Charging and discharging control method and system for photovoltaic energy storage integration

Abstract

The present application relates to photovoltaic energy storage optimization control technical field, especially photovoltaic energy storage integrated charge-discharge control method and system, its method includes: collecting multimodal signal, fusing after time sensitive network synchronization and denoising;Liquid neural-thermal power and electrochemical coupling digital twin system is used to generate heat source power density, battery aging parameter and linear embedded state data;Based on the state data of space-time Gaussian process sampling probability scene, high-dimensional supersymbol coding and mutual information filtering compressed scene set, combined with the aging parameter in quantum-temperature change global optimization model to determine the optimal charge-discharge state, generate reference power trajectory;Neuromorphic chip runs convex model predictive control output real-time power instruction;Inverter executes charge-discharge through pheromone game and self-distillation reinforcement learning and triggers supercapacitor support when frequency is abnormal, executes data write distributed ledger closed loop update.The present application realizes millisecond level power tracking and reduces cycle stress.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic energy storage optimization control technology, and in particular to a charging and discharging control method and system for integrated photovoltaic energy storage. Background Technology

[0002] Photovoltaic power generation is subject to rapid fluctuations due to cloud cover. If the energy storage system cannot compensate within seconds, it will lead to grid connection frequency deviations and accelerate battery degradation. Existing technologies generally employ two methods: First, rolling linear programming based on single irradiance prediction, using conventional cameras to collect frame-by-frame cloud images and performing power planning in a central processing unit; second, using offline impedance models to allocate fixed charge / discharge depths to the batteries. The former, due to cloud image update cycles ≥1s, cannot capture millisecond-level sudden drops in irradiance, resulting in delayed power prediction; the latter ignores real-time aging and temperature coupling, causing current peaks to fail to converge in time, resulting in excessive cycle depths and reduced battery life. Summary of the Invention

[0003] To address the numerous problems existing in the prior art, this invention provides a photovoltaic energy storage integrated charge and discharge control method and system. This invention captures cloud shadows using an event visual flow; a liquid neural network estimates the power density of the additional heat source and couples it with an electrical model to generate an embedded state; a spatiotemporal Gaussian process samples the scene, and after high-dimensional encoding and mutual information filtering to compress the dimension, a binary charge and discharge matrix is ​​determined through quantum annealing and variable-temperature search; a neuromorphic chip outputs power commands within a convex model prediction framework; the inverter coordinates with pheromone concentration and triggers supercapacitor support when the frequency is abnormal, executing on-chain feedback. This achieves millisecond-level power point tracking, cycle depth minimization, and model self-evolution.

[0004] A method for charging and discharging integrated photovoltaic energy storage includes the following steps:

[0005] Multimodal signals from photovoltaic modules, energy storage battery cabinets, grid connection interfaces, and the sky terminal are collected, and then denoised using a deterministic time-sensitive network with unified timestamps. The cloud shadow outline is generated from the integral event visual stream to obtain fused data.

[0006] The fused data is input into a digital twin system coupled with liquid neural-thermal power and electrochemistry to generate heat source power density, battery aging parameters and linear embedding state data, and probabilistic scene data is generated based on the linear embedding state data using a spatiotemporal Gaussian process.

[0007] High-dimensional super-symbol encoding and mutual information filtering are performed on the probabilistic scenario data. The optimal charge and discharge state is determined in the global optimization model in combination with the battery aging parameters to generate a reference power trajectory. The convex model predictive control outputs real-time power commands on the neuromorphic chip.

[0008] Each inverter, based on the real-time power command and local state vector, generates a modulation signal through pheromone game and self-distillation reinforcement learning to control the power converter to perform charging and discharging. When the network frequency deviation reaches the threshold, it triggers supercapacitor discharge support and returns according to the power slope. The execution data is written to the distributed ledger at fixed intervals to update the digital twin system and global optimization model.

[0009] Preferably, the acquisition step synchronizes the event stream output by the event visual sensor with the current signal acquired by the photovoltaic module, the voltage signal acquired by the photovoltaic module, the state of charge signal acquired by the energy storage battery cabinet, the network frequency signal acquired by the grid connection interface, and the cloud image acquired by the sky end. After generating the cloud shadow contour by the event visual stream in the sliding integral manner, it is combined with the other signals and unified with a deterministic time-sensitive network. Then, sliding median filtering and drift correction are sequentially performed in the edge processor to obtain the fused data.

[0010] Preferably, the liquid neural-thermal power coupling model maps the cloud shadow contour to the heat source power density, writes the heat source power density as a source term into the two-dimensional thermal conduction equation, and couples it with the photovoltaic module current signal and photovoltaic module voltage signal to solve the temperature field. The temperature field result is used to correct the state components related to thermal characteristics in the linear embedded state data.

[0011] Preferably, the electrochemical coupling model adopts a joint modeling approach of equivalent circuit and electrochemical kinetics. It constructs a state vector using the state of charge signal and temperature field results collected by the energy storage battery cabinet, and calculates the diffusion coefficient and interface film resistance by recursion through extended Kalman filtering to generate the battery aging parameters.

[0012] Preferably, the spatiotemporal Gaussian process uses the linear embedded state data as the mean function and constructs a covariance structure in the form of the product of spatial kernel function and temporal kernel function to jointly sample the irradiation sequence, temperature sequence and photovoltaic power sequence to form a weighted probabilistic scene data set.

[0013] Preferably, the high-dimensional supersymbolic encoding maps each probability scene data into a fixed-length binary vector through random Gaussian projection, calculates the mutual information score between the binary vector and the linear embedded state data using a hardware operation of bit XOR addition, and selects a preset number of scenes to form the weighted scene subset according to the size of the mutual information score.

[0014] Preferably, the global optimization model discretizes the weighted scenario subset and the battery aging parameters into a binary charge-discharge decision matrix to construct a quadratic unconstrained binary optimization model. Candidate solutions are obtained through a quantum annealing solver. Then, energy consumption values ​​are compared in a variable-temperature stochastic search framework, and the charge-discharge decision with the minimum energy consumption value is retained as the preferred charge-discharge state.

[0015] Preferably, the convex model predictive control uses the linear embedded state data as the system state and the reference power trajectory as the target trajectory. Under linear constraints, it uses the alternating direction multiplier method to iteratively solve the optimization problem to obtain real-time power commands, and re-solves the optimization problem in each control cycle based on the updated state.

[0016] Preferably, each inverter broadcasts its active and reactive pheromone concentrations to neighboring inverters via a pheromone diffusion-evaporation game. The updated pheromone concentrations and local state vectors are input into a policy network trained by self-distillation reinforcement learning to obtain a modulation signal to control the bidirectional power converter to perform charging and discharging. When the network frequency deviation reaches a threshold, the supercapacitor bypass discharge is triggered and returned according to the power slope. At the same time, the timestamp data, output active power data, output reactive power data, and active pheromone concentration data are hashed and written into a distributed ledger at a fixed period. The execution data stored in the distributed ledger is used to synchronously update the digital twin system and the global optimization model.

[0017] A photovoltaic energy storage integrated charging and discharging control system is provided for implementing the aforementioned photovoltaic energy storage integrated charging and discharging control method. The system includes:

[0018] The acquisition and fusion module is used to acquire signals from photovoltaic modules, energy storage battery cabinets, grid connection interfaces, and sky terminals. It unifies timestamps through a deterministic time-sensitive network and performs noise reduction processing. It integrates the visual stream of events into cloud shadow contours to generate fused data.

[0019] The twin generation module is used to input the fused data into a digital twin system coupled with liquid neural-thermal power and electrochemistry to generate heat source power density, battery aging parameters and linear embedding state data, and to generate probabilistic scene data based on the linear embedding state data using a spatiotemporal Gaussian process.

[0020] The scenario optimization module is used to perform high-dimensional super-symbol encoding and mutual information filtering on the probabilistic scenario data, determine the preferred charge and discharge state in the global optimization model in combination with the battery aging parameters, generate a reference power trajectory, and run convex model predictive control on the neuromorphic chip to output real-time power commands.

[0021] The execution feedback module is used by each inverter to generate a modulation signal based on the real-time power command and local state vector through pheromone game and self-distillation reinforcement learning to control the power converter to perform charging and discharging. When the network frequency deviation is detected to reach a threshold, the supercapacitor discharge support is triggered and the output returns according to the power slope. The execution data is written to the distributed ledger at a fixed period to update the twin generation module and the scene optimization module.

[0022] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0023] This invention achieves millisecond-level irradiance prediction by rapidly mapping cloud shadow contours to heat source density using an event vision-liquid neural network; it realizes dynamic lifetime constraints by outputting diffusion coefficients and membrane resistance in real time through an equivalent circuit-extended Kalman filter; it achieves microsecond-level relevant scene retrieval by rapidly reducing scene dimensions through random Gaussian projection and mutual information filtering; it achieves energy cost minimization by obtaining the globally optimal charge / discharge matrix through quantum annealing-variable temperature search; it achieves 20ms instruction refresh through convex model predictive control on a pulse neural chip; it maintains inverter group smoothness and power balance through pheromone diffusion game and self-distillation reinforcement learning; and it achieves closed-loop self-updating of the model through distributed ledger backfeeding of execution logs. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of the method of the present invention;

[0025] Figure 2 This is a structural diagram of the digital twin coupling model in this invention;

[0026] Figure 3 This is the collaborative logic diagram of the pheromone game in this invention;

[0027] Figure 4 This is a structural block diagram of the system of the present invention. Detailed Implementation

[0028] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation.

[0029] like Figure 1 As shown, a photovoltaic energy storage integrated charging and discharging control method includes the following steps:

[0030] Multimodal signals from photovoltaic modules, energy storage battery cabinets, grid connection interfaces, and the sky terminal are collected, and then denoised using a deterministic time-sensitive network with unified timestamps. The cloud shadow outline is generated from the integral event visual stream to obtain fused data.

[0031] The charging and discharging control link of photovoltaic-energy storage integration has extremely high requirements for the time accuracy and physical integrity of input data. This invention uses a three-stage design of "multimodal synchronous acquisition - deterministic time-sensitive network alignment - event visual cloud extraction" to transform heterogeneous information from the photovoltaic side, energy storage side, grid-connected side and sky end into a fused data, providing consistent, low-latency and traceable input for subsequent digital twin modeling and predictive control.

[0032] The multimodal acquisition section installs current and voltage sensors on the photovoltaic modules, state-of-charge sensors and triaxial strain gauges on the energy storage battery cabinet, a synchronous phasor measurement device on the grid connection interface, and an event vision sensor and cloud imaging terminal at a location with a high solar altitude angle. The event vision sensor outputs brightness change events in a microsecond-level triggering manner, enabling it to capture sudden changes in illumination information before the edge of the cloud shadow completely covers the battery modules. Combining these signals can simultaneously reflect the electrical status of the generation, energy storage, and grid ends, as well as the dynamics of sky illumination, making the fused data complete.

[0033] Unified timestamps are achieved through deterministic time-sensitive networking (TSN). TSN uses full-duplex gigabit Ethernet at the physical layer and time-based gating at the link layer, with the transmission windows of all acquisition nodes based on a global clock. This ensures that the jitter of each frame reaching the edge switch is limited to the microsecond level, regardless of node distance. The edge switch inserts nanosecond-level timestamps into all frames before forwarding them to the edge processor, guaranteeing that data from different sources share a common timeline in subsequent calculations.

[0034] The edge processor first performs moving median filtering on the current, voltage, state of charge, strain, and phasor data to remove impulse noise, and then corrects for temperature drift and baseline drift using a 4-hour moving average. For the event visual stream, a fixed integration window is used. Internal pixel-level event counting. Let the positive polarity event count be... The count of negative polarity events is pixel coordinates are The brightness threshold is The cloud shadow contour matrix is ​​given by the formula:

[0035]

[0036] Generate. Here. The value is a pixel label, with a value of 1 indicating that an occlusion edge has been detected. A 5-millisecond integral window balances the speed of cloud movement and the amount of data. The edge processor performs streaming computation through sliding accumulation, without introducing significant latency.

[0037] The fused data consists of timestamps, photovoltaic current, photovoltaic voltage, module backsheet temperature, energy storage state of charge, energy storage strain values, grid-connected phasors, grid frequency, cloud imagery, and the aforementioned cloud shadow contour frames. The cloud shadow contours provide the predictive control module with millisecond-level prior knowledge of irradiance changes, while the state of charge and strain accurately reflect battery health, and the grid phasor and frequency deviations reflect the load dynamics at the grid connection point. Through a unified sampling window, all fields are fully aligned in the time domain, eliminating common errors introduced by multi-source asynchronous processing.

[0038] In a 10 MW photovoltaic-2 MWh energy storage prototype test, the introduction of event-visual cloud shadow contours reduced the 2-second prediction mean square error of photovoltaic active power by approximately 20% compared to using only traditional forecasts. This rapid prediction improved the foresight of battery discharge compensation, and the peak ramp rate of the energy storage inverter in photovoltaic power surge scenarios decreased by approximately 30%, validating the effectiveness of data fusion in mitigating power surges. Another embodiment deployed 50 acquisition nodes in a distributed power station in a mountainous area, using an 8-node ring topology in the time-sensitive network. Verification results showed that the 95th percentile of the end-to-end time jitter distribution did not exceed 2 microseconds, meeting the synchronization accuracy requirements of digital twins.

[0039] Preferably, the acquisition step synchronizes the event stream output by the event visual sensor with the current signal acquired by the photovoltaic module, the voltage signal acquired by the photovoltaic module, the state of charge signal acquired by the energy storage battery cabinet, the network frequency signal acquired by the grid connection interface, and the cloud image acquired by the sky end. After generating the cloud shadow contour by the event visual stream in the sliding integral manner, it is combined with the other signals and unified with a deterministic time-sensitive network. Then, sliding median filtering and drift correction are sequentially performed in the edge processor to obtain the fused data.

[0040] Rapid fluctuations in photovoltaic output power are primarily caused by cloud shadowing, and traditional irradiance prediction methods struggle to capture the instantaneous changes at the edges of cloud shadows on a second-level scale. This invention deploys an event vision sensor in the sky. This sensor possesses microsecond-level temporal resolution and self-triggering output characteristics, enabling it to immediately record pixel coordinates and the polarity of the change when brightness changes exceed a threshold. Compared to frame-based cameras, the event vision sensor does not require a fixed frame rate, thus enabling it to capture brightness changes in real time before cloud shadows reach the photovoltaic array, generating a high-frequency event stream. This event stream is denoted as a quadruple. ,in and For pixel coordinates, The polarity of brightness change This is the event timestamp. To maintain sampling consistency with the electrical signal, this invention selects a fixed time window. The event stream is subjected to sliding integral to generate cloud shadow contour frames.

[0041] The integration process uses a pixel counting method: counting the positive polarity events at the same pixel location within a window. Counting negative polarity events Calculate the brightness difference .like Exceeding the threshold Then, the corresponding pixels are marked as cloud shadow edges, resulting in the cloud shadow contour matrix:

[0042]

[0043] in A value of 1 indicates that there is a cloud-like edge at the pixel, while a value of 0 indicates that the brightness change at the pixel is insufficient to form an edge.

[0044] In parallel with the event visual data, current and voltage signals are acquired at the photovoltaic module side, state-of-charge and triaxial strain signals are acquired at the energy storage battery cabinet side, network frequency signals are acquired at the grid connection interface side, and visible light cloud images are acquired at the sky-end cloud imaging terminal. These multi-source signals need to be strictly aligned to serve as synchronous inputs for the digital twin model. This invention employs a deterministic time-sensitive network (TSN) to connect each acquisition node to an edge processor. The TSN uses time-driven gating at the data link layer, with each transmission queue opening in a preset time slot to ensure all nodes transmit data frames within the same period. The network uses a high-precision synchronous clock as the global time reference, and the edge switches insert nanosecond-level timestamps into each arriving frame, thus ensuring synchronization errors of data from different nodes on a microsecond scale. Experiments show that, under gigabit Ethernet physical layer conditions, the 95th percentile of the end-to-end jitter distribution in the 8-node ring topology does not exceed 2 microseconds.

[0045] After timestamp alignment, the edge processor first performs sliding median filtering on current, voltage, state of charge, strain, and network frequency to remove high-frequency impulse noise, and then uses sliding mean subtraction to eliminate temperature drift and baseline drift. Cloud shadow contour frames are directly entered into subsequent processes after generation without undergoing low-pass filtering to preserve the spatial accuracy of the edges. All fields after noise removal are framed according to a common time axis to form fused data. The fused data fields include: photovoltaic current, photovoltaic voltage, module backsheet temperature, energy storage state of charge, energy storage strain value, grid-connected voltage phasor, grid-connected frequency deviation, cloud imagery, and cloud shadow contour.

[0046] The advantages of data fusion are reflected in the following two aspects. First, cloud shadow contours provide forward-looking information on irradiance changes within a second-level timescale, enabling early guidance for energy storage discharge to compensate for sudden drops in photovoltaic output. Second, state of charge and strain can provide feedback on battery health on the same time reference, avoiding state estimation bias caused by asynchronous sampling. By synchronizing rapid cloud shadow detection with electrical measurements, this invention achieves fully coupled observation of the photovoltaic side, energy storage side, grid-connected side, and sky end, providing complete boundary conditions for the digital twin model.

[0047] The application effect was verified through two examples. In the test of the joint operation of a 10 MW ground-mounted centralized photovoltaic power station and a 2 MWh energy storage system, an integration window was set. The event counting threshold is 5 milliseconds. The data was calibrated before sunrise using an adaptive algorithm. Compared to the baseline scheme that only used irradiance prediction, the mean square error of the 2-second ultra-short-term photovoltaic power prediction was reduced by 20% after adding cloud shadow contours. Within the same test period, the peak ramp rate of the energy storage inverter decreased by 30% in the scenario of sudden drop in photovoltaic output, indicating that the forward-looking information effectively mitigated the power surge. In another embodiment, 50 acquisition nodes were deployed in a distributed photovoltaic system in a mountainous area. Each node sent a 1 kHz sampling frame through a deterministic time-sensitive network, and the edge switch inserted a hardware timestamp before forwarding it to the processor. Verification results showed that during 48 hours of continuous operation, the cumulative missing frame rate of the fused data was less than 0.01%, and the variance of the end-to-end synchronization error remained on the order of 1 microsecond, meeting the high-precision synchronization requirements of the digital twin system.

[0048] like Figure 2 As shown, the fused data is input into a digital twin system that couples liquid neural-thermal power and electrochemistry to generate heat source power density, battery aging parameters and linear embedding state data, and probabilistic scene data is generated based on the linear embedding state data using a spatiotemporal Gaussian process.

[0049] The digital twin system of this invention simultaneously characterizes the thermal-power process of photovoltaic modules and the electrochemical degradation process of energy storage batteries. Its core consists of three parts: a liquid neural network-thermal-power coupling model, an electrochemical coupling model, and a spatiotemporal Gaussian process based on linearly embedded state data. Its principles and applications are as follows.

[0050] A liquid neural network-thermal power coupling model is used to convert cloud shadow contours into instantaneous heat source power density of the photovoltaic module backsheet. A liquid neural network is a recursive network with continuous-time differential equations within its nodes, possessing real-time memory capability for non-stationary excitations. Let... For the hidden state vector, Given an input pixel count vector, the node dynamics can be written as:

[0051]

[0052] in Leakage coefficient, For the input weight matrix, This is the recursive weight matrix. The matrix is ​​read out linearly. The heat source power density is obtained as follows:

[0053]

[0054] In the formula The unit is watts per square meter. As a source term, it can be written into the two-dimensional thermal conductivity-power coupling equation to solve the component temperature distribution in real time. The method couples the current and voltage signals of the photovoltaic module to calculate the output power. This method avoids the lag problem of traditional irradiance-heat coupling models that rely on cloud cover prediction. Experiments show that the temperature field prediction delay can be controlled within 50 milliseconds.

[0055] The electrochemical coupling model is built upon the Doyle-Fuller-Newman dynamic framework. Charge concentration, solution concentration, electrode potential, and electrolyte potential are selected as state variables, and the diffusion coefficient is estimated iteratively using an extended Kalman filter. With interface film resistance These two factors together constitute the battery aging parameters. This real-time estimation result is not only used to limit the allowable current in the charge / discharge strategy, but also provides state variables for lifetime optimization.

[0056] To map nonlinear, high-dimensional multimodal data to a linear subspace suitable for solving convex model predictive control, the system first performs a polynomial-trigonometric mapping on the fused data, and then uses Bayesian-Koopman linearization to obtain linear embedded state data. The mapping result is denoted as... Its evolution equation approximately satisfies:

[0057]

[0058] in To embed the spatial transition matrix, It is Gaussian white noise. Because... With a fixed dimension and approximately linear dynamics, the subsequent model predictive control can be written as a quadratic programming problem, and the solution complexity is linearly related to the time window.

[0059] based on A mean function is constructed, and the system uses a spatiotemporal Gaussian process to sample the joint irradiance and temperature processes in future time periods. The kernel function is chosen to be the product of a spatial radial basis kernel and a temporally exponentially decaying kernel, ensuring spatial smoothness and decreasing temporal correlation. Variational inference allows for the simultaneous sampling of multiple probabilistic scene trajectories, with each trajectory assigned a probability weight. This preserves the uncertainty of the prediction while enabling rapid calculation of the correlation between the scene and the real-time state during subsequent mutual information filtering.

[0060] Combining the above three parts, heat source power density, battery aging parameters, and linearly embedded state data can be obtained. The effects are as follows: First, the liquid neural network directly injects cloud shadow information into the thermal field equation, enabling component temperature prediction to lead cloud shadow arrival time by approximately 120 milliseconds; second, the extended Kalman filter provides real-time aging parameters, allowing the global optimization model to dynamically adjust the cycle depth; third, the linearly embedded state data accelerates the solution of model predictive control, reducing its single-step iteration time from milliseconds to sub-milliseconds.

[0061] Example 1 demonstrates the effectiveness of a 10 MW photovoltaic-2 MWh energy storage system operating for 48 hours. The generated probabilistic scenario data is updated every 30 seconds, with the scenario set containing 128 trajectories. Mutual information filtering takes less than 100 microseconds. Compared to baseline methods without linear mapping, the model predictive control iterations are reduced by 60%, and backup power activation during rapid photovoltaic output declines is 150 milliseconds earlier. Example 2 deploys 50 nodes in a distributed power station, using a liquid neural network with 256 hidden nodes and a Bayesian-Koopman model with a 128-dimensional embedding space. Actual CPU utilization remains below 40%, validating the deployability of the solution in edge computing environments. In summary, this invention, through the combination of digital twins and probabilistic scenario generation, achieves flexible charge and discharge control under uncertain lighting conditions, providing a highly efficient, accurate, and economical solution for deep photovoltaic-energy storage collaboration.

[0062] Preferably, the liquid neural-thermal power coupling model maps the cloud shadow contour to the heat source power density, writes the heat source power density as a source term into the two-dimensional thermal conduction equation, and couples it with the photovoltaic module current signal and photovoltaic module voltage signal to solve the temperature field. The temperature field result is used to correct the state components related to thermal characteristics in the linear embedded state data.

[0063] The liquid neural-thermal power coupling model is a key innovation of this invention in the digital twin layer. Its function is to rapidly convert the cloud shadow contour obtained from event vision into the instantaneous additional heat source power density of the photovoltaic module, and then incorporate this heat source power density, along with the resistive loss power density calculated from current and voltage signals, into the two-dimensional heat conduction equation to solve for the temperature distribution in real time. The global average value of the temperature distribution is then used as a correction factor for the thermal dimension in the linearly embedded state vector, thereby ensuring that the subsequent convex model predictive control remains within the linearly solvable range.

[0064] The principle of mapping cloud shadow contours to heat source power density. An event vision sensor triggers numerous brightness change events when the edge of a cloud shadow passes over a photovoltaic array. The model sparsely encodes the binary cloud shadow contour matrix within an integral window into a fixed-length vector and feeds it into a liquid neural network. The internal states of the liquid neural network nodes follow a first-order differential equation with leakage, and its continuous-time characteristics preserve the high temporal resolution of the event flow. The hidden state vector is denoted as... The input vector is denoted as The node attenuation coefficient is denoted as The node dynamics are as follows:

[0065]

[0066] In the formula For the input weight matrix, For the recursive weight matrix, It is the hyperbolic tangent function. The matrix is ​​read out linearly. The additional heat source power density is obtained:

[0067]

[0068] in The unit is watts per square meter. During the training phase, a time-truncated backpropagation method is used to update the weights, allowing the network to converge on multi-day cloud shadow samples before online inference.

[0069] Next, the method of coupling electrical power to the two-dimensional thermal conductivity equation will be explained. Let the distance from the photovoltaic module's dimension along the step be... and Node coordinates are used This indicates that the time step is... The heat-power equation can be discretized as follows:

[0070]

[0071] For the first Temperature at any given moment Thermal conductivity, For density, For specific heat capacity, To increase the power density of the additional heat source, This is the power density due to resistive losses. The power density due to resistive losses is first calculated by determining the DC power of the component. The result is obtained by subtracting the inverter output power and normalizing it according to the component area. This invention executes the discrete equation in parallel on the graphics processing unit, with a single-frame iteration time of less than [amount missing]. Milliseconds, matched to the sampling period. The spatial average of the backplate temperature is taken after the temperature distribution is obtained. The linearly embedded state vector is denoted as... , of which Each component represents a temperature deviation. This invention utilizes:

[0072]

[0073] Replace it. and These are the offline calibration coefficients. The effect of doing this is to linearly map the nonlinear temperature field change into a state vector increment, so that the embedding space maintains approximately linear dynamics, ensuring that the quadratic programming assumption of the convex model predictive control holds.

[0074] Examples demonstrate the practical value of this model. During a test run at a 10 MW power station in Hami, Xinjiang, fast-moving convective clouds appeared in the sky. The liquid neural-thermal-power coupling model controlled the temperature field update delay to 50 milliseconds, a 60% reduction compared to the gray-box model. Simultaneously, the energy storage system adjusted its discharge power in advance according to the updated state vector. When photovoltaic power dropped, the peak system frequency deviation decreased from the baseline of 0.05 Hz to 0.03 Hz, verifying the gain of the thermal-power coupling input on frequency support. Another example involved deploying 50 acquisition nodes in a distributed mountain power station. The liquid neural network could operate with only 256 nodes, and the overall central processing unit occupancy rate remained below 40%, indicating that the algorithm's load is suitable for edge deployment.

[0075] In summary, the liquid neural-thermal power coupling model drives the heat conduction equation through event visual contours, generating temperature fields and correcting embedded states in milliseconds, achieving real-time closed-loop control of thermal, electrical, and cloud shadow dynamics. This mechanism significantly improves the response speed and robustness of subsequent scenario prediction and control planning, demonstrating replicable and scalable application value in industrial and commercial energy storage scenarios.

[0076] Preferably, the electrochemical coupling model adopts a joint modeling approach of equivalent circuit and electrochemical kinetics. It constructs a state vector using the state of charge signal and temperature field results collected by the energy storage battery cabinet, and calculates the diffusion coefficient and interface film resistance by recursion through extended Kalman filtering to generate the battery aging parameters.

[0077] The electrochemical coupling model of this invention aims to identify the internal diffusion process and interfacial film growth process of energy storage batteries online, so as to provide real-time aging parameters for subsequent global optimization models. This model adopts a joint modeling approach of equivalent circuit and electrochemical kinetics, incorporating the time-domain characteristics of the voltage response into the equation of state while retaining the physical meaning of anodic diffusion and solid electrolyte film growth. Its workflow can be summarized as three parts: state vector construction, extended Kalman filter recursion, and aging parameter output.

[0078] State vector construction. The acquisition module has provided the energy storage battery's state of charge signal and the average temperature of the temperature field. Based on the equivalent circuit concept, the battery terminal voltage, battery current, and the voltage of a parallel capacitor branch are first selected as electrical state variables; secondly, two types of slow variables from electrochemical kinetics are introduced: the diffusion coefficient of anode active particles and the resistance of the solid electrolyte membrane. Finally, a comprehensive state vector containing both electrical and electrochemical sub-vectors is formed.

[0079]

[0080] in Indicates the voltage of the capacitor branch. This indicates the voltage drop corresponding to the ohmic internal resistance. Indicates battery current. Indicates the anodic diffusion coefficient. This represents the resistance of the solid electrolyte membrane. The state-of-charge signal is used to limit the feasible range of the diffusion coefficient, and the temperature field-averaged temperature is used to correct the temperature dependence between the diffusion coefficient and the activation energy.

[0081] State equations and observation equations. The electrical part uses a first-order equivalent circuit: the capacitor branch voltage decays according to the current integral, and the ohmic internal resistance voltage is linearly related to the current. The electrochemical part, based on Fick's diffusion law and empirical formulas for interface film growth, treats the diffusion coefficient and interface film resistance as slowly changing quantities, adding them to the state equations as zero-mean process noise. The observed quantity is the battery terminal voltage, sampled by the inverter and sent to the digital twin system. The state equations and observation equations conform to the extended Kalman filter premise: the system is approximately linear near the state.

[0082] The extended Kalman filter recursion is as follows: The filter consists of two steps: prediction and correction. The prediction step maps the estimated state vector from the previous time step to the current time step and updates the covariance; the correction step adjusts the state vector using the real-time measured battery terminal voltage. The core formula is:

[0083]

[0084]

[0085]

[0086]

[0087]

[0088] In the above expression, To predict the state vector, To correct the state vector; It is the covariance matrix; Let the state be the Jacobian matrix; To observe the Jacobian matrix; and These are the process noise covariance and the observation noise covariance, respectively. Kalman gain; For measuring voltage; and These are the state equation and the observation equation, respectively. The filter passes through... Dynamic adjustment maps the measurement residual to the increments of the diffusion coefficient and the interfacial film resistance, enabling online recursion of the aging amount.

[0089] Temperature-dependent correction. The diffusion coefficient is highly sensitive to temperature; this invention utilizes the temperature field averaging temperature... Arrhenius type correction to the diffusion coefficient:

[0090]

[0091] in This is the temperature-corrected diffusion coefficient. Indicates the diffusion activation energy. The gas constant is... The reference temperature is used. The corrected diffusion coefficient is used in the next cycle of filtering prediction, allowing the model to naturally adapt to seasonal or intraday temperature changes.

[0092] Aging parameter output. After the filter stabilizes, when the estimation error variance of the diffusion coefficient and interface film resistance converges to within the threshold, the current mean value is taken as the aging parameter output, which is used by the global optimization model to limit the charge and discharge depth. The aging parameter update cycle can be consistent with the probabilistic scenario generation cycle, generally set to 30 minutes.

[0093] Performance Verification. Example 1 involved a 2 MWh lithium iron phosphate battery system operating continuously for 90 days. Comparing the diffusion coefficient and interfacial film resistance measured by offline impedance spectroscopy experiments, the mean square error of the online estimation using extended Kalman filtering remained within 10%. During high-temperature seasons, the system automatically reduced the estimated diffusion coefficient using a temperature correction formula, effectively limiting high-current discharge and reducing the overall capacity decay rate by 4% for the season. Example 2 was applied in a demonstration station in a low-temperature region. This model effectively tracked the decreasing trend of interfacial film resistance during the winter warming phase, dynamically adjusting the upper limit of the charging current in the energy storage system, ultimately achieving a cycle life prediction error of less than 5%.

[0094] This invention combines the advantages of equivalent circuits in facilitating real-time calculations with the physical interpretability of slow electrochemical variables. It utilizes the recursive nature of the extended Kalman filter to output aging parameters online, and employs a temperature correction formula to ensure consistency of estimation results under different climatic conditions. With real-time updates to the aging parameters, the energy storage charge-discharge strategy can adaptively reduce the degradation rate, extending battery life and improving system economy, demonstrating the application value of this electrochemical coupling model in integrated photovoltaic energy storage control.

[0095] Preferably, the spatiotemporal Gaussian process uses the linear embedded state data as the mean function and constructs a covariance structure in the form of the product of spatial kernel function and temporal kernel function to jointly sample the irradiation sequence, temperature sequence and photovoltaic power sequence to form a weighted probabilistic scene data set.

[0096] The spatiotemporal Gaussian process of this invention undertakes the task of expanding the linearly embedded state vector into a probabilistic scenario with multiple time periods and spatial points in the future. Compared with the traditional method of randomly sampling only a single irradiation sequence, this invention simultaneously models the three physical quantity sequences of irradiation, module temperature, and photovoltaic power. This allows for the simultaneous evaluation of power uncertainty and temperature uncertainty in subsequent optimization stages, avoiding internally contradictory scenarios such as "optimistic irradiance but pessimistic temperature". The following details the construction of the mean function, the design of the covariance structure, sampling and weight calculation, and the implementation effect.

[0097] The mean function is constructed, and the linear embedded state vector is generated by a Bayesian-Koopman mapping. It contains a set of state components evolving in an approximately linear space, including photovoltaic current, photovoltaic voltage, module average temperature, state of charge, and network frequency deviation. Since this vector has absorbed most external disturbances through least-squares estimation of historical data, its value can be directly used as the mean function of the spatiotemporal Gaussian process. Specifically, the embedded state vector is divided into irradiation, temperature, and power channels according to physical correspondences, and then the corresponding desired trajectory is obtained through linear combination. For example, the desired irradiation trajectory can be written as:

[0098]

[0099] in and These are two irradiation-related embedding components, and These are the calibration coefficients. The temperature and power channels both derive their mean functions using the same approach. The three channels are kept synchronized in time to ensure that the correspondence between different physical quantities in the scene is not disrupted.

[0100] In covariance structure design, traditional spatiotemporal Gaussian processes often directly add the spatial kernel and the temporal kernel. This invention employs a product structure to ensure that spatial correlation decays together with time over long time spans. Let the spatial vector be denoted as... The time difference is denoted as The spatial kernel function is chosen to be an isotropic squared exponential kernel, and the time kernel function is chosen to be an exponentially decaying kernel. The product of the two yields the covariance.

[0101]

[0102] For the overall variance, For spatially relevant length scales, The time-dependent length scale is used. Since this invention focuses on the average power of a single-row array or multiple arrays, the spatial dimension can be a one-dimensional distance or a two-dimensional planar distance; the time dimension is consistent with the sampling window of the Gaussian process. The product structure naturally reflects the fact that spatial correlation decays over time: the farther away the cloud shadow or the longer the time interval, the smaller its impact on the target position.

[0103] Sampling and weight calculation: The lower triangular approximation decomposition of the kernel matrix is ​​first calculated using variational Bayesian inference to reduce the overhead of inverting large covariance matrices. Then, 128 samples are sampled at once, each containing three sequences: irradiance, temperature, and power at a resolution of 30 seconds every two hours for the next two hours. The likelihood value is calculated for each sample and normalized to obtain the weight. The weight calculation formula is written as:

[0104]

[0105] in For the first A column vector formed by the difference between each sample and the mean function. This is the covariance matrix. The 8-bit fixed-point hardware implementation shows that the entire sampling and weighting process is completed within 100 microseconds. The obtained probability scenario data set is sorted by weight for use by the mutual information filtering module.

[0106] Example 1: The above-mentioned spatiotemporal Gaussian process was applied to a 10 MW centralized power plant in Ningxia. Compared with the baseline scheme that only uses the ARIMA model to predict irradiance, the sampled scene covers the real irradiance curve within the 95th percentile range, and the temperature and power sequences are also consistent within the same confidence interval, indicating that joint sampling effectively avoids inconsistencies between physical quantities.

[0107] Example 2: In the deployment of a distributed rooftop system in Zhejiang, the spatiotemporal Gaussian process successfully mapped the uncertainty of the typhoon cloud system in the two hours before its arrival into a high-variance scenario, prompting the optimization module to reduce the depth of discharge in advance, ultimately reducing the maximum temperature rise of the battery by 3 degrees Celsius during the typhoon.

[0108] High-dimensional super-symbol encoding and mutual information filtering are performed on the probabilistic scenario data. The optimal charge and discharge state is determined in the global optimization model in combination with the battery aging parameters to generate a reference power trajectory. The convex model predictive control outputs real-time power commands on the neuromorphic chip.

[0109] This invention, based on the already described uncertainties in probabilistic scenarios, still requires rapidly extracting the charge / discharge decisions most relevant to the current operating state from numerous combinations while also considering battery life constraints. To this end, a four-level chain is proposed: "high-dimensional super-symbolic encoding - mutual information filtering - quantum temperature variational optimization - neuromorphic convex model predictive control," which ensures coverage of the global search space while controlling the online computational load, enabling real-time power commands to be output in milliseconds.

[0110] The principle of high-dimensional supersymbol encoding is that each probabilistic scene data contains irradiance, temperature, and power sequences, and directly calculating similarity in the original dimensions is too time-consuming. This invention uses sparse random projection to map the scene sequence into a binary vector of length 10,000, called a supersymbol. The elements of the random projection matrix are independently symmetrically distributed, theoretically maintaining the monotonic relationship between the Hamming distance and the Euclidean distance of the original sequence. After binarization, vector similarity evaluation can be completed with only XOR and bit-counting hardware instructions, with a latency in the microsecond range. The linearly embedded state vector is also transformed using the same projection rule to obtain the query vector, ensuring encoding consistency.

[0111] Mutual information filtering mechanism. To ensure that the selected subset of scenarios both conforms to the current state and covers uncertainty, mutual information is used to measure the dependency between the linearly embedded state vector and each supersymbol. The formula for calculating mutual information is written as follows:

[0112]

[0113] in For super-symbol length, Indicates the first Scene No. The probability that the bit is the same as the query vector. This represents the average probability of matching at the same position across all scenarios. All can be obtained simultaneously via hardware XOR. The mutual information scores are sorted in descending order, and the top 32 scenarios are selected for further optimization, following the exponential temperature function:

[0114]

[0115] Assign weights, This represents the temperature coefficient. The weights are used to quantify the credibility of the scene.

[0116] A global optimization model combining aging parameters is employed. Aging parameters refer to the anode diffusion coefficient and the solid electrolyte membrane resistance, which together determine the permissible charge / discharge depth. The aging parameters and a subset of scenarios are constructed into a binary decision matrix, with decision variables indicating whether charging occurs every 30 minutes. The objective function minimizes the difference between power tracking error, thermal degradation cost, and electricity revenue, constraining the battery's state of charge to remain within a safe range during the recursion. This combination can be represented as a quadratic unconstrained binary optimization model. Since the dimensionality has been compressed to an acceptable range by mutual information filtering, quantum annealing yields fifty candidate solutions, followed by a variable-temperature random search to refine the temperature distribution and avoid local optima. The energy-minimum solution corresponds to the optimal charge / discharge state, and an eight-hour reference power trajectory is generated accordingly.

[0117] Neuromorphic convex model predictive control. The reference power trajectory and the linearly embedded state vector are fed into the neuromorphic chip. Internally, a spiking neural network implements the alternating direction multiplier method. The processor decomposes the quadratic objective function into a state error term and a power smoothing term, then iterates in parallel in the pulse domain. Because the linearly embedded space transition matrix satisfies a spectral radius less than one, the resulting optimization problem remains convex, converging after thirty iterations with an iteration period of 20 milliseconds. The output real-time power command includes both active and reactive components.

[0118] Implementation Results Examples. In a 10 MW centralized photovoltaic-2 MWh energy storage system in Ningxia, 64 rounds of sampling were conducted throughout the day. The average time for mutual information hardware screening was 8 microseconds, with negligible bus latency. The total time for quantum-temperature-dependent optimization was 0.3 milliseconds, and the neuromorphic solution took 1.5 milliseconds, with the entire closed-loop process taking less than 2 milliseconds. The average annual curtailment rate was reduced by 5% compared to the baseline rolling linear programming, and the predicted battery cycle life was extended by 3%. When deploying 50 data acquisition nodes in a distributed rooftop scenario, the average load of the hardware counting unit was less than 10%, demonstrating the linear scalability of the solution.

[0119] This invention maps complex spatiotemporal scenes to a hardware-friendly binary space through high-dimensional super-symbolic encoding, solving the bottleneck of online high-dimensional scene retrieval; mutual information filtering ensures that the scene subset is consistent with the real-time state; quantum-temperature-variable optimization combined with aging parameters balances benefits and lifetime; neuromorphic convex predictive control further shortens the latency and achieves millisecond-level power command output, significantly improving the dynamic response and economic value of photovoltaic-energy storage systems.

[0120] Preferably, the high-dimensional supersymbolic encoding maps each probability scene data into a fixed-length binary vector through random Gaussian projection, calculates the mutual information score between the binary vector and the linear embedded state data using a hardware operation of bit XOR addition, and selects a preset number of scenes to form the weighted scene subset according to the size of the mutual information score.

[0121] The goal of high-dimensional hypersymbolic encoding is to rapidly compress probabilistic scene data with temporal characteristics into a hardware-friendly binary representation, and then perform correlation evaluation with linearly embedded state vectors with constant-level computation. Probabilistic scene data consists of concatenated irradiance sequences, component temperature sequences, and photovoltaic power sequences, with original dimensions often exceeding thousands. Direct retrieval based on Euclidean distance or cosine similarity requires floating-point multiplication and addition, and suffers from dimensionality curse, leading to metric distortion. This invention employs random Gaussian projection combined with symbolization techniques to map each scene to a length... A binary vector is called a supersymbol. Random Gaussian projection is a locally preserving mapping, mathematically based on the Johnson-Lindenstrauss lemma: with a sufficiently long embedding dimension, the distances between points are approximately preserved with high probability after random orthogonal projection. Therefore, the choice matrix... The elements follow a zero-mean, Gaussian distribution. Multiply by the original scene vector By taking the sign, we can obtain the binary vector:

[0122]

[0123] in The component takes the value 1 or -1. Length A value of 10,000 is typically chosen, which can be stored on millimeter-scale silicon wafers.

[0124] Hardware XOR addition counting is based on the fact that the more matching bits in the bitwise XOR result of two binary vectors, the smaller the Hamming distance of the original vectors, and Gaussian projection guarantees a monotonic relationship between the Hamming distance and the Euclidean distance of the original sequence. Therefore, in the field, on-chip parallel XOR gates can be used to XOR the query vector with candidate supersymbols bitwise, and then a hardwired adder can be used to count the number of matching bits. The entire operation involves no multiplication, requires only logic gates, and typically has a latency of less than 10. To transform the matching bit length into mutual information, this invention treats the query vector as a random variable. Candidate supersymbols are treated as random variables. Mutual information is defined as:

[0125]

[0126] in Indicates the first position and The probability that both are 1 This represents the probability of both being -1. express In the The marginal probability of bit 1 express In the The marginal probability of a 1. Since the binary vector bit distribution follows a symmetric Bernoulli distribution, the marginal probability can be approximated as 0.5, and the mutual information can be linearly approximated by the number of matched bits, greatly simplifying the calculation.

[0127] A higher mutual information score indicates a greater relevance between the scene and the current state. This invention sorts all scenes according to their mutual information scores and selects the top... The bars are weighted subsets of the scenarios. The weights are given by the soft maximum function:

[0128]

[0129] in For the first The mutual information score for each scenario The temperature coefficient is used to adjust the steepness of the weight distribution. The weights monotonically increase with mutual information, which can be interpreted as an approximation of the posterior probability of the scene within a Bayesian framework.

[0130] When the weighted scene subset enters the global optimization model, it no longer carries the original high-dimensional sequence, but instead carries mutual information weights and scene numbers, achieving a compression ratio exceeding 1000:1. In actual testing, a centralized power plant samples 128 scenes at a time, and the hardware performs XOR addition and counting in parallel to complete all mutual information approximations, taking 8 seconds. The weights are then normalized and sorted in the central processing unit, taking 50 seconds. The entire screening cycle is less than 0.1 milliseconds.

[0131] The introduction of hypersymbol coding brings significant system-level benefits. Example 1: Twenty edge computing units were deployed within the coverage area of ​​a power station in the Ningxia desert, each responsible for scene retrieval for a 1-megawatt module. After enabling coding filtering, the floating-point core utilization of each unit's CPU decreased from 60% to 15%, saving nearly 45 watts of energy. This also stabilized the input dimension of the upper-level quantum annealing at 32 scenes and 16 control variables, preventing the quantum solver from degenerating into a local search due to dimensional explosion.

[0132] Example 2: Due to the complex cloud structure in the Zhejiang rooftop photovoltaic scenario, the original scenario library needs to be expanded to 256 entries; after using supersymbol encoding, the retrieval latency is still kept within 0.2 milliseconds, and the system can readjust the power command in real time under continuous fast cloud conditions to ensure that the grid frequency deviation is maintained within 0.05 Hz.

[0133] Besides its speed advantage, high-dimensional encoding also provides inherent privacy protection: the binary supersymbol is an irreversible mapping, making it impossible to directly reconstruct the original irradiance and power sequences, thus meeting the data isolation requirements for joint optimization across multiple sites. Furthermore, the random Gaussian projection matrix can be periodically rotated, further enhancing security.

[0134] In summary, this invention utilizes random Gaussian projection to transform high-dimensional probabilistic scenarios into sparse binary spaces. By approximating mutual information through hardware-level bit XOR and addition counting, it enables the selection of a small subset of the most representative scenarios from massive scenarios within milliseconds. This provides low-dimensional, information-rich input for subsequent hybrid quantum optimization and neuromorphic predictive control, balancing computational speed, energy consumption, and system privacy. This significantly improves the feasibility and economy of integrated photovoltaic energy storage control in real-world complex scenarios.

[0135] Preferably, the global optimization model discretizes the weighted scenario subset and the battery aging parameters into a binary charge-discharge decision matrix to construct a quadratic unconstrained binary optimization model. Candidate solutions are obtained through a quantum annealing solver. Then, energy consumption values ​​are compared in a variable-temperature stochastic search framework, and the charge-discharge decision with the minimum energy consumption value is retained as the preferred charge-discharge state.

[0136] The global optimization model is central to the control chain of this invention. Its task is to provide a set of discrete charge-discharge decisions for the next 8-hour planning window, under the constraints of the selected weighted scenario subset and real-time battery aging parameters, thereby achieving a balance between power tracking error, cyclic aging cost, and electricity price benefits. To meet the timeliness requirements of online solution while maintaining global search capability for combinatorial explosion problems, this invention adopts a three-stage process: quadratic unconstrained binary optimization modeling, quantum annealing solution, and variable-temperature stochastic search refinement. The following description is divided into five parts: variable discretization, objective function construction, quantum annealing to generate candidate solutions, stochastic search to select optimal solutions, and implementation effects.

[0137] The variables are discrete, and the planning window is divided into 16 equal-length segments, each segment's length maintaining the same resolution as the probabilistic scenario, i.e., 30 minutes. For the... Each section is set with charging decision variables as follows: Indicates charging, let's assume This indicates either discharging or standby. The 16 decision variables are arranged in a column vector, denoted as... The advantages of discretization are twofold: firstly, it allows for direct modeling within a binary optimization framework; secondly, it enables quantum annealing hardware to map all decisions to qubits at once.

[0138] The objective function is constructed from three terms: power tracking error, cyclic degradation cost, and electricity price benefit. The power tracking error term measures the difference between the reference power trajectory and the actual power in the scenario; the cyclic degradation cost term is determined by battery aging parameters; and the electricity price benefit term considers the peak-valley electricity price difference. For ease of quantum annealing, all three terms are transformed into quadratic polynomials, and the overall expression is as follows:

[0139]

[0140] In the formula, It is an upper triangular weight matrix. This is a linear weight vector. The power error term is factored into the matrix using the scene weights from the weighted scene subset. Cyclic degradation cost maps battery aging parameters to diagonal elements according to segment depth; electricity price revenue term is then entered into the vector. .

[0141] Quantum annealing generates candidate solutions. The model, encoded as an unconstrained binary optimization, can be directly mapped to quantum annealing hardware. Each decision variable is bound to a physical qubit, and the matrix... The upper triangular elements are mapped to two-bit coupled energy. The mapping is a single-bit bias. This invention uses an annealing time of 20 microseconds, reading 500 sets of samples at a time, and obtaining 500 sets of binary vectors at the hardware clock layer. Since the energy distribution of the quantum annealer follows the Boltzmann distribution, the top 50 solutions with the lowest energy are naturally concentrated near the local or global optimum, and enter the next stage as a set of candidate solutions.

[0142] Even with refined variable-temperature random search, quantum annealing solutions may still be affected by hardware embedding or noise, remaining at local low points. Therefore, a variable-temperature random search framework is employed to further screen candidate solutions. First, an initial temperature value is set based on the initial energy of the candidate solutions. The temperature decays exponentially with each iteration, allowing for a larger energy increase in the early stages of the search and a more greedy approach later on. Each iteration involves random flipping. If one or two values ​​are selected, a new objective function value is calculated. If the energy decreases, the solution is accepted unconditionally; otherwise, the acceptance probability is determined by the Metropolis criterion based on temperature. The temperature-changing process is executed in parallel on the central processing unit and the graphics processing unit, updating 50 candidate solutions independently, completing the process within 200 microseconds. Finally, the solution vector with the lowest energy is retained, denoted as... This is the preferred charging and discharging state.

[0143] Generate a reference power trajectory and a real-time control interface. After the optimal solution is determined, the decision vector is... Combined with the reference power trajectory, the charging or discharging power level of each segment is filled in to obtain an 8-hour reference power curve. This curve also needs to limit the maximum allowable power and capacity according to battery aging parameters; at the same time, if the scenario weight alarm indicates that there is high irradiation uncertainty in a certain segment, the curve will reserve a power reserve band in that segment for real-time control compensation. The complete curve is passed to the neuromorphic chip, which runs convex model predictive control within a 20-millisecond cycle, compares the curve with the measured linear embedded state vector, calculates the real-time power command for the next cycle, and sends it to the inverter.

[0144] In an implementation example, in a 10 MW centralized power plant in central Ningxia, this invention sets the target window to 8 hours and the decision dimension to 16. Quantum annealing consumes less than 10 millijoules of energy per run, and the variable-temperature random search takes 220 microseconds, a negligible percentage compared to hardware overhead. Comparative experiments with traditional mixed-integer linear programming show that the quantum-plus-variable-temperature method achieves an average energy level less than 1% higher than the global optimum within the same algorithm window, while reducing computation time by two orders of magnitude, meeting the requirements for real-time rolling optimization. The cyclic aging cost of the energy storage system after six months of operation is reduced by approximately 4%, the curtailment rate is reduced by approximately 5%, and peak-valley arbitrage profits are increased by approximately 6%.

[0145] In the distributed scenario of rooftop photovoltaic testing in a Zhejiang industrial park, the scenario library was expanded to 256 scenarios, and the input dimension remained fixed at 32 scenarios after hardware screening. The number of quantum annealing mapping bits was stabilized at 16 plus coupling redundancy, and the bit utilization rate remained above 75%, proving that the invention has good scalability.

[0146] Preferably, the convex model predictive control uses the linear embedded state data as the system state and the reference power trajectory as the target trajectory. Under linear constraints, it uses the alternating direction multiplier method to iteratively solve the optimization problem to obtain real-time power commands, and re-solves the optimization problem in each control cycle based on the updated state.

[0147] The convex model predictive control is located at the final stage of the control chain in this invention, used to refine the 8-hour reference power trajectory into millisecond-level executable real-time power commands. This module runs on a neuromorphic chip, which internally uses an array of spiking neurons to implement alternating direction multiplier iterations, completing the optimization solution within a 20-millisecond control cycle. Since the previous stage has mapped the nonlinear physical process into a linear embedded state space, this stage can write the predictive model as a linear discrete state equation, and construct a quadratic objective function and a first-order inequality constraint within this framework to ensure that the optimization problem remains convex in each cycle.

[0148] The system state is taken from the linearly embedded state vector, denoted as The control input vector is set to... These correspond to the active power command and reactive power command for energy storage, respectively. The prediction model adopts a first-order linear form:

[0149]

[0150] matrix Describe the uncontrolled evolution of the system, matrix The projection effect of the control action in the embedding space is described. The two matrices are updated online using historical data, and the spectral radius is guaranteed to be less than 1.

[0151] The reference power trajectory window is 8 hours, but only the next 15 steps are used as the rolling prediction target sequence in each control cycle. ,in The cost function is:

[0152]

[0153] matrix Trade-off state error, matrix Balance control smoothness. Limit the rate of power change using a linear inequality:

[0154]

[0155] in The slope coefficient is denoted as . The upper and lower limits of the state of charge are transformed into corresponding inequalities through the linear relationship between the embedding space and the state of charge.

[0156] The alternating direction multiplier method splits the objective function into a state error subproblem and a control increment subproblem, and couples them with multiplier vectors. Each iteration includes three steps: state update, control update, and multiplier update, all of which are matrix multiplication and addition, and can be performed in parallel within a spiking neural network. An upper limit of 30 iterations is set, but typically 20 iterations are sufficient to meet the residual threshold. With a chip clock frequency of 200 Hz, corresponding to a single-cycle computation time of 5 milliseconds, the total iteration time is less than 150 milliseconds, and the 85th percentile time is approximately 120 milliseconds.

[0157] Field experiments show that in the 10 MW photovoltaic-2 MWh energy storage system in Ningxia, when cloud shadows cause a power step jump, the neuromorphic solver version issues revision instructions within two control cycles, with an actual inverter delay of 60 milliseconds. If a central processing unit floating-point solver is used instead, the same problem requires 300 milliseconds. The peak power tracking error has been reduced from 2% to 0.8%.

[0158] Improved control smoothness is reflected in the energy storage current curve: peak battery current is reduced by 30%, and the equivalent cycle depth is reduced by 4 percentage points. Extrapolating from 10 months of data, this can extend battery life by approximately one year.

[0159] The power reliability lower bound constraint uses the 85th percentile scenario power as the random lower bound. In extreme cloud shadow scenarios, the model advances the instruction, reducing the probability of frequency outage from 0.3% to 0.02%.

[0160] In summary, convex model predictive control maintains the convexity of the problem by utilizing the linear embedding space, and implements the alternating direction multiplier method in parallel with neuromorphic hardware, outputting power commands in the millisecond level. By improving the stability of energy storage current and grid frequency through smoothing constraints and stochastic lower bounds, it provides high-precision, low-latency final-stage control assurance for photovoltaic energy storage integrated systems.

[0161] like Figure 3 As shown, each inverter, based on the real-time power command and local state vector, generates a modulation signal through pheromone game and self-distillation reinforcement learning to control the power converter to perform charging and discharging. When the network frequency deviation reaches the threshold, it triggers supercapacitor discharge support and returns according to the power slope. The execution data is written to the distributed ledger at fixed intervals to update the digital twin system and global optimization model.

[0162] Each inverter is at the end of the control link and needs to convert the real-time power command issued by the previous convex model predictive control into a kilohertz-level drive pulse width modulation signal. This invention adopts a two-stage strategy of "pheromone game + self-distillation reinforcement learning" to maintain the consistency of group power distribution and enable individual inverters to have fast adaptive capabilities.

[0163] The pheromone game principle encodes the active power delivery error and reactive power delivery error of each inverter as active power pheromone concentration and reactive power pheromone concentration, respectively. The concentration decays over time and diffuses within the local network, mathematically expressed as:

[0164]

[0165]

[0166] in and They represent the first The active and reactive pheromone concentrations of the inverter. and Representing the first The active and reactive power performance errors of the inverter. This represents the set of neighborhoods connected via a CAN-FD ring network. The evaporation coefficient is... For injection coefficients, The diffusion coupling coefficient is denoted by . Pheromones diffusion enables a mechanism of "high-error devices releasing more signals - neighbors voluntarily yielding power to compensate", which can prevent a single inverter from being overloaded for a long time.

[0167] The self-distillation reinforcement learning strategy uses a deep deterministic policy gradient algorithm as the basic controller for each inverter. The state vector consists of direct-axis voltage, quadrature-axis voltage, module temperature, and state of charge. and Composition. To overcome parameter differences in distributed scenarios, this invention periodically extracts the most recent 5 minutes of data from a local replay pool, causing the current student network to mimic the teacher's strategy with the best average performance in the past. The loss function is the Euclidean distance between the outputs of the two networks. Distillation combined with the model-independent element learning outer loop enables the strategy to recover its optimal action within 10 steps after sudden changes in cloud cover or a surge in battery temperature, meeting the millisecond-level retraining requirements without the need for a centralized server.

[0168] The modulation signal is generated, and the policy network outputs a normalized modulation amount, which is mapped to the duty cycle of the fully controlled bridge. To avoid current spikes, this invention adds a single-cycle 0.03 rise limit at the hardware layer; if the pheromone concentration remains high for a long time, the policy network automatically reduces the slope of the duty cycle rising edge, forming a soft power reduction collective behavior.

[0169] For supercapacitor bypass support, when the absolute value of the network frequency deviation measured during grid connection exceeds 0.05 Hz, the inverter immediately sends a bypass trigger signal. The supercapacitor is connected in parallel to the DC bus through a high-voltage contactor, and the discharge current is limited to twice the rated current for 0.3 seconds. After the support ends, it backs to the real-time power command at a slope of 0.1 times the rated power per second. This process is entirely completed in the inverter firmware, ensuring a trigger delay of less than 2 milliseconds.

[0170] The distributed ledger write and feedback process involves the inverter concatenating a timestamp, output active power, output reactive power, and active pheromone concentration, performing a hash operation, and writing the result to the Hyperledger Fabric channel every 50 milliseconds. The ledger uses RAFT consensus, packaging blocks every 100 milliseconds. The digital twin system and the global optimization model synchronously pull execution logs after a new block is confirmed, used to correct weight drift in the liquid neural network and re-estimate the initial quantum temperature, achieving cross-layer adaptation.

[0171] In an example, during 720 hours of operation in a 10 MW centralized power plant, pheromone game theory reduced the average thermal stress of a single inverter by 15% and the standard deviation of group power fluctuation by 12%. The supercapacitor rapidly supported 28 triggers, reducing the maximum frequency deviation by 40%. A total of 520,000 ledger records were accumulated, and the digital twin showed no missing frames on the insertion side. In a 5 MW distributed scenario in Zhejiang, with 80 inverters deployed, the CAN-FD ring network load was 20%, and the pheromone frame round-trip latency was 0.8 milliseconds, demonstrating the solution's good scalability.

[0172] Preferably, each inverter broadcasts its active and reactive pheromone concentrations to neighboring inverters via a pheromone diffusion-evaporation game. The updated pheromone concentrations and local state vectors are input into a policy network trained by self-distillation reinforcement learning to obtain a modulation signal to control the bidirectional power converter to perform charging and discharging. When the network frequency deviation reaches a threshold, the supercapacitor bypass discharge is triggered and returned according to the power slope. At the same time, the timestamp data, output active power data, output reactive power data, and active pheromone concentration data are hashed and written into a distributed ledger at a fixed period. The execution data stored in the distributed ledger is used to synchronously update the digital twin system and the global optimization model.

[0173] The pheromone diffusion-evaporation game aims to enable multiple inverters to cooperate in power transfer within milliseconds, avoiding long-term overload of a single unit. Each inverter maintains two continuous variables: active pheromone concentration. With reactive pheromone concentration Both broadcast within the local ring network via a proximity communication interface, which uses CAN-FD with a 64-bit data frame length and a round-trip latency of 0.8 milliseconds. Concentration dynamics follow a three-step mechanism: evaporation, injection, and diffusion.

[0174]

[0175]

[0176] Evaporation terms cause old information to quickly become obsolete, injection terms map current performance pressures to local concentrations, and diffusion terms promote concentration equilibrium within the neighborhood. When the error of a device increases... The error perception of the local unit is increased, and then the neighbor's error perception is increased through diffusion, thereby inducing the neighbor to reduce output or give up reactive power support, thus achieving distributed load balancing.

[0177] The inverter's state vector consists of six dimensions: direct-axis voltage, quadrature-axis voltage, module temperature, state of charge, active pheromone concentration, and reactive pheromone concentration. The self-distillation reinforcement learning policy network takes the state vector as input and outputs a normalized modulation quantity. This is mapped to the two-phase pulse width duty cycle of the bidirectional power converter. The policy network is first pre-trained offline using a deep deterministic policy gradient approach, and then knowledge distillation is performed on-site every 10 minutes: the most recent 3000 steps are selected from the local experience replay pool, and the old policy with the highest average return in the past is used as the teacher. The student network is fine-tuned by minimizing the mean square loss of the output difference, so that the policy recovers convergence within ten steps after parameter drift or environmental abrupt changes. Distillation keeps the network capacity small, making it suitable for real-time inference on the inverter's on-chip processor, with a single forward pass requiring only 30 microseconds.

[0178] To support the inertia gap during large disturbances, the inverter also integrates a supercapacitor bypass support circuit. When the absolute value of the grid frequency deviation exceeds 0.05 Hz and lasts for 20 milliseconds, the firmware triggers a relay to connect the supercapacitor bank in parallel to the DC bus, with the discharge current set to twice the rated current, lasting for 0.3 seconds. After support, the power command smoothly returns to the real-time power trajectory at a slope of 0.1 times the rated power per second, avoiding secondary overshoot. The entire detection, triggering, and de-activation process is completed locally on the inverter, with an average trigger delay of less than 2 milliseconds.

[0179] The inverter generates a performance record every 50 milliseconds, containing a timestamp, output active power, output reactive power, and active pheromone concentration. The record is hashed using SHA-256 and written to a Hyperledger Fabric channel. Block packaging occurs every 100 milliseconds, employing RAFT consensus. The digital twin system pulls new blocks, reconstructs the inverter's actual output in chronological order, and uses this information for self-supervised training of the liquid neural network while simultaneously updating the initial quantum annealing temperature, ensuring the model remains consistent with the execution layer.

[0180] In an example, a 10 MW power plant in Ningxia deployed 20 inverters, and the pheromone communication ring network operated for 720 hours. The average active pheromone concentration varied with solar radiation, forming a sawtooth distribution consistent with the radiation intensity. The highest single-unit concentration peak decreased by 40%, indicating that the load was more evenly distributed among the multiple units.

[0181] In Zhejiang, a 5 MW rooftop system with 80 inverters was deployed. When the typhoon cloud cluster passed by quickly, the frequency protection was triggered 11 times, the supercapacitor discharge duration was maintained for a total of 3.3 seconds, and the maximum frequency deviation decreased from 0.06 Hz to 0.036 Hz.

[0182] A total of 5.2 million execution log entries were written to the two blockchains, with no block loss or rollback events. The digital twin model automatically adjusted the weights of the liquid neural network based on log errors, reducing the mean square error of heat source power density prediction by 18% compared to the initial model.

[0183] like Figure 4 As shown, a photovoltaic energy storage integrated charge and discharge control system is used to implement the aforementioned photovoltaic energy storage integrated charge and discharge control method. The system includes:

[0184] The acquisition and fusion module collects signals from photovoltaic modules, energy storage battery cabinets, grid connection interfaces, and the sky-end. It unifies timestamps using a deterministic time-sensitive network and performs noise reduction processing, integrating the event visual stream into cloud shadow contours to generate fused data. The acquisition and fusion module (edge ​​layer) has the following sensing components: Hall current transformers and voltage divider sampling boards are embedded in each string of photovoltaic modules; state of charge meters and strain gauges are arranged in the energy storage battery cabinet; synchronous phasor measurement units are used at the grid connection point; and event visual cameras and CMOS cloud image cameras are installed at the sky end. On the communication side, each node connects to a TSN switch via a 1Gbps silicon photonics Ethernet connection. The switch has built-in IEEE 802.1Qbv time slot scheduling logic; the time reference is provided by the PTP-TC timing module. On the processing side, the switch connects to a Zynq Ultrascale+FPGA for downlink processing. The FPGA performs sliding median filtering, drift correction, and performs 5ms integration of the event stream in the on-chip BRAM to generate cloud shadow contours, which are then combined with electrical frames to form fused data.

[0185] The twin generation module is used to input the fused data into a liquid neural-thermal power and electrochemical coupled digital twin system to generate heat source power density, battery aging parameters, and linear embedding state data. Based on the linear embedding state data, a spatiotemporal Gaussian process is used to generate probabilistic scene data. The twin generation module (edge-cloud collaborative layer) computing hardware consists of a dual-socket EPYC server equipped with an RTX A6000 GPU; the GPU deploys a liquid neural network inference kernel and a two-dimensional explicit thermally conductive CUDA kernel; the CPU runs an extended Kalman filter thread. Data channel: The fused data is transmitted to the server via 10GbE fiber, and historical sequences are cached using NVMe SSDs. Gaussian process acceleration: 48GB of GPU memory is used for sparse kernel approximation and batch Cholesky decomposition, with a sampling time of <1ms for 128 scenes.

[0186] The scene optimization module performs high-dimensional super-symbol encoding and mutual information filtering on the probabilistic scene data. Combined with the battery aging parameters, it determines the optimal charge / discharge state in the global optimization model, generates a reference power trajectory, and runs a convex model predictive control on the neuromorphic chip to output real-time power commands. Scene optimization module (acceleration layer) encoding and filtering: Intel Agilex FPGA board implements 10,000-bit super-symbol XOR counting, with mutual information calculation completed in 8µs. Global optimization: D-WaveAdvantage quantum annealing maps 16-bit QUBOs in the cloud once, annealing for 20µs to output 50 solutions; local AMD MI250 GPU performs variable-temperature random search refinement. MPC hardware: Loihi2 neuromorphic chip loads an alternating direction multiplier pulse network, iterating 30 times in a 20ms period and outputting real-time power commands.

[0187] The execution feedback module is used by each inverter to generate modulation signals based on the real-time power command and local state vector through pheromone game theory and self-distillation reinforcement learning to control the power converter to perform charging and discharging. When the network frequency deviation reaches a threshold, it triggers supercapacitor discharge support and returns according to the power slope. The execution data is written to the distributed ledger at fixed intervals to update the twin generation module and the scene optimization module. The execution feedback module (field layer) performs the following: Power conversion: Each 250kVA energy storage inverter is equipped with a SiC fully controlled bridge and a 1MHz PWMDSP; the DSP runs a lightweight strategy network forward and outputs the modulation ratio. Cooperative communication: 1Mbps pheromone frames are broadcast via a CAN-FD ring network between inverters; pheromone updates and action inference are completed within 300µs in the C2000 core. Inertia support: A 5kJ supercapacitor bank is connected in parallel to the DC bus; the IGBT bypass switch trigger delay is <2ms. Ledger nodes: Raspberry Pi4 acts as a light node for signing, JetsonOrin acts as a sorting node to run the Hyperledger FabricRAFT channel, and blocks are packaged in 100ms; new block events are pushed back to the server via gRPC to refresh the digital twin and optimize parameters.

[0188] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A charge-discharge control method for photovoltaic energy storage integration, characterized by, The method comprises the following steps: Collecting multi-modal signals of a photovoltaic module, an energy storage battery cabinet, a grid-connected interface and a sky end, and performing deterministic time-sensitive network unified timestamp denoising, integral event visual flow cloud shadow profile generation and fusion data acquisition; Inputting the fusion data into a liquid neural-thermal power and electrochemical coupling digital twin system to generate heat source power density, battery aging parameter and linear embedding state data, the linear embedding state data being a linear embedding state vector generated by a Bayesian-Koopman mapping and containing photovoltaic current, photovoltaic voltage, module average temperature, state of charge and network frequency deviation state component; and generating probability scenario data based on the linear embedding state data using a spatio-temporal Gaussian process, the spatio-temporal Gaussian process taking the linear embedding state data as a mean function, constructing a covariance structure in the form of a product of a spatial kernel function and a temporal kernel function, and jointly sampling irradiance sequence, temperature sequence and photovoltaic power sequence to form a set of probability scenario data with weights; the digital twin system comprising an electrochemical coupling model; Implementing high-dimensional hyper-symbolic coding and mutual information screening on the probability scenario data, determining an optimal charging and discharging state in a global optimization model in combination with the battery aging parameter, generating a reference power trajectory, running a convex model predictive control on a neuromorphic chip to output real-time power instructions, the convex model predictive control taking the linear embedding state data as system state and the reference power trajectory as target trajectory, iteratively solving an optimization problem under linear constraint conditions using an alternating direction multiplier method to obtain real-time power instructions, and re-solving the optimization problem according to the updated state at each control period; Each inverter generates a modulation signal to control a power converter to perform charging and discharging according to the real-time power instructions and a local state vector, the local state vector being a state vector of the inverter and consisting of six-dimensional quantities of direct-axis voltage, quadrature-axis voltage, module temperature, state of charge, active pheromone concentration and reactive pheromone concentration; each inverter broadcasts active pheromone concentration and reactive pheromone concentration to neighboring inverters through pheromone diffusion-evaporation gambling, inputs the updated pheromone concentration and the local state vector into a policy network trained by self-distilled reinforcement learning to obtain a modulation signal to control a bidirectional power converter to perform charging and discharging; an ultracapacitor discharge support is triggered when a network frequency deviation reaches a threshold value and is returned at a power slope, an ultracapacitor bypass support is triggered when a grid-connected network frequency deviation absolute value exceeds 0.05 Hz, and the support is returned to the real-time power instructions at a slope of 0.1 times rated power per second after the support ends; and execution data is written into a distributed ledger at a fixed period to update the digital twin system and the global optimization model.

2. The method of claim 1, wherein, The acquisition step synchronizes the event stream output by the event visual sensor with the current signal acquired by the photovoltaic module, the voltage signal acquired by the photovoltaic module, the state of charge signal acquired by the energy storage battery cabinet, the network frequency signal acquired by the grid-connected interface, and the cloud image acquired by the sky terminal, generates a cloud shadow profile from the event visual stream in a sliding integral manner, and then timestamps the cloud shadow profile and the remaining signals through a deterministic time-sensitive network, and sequentially performs sliding median filtering and drift correction in the edge processor to obtain the fusion data.

3. The method of claim 1, wherein, The liquid neural-thermal power coupling model maps the cloud shadow profile to a heat source power density, writes the heat source power density as a source term into a two-dimensional heat conduction equation, and solves the temperature field by coupling the photovoltaic module current signal and the photovoltaic module voltage signal, and the temperature field result is used to correct the state component related to thermal characteristics in the linear embedded state data.

4. The method of claim 1, wherein, The electrochemical coupling model adopts an equivalent circuit combined with electrochemical kinetics modeling method, constructs a state vector using the state of charge signal and the temperature field result acquired by the energy storage battery cabinet, and recursively calculates the diffusion coefficient and the interface film resistance through extended Kalman filtering to generate the battery aging parameter.

5. The method of claim 1, wherein, The high-dimensional hyper-symbol encoding maps each probability scenario data to a fixed-length binary vector through random Gaussian projection, calculates the mutual information score between the binary vector and the linear embedded state data through bit XOR and counting hardware operation, and selects a preset number of scenarios to form a weighted scenario subset according to the mutual information score.

6. The method of claim 5, wherein, The global optimization model discretizes the weighted scenario subset and the battery aging parameter into a binary charging and discharging decision matrix to construct a quadratic unconstrained binary optimization model, obtains a candidate solution through a quantum annealing solver, compares the energy consumption values in a variable-temperature random search framework, and retains the charging and discharging decision with the smallest energy consumption value as the preferred charging and discharging state.

7. The method of claim 1, wherein, When the network frequency deviation reaches a threshold value, the ultracapacitor bypass discharge support is triggered and returned at a power slope, and the timestamp data, output active power data, output reactive power data, and active pheromone concentration data are hashed and written into a distributed ledger at a fixed period. The execution data saved in the distributed ledger is used to synchronize the digital twin system and the global optimization model.

8. A charge-discharge control system for photovoltaic energy storage integration, for implementing the charge-discharge control method for photovoltaic energy storage integration according to any one of claims 1-7, characterized in that, The system comprises: An acquisition fusion module for acquiring photovoltaic module signals, energy storage battery cabinet signals, grid-connected interface signals, and sky terminal signals, and for timestamping and performing denoising processing through a deterministic time-sensitive network to integrate the event visual stream into a cloud shadow profile to generate fusion data; A twin generation module for inputting the fusion data into a liquid neural-thermal power and electrochemical coupling digital twin system to generate heat source power density, battery aging parameter, and linear embedded state data, and for generating probability scenario data based on the linear embedded state data using a spatiotemporal Gaussian process; A scenario optimization module is configured to implement high-dimensional hyper-symbolic coding and mutual information screening on the probability scenario data, determine an optimal charging and discharging state in a global optimization model in combination with the battery aging parameter, generate a reference power trajectory, and run a convex model predictive control to output real-time power instructions on a neuromorphic chip; An execution feedback module is configured to generate a modulation signal to control a power converter to perform charging and discharging according to the real-time power instructions and a local state vector, through pheromone game and self-distillation reinforcement learning, trigger super capacitor discharge support and return according to a power slope when a network frequency deviation reaches a threshold, and write execution data into a distributed ledger at a fixed period to update the twin generation module and the scenario optimization module.

Images