Photovoltaic energy storage integrated charging and discharging control method and system

Abstract

The invention relates to the technical field of photovoltaic energy storage optimization control, in particular to a photovoltaic energy storage integrated charging and discharging control method and system, and the method comprises the steps: collecting a multi-mode signal, and carrying out the synchronization and denoising of a time-sensitive network, and then carrying out the fusion; generating heat source power density, battery aging parameters and linear embedding state data by using a liquid nerve-thermal power and electrochemical coupled digital twin system; on the basis of a space-time Gaussian process sampling probability scene of the state data, high-dimensional super-symbol coding and mutual information screening compression scene set, determining an optimal charging and discharging state in a quantum-variable temperature global optimization model by combining aging parameters, and generating a reference power trajectory; the neuromorphic chip operates convex model prediction control to output a real-time power instruction; and the inverter executes charging and discharging through pheromone game and self-distillation reinforcement learning, triggers the super capacitor support when the frequency is abnormal, and executes data writing-in distributed account book closed-loop updating. According to the invention, millisecond-level power tracking is realized, and cyclic stress is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic energy storage optimization control, and in particular to a charging and discharging control method and system for integrated photovoltaic energy storage. Background Art

[0002] Photovoltaic power generation is subject to rapid fluctuations due to cloud shadows. If the energy storage system cannot compensate within seconds, this will lead to grid frequency deviations and accelerated battery degradation. Existing technologies generally use two approaches: first, rolling linear programming based on a single irradiance forecast, using conventional cameras to capture frame-based cloud images and perform power planning in the central processing unit; second, using an offline impedance model to assign fixed charge and discharge depths to batteries. The former, with a cloud image update cycle of ≥1s, cannot capture millisecond-level sudden drops in sunlight, resulting in delayed power prediction; the latter ignores real-time aging and temperature coupling, preventing the current peak from converging in time, resulting in excessive cycle depth and reduced battery life. Summary of the Invention

[0003] To address the numerous issues with the aforementioned existing technologies, the present invention provides a charge-discharge control method and system for integrated photovoltaic energy storage. This method uses event visual streams to capture cloud shadows, a liquid neural network to estimate the power density of additional heat sources, and couples this with an electrical model to generate an embedded state. A spatiotemporal Gaussian process samples the scene, compresses the dimensions through high-dimensional encoding and mutual information filtering, and then determines the binary charge-discharge matrix through quantum annealing and variable temperature search. A neuromorphic chip outputs power commands within a convex model prediction framework. The inverter uses pheromone concentration to trigger ultracapacitor support when the frequency is abnormal, providing feedback on the data link. This method achieves millisecond-level power tracking, cycle depth minimization, and model self-evolution.

[0004] A photovoltaic energy storage integrated charge and discharge control method comprises the following steps: Multimodal signals from photovoltaic panels, energy storage battery cabinets, grid-connected interfaces, and the sky terminal are collected, and unified timestamps are used for denoising via a deterministic time-sensitive network. Cloud shadow outlines are generated by integrating event visual streams to obtain fused data. Inputting 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 embedded state data, and generating probabilistic scenario data based on the linear embedded state data using a spatiotemporal Gaussian process; Implementing high-dimensional hypersymbolic encoding and mutual information screening on the probabilistic scenario data, determining the optimal charge and discharge state in a global optimization model in combination with the battery aging parameters, generating a reference power trajectory, and running a convex model predictive control on a neuromorphic chip to output real-time power instructions; Each inverter generates a modulation signal based on the real-time power instruction and local state vector through pheromone game and self-distillation reinforcement learning to control the power converter to perform charging and discharging. When it detects that the network frequency deviation reaches the threshold, the ultracapacitor discharge support is triggered and returns according to the power slope. The execution data is written to the distributed ledger at a fixed period to update the digital twin system and the global optimization model.

[0005] Preferably, the acquisition step synchronizes the event stream output by the event vision sensor with the current signal collected by the photovoltaic module, the voltage signal collected by the photovoltaic module, the charge state signal collected by the energy storage battery cabinet, the network frequency signal collected by the grid-connected interface, and the cloud image collected by the sky end. The event vision stream is used to generate a cloud shadow outline in a sliding integration manner, and then the cloud shadow outline is unified with the remaining signals through a deterministic time-sensitive network, and then sliding median filtering and drift correction are performed in sequence in the edge processor to obtain the fused data.

[0006] Preferably, the liquid neural-thermal power coupling model maps the cloud shadow contour into heat source power density, writes the heat source power density as a source term into the two-dimensional heat conduction equation, and couples it with the photovoltaic module current signal and the photovoltaic module voltage signal to solve the temperature field. The temperature field result is used to correct the state component related to the thermal characteristics in the linear embedded state data.

[0007] Preferably, the electrochemical coupling model adopts a joint modeling method of equivalent circuit and electrochemical kinetics, constructs a state vector using the state of charge signal and temperature field results collected by the energy storage battery cabinet, and recursively calculates the diffusion coefficient and interface film resistance through extended Kalman filtering to generate the battery aging parameters.

[0008] Preferably, the spatiotemporal Gaussian process uses the linear embedded state data as the mean function, constructs a covariance structure in the form of the product of the spatial kernel function and the temporal kernel function, and jointly samples the irradiation sequence, temperature sequence and photovoltaic power sequence to form a weighted probability scenario data set.

[0009] Preferably, the high-dimensional super-symbol encoding maps each probability scene data into a binary vector of fixed length through random Gaussian projection, calculates the mutual information score between the binary vector and the linear embedded state data by a hardware operation method of bitwise XOR plus counting, and selects a preset number of scenes according to the size of the mutual information score to form the weighted scene subset.

[0010] Preferably, the global optimization model discretizes the weighted scenario subset and the battery aging parameter into a binary charge and discharge decision matrix, constructs a quadratic unconstrained binary optimization model, obtains candidate solutions through a quantum annealing solver, and then compares energy consumption values ​​in a variable temperature random search framework and retains the charge and discharge decision with the smallest energy consumption value as the preferred charge and discharge state.

[0011] 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, and uses the alternating direction multiplier method to iteratively solve the optimization problem under linear constraints to obtain real-time power instructions, and re-solves the optimization problem according to the updated state in each control cycle.

[0012] Preferably, each inverter broadcasts the active pheromone concentration and reactive pheromone concentration to neighboring inverters through the pheromone diffusion-evaporation game, and inputs the updated pheromone concentration and the local state vector into the 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 it is detected that the network frequency deviation reaches the threshold, the ultracapacitor bypass discharge support is triggered and returned according to the power slope, and 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 the distributed ledger according to a fixed period, and the execution data stored in the distributed ledger is used to synchronously update the digital twin system and the global optimization model.

[0013] A photovoltaic energy storage integrated charge and discharge control system, used to implement the photovoltaic energy storage integrated charge and discharge control method, the system comprising: The acquisition and fusion module collects signals from photovoltaic modules, energy storage battery cabinets, grid-connected interfaces, and the sky unit. It unifies timestamps and performs denoising via a deterministic time-sensitive network, integrating the event visual stream into cloud shadow outlines to generate fused data. a twin generation module, configured 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 embedded state data, and to generate probabilistic scenario data based on the linear embedded state data using a spatiotemporal Gaussian process; a scenario optimization module for performing high-dimensional hypersymbolic encoding and mutual information screening on the probabilistic scenario data, determining the optimal charge and discharge state in a global optimization model in combination with the battery aging parameters, generating a reference power trajectory, and running a convex model predictive control on a neuromorphic chip to output real-time power instructions; The execution feedback module is used for each inverter to generate a modulation signal based on the real-time power instruction and the local state vector through pheromone game and self-distillation reinforcement learning to control the power converter to perform charging and discharging. When it is detected that the network frequency deviation reaches the threshold, the ultracapacitor discharge support is triggered and returned according to the power slope, and the execution data is written into the distributed ledger according to a fixed period to update the twin generation module and the scenario optimization module.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are: The present invention uses event vision-liquid neural network to quickly map cloud shadow contours into heat source density, achieving millisecond-level irradiation prediction; uses equivalent circuit-extended Kalman filtering to output diffusion coefficient and membrane resistance in real time, realizing dynamic lifetime constraint; uses random Gaussian projection and mutual information screening to quickly reduce scene dimensions and realize microsecond-level related scene retrieval; uses quantum annealing-variable temperature search to obtain the global optimal charge and discharge matrix and minimize energy costs; uses convex model predictive control on the pulse neural chip to achieve 20ms instruction refresh; uses pheromone diffusion game and self-distillation reinforcement learning to keep the inverter group smooth and balance power; and uses distributed ledger to re-inject execution logs to achieve model closed-loop self-update. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the process of the present invention; Figure 2 This is a structural diagram of the digital twin coupling model in the present invention; Figure 3 This is the collaborative logic diagram of the pheromone game in the present invention; Figure 4 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0016] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure.

[0017] like Figure 1 As shown, a photovoltaic energy storage integrated charging and discharging control method includes the following steps: Multimodal signals from photovoltaic panels, energy storage battery cabinets, grid-connected interfaces, and the sky terminal are collected, and unified timestamps are used for denoising via a deterministic time-sensitive network. Cloud shadow outlines are generated by integrating event visual streams to obtain fused data. The integrated photovoltaic-energy storage charging and discharging control chain places extremely high demands on the temporal accuracy and physical integrity of input data. This invention, through a three-step design consisting of "multimodal synchronous acquisition - deterministic time-sensitive network alignment - and event-based cloud shadow extraction," transforms heterogeneous information from the photovoltaic, energy storage, grid-connected, and airside systems into a single, fused data source. This provides consistent, low-latency, and traceable input for subsequent digital twin modeling and predictive control.

[0018] The multimodal data collection system utilizes current and voltage sensors installed on photovoltaic panels, state-of-charge sensors and triaxial strain gauges on energy storage battery cabinets, synchronized phasor measurement devices at grid-connected interfaces, and event vision sensors and cloud imaging terminals at locations with high solar altitudes. The event vision sensors output brightness change events with microsecond triggering, enabling them to capture sudden changes in illumination before the edge of a cloud shadow completely covers the battery panels. The combined signals simultaneously reflect the electrical status of the power generation, energy storage, and grid terminals, as well as sky illumination dynamics, ensuring comprehensive fusion data.

[0019] Unified timestamps are achieved through deterministic time-sensitive networking (TSN). TSN uses full-duplex Gigabit Ethernet at the physical layer and schedule-based gating at the link layer. The sending windows of all acquisition nodes are referenced to a global clock. This ensures that the jitter of each frame arriving at the edge switch is limited to microseconds, regardless of the distance between nodes. The edge switch inserts nanosecond timestamps into all frames before forwarding them to the edge processor, ensuring that data from different sources shares a common timeline for subsequent computations.

[0020] The edge processor first performs a sliding median filter on the current, voltage, state of charge, strain and phasor data to remove impulse noise, and then corrects temperature drift and baseline drift with a 4-hour sliding average. For event visual flow, the fixed integration window is used. Count pixel-level events. Assume that the positive polarity event count is , the negative polarity event count is , the pixel coordinates are , the brightness threshold is , then the cloud shadow outline matrix is ​​given by the formula:

[0021] Generate. Here is the pixel label, with a value of 1 indicating that an occluded edge has been detected. A 5-millisecond integration window balances the speed of cloud shadow motion and the amount of data. The edge processor implements streaming operations through sliding accumulation, without introducing significant latency.

[0022] The fused data consists of timestamps, PV current, PV voltage, module backplane temperature, energy storage state of charge (SOC), energy storage strain, grid-connected phasors, grid frequency, cloud imagery, and the aforementioned cloud shadow outline frames. The cloud shadow outlines provide millisecond-level irradiance change priors for the predictive control module. The SOC and strain accurately reflect battery health, while grid phasors and frequency deviations reflect load dynamics at the grid connection point. By using a unified sampling window, all fields are fully aligned in the time domain, eliminating common errors introduced by asynchronous multi-source data.

[0023] In a 10 MW PV-2 MWh energy storage prototype test, the introduction of event vision-cloud shadow contours reduced the mean squared error of the 2-second PV active power forecast by approximately 20% compared to using only traditional forecasts. Rapid predictions improved the foresight of battery discharge compensation, reducing the peak ramp rate of the energy storage inverter by approximately 30% in a PV power dip scenario, validating the effectiveness of fused data in mitigating power surges. Another implementation deployed 50 data collection nodes in a mountainous distributed power station, using an 8-node ring topology for 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.

[0024] Preferably, the acquisition step synchronizes the event stream output by the event vision sensor with the current signal collected by the photovoltaic module, the voltage signal collected by the photovoltaic module, the charge state signal collected by the energy storage battery cabinet, the network frequency signal collected by the grid-connected interface, and the cloud image collected by the sky end. The event vision stream is used to generate a cloud shadow outline in a sliding integration manner, and then the cloud shadow outline is unified with the remaining signals through a deterministic time-sensitive network, and then sliding median filtering and drift correction are performed in sequence in the edge processor to obtain the fused data.

[0025] The rapid fluctuation of photovoltaic output power is mainly caused by cloud shadows, and traditional irradiance prediction methods are difficult to capture the instantaneous changes of cloud shadow edges at the second scale. The present invention deploys an event vision sensor at the sky end. The event vision sensor has microsecond time resolution and self-triggering output characteristics, and can immediately record pixel coordinates and change polarity when the brightness change exceeds the threshold. Compared with frame cameras, event vision sensors do not require a fixed frame rate, so they can capture brightness changes in real time before the cloud shadow reaches the photovoltaic array and generate a high-frequency event stream. The event stream is recorded as a quadruple ,in and is the pixel coordinate, is the brightness change polarity, In order to maintain sampling consistency with the electrical signal, the present invention selects the fixed time window The event stream is integrated internally to generate cloud shadow outline frames.

[0026] The integration process uses pixel counting method: the number of positive polarity events at the same pixel position in a window is counted Counting negative polarity events , calculate the brightness difference .like Exceeding the threshold , then mark the corresponding pixels as cloud shadow edges and get the cloud shadow contour matrix:

[0027] in A value of 1 indicates that there is a cloud edge at the pixel, and a value of 0 indicates that the brightness change at the pixel is not enough to constitute an edge.

[0028] In parallel with event visual data, the photovoltaic panel side collects current and voltage signals, the energy storage battery cabinet side collects state-of-charge signals and triaxial strain signals, the grid-connected interface side collects network frequency signals, and the sky-side cloud imaging terminal collects visible light cloud images. These multi-source signals require strict alignment to serve as synchronous inputs for the digital twin model. This invention uses a deterministic time-sensitive network to connect each acquisition node to the edge processor. The deterministic time-sensitive network uses schedule-driven gating at the data link layer. Each transmit queue opens at a preset time slot, ensuring that all nodes transmit data frames within the same period. The network uses a high-precision synchronized clock as the global time reference. Edge switches insert nanosecond-level timestamps into each incoming frame, ensuring microsecond-scale synchronization errors between data at different nodes. Experiments show that under Gigabit Ethernet physical layer conditions, the 95th percentile of the end-to-end jitter distribution of an 8-node ring topology does not exceed 2 microseconds.

[0029] After timestamp alignment, the edge processor first performs a sliding median filter on the current, voltage, state of charge, strain, and network frequency to remove high-frequency impulse noise. Then, a sliding mean subtraction is used to eliminate temperature and baseline drift. After generation, the cloud shadow outline frame enters the subsequent process directly without undergoing low-pass filtering to preserve the spatial accuracy of the edge. All fields after noise removal are framed along a common timeline to form fused data. The fused data fields include: PV current, PV voltage, module backsheet temperature, energy storage state of charge, energy storage strain, grid voltage phasor, grid frequency deviation, cloud image, and cloud shadow outline.

[0030] The advantages of fused data lie in two key aspects. First, cloud shadow contours provide forward-looking information on irradiance changes on a timescale of seconds, enabling proactive guidance of energy storage discharge to compensate for sudden drops in PV output. Second, state of charge and strain provide feedback on battery health on the same timescale, avoiding state estimation bias caused by asynchronous sampling. By synchronizing rapid cloud shadow detection with electrical measurements, this method achieves fully coupled observations across the PV, energy storage, grid, and sky systems, providing complete boundary conditions for the digital twin model.

[0031] The application effect is verified through two examples. In the test of the joint operation of a 10MW ground-based centralized photovoltaic power station and a 2MWh energy storage system, the integral window is set. The event counting threshold is 5 milliseconds Calibrated before sunrise by an adaptive algorithm. Compared with the baseline solution that only uses irradiance prediction, the mean square error of the 2-second ultra-short-term photovoltaic power prediction is reduced by 20% after adding cloud shadow outlines. During the same test cycle, the peak ramp rate of the energy storage inverter in the photovoltaic output sudden drop scenario decreased by 30%, indicating that the forward-looking information effectively alleviated the power shock. Another embodiment deploys 50 acquisition nodes in a mountain distributed photovoltaic system. Each node sends a 1 kHz sampling frame through a deterministic time-sensitive network. The edge switch inserts a hardware timestamp and forwards it to the processor. The verification results show that the cumulative frame loss rate of the fused data is less than 0.01% during 48 hours of continuous operation, and the variance of the end-to-end synchronization error is maintained at the order of 1 microsecond, meeting the digital twin system's requirements for high-precision synchronization.

[0032] like Figure 2 As shown, the fusion data is input into the digital twin system of liquid neural-thermal power and electrochemical coupling to generate heat source power density, battery aging parameters and linear embedded state data, and based on the linear embedded state data, probabilistic scenario data is generated using a spatiotemporal Gaussian process; The digital twin system of this invention simultaneously depicts the thermal-power process of photovoltaic modules and the electrochemical decay process of energy storage batteries. Its core consists of three components: 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.

[0033] The liquid neural network-thermal power coupling model is used to convert cloud shadow contours into instantaneous heat source power density on the back of photovoltaic modules. The liquid neural network is a recursive network with continuous time differential equations inside the nodes, which has the ability to remember non-stationary excitation in real time. is the hidden state vector, is the input pixel count vector, then the node dynamics is written as:

[0034] in is the leakage coefficient, is the input weight matrix, is the recursive weight matrix. The matrix is ​​read out linearly Get the heat source power density:

[0035] In the formula The unit is watts per square meter. Write the two-dimensional heat conduction-power coupling equation as the source term to solve the component temperature distribution in real time The method is coupled with the current and voltage signals of the photovoltaic modules to calculate the output power. This method avoids the lag problem of traditional radiation-thermal coupling models that rely on cloud cover prediction. Experiments have shown that the temperature field prediction delay can be controlled within 50 milliseconds.

[0036] The electrochemical coupling model is based on the Doyle-Fuller-Newman dynamic framework. The charge concentration, solution concentration, electrode potential, and electrolyte potential are selected as state variables, and the diffusion coefficient is estimated by iterative extended Kalman filtering. and interface film resistance , which together constitute the battery aging parameters. This real-time estimation result is not only used to limit the allowable current in the charge and discharge strategy, but also provides a state variable for life optimization.

[0037] In order to map the nonlinear high-dimensional multimodal data into a linear subspace suitable for convex model predictive control, the system first performs polynomial-triangular mapping on the fused data, and then uses Bayesian-Koopman linearization to obtain the linear embedded state data. The mapping result is recorded as , its evolution equation approximately satisfies:

[0038] in is the embedding space transfer matrix, is Gaussian white noise. The dimension is fixed and the dynamics are approximately linear. The subsequent model predictive control can be written as a quadratic programming problem, and the solution complexity is linearly related to the time window.

[0039] based on The system constructs a mean function and uses a spatiotemporal Gaussian process to sample the combined radiation and temperature processes for future time periods. The kernel function is the product of a spatial radial basis kernel and a temporal exponential decay kernel to ensure spatial smoothness and decreasing correlation over time. Variational inference allows for the simultaneous sampling of multiple probabilistic scenario trajectories, each with a probability weight. This preserves the uncertainty of the prediction while enabling rapid calculation of the correlation between the scenario and the real-time state during subsequent mutual information screening.

[0040] Combining these three components yields heat source power density, battery aging parameters, and linearly embedded state data. The results are as follows: First, the liquid neural network directly injects cloud shadow information into the thermal field equation, enabling component temperature prediction to precede cloud shadow arrival by approximately 120 milliseconds; second, the extended Kalman filter provides real-time aging parameters, enabling the global optimization model to dynamically adjust loop depth; and third, the linearly embedded state data accelerates the solution of model predictive control, reducing single-step iteration time from milliseconds to sub-milliseconds.

[0041] Example 1: The 10 MW photovoltaic-2 MWh energy storage system runs for 48 hours. The generated probabilistic scenario data is updated every 30 seconds. The scenario set contains 128 trajectories, and the mutual information screening takes less than 100 microseconds. Compared with the baseline method that does not use linear mapping, the number of model predictive control iterations is reduced by 60%, and the backup power call is advanced by 150 milliseconds when the photovoltaic output drops rapidly. Example 2: 50 nodes are arranged in the distributed power station, the liquid neural network uses 256 hidden nodes, and the Bayesian-Koopman model uses a 128-dimensional embedding space. The actual CPU utilization rate is maintained below 40%, verifying the deployability of the solution in the edge computing environment. In summary, the present invention realizes flexible charging and discharging control for uncertain light through the combination of digital twins and probabilistic scenario generation, providing a high-efficiency, high-precision and economical solution path for deep collaboration between photovoltaics and energy storage.

[0042] Preferably, the liquid neural-thermal power coupling model maps the cloud shadow contour into heat source power density, writes the heat source power density as a source term into the two-dimensional heat conduction equation, and couples it with the photovoltaic module current signal and the photovoltaic module voltage signal to solve the temperature field. The temperature field result is used to correct the state component related to the thermal characteristics in the linear embedded state data.

[0043] The liquid neural-thermal power coupling model is a key innovation in the digital twin layer of this invention. Its function is to rapidly convert cloud shadow outlines obtained by event vision into the instantaneous additional heat source power density of the photovoltaic module. This heat source power density is then integrated into the two-dimensional heat conduction equation along with the resistance loss power density calculated from the current and voltage signals to solve the temperature distribution in real time. The global average of the temperature distribution is then used as a correction factor for the thermal dimension in the linearly embedded state vector, ensuring that the subsequent convex model predictive control remains within the linear solvable range.

[0044] The mapping principle of cloud shadow outline to heat source power density. When the edge of a cloud shadow passes over the photovoltaic array, the event vision sensor triggers a large number of brightness change events. The model sparsely encodes the binary cloud shadow outline matrix within an integration window into a fixed-length vector and feeds it into the liquid neural network. The internal state of the liquid neural network node follows a leaky first-order differential equation, and its continuous-time characteristics can preserve the high temporal resolution of the event stream. The hidden state vector is denoted as , the input vector is denoted as , the node attenuation coefficient is recorded as The node dynamics are:

[0045] In the formula is the input weight matrix, is the recursive weight matrix, is the hyperbolic tangent function. By linearly reading out the matrix The additional heat source power density is obtained:

[0046] in The unit is watts per square meter. During the training phase, a time-truncated backpropagation method is used to update the weights, so that the network can be used for online inference after convergence on multi-day cloud shadow samples.

[0047] Next, we will explain the method of coupling the two-dimensional heat conduction equation with the electrical power. Assume that the discrete step length in the dimensional direction of the photovoltaic module is and , the node coordinates are Indicates that the time step is The heat-power equation is discretized as:

[0048] For the The node temperature at the moment, is the thermal conductivity, is the density, is the specific heat capacity, is the additional heat source power density, is the resistance loss power density. The resistance loss power density is first calculated to calculate the component DC power , then subtract the inverter output power and normalize it according to the component area to get the value. The present invention executes the discrete equation in parallel in the graphics processing unit, and the single frame iteration time is less than milliseconds, matching the sampling period. After the temperature distribution is obtained, the spatial average value of the backplane temperature is taken. The linear embedding state vector is denoted as , among which The components represent temperature deviations, and the present invention adopts:

[0049] Make a replacement, and The effect of this is to linearly map the nonlinear temperature field changes into state vector increments, so that the embedding space maintains approximately linear dynamics and ensures that the quadratic programming assumption of convex model predictive control holds.

[0050] The practical value of this model is demonstrated in the examples. On the inspection day of the 10MW power station in Hami, Xinjiang, fast-moving convective clouds appeared in the sky. The liquid neural-thermal power coupling model controls the temperature field update delay to 50 milliseconds, which is 60% shorter than the gray box model. At the same time, the energy storage system adjusts the discharge power in advance according to the updated state vector. When the photovoltaic power drops, the peak value of the system frequency deviation decreases from the baseline of 0.05 Hz to 0.03 Hz, verifying the gain of the thermal-power coupling input on the frequency support. Another example deploys 50 acquisition nodes in a mountain distributed power station. The liquid neural network can operate at a scale of 256 nodes, and the overall central processing unit occupancy rate of the system is maintained below 40%, indicating that the algorithm load of the scheme is suitable for edge deployment.

[0051] In summary, the liquid neural-thermal power coupling model, driven by event visual contours, generates temperature fields and corrects embedded states within milliseconds, achieving a real-time closed-loop for thermal, electrical, and cloud shadow dynamics. This mechanism significantly improves the response speed and robustness of subsequent scenario prediction and control planning, and has replicable and scalable application value in industrial and commercial energy storage scenarios.

[0052] Preferably, the electrochemical coupling model adopts a joint modeling method of equivalent circuit and electrochemical kinetics, constructs a state vector using the state of charge signal and temperature field results collected by the energy storage battery cabinet, and recursively calculates the diffusion coefficient and interface film resistance through extended Kalman filtering to generate the battery aging parameters.

[0053] The electrochemical coupling model proposed in this paper aims to online identify internal diffusion processes and interfacial film growth in energy storage batteries, thereby providing real-time aging parameters for subsequent global optimization models. This model utilizes a combined modeling approach using equivalent circuits and electrochemical kinetics. This model incorporates the time-domain characteristics of the voltage response into the equation of state while retaining the physical meaning of anode diffusion and solid electrolyte film growth. Its workflow can be summarized as follows: state vector construction, extended Kalman filter recursion, and aging parameter output.

[0054] State vector construction. The acquisition module already provides the energy storage battery 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 a parallel capacitor branch voltage are first selected as electrical state variables. Secondly, two slow variables in electrochemical kinetics are introduced: the anode active particle diffusion coefficient and the solid electrolyte membrane resistance. Ultimately, a comprehensive state vector is formed, which includes both electrical and electrochemical subvectors:

[0055] in represents the capacitor branch voltage, Indicates the voltage drop corresponding to the ohmic internal resistance, Indicates the battery current, represents the anode diffusion coefficient, represents the solid electrolyte membrane resistance. The state of charge signal is used to constrain the feasible range of the diffusion coefficient, and the average temperature of the temperature field is used to correct the temperature dependence of the diffusion coefficient and activation energy.

[0056] The state equation and observation equations are described. The electrical component 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 component, based on Fick's diffusion law and empirical formulas for interfacial film growth, treats the diffusion coefficient and interfacial film resistance as slowly varying quantities and incorporates them into the state equation as zero-mean process noise. The observation variable is the battery terminal voltage, sampled by the inverter, and transmitted to the digital twin system. The state equation and observation equations meet the extended Kalman filter premise: the system can be approximately linear near the state.

[0057] Extended Kalman filter recursion. The filter consists of two steps: prediction and correction. The prediction step maps the state vector estimated at the previous moment to the current moment and updates the covariance; the correction step adjusts the state vector using the real-time measured battery terminal voltage. The core formula is:

[0058]

[0059]

[0060]

[0061]

[0062] In the above expression, is the predicted state vector, is the correction state vector; is the covariance matrix; is the state Jacobian matrix; is the observation Jacobian matrix; and are the process noise covariance and observation noise covariance respectively; is the Kalman gain; To measure voltage; and are the state equation and the observation equation respectively. The filter is passed Dynamic adjustment maps the measurement residual to the increment of diffusion coefficient and interface film resistance to achieve online recursion of aging amount.

[0063] Temperature dependent correction. The diffusion coefficient is highly sensitive to temperature. The present invention uses the average temperature of the temperature field Make an Arrhenius-type correction to the diffusion coefficient:

[0064] in is the temperature-corrected diffusion coefficient, represents the diffusion activation energy, is the gas constant, The modified diffusion coefficient is used in the next cycle filtering prediction, so that the model can naturally adapt to seasonal or daily temperature changes.

[0065] Aging parameter output. After the filter stabilizes and the estimated error variance of the diffusion coefficient and interface film resistance converges to within a threshold, the current mean 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 probability scenario generation cycle, generally set to 30 minutes.

[0066] Effect verification. Example 1: A 2-MWh lithium iron phosphate battery system was operated continuously for 90 days. The diffusion coefficient and interface film resistance measured by the offline impedance spectroscopy experiment were compared. The mean square error of the extended Kalman filter online estimation was kept within 10%. The system automatically reduced the estimated diffusion coefficient through the temperature correction formula in the high temperature season, timely limited the large current discharge, and reduced the capacity attenuation rate for the entire season by 4%. Example 2 was applied in a demonstration station in a low temperature area. This model has a good tracking effect on the downward trend of the interface film resistance during the winter warming stage. The energy storage system maintains dynamic adjustment of the upper limit of the charging current, and ultimately achieves a cycle life prediction error of less than 5%.

[0067] This method combines the real-time computational advantages of equivalent circuits with the physical interpretability of slow electrochemical variables. It leverages the recursive nature of the extended Kalman filter to output aging parameters online, and uses a temperature correction formula to ensure consistent estimation results under varying climate conditions. With the aging parameters updated in real time, the energy storage charge and discharge strategy can adapt to the decay rate, extending battery life and improving system economics. This demonstrates the application value of this electrochemical coupling model in integrated photovoltaic energy storage control.

[0068] Preferably, the spatiotemporal Gaussian process uses the linear embedded state data as the mean function, constructs a covariance structure in the form of the product of the spatial kernel function and the temporal kernel function, and jointly samples the irradiation sequence, temperature sequence and photovoltaic power sequence to form a weighted probability scenario data set.

[0069] The spatiotemporal Gaussian process of this invention is responsible for expanding the linearly embedded state vector into probabilistic scenarios across multiple future time periods and spatial points. Compared to traditional methods that randomly sample only a single irradiance sequence, this invention jointly models three physical quantity sequences: irradiance, module temperature, and photovoltaic power. This allows for simultaneous evaluation of power and temperature uncertainties in subsequent optimization steps, avoiding internal contradictions such as "optimistic for light but pessimistic for temperature." The following details the construction of the mean function, the design of the covariance structure, sampling and weight calculation, and the implementation results.

[0070] The mean function is constructed, and the linear embedded state vector is generated by the Bayesian-Koopman mapping, which contains a set of state components that evolve in an approximately linear space, such as photovoltaic current, photovoltaic voltage, average module temperature, state of charge, network frequency deviation, etc. Since this vector has absorbed most external disturbances through the least squares estimation of historical data, its value can be directly used as the mean function of the spatiotemporal Gaussian process. The specific approach is: the embedded state vector is divided into irradiation channel, temperature channel and power channel according to the physical correspondence, and then the corresponding expected trajectory is obtained through linear combination. For example, the expected irradiation trajectory can be written as:

[0071] in and are the two embedded components related to irradiance, and is the calibration coefficient. The mean function for both the temperature and power channels follows the same principle. The three channels are synchronized in time to ensure that the corresponding relationships between different physical quantities in the scene are not disrupted.

[0072] Covariance structure design: Traditional spatiotemporal Gaussian processes often directly add the spatial kernel and the temporal kernel. This invention adopts a product structure to keep the spatial correlation decaying over time over a long time span. Let the spatial vector be , the time difference is recorded as The spatial kernel function selects the isotropic square exponential kernel, and the temporal kernel function selects the exponential decay kernel. The product of the two is the covariance:

[0073] is the overall variance, is the spatial correlation length scale, is the temporal correlation length scale. Since this paper focuses on the average power of a single-row array or multiple arrays, the spatial dimension can be either one-dimensional distance or two-dimensional planar distance; the temporal dimension remains consistent with the sampling window of the Gaussian process. The product structure naturally reflects the fact that spatial correlation decays over time: the farther the cloud shadow or the longer the time interval, the less impact it has on the target location.

[0074] For sampling and weight calculation, we first compute the lower triangular approximate decomposition of the kernel matrix using the variational Bayesian inference method to reduce the overhead of inverting the large covariance matrix. We then sample 128 samples at a time, each containing three series of irradiance, temperature, and power for the next two hours at a 30-second resolution. For each sample, we calculate the likelihood and normalize it to obtain a weight. The weight calculation formula is:

[0075] in For the The column vector consisting of the difference between the samples and the mean function, is the covariance matrix. 8-bit fixed-point hardware implementation demonstrates that sampling and weighting are completed in 100 microseconds. The obtained probability scenario data set is sorted by weight and then used by the mutual information filtering module.

[0076] In Example 1, the above-mentioned spatiotemporal Gaussian process is applied to a 10-megawatt centralized power station in Ningxia. Compared with the baseline scheme that only uses the ARIMA model to predict irradiation, the sampled scenarios cover the real irradiation curve within the 95th percentile range, and the temperature and power series are also consistent within the same confidence interval, indicating that joint sampling effectively avoids inconsistencies between physical quantities.

[0077] In Example 2, a distributed rooftop system was deployed in Zhejiang. The spatiotemporal Gaussian process successfully mapped the uncertainty two hours before the arrival of the typhoon cloud system into a high-variance scenario, prompting the optimization module to reduce the discharge depth in advance, ultimately reducing the maximum temperature rise of the battery by 3 degrees Celsius during the typhoon.

[0078] Implementing high-dimensional hypersymbolic encoding and mutual information screening on the probabilistic scenario data, determining the optimal charge and discharge state in a global optimization model in combination with the battery aging parameters, generating a reference power trajectory, and running a convex model predictive control on a neuromorphic chip to output real-time power instructions; While probabilistic scenarios already describe uncertainty, this invention still requires rapidly extracting from numerous combinations the charging and discharging decisions most relevant to the current operating state while also taking into account battery life constraints. To this end, a four-stage chain of "high-dimensional hypersymbolic encoding - mutual information screening - quantum temperature variational optimization - neuromorphic convex model predictive control" is proposed. This ensures global search space coverage while controlling online computational load, enabling real-time power command output within milliseconds.

[0079] According to the principle of high-dimensional hypersymbol coding, each probabilistic scene data contains an irradiation sequence, a temperature sequence, and a power sequence. It takes too much time to calculate the similarity directly in the original dimension. The present invention uses sparse random projection to map the scene sequence into a binary vector of length 10,000, called a hypersymbol. The elements of the random projection matrix are independently and symmetrically distributed, which can theoretically maintain the monotonic relationship between the Hamming distance and the Euclidean distance of the original sequence. After binarization, only XOR and bit-counting hardware instructions are required to complete the vector similarity evaluation, with a delay of microseconds. The linear embedded state vector is also transformed by the same projection rule to obtain the query vector to ensure coding consistency.

[0080] Mutual information screening mechanism. In order to make the selected scenario subset both consistent with the current state and covering uncertainty, mutual information is used to measure the degree of dependence between the linear embedded state vector and each hyper-symbol. The mutual information calculation formula is written as:

[0081] in is the super symbol length, Indicates the The scene in The probability that the bit is the same as the query vector, is the average probability of all scenarios matching at the same bit. All Sort the mutual information scores in descending order, take the first 32 scenarios for subsequent optimization, and use the exponential temperature function:

[0082] Assign weight, is the temperature coefficient. The weight is used to quantify the credibility of the scene.

[0083] A global optimization model that combines aging parameters. Aging parameters refer to the anode diffusion coefficient and the solid electrolyte membrane resistance, which together determine the allowable depth of charge and discharge. The aging parameters and the scenario subset are constructed into a binary decision matrix, with the decision variable marking whether to charge every thirty minutes. The objective function minimizes the difference between the power tracking error, the thermal decay cost, and the electricity price revenue, and constrains the battery state of charge to not exceed the safe range during recursion. This combination can be expressed as a quadratic unconstrained binary optimization model. Since the dimension has been compressed to an acceptable range by mutual information screening, quantum annealing is used to obtain fifty candidate solutions, and then variable temperature random search is used to refine the temperature distribution to avoid local optimality. The energy minimum solution corresponds to the preferred charge and discharge state, and an eight-hour reference power trajectory is generated based on this.

[0084] 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 is used to implement the alternating direction multiplier method. The processor decomposes the quadratic objective function into a state error term and a power smoothing term, which are then iterated in parallel in the pulse domain. Because the linear embedding space transfer matrix satisfies the spectral radius less than one, the resulting optimization problem maintains convexity. Convergence is achieved after thirty iterations, with an iteration period of 20 milliseconds. The output is a real-time power command containing both active and reactive components.

[0085] Implementation results: In a 10 MW centralized photovoltaic-2 MWh energy storage system in Ningxia, 64 sampling cycles were performed throughout the day. Mutual information hardware screening took an average of 8 microseconds, with negligible bus latency. Quantum-variable temperature optimization took a total of 0.3 milliseconds, neuromorphic solution took 1.5 milliseconds, and the entire link closed loop took less than 2 milliseconds. The annual average curtailment rate was reduced by 5% compared to the baseline rolling linear programming, and the battery cycle life prediction was extended by 3%. When 50 collection nodes were deployed in a distributed rooftop scenario, the average hardware counting unit load was less than 10%, demonstrating the linear scalability of the solution.

[0086] The present invention maps complex spatiotemporal scenes into a hardware-friendly binary space through high-dimensional hypersymbolic coding, thus solving the bottleneck of online high-dimensional scene retrieval; mutual information screening ensures that the scene subset is consistent with the real-time state; quantum-variable temperature optimization combined with aging parameters takes into account both benefits and lifespan; neuromorphic convex predictive control further shortens latency and achieves millisecond-level power command output, significantly improving the dynamic response and economic value of the photovoltaic-energy storage system as a whole.

[0087] Preferably, the high-dimensional super-symbol encoding maps each probability scene data into a binary vector of fixed length through random Gaussian projection, calculates the mutual information score between the binary vector and the linear embedded state data by a hardware operation method of bitwise XOR plus counting, and selects a preset number of scenes according to the size of the mutual information score to form the weighted scene subset.

[0088] The goal of high-dimensional hypersymbolic coding is to quickly compress probabilistic scenario data with time series characteristics into a hardware-friendly binary expression, and then complete the correlation evaluation between the linear embedded state vector with a constant level of computation. The probabilistic scenario data is composed of the irradiation sequence, the component temperature sequence and the photovoltaic power sequence, and the original dimension is often thousands; if it is directly retrieved based on Euclidean distance or cosine similarity, floating-point multiplication and addition are required, and the dimensionality disaster will also cause measurement distortion. The present invention uses random Gaussian projection combined with symbolization technology to map each scene to a length of The binary vector of is called a hypersymbol. Random Gaussian projection is a locality preserving mapping, and its mathematical basis is the Johnson–Lindenstrauss lemma: under sufficiently long embedding dimensions, the distance between points is approximately preserved with high probability after random orthogonal projection. Therefore, the matrix is ​​chosen The elements of obey zero mean, Gaussian distribution, when using Multiply by the original scene vector And taking the sign, we can get a binary vector:

[0089] in The components of are either 1 or -1. Ten thousand is usually chosen, which can be stored on millimeter-scale silicon wafers.

[0090] Hardware XOR counting is based on the fact that the more matching bits in the bitwise XOR result between two binary vectors, the smaller the Hamming distance of the original vectors. Gaussian projection guarantees a monotonic relationship between the Hamming distance and the Euclidean distance of the original sequence. Therefore, on-chip parallel XOR gates can be used to perform bitwise XOR of the query vector and the candidate super-symbols, and then a hard-wired adder is used to count the number of matching bits. The entire operation requires no multiplications, only logic gates, and a typical latency of less than 10 In order to convert the number of matching bits into mutual information, the present invention regards the query vector as a random variable , the candidate hypersymbols are considered as random variables , the mutual information is defined as:

[0091] in Indicates in On the position and The probability of being 1 at the same time, represents the probability of being -1 at the same time, express In the The marginal probability of a bit being 1, express In the The marginal probability of a bit being 1. Since the bit distribution of binary vectors follows a symmetric Bernoulli distribution, the marginal probability can be approximately 0.5, and the mutual information can be linearly approximated by the number of matching bits, greatly simplifying the calculation.

[0092] The higher the mutual information score, the more relevant the scene is to the current state. The strips are treated as weighted scene subsets. The weights are given by the softmax function:

[0093] in For the The mutual information score of the scene, is the temperature coefficient, which is used to adjust the steepness of the weight distribution. The weight increases monotonically with the mutual information and can be interpreted as an approximation of the posterior probability of the scene in a Bayesian framework.

[0094] When the weighted scenario subset enters the global optimization model, it no longer carries the original high-dimensional sequence, but carries the mutual information weight and scenario number, with a compression ratio of over 1000:1. In actual tests, a centralized power station samples 128 scenarios at a time, and the hardware XOR plus counting completes all mutual information approximations in parallel, which takes 8 The weight normalization and sorting are then completed in the central processing unit, which takes 50 seconds. ; The entire screening cycle is less than 0.1 millisecond.

[0095] The introduction of hypersymbolic encoding has brought significant system-level benefits. Example 1: 20 edge computing units were deployed within the coverage area of ​​the Ningxia Desert Power Station, with each unit responsible for scene retrieval of a 1-megawatt module. When encoding filtering was enabled, the floating-point core utilization of each unit's CPU dropped from 60% to 15%, saving nearly 45 watts of energy. It also stabilized the input dimension of the upper-layer quantum annealing at 32 scenarios and 16 control variables, preventing the quantum solver from degenerating into a local search due to dimensionality explosion.

[0096] Example 2: Due to the complex cloud system structure of Zhejiang rooftop photovoltaic scenarios, the original scenario library needs to be expanded to 256 items. After using hypersymbol encoding, the retrieval delay is still maintained within 0.2 milliseconds. The system can readjust power instructions in real time under continuous fast cloud conditions to ensure that the grid frequency deviation is maintained within 0.05 Hz.

[0097] In addition to its speed advantage, high-dimensional encoding also provides inherent privacy protection: the binary hypersymbols are irreversible, making it impossible to directly restore the original irradiance and power sequences, thus meeting the data isolation requirements for multi-site joint optimization. Furthermore, the random Gaussian projection matrix can be periodically rotated, further enhancing security.

[0098] In summary, the present invention uses random Gaussian projection to transform high-dimensional probabilistic scenarios into sparse binary spaces. Through hardware-level bitwise XOR and adder counting to approximate mutual information, it is possible to screen a small number of the most representative subsets from massive scenarios within milliseconds, providing low-dimensional, information-rich input for subsequent hybrid quantum optimization and neuromorphic predictive control, while taking into account computing speed, energy consumption and system privacy, and significantly improving the feasibility and economy of integrated photovoltaic energy storage control in real complex scenarios.

[0099] Preferably, the global optimization model discretizes the weighted scenario subset and the battery aging parameter into a binary charge and discharge decision matrix, constructs a quadratic unconstrained binary optimization model, obtains candidate solutions through a quantum annealing solver, and then compares energy consumption values ​​in a variable temperature random search framework and retains the charge and discharge decision with the smallest energy consumption value as the preferred charge and discharge state.

[0100] The global optimization model is located at the core of the control link of the present invention. Its task is to provide a set of discrete charge and discharge decisions for the next 8-hour planning window under the joint constraints of the selected weighted scenario subset and the real-time battery aging parameters, so as to achieve a balance between power tracking error, cycle aging cost and electricity price benefits. In order to meet the timeliness of online solution while maintaining the global search capability for combinatorial explosion problems, the present invention adopts a three-stage process of quadratic unconstrained binary optimization modeling, quantum annealing solution and variable temperature random search refinement. The following is an explanation of the five parts of variable discretization, objective function construction, quantum annealing to generate candidate solutions, random search to screen the best solution and implementation effect.

[0101] The variable is discrete, and the planning window is divided into 16 equal-length segments, with the length of each segment being consistent with the probability scenario resolution, i.e., 30 minutes. The charging decision variables for each section are set as Indicates charging, Indicates discharge or standby. The 16 decision variables are arranged into column vectors and denoted as The benefits of discretization are: on the one hand, it can be directly modeled in the binary optimization framework, and on the other hand, it enables quantum annealing hardware to map all decisions to quantum bits at once.

[0102] The objective function is constructed by a power tracking error term, a cyclic decay cost term, and an electricity price benefit term. The power tracking error term measures the difference between the reference power trajectory and the actual power of the scenario; the cyclic decay cost term is determined by the battery aging parameters; and the electricity price benefit term considers the difference between peak and valley electricity prices. To facilitate quantum annealing, all three terms are converted into quadratic polynomial form, which can be written as follows:

[0103] In the formula, is the upper triangular weight matrix, is a linear weight vector. The power error term is converted into the matrix by the scene weight in the weighted scene subset ; Cycle decay cost maps battery aging parameters to diagonal elements according to segment depth; the electricity price benefit term enters the vector .

[0104] Quantum annealing generates candidate solutions. After the model is encoded as an unconstrained binary optimization, it can be directly mapped to quantum annealing hardware. Each decision variable is bound to a physical quantum bit, and the matrix The upper triangular elements of are mapped to the two-bit coupling energy, Mapped to a single-bit bias. This invention uses an annealing time of 20 microseconds, reading 500 samples at a time, and obtaining 500 binary vectors at the hardware clock level. Because the quantum annealer's energy distribution conforms to the Boltzmann distribution, the top 50 lowest-energy solutions naturally cluster near the local or global optimum and enter the next stage as candidate solutions.

[0105] When the variable temperature random search is refined, the quantum annealing solution may still be affected by hardware embedding or noise and stay at a local low point. Therefore, a variable temperature random search framework is used to re-screen the candidate solutions. First, the initial temperature value is set according to the initial energy of the candidate solution. The temperature decays exponentially with iterations, so that the search accepts a large energy increase in the early stage and tends to be greedy in the later stage. Each iteration randomly flips The new objective function value is calculated. If the energy decreases, it is accepted unconditionally. Otherwise, the acceptance probability is determined by the Metropolis criterion based on the temperature. The temperature change process is executed in parallel on the central processing unit and the graphics processing unit, and 50 candidate solutions are updated independently. The process is completed within 200 microseconds, and the solution vector with the lowest energy is finally retained, which is recorded as , as the preferred charge and discharge state.

[0106] Generate reference power trajectory and real-time control interface, after the optimal solution is determined, the decision vector Combined with the reference power trajectory, the charging or discharging power level for each segment is filled in to generate an 8-hour reference power curve. This curve also limits the maximum allowable power and capacity based on battery aging parameters. Furthermore, if the scenario weights indicate high irradiation uncertainty in a particular segment, the curve will reserve a power reserve band in that segment for real-time control compensation. The complete curve is transmitted to the neuromorphic chip, which runs convex model predictive control within a 20-millisecond cycle, comparing the curve with the measured linear embedded state vector, calculating the real-time power command for the next cycle, and issuing it to the inverter.

[0107] As an implementation example, in a 10-megawatt centralized power station in central Ningxia, the present invention set the target window to 8 hours and the decision dimension to 16. Quantum annealing consumes less than ten millijoules of energy per run, and the variable temperature random search takes 220 microseconds, which is extremely low compared to the hardware overhead. Experiments comparing the quantum plus variable temperature method with traditional mixed integer linear programming show that the average energy within the same algorithm window is less than 1% higher than the global optimal, but the computation time is shortened by two orders of magnitude, meeting the needs of real-time rolling optimization. The cycle 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 the peak-valley arbitrage income is increased by approximately 6%.

[0108] In the rooftop photovoltaic test of the Zhejiang Park in a distributed scenario, the scenario library was expanded to 256, and the input dimension was still 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 present invention has good scalability.

[0109] 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, and uses the alternating direction multiplier method to iteratively solve the optimization problem under linear constraints to obtain real-time power instructions, and re-solves the optimization problem according to the updated state in each control cycle.

[0110] Convex model predictive control, located at the final stage of the control chain, is used to refine the 8-hour reference power trajectory into real-time power instructions executable in milliseconds. This module runs on a neuromorphic chip, which uses an array of spiking neurons to implement alternating direction multiplier iterations, completing the optimization solution within a 20-millisecond control cycle. Because the previous step mapped the nonlinear physical process into a linear embedded state space, this step can write the predictive model as a linear discrete state equation. Within this framework, a quadratic objective function and linear inequality constraints are constructed to ensure that the optimization problem maintains convexity in each cycle.

[0111] The system state is taken from the linear embedding state vector, denoted as The control input vector is set to , corresponding to the energy storage active power command and reactive power command respectively. The prediction model adopts the first-order linear form:

[0112] matrix Describing the uncontrolled evolution of the system, the matrix To describe the projection effect of the control action in the embedding space, the two matrices are updated online through historical data, and the spectral radius is guaranteed to be less than 1.

[0113] The reference power trajectory window is 8 hours, but only 15 steps into the future are used as the rolling forecast target sequence in each control cycle. ,in The cost function is:

[0114] matrix Trade-off state error, matrix Trade-off control smoothness. The power rate of change is limited by a linear inequality:

[0115] in is the slope coefficient. 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.

[0116] The alternating direction multiplier method splits the objective function into a state error subproblem and a control increment subproblem, and couples constraints using multiplier vectors. One iteration consists of three steps: state update, control update, and multiplier update, all of which are matrix multiplications and additions, which can be completed in parallel within a spiking neural network. The iteration limit is set at 30, and 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, all iterations take less than 150 milliseconds, with the 85th percentile taking approximately 120 milliseconds.

[0117] Field experiments showed that in a 10 MW PV-2 MWh energy storage system in Ningxia, when cloud shadows caused a power step, the neuromorphic solution issued revised instructions within two control cycles, resulting in an actual delay of 60 milliseconds at the inverter end. Using a floating-point solution on a central processing unit would have required 300 milliseconds for the same problem. Peak power tracking error was reduced from 2% to 0.8%.

[0118] The improved control smoothness is reflected in the energy storage current curve: the battery current peak is reduced by 30%, and the equivalent cycle depth is reduced by 4 percentage points. Extrapolating 10 months of data, this can extend the battery life by approximately one year.

[0119] The power reliability lower bound constraint uses the 85th percentile scenario power as a random lower bound. In extreme cloud shadow scenarios, the model preemptively increases instructions, reducing the probability of frequency dropout from 3 / 1000 to 2 / 10,000.

[0120] In summary, convex model predictive control uses linear embedding space to maintain the convexity of the problem, and uses neuromorphic hardware to parallelize the alternating direction multiplier method to output power commands in milliseconds. It also improves the stability of energy storage current and grid frequency through smoothing constraints and random lower bounds, providing high-precision, low-latency final-stage control guarantees for the integrated photovoltaic energy storage system.

[0121] like Figure 3 As shown, each inverter generates a modulation signal based on the real-time power instruction and local state vector through pheromone game and self-distillation reinforcement learning to control the power converter to perform charging and discharging. When it is detected that the network frequency deviation reaches the threshold, the ultracapacitor discharge support is triggered and returns according to the power slope, and the execution data is written to the distributed ledger at a fixed period to update the digital twin system and the global optimization model.

[0122] Each inverter, at the end of the control chain, needs to convert the real-time power commands issued by the previous convex model predictive control into kilohertz-level drive pulse-width modulation signals. This invention employs a two-stage strategy of "pheromone game + self-distillation reinforcement learning" to maintain consistent power allocation within the group while enabling rapid self-adaptation for individual inverters.

[0123] Based on the pheromone game principle, the active power performance error and reactive power performance error of each inverter are encoded as active power pheromone concentration and reactive power pheromone concentration, respectively. The concentration decays over time and diffuses within the local network. The mathematical form can be written as:

[0124]

[0125] in and Respectively represent The active and reactive pheromone concentrations of the inverters, and Representing the The active and reactive performance errors of the inverters, Represents a set of neighborhoods connected via a CAN-FD ring network, is the evaporation coefficient, is the injection coefficient, is the diffusion coupling coefficient. Pheromone diffusion implements a mechanism whereby high-error devices release more signals, and neighbors voluntarily compensate by giving up power, thus preventing a single inverter from being overloaded for extended periods of time.

[0126] Self-distillation reinforcement learning strategy, each inverter uses a deep deterministic policy gradient algorithm as the basic controller, and the state vector consists of direct-axis voltage, quadrature-axis voltage, module temperature, state of charge, and To overcome parameter discrepancies in distributed scenarios, this method regularly extracts the last 5 minutes of data from a local replay pool, forcing the current student network to mimic the teacher policy with the best average performance in the past. The loss function takes the Euclidean distance between the two network outputs. Distillation combined with a model-independent meta-learning outer loop enables the policy to recover the optimal action within 10 steps after a sudden cloud shadow change or battery temperature surge, meeting millisecond-level retraining requirements without the need for a centralized server.

[0127] The modulation signal is generated, and the strategy network outputs a normalized modulation value, which is mapped to the duty cycle of the fully controlled bridge. To avoid current spikes, the present invention adds a single-cycle 0.03 rising limit at the hardware level. If the pheromone concentration remains high for a long time, the strategy network automatically reduces the duty cycle's rising slope, forming a group behavior of soft power reduction.

[0128] With ultracapacitor bypass support, when the absolute value of the measured network frequency deviation exceeds 0.05 Hz, the inverter immediately sends a bypass trigger signal. The ultracapacitor is connected in parallel to the DC bus via a high-voltage contactor, with the discharge current limited to twice the rated power for 0.3 seconds. After the bypass is complete, the system falls back to the real-time power command at a rate of 0.1 times the rated power per second. This process is implemented entirely in the inverter firmware, ensuring a trigger delay of less than 2 milliseconds.

[0129] Every 50 milliseconds, the inverter concatenates the timestamp, output active power, output reactive power, and active pheromone concentration, performs a hash operation, and writes the data to the Hyperledger Fabric channel. The ledger uses RAFT consensus, with blocks packaged every 100 milliseconds. The digital twin system and global optimization model synchronously pull execution logs after a new block is confirmed. This is used to correct for liquid neural network weight drift and re-estimate the initial quantum temperature, achieving cross-layer adaptation.

[0130] In a practical example, during 720 hours of operation at a 10 MW centralized power station, pheromone gaming reduced the average thermal stress of a single inverter by 15% and the standard deviation of group power fluctuation by 12%. Ultracapacitor rapid support was triggered 28 times, reducing the maximum frequency deviation by 40%. A total of 520,000 ledger records were accumulated, and the digital twin had no frame misses on the insertion side. When 80 inverters were deployed in a 5 MW distributed scenario in Zhejiang, the CAN-FD ring network load was 20%, and the round-trip latency of pheromone frames was 0.8 milliseconds, demonstrating the solution's excellent scalability.

[0131] Preferably, each inverter broadcasts the active pheromone concentration and reactive pheromone concentration to neighboring inverters through the pheromone diffusion-evaporation game, and inputs the updated pheromone concentration and the local state vector into the 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 it is detected that the network frequency deviation reaches the threshold, the ultracapacitor bypass discharge support is triggered and returned according to the power slope, and 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 the distributed ledger according to a fixed period, and the execution data stored in the distributed ledger is used to synchronously update the digital twin system and the global optimization model.

[0132] The pheromone diffusion-evaporation game aims to enable multiple inverters to achieve power transfer cooperation within milliseconds to avoid long-term overload of a single machine. Each inverter maintains two continuous variables: active pheromone concentration and Reactive pheromone concentration The two are broadcast in the local ring network through the proximity communication interface. The communication interface adopts CAN-FD, with a data frame length of 64 bits and a round-trip delay of 0.8 milliseconds. The concentration dynamics follows the three mechanisms of "evaporation-injection-diffusion":

[0133]

[0134] The evaporation term makes old information invalid quickly, the injection term maps the current compliance pressure to local concentration, and the diffusion term makes the concentration tend to equilibrium in the neighborhood. The local error is increased and the neighboring error perception is pushed up through diffusion, thereby inducing the neighboring to reduce output or give up reactive support to achieve distributed load balancing.

[0135] The state vector of the inverter 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 strategy network takes the state vector as input and outputs the normalized modulation quantity. , mapped to the two-phase pulse width duty cycle of the bidirectional power converter. The policy network is pre-trained offline using a deep deterministic policy gradient approach. Knowledge distillation is then performed on-site every 10 minutes. The most recent 3000 steps are selected from the local experience replay pool, and the previous policy with the highest average return is used as the teacher. The student network is fine-tuned by minimizing the mean squared loss of the output differences, ensuring that the policy converges within ten steps after parameter drift or sudden environmental 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 taking only 30 microseconds.

[0136] To support the inertia gap during large disturbances, the inverter also integrates an ultracapacitor bypass support circuit. When the absolute value of the grid frequency deviation exceeds 0.05 Hz and persists for 20 milliseconds, the firmware triggers a relay to connect the ultracapacitor bank in parallel to the DC bus, setting the discharge current to twice the rated current for 0.3 seconds. After support, the power command is smoothly returned to the real-time power trajectory at a slope of 0.1 times the rated power per second to avoid secondary overshoot. The entire detection, triggering, and exit process is performed locally in the inverter, with an average trigger delay of less than 2 milliseconds.

[0137] The inverter generates a performance record every 50 milliseconds, including a timestamp, output active power, output reactive power, and active pheromone concentration. This record is hashed with SHA-256 and written to a Hyperledger Fabric channel. Blocks are packaged every 100 milliseconds using RAFT consensus. The digital twin system pulls the new block and reconstructs the actual inverter output in chronological order. This is used for self-supervised training of the liquid neural network and updates the initial quantum annealing temperature to ensure consistency between the model and the execution layer.

[0138] In this example, a 10-megawatt power station in Ningxia deployed 20 inverters, and the pheromone communication ring network operated for 720 hours. The average active pheromone concentration formed a sawtooth-shaped distribution consistent with the irradiation intensity as the solar insolation varied. The peak concentration of the highest single unit decreased by 40%, indicating that the load was more evenly distributed across the units.

[0139] A 5-megawatt rooftop system in Zhejiang deployed 80 inverters. When a typhoon cloud passed quickly, the frequency protection action was triggered 11 times, the ultracapacitor discharge maintenance time accumulated for 3.3 seconds, and the maximum frequency deviation dropped from 0.06 Hz to 0.036 Hz.

[0140] A total of 5.2 million execution logs were written to the two stations' chains, 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.

[0141] like Figure 4 As shown, a photovoltaic energy storage integrated charge and discharge control system is used to implement the photovoltaic energy storage integrated charge and discharge control method, and the system includes: The acquisition and fusion module collects signals from PV panels, energy storage battery cabinets, grid-connected interfaces, and the sky unit. Using a deterministic time-sensitive network (DTSN), it unifies timestamps and performs denoising, integrating the event visual stream into cloud shadows to generate fused data. The acquisition and fusion module (edge ​​layer) includes: On the sensor side, Hall effect current transformers and voltage sampling boards are embedded in each PV string, and state-of-charge meters and strain gauges are deployed in the energy storage battery cabinets. Synchronized phasor measurement units are used at the grid connection point. The sky unit is equipped with an event visual camera and a CMOS cloud image camera. On the communication side, each node is connected to a TSN switch via 1Gbps silicon photonic Ethernet. The switch has built-in IEEE 802.1Qbv time slot scheduling logic. The time base is provided by a PTP-TC timing module. On the processing side, the switch is connected downstream to a Zynq Ultrascale+ FPGA, which performs sliding median filtering, drift correction, and 5ms integration of the event stream in on-chip BRAM. Cloud shadows are generated and then combined with the electrical frame to form fused data.

[0142] 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 embedded state data. Based on this linear embedded state data, it uses a spatiotemporal Gaussian process to generate probabilistic scenario data. The twin generation module (edge-cloud collaboration layer) compute hardware consists of a dual-socket EPYC server with RTXA6000 GPUs. The GPUs deploy a liquid neural network inference kernel and a two-dimensional explicit thermal CUDA kernel, while the CPU runs an extended Kalman filter thread. Data pipeline: The fused data is transmitted to the server via 10GbE fiber, with historical data cached on an NVMe SSD. Gaussian process acceleration utilizes 48GB of GPU memory for sparse kernel approximation and batch Cholesky decomposition, achieving a sampling time of less than 1ms for 128 scenarios.

[0143] The scenario optimization module performs high-dimensional hypersymbolic encoding and mutual information filtering on the probabilistic scenario data. Combined with the battery aging parameters, it determines the optimal charge and discharge state within a global optimization model, generates a reference power trajectory, and runs convex model predictive control on the neuromorphic chip to output real-time power commands. The scenario optimization module (acceleration layer) performs encoding and filtering using an Intel Agilex FPGA board, implementing 10,000-bit hypersymbolic XOR counting and calculating mutual information in 8µs. Global optimization uses a D-Wave Advantage quantum annealing cloud to map 16-bit QUBOs at a time, annealing in 20µs and outputting 50 solutions. A local AMD Mi250 GPU performs variable-temperature random search refinement. MPC hardware uses a Loihi2 neuromorphic chip loaded with an alternating direction multiplier pulse network, iterating 30 times in a 20ms cycle and outputting real-time power commands.

[0144] The execution feedback module, based on the real-time power command and local state vector, generates modulation signals for each inverter through pheromone game theory and self-distillation reinforcement learning to control power converter charging and discharging. When the network frequency deviation reaches a threshold, it triggers ultracapacitor discharge support and returns according to the power slope. Execution data is periodically written to the distributed ledger to update the twin generation module and the scenario optimization module. The execution feedback module (field layer) includes power conversion: Each 250kVA energy storage inverter is equipped with a SiC fully controlled bridge and a 1MHz PWM DSP. The DSP runs a lightweight strategy network forward and outputs the modulation ratio. Collaborative communication: 1Mbps pheromone frames are broadcast between inverters over a CAN-FD ring network. Pheromone updates and action inference are completed within 300µs within the C2000 core. Inertia support: A 5kJ ultracapacitor bank is connected in parallel to the DC bus, and the IGBT bypass switch trigger delay is less than 2ms. Ledger node: Raspberry Pi 4 is used as a light node signature, Jetson Orin is used as a sorting node to run the Hyperledger Fabric RAFT 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.

[0145] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included within the scope of the claims of the present application.

Claims

1. A photovoltaic energy storage integrated charge and discharge control method, characterized in that: The following steps are involved: Multimodal signals from photovoltaic panels, energy storage battery cabinets, grid-connected interfaces, and the sky terminal are collected, and unified timestamps are used for denoising via a deterministic time-sensitive network. Cloud shadow outlines are generated by integrating event visual streams to obtain fused data. Inputting 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 embedded state data, and generating probabilistic scenario data based on the linear embedded state data using a spatiotemporal Gaussian process; Implementing high-dimensional hypersymbolic encoding and mutual information screening on the probabilistic scenario data, determining the optimal charge and discharge state in a global optimization model in combination with the battery aging parameters, generating a reference power trajectory, and running a convex model predictive control on a neuromorphic chip to output real-time power instructions; Each inverter generates a modulation signal based on the real-time power instruction and local state vector through pheromone game and self-distillation reinforcement learning to control the power converter to perform charging and discharging. When it detects that the network frequency deviation reaches the threshold, the ultracapacitor discharge support is triggered and returns according to the power slope. The execution data is written to the distributed ledger at a fixed period to update the digital twin system and the global optimization model.

2. The photovoltaic energy storage integrated charge and discharge control method according to claim 1, characterized in that: The acquisition step synchronizes the event stream output by the event vision sensor with the current signal collected by the photovoltaic module, the voltage signal collected by the photovoltaic module, the charge state signal collected by the energy storage battery cabinet, the network frequency signal collected by the grid-connected interface, and the cloud image collected by the sky end. The event vision stream is used to generate a cloud shadow outline in a sliding integration manner, and then the cloud shadow outline is unified with the remaining signals through a deterministic time-sensitive network. Then, sliding median filtering and drift correction are sequentially performed in the edge processor to obtain the fused data.

3. The method according to claim 1, characterized in that The liquid neural-thermal power coupling model maps the cloud shadow contour into heat source power density, writes the heat source power density into the two-dimensional heat conduction equation as the source term, and couples it with the photovoltaic module current signal and the photovoltaic module voltage signal to solve the temperature field. The temperature field result is used to correct the state component related to the thermal characteristics in the linear embedded state data.

4. The method according to claim 1, wherein The electrochemical coupling model adopts a joint modeling method of equivalent circuit and electrochemical kinetics, uses the state of charge signal and temperature field results collected by the energy storage battery cabinet to construct a state vector, and recursively calculates the diffusion coefficient and interface film resistance through extended Kalman filtering to generate the battery aging parameters.

5. The method according to claim 1, wherein The spatiotemporal Gaussian process uses the linear embedded state data as the mean function, constructs a covariance structure in the form of the product of a spatial kernel function and a temporal kernel function, and jointly samples the irradiation sequence, the temperature sequence, and the photovoltaic power sequence to form a weighted probability scenario data set.

6. The method according to claim 1, wherein The high-dimensional super-symbol coding maps each probability scene data into a binary vector of fixed length through random Gaussian projection, calculates the mutual information score between the binary vector and the linear embedded state data using a hardware operation method of bitwise XOR plus counting, and selects a preset number of scenes according to the size of the mutual information score to form the weighted scene subset.

7. The method according to claim 6, characterized in that The global optimization model discretizes the weighted scenario subset and the battery aging parameter into a binary charge and discharge decision matrix, then constructs a quadratic unconstrained binary optimization model. Candidate solutions are obtained through a quantum annealing solver, and energy consumption values ​​are compared in a variable temperature random search framework. The charge and discharge decision with the minimum energy consumption value is retained as the preferred charge and discharge state.

8. The method according to claim 1, characterized in that The convex model predictive control uses the linearly embedded state data as the system state and the reference power trajectory as the target trajectory, and uses the alternating direction multiplier method to iteratively solve the optimization problem under linear constraints to obtain real-time power instructions, and re-solves the optimization problem according to the updated state in each control cycle.

9. The method according to claim 1, characterized in that Each inverter broadcasts the active pheromone concentration and reactive pheromone concentration to neighboring inverters through the pheromone diffusion-evaporation game, and inputs the updated pheromone concentration and the local state vector into the policy network trained by self-distillation reinforcement learning to obtain the modulation signal to control the bidirectional power converter to perform charging and discharging; when it is detected that the network frequency deviation reaches the threshold, the ultracapacitor bypass discharge support is triggered and returned according to the power slope, and 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 the distributed ledger according to 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.

10. A photovoltaic energy storage integrated charge and discharge control system, used to implement the photovoltaic energy storage integrated charge and discharge control method according to any one of claims 1 to 9, characterized in that: The system includes: The acquisition and fusion module collects signals from photovoltaic modules, energy storage battery cabinets, grid-connected interfaces, and the sky unit. It unifies timestamps and performs denoising via a deterministic time-sensitive network, integrating the event visual stream into cloud shadow outlines to generate fused data. a twin generation module, configured 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 embedded state data, and to generate probabilistic scenario data based on the linear embedded state data using a spatiotemporal Gaussian process; a scenario optimization module for performing high-dimensional hypersymbolic encoding and mutual information screening on the probabilistic scenario data, determining the optimal charge and discharge state in a global optimization model in combination with the battery aging parameters, generating a reference power trajectory, and running a convex model predictive control on a neuromorphic chip to output real-time power instructions; The execution feedback module is used for each inverter to generate a modulation signal based on the real-time power instruction and the local state vector through pheromone game and self-distillation reinforcement learning to control the power converter to perform charging and discharging. When it is detected that the network frequency deviation reaches the threshold, the ultracapacitor discharge support is triggered and returned according to the power slope, and the execution data is written into the distributed ledger according to a fixed period to update the twin generation module and the scenario optimization module.

Images