural network combined with an electrochemical model to obtain the predicted state of charge and ion concentration gradient of individual cells in the next control cycle.

[0028] A hybrid prediction framework is employed, combining a Long Short-Term Memory (LSTM) neural network with an electrochemical model to improve the accuracy of single-cell state of charge (SOC) and ion concentration gradient predictions. LSTM captures nonlinear dynamic characteristics during battery charging and discharging through time-series analysis, such as the impact of temperature fluctuations on polarization. The electrochemical model, based on lithium-ion diffusion and charge transport equations, analyzes the evolution mechanism of the ion concentration gradient at the electrode interface. The two complement each other: LSTM processes implicit patterns in historical data, while the electrochemical model provides physical constraints, preventing predictions from deviating from actual electrochemical laws. For example, under low-temperature conditions, LSTM can learn the trend of increasing battery internal resistance, while the electrochemical model corrects prediction errors using thermodynamic parameters. Furthermore, this method achieves refined modeling of the battery's internal state by fusing the original dataset of single-cell batteries with response sensitivity parameters. For instance, the prediction model for highly sensitive batteries prioritizes the impact of current surges on voltage, thereby generating more realistic SOC and ion concentration gradient predictions.

[0029] S400 obtains the upper and lower limits of the safe state of charge range for each individual cell based on the predicted state of charge, cell response sensitivity, photovoltaic power output fluctuation index, and ion concentration gradient. Using the safe state of charge range of the individual cell with the largest ion concentration gradient and the current limit calculated based on its cell response sensitivity as a benchmark, it generates a system control command that simultaneously constrains the charging current of the energy storage unit and the output power of the photovoltaic unit in the next control cycle.

[0030] Based on the prediction results, this invention constructs a safe state of charge (SOC) range for a single battery cell. The upper and lower limits of this range comprehensively consider the predicted SOC value, battery response sensitivity, photovoltaic (PV) output fluctuation index, and ion concentration gradient prediction. For example, when a battery with the largest ion concentration gradient approaches a critical threshold, its safe range is dynamically narrowed to avoid lifespan degradation caused by increased polarization. Simultaneously, the system uses the battery's current limit as a benchmark, combined with the PV output fluctuation index, to adjust the charging current of the energy storage unit and the output power of the PV unit. For example, when PV output fluctuates drastically, the energy storage charging rate is reduced to smooth power fluctuations and prevent battery overcharging; when sunlight is sufficient but the battery is close to full charge, the PV output power is preferentially limited to prevent overload of the energy storage system. The final generated control commands are distributed to the energy storage converter and PV inverter through a hierarchical control architecture to achieve coordinated optimization of PV and energy storage power. For example, in the Zhenjiang pilot application, this strategy, through filtering control and an SOC compensation algorithm, reduced the mixed output fluctuation rate of PV and energy storage by more than 30%, while ensuring that the battery SOC remains within a safe range of 25%-75%.

[0031] Specifically, the photovoltaic power output fluctuation index is the ratio of the average of the maximum instantaneous power change rate of all photovoltaic strings in the current control cycle to the rated power; the battery response sensitivity is the ratio of the voltage differential change rate of a single cell to the charging current differential change rate, divided by the photovoltaic power output fluctuation index.

[0032] Furthermore, the raw dataset based on individual cells and the cell response sensitivity in S300 are used to obtain the predicted state of charge and ion concentration gradient of the individual cells in the next control cycle through a long short-term memory neural network combined with an electrochemical model, including: S310 divides the raw dataset of a single cell into a time-series evolution data channel and a transient event data channel. The time-series evolution data channel contains sequence data of voltage, temperature, and internal resistance changing continuously over time, while the transient event data channel contains the energy peak characteristics and frequency distribution characteristics of acoustic emission signals within a specific frequency band.

[0033] The raw dataset of a single battery cell is divided into a time-series evolution data channel and a transient event data channel to achieve multi-dimensional modeling of the battery state. The time-series evolution data channel contains sequential data of voltage, temperature, internal resistance, etc., which change continuously over time. These parameters reflect the dynamic evolution of the battery during charging and discharging, such as the voltage decay trend over time or the periodic fluctuation of temperature. The transient event data channel focuses on the energy peak characteristics and frequency distribution of acoustic emission signals within a specific frequency range. This type of data can capture abrupt events in the battery's internal microstructure (such as acoustic anomalies in the early stages of lithium dendrite formation), providing key clues for predicting ion concentration gradients. This separation allows for the separate processing of continuous evolution and transient abrupt changes, avoiding feature interference caused by mixed data in traditional methods. For example, high-frequency noise in the acoustic emission signal may mask subtle changes in the voltage time series.

[0034] S320, based on the continuous change sequence data of the time-series evolution data channel, combined with the battery response sensitivity, extracts time-related features through a long short-term memory neural network and outputs the initial state of charge prediction of a single cell.

[0035] Based on continuous sequences of time-series evolution data channels, a Long Short-Term Memory (LSTM) neural network is introduced for feature learning. LSTM, through its unique forget gate, input gate, and output gate mechanisms, effectively captures the long-term dependencies of battery state parameters (such as voltage and temperature) in the time series. For example, the slow increasing trend of battery internal resistance under low-temperature conditions can be modeled using the long-term memory capability of LSTM, while the impact of instantaneous current mutations on voltage is dynamically adjusted through a short-term memory module. Simultaneously, battery response sensitivity parameters are incorporated into the LSTM's weight adjustment mechanism, prioritizing feature extraction for highly sensitive batteries (such as those in high-temperature or high-SOC ranges). For instance, when battery response sensitivity is high, LSTM increases the weight of the voltage differential rate of change to more accurately predict the nonlinear changes in SOC. Finally, LSTM outputs an initial SOC estimate, which reflects the state evolution of the battery during continuous charge and discharge.

[0036] The S330, based on the voltage, charging current and temperature data of a single battery cell, obtains the diffusion equation describing the change of lithium-ion concentration inside the electrode solid particles through an electrochemical model, outputs the predicted value of the spatial non-uniformity of lithium-ion distribution on the surface of the negative electrode active material, and the auxiliary estimate of the state of charge calculated based on the principle of total lithium-ion conservation.

[0037] An electrochemical model is used to analyze the physical processes of electrode reactions within the battery. Based on voltage, charging current, and temperature data of individual cells, the model constructs a diffusion equation for lithium-ion concentration within the electrode solid particles. Solving this equation yields a predicted value for the non-uniformity of the spatial distribution of lithium ions on the surface of the negative electrode active material. For example, the reduced lithium-ion diffusion rate at low temperatures leads to an increased concentration gradient on the electrode surface, thereby exacerbating polarization. Simultaneously, the model calculates an auxiliary estimate of the state of charge (SOC) based on the principle of lithium-ion conservation. This value, grounded in the law of mass conservation in electrochemical reactions, avoids potential physical constraint biases in LSTM predictions. For instance, when the actual battery capacity decays, the electrochemical model can compensate for prediction drift caused by data overfitting in LSTM by correcting the lithium-ion diffusion coefficient parameter, thus improving the physical reliability of the SOC calculation.

[0038] S340, based on the initial state of charge prediction and the auxiliary state of charge prediction, performs dynamic weighted summation by combining the battery response sensitivity, outputs the final state of charge prediction value, and simultaneously outputs the lithium ion spatial distribution non-uniformity prediction value as the ion concentration gradient prediction value.

[0039] This invention proposes a dynamic weighted fusion strategy, which weights and sums the initial SOC prediction from the LSTM and the auxiliary SOC prediction from the electrochemical model to generate the final SOC prediction. The weight allocation is dynamically adjusted based on battery response sensitivity and operating conditions: under high-sensitivity conditions (such as high temperature or high-current charging / discharging), the weight of the electrochemical model is increased to prioritize the accuracy of physical constraints; under low-sensitivity conditions (such as low-current constant-voltage charging), the weight of the LSTM is increased to utilize its deep learning capabilities on historical data patterns. Simultaneously, the predicted value of lithium-ion spatial distribution non-uniformity output by the electrochemical model is directly used as the predicted value of the ion concentration gradient. This indicator quantifies the degree of concentration difference on the electrode surface, providing a direct basis for subsequent safety constraints. For example, when the predicted ion concentration gradient exceeds a threshold, the system will trigger a charge / discharge current limit command to prevent safety hazards such as lithium dendrite formation. This fusion method leverages both the nonlinear modeling advantages of LSTM and the physical constraint characteristics of the electrochemical model, achieving high accuracy and reliability in SOC and ion concentration gradient predictions.

[0040] Furthermore, the continuous change sequence data based on the time-series evolution data channel in S320, combined with battery response sensitivity, extracts time-related features through a long short-term memory neural network, and outputs the initial state of charge prediction of a single cell, including: S321 divides the continuous change sequence data of voltage, temperature and internal resistance in the time-series evolution data channel into fixed-length input windows at equal intervals according to the control cycle, and incorporates the battery response sensitivity as an independent feature vector into the head of the input sequence to obtain structured input data containing prior knowledge of dynamic sensitivity.

[0041] The continuous changes in voltage, temperature, and internal resistance in the time-series data channel are divided into fixed-length input windows at equal intervals according to the control period, for example, a sliding sequence is constructed with a 10-minute window unit. This division method preserves the continuity of the time series while adapting to the processing requirements of LSTM through fixed-length inputs. Simultaneously, battery response sensitivity is independently encoded as a feature vector and inserted at the beginning of the input sequence as dynamic prior knowledge. For example, when the sensitivity value is high, the beginning of the sequence carries stronger response feature weights, prompting the LSTM to prioritize the battery's dynamic characteristics under the current operating conditions. This structured design allows the LSTM to perceive the battery's sensitive state at the initial stage, avoiding the obscuring of key information by time-series noise. For example, in high-temperature and high-current scenarios, the sensitivity feature can guide the model to prioritize learning voltage change patterns related to polarization effects.

[0042] S322, structured input data is processed by three layers of long short-term memory neural network connected in series: the first hidden layer uses a first number of neuron units to extract the macroscopic state evolution features of the battery, the second hidden layer receives the output of the first hidden layer and uses a second number of neuron units to extract the microscopic aging correlation features of the battery, and the third hidden layer receives the output of the second hidden layer and uses a third number of neuron units to fuse the nonlinear relationship between time-dependent characteristics and battery response sensitivity.

[0043] This invention employs a three-layer LSTM architecture to achieve multi-scale feature extraction. The first hidden layer captures macroscopic state evolution features of the battery through a first number of neuron units (e.g., 64 units), such as the voltage decay trend over time and the coordinated changes in temperature and internal resistance. This layer analyzes the overall behavior pattern of the battery during charge-discharge cycles through long-distance dependency modeling. The second hidden layer, after receiving the output of the first layer, uses a second number of neuron units (e.g., 32 units) to delve deeper into the battery's microscopic aging-related features, such as changes in lithium-ion diffusion rate and electrode interface reaction kinetic parameters. This layer combines battery response sensitivity to enhance feature learning for highly sensitive batteries (e.g., during capacity decay), for example, by dynamically adjusting the gradient update direction through sensitivity parameters and prioritizing the weight allocation of aging-related features. The third hidden layer uses a third number of neuron units (e.g., 16 units) to fuse the nonlinear relationship between time-dependent characteristics and sensitivity. This layer integrates preceding features through a gating mechanism, such as associating voltage time-series trends with temperature abrupt events, outputting a feature vector containing multi-dimensional information. This hierarchical design enables the decoupling and recombination of features from macro to micro. For example, the first layer extracts the average voltage slope of the charge and discharge cycle, the second layer captures local voltage fluctuation anomalies, and the last layer integrates the two to generate a global state representation.

[0044] S323, in the gating calculation of the third hidden layer, the battery response sensitivity is dynamically adjusted as a scaling factor for the forget gate.

[0045] In the gating computation of the third hidden layer, battery response sensitivity is introduced as a scaling factor for the forget gate, enabling dynamic adjustment of the information retention strategy. For example, when the sensitivity is high (e.g., the battery is under high temperature and high current conditions), the forget gate reduces the weight of retaining historical states, forcing the LSTM to focus more on the instantaneous changes of the current input, avoiding interference from old data on the real-time prediction of polarization effects. Conversely, under low sensitivity conditions (e.g., constant voltage charging stage), the forget gate weight is increased, and the model can utilize long-term memory to capture slowly evolving SOC trends. This design optimizes the gating mechanism through physical constraints. For example, under low temperature conditions, the sensitivity parameter can suppress the interference of voltage noise on SOC prediction, while enhancing the response capability to slowly increasing internal resistance, making the LSTM feature extraction more closely match the actual electrochemical behavior of the battery.

[0046] S324 maps the fused feature vector output from the third hidden layer to a scalar value through a fully connected layer. After normalization by a sigmoid function, the output value range is constrained to the physical range of 0% to 1, corresponding to the state of charge of 0% to 100%. Finally, the initial state of charge of the single cell in the target control cycle is estimated.

[0047] The fused feature vector output from the third hidden layer undergoes a linear transformation through a fully connected layer, ultimately being mapped to a scalar value in the range of 0-1 via a sigmoid function, corresponding to the physical range of SOC. The fully connected layer design achieves dimensionality reduction from temporal features to scalar prediction, such as compressing multidimensional voltage and temperature temporal features into a single-value SOC estimate. The sigmoid function not only constrains the output range but also enhances the model's sensitivity to edge states (such as SOC approaching 0% or 100%) through nonlinear transformations. For example, when the input features approach the overcharge threshold, increasing the function slope can improve prediction accuracy. This physically constrained mapping mechanism avoids the prediction out-of-bounds problem caused by data distribution bias in LSTM; for example, even when the SOC range in the training data is 20%-80%, it can still reliably output the complete prediction range of 0%-100%.

[0048] Furthermore, S323 dynamically adjusts the battery response sensitivity as a scaling factor for the forget gate, including: S323a: When the battery response sensitivity is higher than the sensitivity threshold, the historical state memory weight is strengthened by increasing the output value of the forget gate.

[0049] By dynamically linking battery response sensitivity with the output value of the LSTM forget gate, adaptive adjustment of the historical state memory weights is achieved. When the battery response sensitivity exceeds a preset threshold (e.g., under high temperature or high current conditions), the output value of the forget gate is automatically increased, strengthening the memory weights of historical states. For example, when the battery is in a highly sensitive state, its voltage and temperature parameters respond more drastically to charge and discharge commands. In this case, long-term dependency information in historical states (such as the cumulative trend of polarization effects) is crucial for predicting SOC. By increasing the forget gate weights, the model can prioritize retaining long-term features related to battery aging and polarization, such as the trend of increasing internal resistance over time or the duration of the voltage plateau region, thereby avoiding prediction bias caused by short-term noise interference. Introducing dynamic memory depth into the LSTM allows the network to focus more on historical evolution patterns rather than instantaneous fluctuations under complex operating conditions.

[0050] S323b, when the battery response sensitivity is lower than the sensitivity threshold, enhances the current input update weight by reducing the forget gate output value.

[0051] Conversely, when the battery's response sensitivity is below a threshold (such as during low-temperature or low-current constant-voltage charging), the forget gate output value is reduced, and the update weight of the current input data is increased. At this time, the battery state changes gradually, and the current input parameters such as voltage and temperature better reflect real-time SOC changes. For example, during low-temperature constant-voltage charging, the battery's internal resistance changes slowly and the response linearity is high. The model can reduce the weight of the forget gate to decrease its reliance on distant historical states and prioritize the use of precise features of the current input (such as small changes in the voltage slope) for prediction. This dynamic adjustment mechanism allows LSTM to flexibly switch its memory strategy under different operating conditions: maintaining long-term memory to cope with nonlinear changes in high-sensitivity scenarios, and focusing on short-term inputs to improve response sensitivity in low-sensitivity scenarios. For example, when the battery enters the final stage of capacity decay, its response sensitivity may remain high. In this case, the dynamic increase in the forget gate weight can effectively capture long-term features such as the shortening of the voltage plateau caused by capacity decay, thereby maintaining prediction accuracy.

[0052] Furthermore, the S400 uses the predicted state of charge (SOC) value of a single cell, cell response sensitivity, photovoltaic power output fluctuation index, and ion concentration gradient prediction value to obtain the upper and lower limits of the safe SOC range for each single cell, including: S411, based on the preset lower limit threshold, superimposes the linear compensation amount of the predicted ion concentration gradient value to obtain the calculated value of the lower limit of the safe range of the state of charge.

[0053] Using a preset lower limit threshold (e.g., 20%) as a baseline, a linear compensation amount is added to the predicted ion concentration gradient to construct the lower limit of the safe state-of-charge range. The predicted ion concentration gradient reflects the non-uniformity of lithium-ion distribution inside the battery; a larger gradient indicates a more significant polarization effect on the electrode surface, which can easily lead to safety hazards such as lithium dendrite growth. For example, when the predicted value exceeds the threshold, it indicates that the battery is in a high-risk state, requiring further constraint on the depth of discharge through lower limit compensation. The compensation amount is calculated based on the linear relationship between the gradient and the safety margin; for example, for every 10% increase in the gradient, the lower limit threshold decreases by 2%. This design quantifies the internal electrochemical risks, dynamically adjusts the safety boundary, and avoids irreversible battery damage caused by polarization effects.

[0054] S412, based on the preset basic upper limit threshold, subtracts the linear compensation amount of the photovoltaic power output fluctuation index to obtain the calculated value of the upper limit of the safe state of charge range.

[0055] The upper limit of the safe state-of-charge range is calculated by subtracting a linear compensation amount from the photovoltaic power output fluctuation index based on a preset upper limit threshold (e.g., 90%). The photovoltaic power output fluctuation index reflects the severity of sudden power fluctuations in the system; greater fluctuations mean that the energy storage system needs to charge and discharge more frequently to smooth out power fluctuations. For example, when the photovoltaic power output fluctuation index exceeds the threshold, it indicates that the system faces high dynamic charging and discharging demands, requiring a safety buffer to prevent overcharging. The compensation amount calculation combines the fluctuation index with the battery capacity margin; for example, for every unit increase in the index, the upper limit threshold decreases by 3%. This design, by coupling photovoltaic fluctuation characteristics with battery safety requirements, ensures that the battery always operates within a safe range under dynamic conditions, avoiding electrochemical reaction runaway caused by instantaneous overcharging.

[0056] S413 When the battery response sensitivity exceeds the sensitivity threshold, the gradient compensation coefficient is adaptively amplified to obtain the corrected gradient compensation coefficient.

[0057] When the battery's response sensitivity exceeds a preset threshold, the system adaptively amplifies the gradient compensation coefficient. High battery response sensitivity indicates a significant sensitivity to sudden current changes; for example, under high temperature or high SOC conditions, even small current variations can easily trigger drastic voltage fluctuations. In this case, amplifying the gradient compensation coefficient can enhance the response to ion concentration gradient risks, for example, increasing the compensation coefficient from 1.0 to 1.5. This dynamic adjustment mechanism optimizes the safety boundary through physical constraints. For example, when the battery is near full charge and the response sensitivity is high, the lower threshold is further tightened to prevent over-discharge risks caused by intensified polarization effects. This design achieves real-time coupling between safety constraints and battery state, avoiding the insufficient adaptability of traditional fixed threshold strategies under complex operating conditions.

[0058] S414: When the difference between the upper and lower calculated values ​​of the state of charge safety range is less than the preset ratio of the state of charge of a single cell, the state of charge safety range of the single cell is set to the default value.

[0059] When the difference between the upper and lower limits of the safe state of charge range is less than a preset percentage (e.g., 15%), the safe range is forcibly set to the default value (15%-85%). This mechanism aims to prevent the range from being too narrow due to dynamic calculations, which could affect the normal operation of the system. For example, when the battery is in a deep aging state, and the ion concentration gradient and photovoltaic fluctuation index reach extreme values ​​simultaneously, dynamic calculations may generate a range of 10%-80%, but in practical applications, at least 15% of the usable capacity must be guaranteed to maintain basic functionality. The default value setting of 15%-85% combines the requirements of battery safety and system availability. For example, within this range, lithium iron phosphate batteries can balance life protection and energy utilization while avoiding range loss due to excessive conservatism. This design balances dynamic optimization and system robustness by setting a minimum effective range, ensuring stable operation even under extreme conditions.

[0060] Furthermore, the preset percentage of the state of charge (SOC) of a single cell is 15%; the default value of the safe range of SOC is 15%-85% of the SOC of a single cell.

[0061] Furthermore, based on the state-of-charge safety range of the single cell with the largest predicted ion concentration gradient and the current limit calculated based on its cell response sensitivity, the S400 generates system control commands that simultaneously constrain the charging current of the energy storage unit and the output power of the photovoltaic unit in the next control cycle, including: S421 calculates the current limit value based on the battery response sensitivity of a single cell using the negative power of the natural exponential function. The current limit value is equal to the maximum sustainable charging current value of the battery multiplied by the negative power of the natural exponential function for battery response sensitivity.

[0062] A non-linear relationship between the current limit and battery response sensitivity is constructed using a natural exponential function, enabling dynamic adaptation of the charge and discharge current. Specifically, the current limit is equal to the battery's maximum sustainable charging current (e.g., C / 2 rate) multiplied by the negative power of the natural exponential function's sensitivity, i.e.: ,in Let k be the battery response sensitivity and k be the decay coefficient. This formula utilizes the exponential decay characteristic to ensure that higher sensitivity corresponds to lower current limits, thereby actively suppressing charge and discharge intensity under high-risk operating conditions. For example, when the battery is at high temperature or nearing the end of its aging process (when sensitivity increases significantly), the exponential term... The value will quickly approach zero, causing The current is reduced below the safety threshold to avoid the risk of lithium plating, such as lithium dendrite formation. This design directly links battery physical characteristics with safety constraints. For example, in low-temperature environments, sensitivity increases due to increased electrolyte viscosity, and the automatic reduction of the current limit can reduce polarization effects and extend battery life.

[0063] S422, take the upper limit of the safe range of the state of charge of a single cell multiplied by the rated output power of the photovoltaic unit as the photovoltaic output power limit threshold.

[0064] A dynamic threshold for photovoltaic (PV) output power is generated by multiplying the upper limit of the safe state of charge (SOC) range of a single battery cell (e.g., 85%) by the rated output power of the PV unit. For example, if the rated PV power is 100kW and the upper limit of the safe range is 85%, the power threshold is 85kW. This threshold is used to constrain the impact of PV output fluctuations on battery charging and discharging, preventing battery SOC from exceeding limits due to sudden power changes. For example, when PV output drops sharply due to cloud cover, the power threshold is lowered to limit the energy storage discharge power, preventing the battery from entering the over-discharge safe range. Simultaneously, this threshold is dynamically linked to the battery SOC safe range. When the safe range shrinks due to changes in ion concentration gradients or sensitivity, the power threshold is adjusted accordingly. For example, if the upper limit of the safe range drops to 80%, the power threshold becomes 80kW, forming a dual constraint mechanism.

[0065] S423: When the photovoltaic power output fluctuation index exceeds the fluctuation threshold and the ion concentration gradient prediction value exceeds the lithium plating threshold at the same time, the current limiting instruction is executed first. If only the photovoltaic fluctuation index exceeds the fluctuation threshold, the power constraint instruction is executed. If only the ion concentration gradient exceeds the lithium plating threshold, both instructions are executed at the same time.

[0066] Based on the exceedance of the photovoltaic power output fluctuation index and the predicted ion concentration gradient, a tiered control strategy is formulated: First, a dual-threshold exceedance scenario: When the photovoltaic fluctuation index exceeds the threshold and the predicted ion concentration gradient simultaneously exceeds the lithium plating threshold (e.g., gradient prediction > 0.2), the system prioritizes executing a current limiting command. In this case, the battery faces dual risks (power surge and lithium plating), requiring strict limitation of the charging and discharging current (e.g., reducing it to 30% of the maximum value) to prevent runaway electrochemical reactions. For example, on a sunny afternoon with drastic photovoltaic power output fluctuations, the battery simultaneously experiences high-current charging and discharging and a surge in internal concentration gradient; current limiting can significantly reduce the probability of lithium dendrite formation. Second, a single-threshold exceedance scenario: If only the photovoltaic fluctuation index exceeds the limit (e.g., index > 0.5), a power constraint command is executed. By adjusting the photovoltaic output power (e.g., limiting it to 70% of the rated value), fluctuations are smoothed, reducing the dynamic adjustment pressure on the energy storage system. If only the ion concentration gradient exceeds the limit, two commands are executed simultaneously: current limiting reduces charging and discharging intensity, and power constraint prevents sudden power changes from exacerbating gradient inhomogeneity. For example, when the ion diffusion rate of a battery decreases due to aging, even if the photovoltaic fluctuations are small, the risk of lithium plating still needs to be suppressed through both current limiting and power constraint.

[0067] S424 sends digital control commands to the power converter, including the target current value for the current limiting command and the target power value for the power constraint command.

[0068] Precise execution of current and power commands is achieved through a power converter. These commands include current limits (e.g., 200A) and power constraints (e.g., 85kW), and are sent digitally to the PCS (Power Conversion System) and the photovoltaic inverter. For example, when an excessive ion concentration gradient is detected, a current limit command is immediately generated, requiring the PCS to limit the charging current to 50% of its current maximum value. Simultaneously, a power constraint command is issued to the photovoltaic inverter, requiring its output power to not exceed a safety threshold. This device-specific command mode achieves multi-level control: the PCS directly adjusts the charging and discharging power, and the photovoltaic inverter adjusts the power generation; the two work together to suppress system fluctuations. Command issuance uses digital communication protocols (e.g., Modbus or IEC 61850) to ensure real-time and reliable command transmission, such as completing command parsing and execution within millisecond response time, avoiding safety risks caused by delays.

[0069] Accordingly, please refer to Figure 2 A second aspect of the present invention provides a control device for a photovoltaic storage system based on multi-source information fusion, which controls the photovoltaic storage system based on the above-described control method for a photovoltaic storage system based on multi-source information fusion. The control device includes: Data acquisition module 1 is used to acquire the state parameters and environmental parameters of each photovoltaic string in the photovoltaic unit and each individual cell in the energy storage unit in the current control cycle to obtain the original dataset; Data calculation module 2 is used to calculate the photovoltaic output fluctuation index of the photovoltaic unit and the cell response sensitivity of each individual cell based on the original dataset. The state prediction module 3 is used to obtain the predicted state of charge and ion concentration gradient of a single cell in the next control cycle based on the original dataset of the single cell and the cell response sensitivity, through a long short-term memory neural network combined with an electrochemical model. System control module 4 is used to obtain the upper and lower limits of the safe state of charge range for each individual cell based on the predicted state of charge, cell response sensitivity, photovoltaic power output fluctuation index and ion concentration gradient. Based on the safe state of charge range of the individual cell with the largest ion concentration gradient and the current limit calculated based on its cell response sensitivity, it generates system control commands that simultaneously constrain the charging current of the energy storage unit and the output power of the photovoltaic unit in the next control cycle.

[0070] A third aspect of the present invention provides an electronic device, including: at least one processor; and a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the above-described optical storage system control method based on multi-source information fusion.

[0071] A fourth aspect of the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the above-described control method for a photoelectric storage system based on multi-source information fusion.

[0072] The embodiments of the present invention aim to protect a control method and device for a photovoltaic energy storage system based on multi-source information fusion, which has the following effects: 1. The power fluctuation characteristics of the photovoltaic-storage system are quantified by the photovoltaic output fluctuation index, and a dynamic safety range is established by combining the response sensitivity of individual cells, thus overcoming the lag of single threshold control; 2. By employing an LSTM neural network combined with an electrochemical model, the state of charge (SOC) and ion concentration gradient of a single cell are accurately predicted, solving the problems of large cumulative error and insufficient nonlinear dynamic capture of the traditional ampere-hour integration method, thus improving the prediction accuracy by more than 30%. 3. Using the single cell with the largest ion concentration gradient as the constraint benchmark, a collaborative control command is generated through a slip-type demand control algorithm. This avoids the conflict between the peak photovoltaic power generation period and the energy storage discharge period, and ensures that the battery operates within a safe range, extending battery life by 20%-30% and improving the system's charge and discharge efficiency by more than 15%.

[0073] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A control method for a photovoltaic energy storage system based on multi-source information fusion, characterized in that, The photovoltaic-energy storage system includes an energy storage unit and several photovoltaic units, and the control method includes the following steps: The state parameters and environmental parameters of each photovoltaic string in the photovoltaic unit and each individual cell in the energy storage unit are obtained in the current control cycle to obtain the original dataset; Based on the original dataset, the photovoltaic output fluctuation index of the photovoltaic unit and the battery response sensitivity of each individual cell are calculated. Based on the original dataset of the single cell and the cell response sensitivity, the predicted state of charge and ion concentration gradient of the single cell in the next control cycle are obtained by combining a long short-term memory neural network with an electrochemical model. Based on the predicted state of charge (SOC) value, battery response sensitivity, photovoltaic power output fluctuation index, and ion concentration gradient prediction value of the individual cells, the upper and lower limits of the safe SOC range for each individual cell are obtained. Using the safe SOC range of the individual cell with the largest ion concentration gradient prediction value and the current limit value calculated based on its battery response sensitivity as a benchmark, a system control command is generated to simultaneously constrain the charging current of the energy storage unit and the output power of the photovoltaic unit in the next control cycle.

2. The control method for a photovoltaic energy storage system based on multi-source information fusion according to claim 1, characterized in that, The photovoltaic output fluctuation index is the ratio of the average of the maximum instantaneous power change rate of all photovoltaic strings in the current control cycle to the rated power. The battery response sensitivity is the ratio of the differential voltage change rate of the individual cell to the differential charging current change rate, divided by the photovoltaic output fluctuation index.

3. The control method for a photovoltaic energy storage system based on multi-source information fusion according to claim 1, characterized in that, The process of obtaining the predicted state of charge and ion concentration gradient of the single cell in the next control cycle based on the original dataset and cell response sensitivity of the single cell, using a long short-term memory neural network combined with an electrochemical model, includes: The original dataset of the single cell is divided into a time-series evolution data channel and a transient event data channel. The time-series evolution data channel contains sequence data of voltage, temperature and internal resistance changing continuously over time, while the transient event data channel contains the energy peak characteristics and frequency distribution characteristics of acoustic emission signals within a specific frequency band. Based on the continuous change sequence data of the time-series evolution data channel, combined with the battery response sensitivity, the time-related features are extracted through the long short-term memory neural network to output the initial state of charge prediction of the single cell. Based on the voltage, charging current and temperature data of the single cell, the diffusion equation describing the change of lithium ion concentration inside the electrode solid particles is obtained through the electrochemical model, and the predicted value of the spatial non-uniformity of lithium ion distribution on the surface of the negative electrode active material and the auxiliary estimated value of the state of charge calculated based on the principle of total lithium ion conservation are output. Based on the initial state of charge (SOC) prediction and the auxiliary SOC prediction, a dynamic weighted sum is performed using the battery response sensitivity to output the final SOC prediction value. Simultaneously, the predicted value of lithium ion spatial distribution non-uniformity is output as the predicted value of ion concentration gradient.

4. The control method for a photovoltaic energy storage system based on multi-source information fusion according to claim 3, characterized in that, The continuous change sequence data based on the time-series evolution data channel, combined with the battery response sensitivity, extracts time-related features through the long short-term memory neural network to output the initial state of charge prediction of the single cell, including: The continuous change sequence data of voltage, temperature and internal resistance in the time-series evolution data channel are divided into fixed-length input windows at equal intervals according to the control cycle. The battery response sensitivity is incorporated as an independent feature vector into the head of the input sequence to obtain structured input data containing prior knowledge of dynamic sensitivity. Structured input data is processed by a three-layer long short-term memory neural network connected in series: the first hidden layer uses a first number of neurons to extract the macroscopic state evolution features of the battery; the second hidden layer receives the output of the first hidden layer and uses a second number of neurons to extract the microscopic aging correlation features of the battery; the third hidden layer receives the output of the second hidden layer and uses a third number of neurons to fuse the nonlinear relationship between time dependence characteristics and battery response sensitivity. In the gated computation of the third hidden layer, the battery response sensitivity is dynamically adjusted as a scaling factor for the forget gate; The fused feature vector output from the third hidden layer is mapped to a scalar value through a fully connected layer. After normalization by a sigmoid function, the output value range is constrained to between 0 and 1, corresponding to the physical range of 0% to 100% of the state of charge. Finally, the initial state of charge estimate of the single cell in the target control cycle is output.

5. The control method for a photovoltaic energy storage system based on multi-source information fusion according to claim 4, characterized in that, The method of dynamically adjusting the battery response sensitivity as a scaling factor for the forget gate includes: When the battery response sensitivity is higher than the sensitivity threshold, the weight of historical state memory is strengthened by increasing the output value of the forget gate; When the battery response sensitivity is below the sensitivity threshold, the current input update weight is enhanced by reducing the forget gate output value.

6. The control method for a photovoltaic energy storage system based on multi-source information fusion according to claim 1, characterized in that, The method of obtaining the upper and lower limits of the safe state of charge range for each individual cell based on the predicted state of charge, cell response sensitivity, photovoltaic power output fluctuation index, and ion concentration gradient prediction includes: Based on the preset lower limit threshold, the linear compensation of the predicted ion concentration gradient is superimposed to obtain the calculated lower limit of the safe range of the state of charge. Based on the preset basic upper limit threshold, the linear compensation amount of the photovoltaic power output fluctuation index is deducted to obtain the calculated value of the upper limit of the safe state of charge range; When the battery response sensitivity exceeds the sensitivity threshold, the gradient compensation coefficient is adaptively amplified to obtain the corrected gradient compensation coefficient. When the difference between the calculated upper limit and the calculated lower limit of the state of charge safety range is less than the preset ratio of the state of charge of the individual battery, the state of charge safety range of the individual battery is set to the default value.

7. The control method for a photovoltaic energy storage system based on multi-source information fusion according to claim 6, characterized in that, The preset percentage of the state of charge of the individual battery cell is 15%; The default value for the safe state of charge range is 15%-85% of the state of charge of the individual battery cell.

8. The control method for a photovoltaic energy storage system based on multi-source information fusion according to claim 1, characterized in that, The system control command, based on the safe state-of-charge range of the single cell with the largest predicted ion concentration gradient and the current limit calculated based on its cell response sensitivity, generates a system control command that simultaneously constrains the charging current of the energy storage unit and the output power of the photovoltaic unit in the next control cycle, including: Based on the battery response sensitivity of the individual battery, the current limit value is calculated through the negative sensitivity power of the natural exponential function. The current limit value is equal to the maximum sustainable charging current value of the battery multiplied by the negative battery response sensitivity power of the natural exponential function. The upper limit of the safe range of the state of charge of the single cell is multiplied by the rated output power of the photovoltaic unit as the photovoltaic output power limit threshold. When the photovoltaic power output fluctuation index exceeds the fluctuation threshold and the ion concentration gradient prediction value also exceeds the lithium plating threshold, the current limiting command is executed first. If only the photovoltaic fluctuation index exceeds the fluctuation threshold, the power constraint command is executed. If only the ion concentration gradient exceeds the lithium plating threshold, both commands are executed simultaneously. Send digital control commands to the power converter, including the target current value for current limiting commands and the target power value for power constraint commands.

9. A control device for a photovoltaic energy storage system based on multi-source information fusion, characterized in that, The photovoltaic storage system is controlled based on the multi-source information fusion-based photovoltaic storage system control method according to any one of claims 1-8, and the control device includes: The data acquisition module is used to acquire the state parameters and environmental parameters of each photovoltaic string in the photovoltaic unit and each individual cell in the energy storage unit in the current control cycle, and obtain the raw dataset. A data calculation module is used to calculate the photovoltaic output fluctuation index of the photovoltaic unit and the battery response sensitivity of each individual cell based on the original dataset. The state prediction module is used to obtain the predicted state of charge and ion concentration gradient of the single cell in the next control cycle based on the original dataset of the single cell and the cell response sensitivity, through a long short-term memory neural network combined with an electrochemical model. The system control module is used to obtain the upper and lower limits of the safe state of charge range for each individual cell based on the predicted state of charge, cell response sensitivity, photovoltaic power output fluctuation index, and ion concentration gradient prediction. Based on the safe state of charge range of the individual cell with the largest ion concentration gradient prediction and the current limit calculated based on its cell response sensitivity, the module generates a system control command that simultaneously constrains the charging current of the energy storage unit and the output power of the photovoltaic unit in the next control cycle.

10. An electronic device, characterized in that, include: At least one processor; The system includes a memory connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to cause the at least one processor to perform the optical storage system control method based on multi-source information fusion as described in any one of claims 1-8.

Images