Impulse neural network-based partial shading maximum power point tracking method and system

Abstract

The application discloses a partial shading maximum power point tracking method and system based on a pulse neural network, and relates to the technical field of new energy power generation and artificial intelligence cross control. The method only obtains photovoltaic array electrical signals through voltage and current sensors, extracts voltage, current and power variation as dynamic gradient features, maps local shading multi-peak distortion to the step rate of current variation relative to voltage variation, encodes the step rate into a multi-dimensional space-time pulse sequence through a Delta modulator, takes the pulse sequence as input, performs event-driven forward reasoning on the pulse neural network by using a leakage integral trigger model, completes weight accumulation by a pure adder tree, updates the weight on line based on reward modulation pulse timing dependence plasticity rules, triggers an exploratory penalty signal to cross a power valley when falling into a local extreme value, and maps the output layer pulse frequency to a transformer duty cycle increment. The application eliminates environmental sensors, realizes low-power consumption, and achieves microsecond-level response global maximum power point tracking.

Description

Technical Field

[0001] This invention relates to the field of cross-control technology of new energy power generation and artificial intelligence, and in particular to a method and system for tracking maximum power point in partial shading based on a pulse neural network. Background Technology

[0002] Photovoltaic (PV) systems have become a widely used renewable energy source due to their sustainability and continuously decreasing installation and maintenance costs. However, under partial shading conditions, the output power of PV systems exhibits a multi-peak characteristic curve with multiple local maxima. Traditional maximum power point tracking algorithms, such as the perturbation-observation method and the incremental conductance method, tend to converge to local peaks, resulting in significant power loss.

[0003] To address these issues, maximum power point tracking (MPPT) technology based on machine learning has been extensively studied. However, the engineering deployment of existing machine learning models on distributed photovoltaic (PV) edge devices is still constrained by the following bottlenecks: First, high dependence on environmental sensors—existing high-precision models typically require solar irradiance and module temperature as input features, which not only increases hardware costs but also makes it difficult for macroscopic environmental data to accurately characterize the non-uniform local micro-shading state in large-area PV arrays; Second, decay model mismatch—PV arrays are inherently highly nonlinear time-varying systems. As modules age and operating conditions become more complex, their output characteristic curves exhibit unpredictable dynamic drift, which static models relying on offline data pre-training cannot effectively adapt to; Third, high computational latency restricts transient optimization efficiency—when the environment changes rapidly, the intensive multiply-accumulate operations of traditional deep models generate non-negligible computational latency, prolonging the transient adjustment period that deviates from the global maximum power point.

[0004] Therefore, there is an urgent need for a maximum power point tracking scheme that can achieve low power consumption, high dynamic response, and full lifecycle robustness on resource-constrained edge embedded devices. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a partial shading maximum power point tracking method and system based on a spiking neural network. Starting from the implementation method of maximum power point tracking itself, it solves the problems of strong dependence on environmental sensors, difficulty of offline models to adapt online, and high computational latency restricting transient optimization efficiency in the prior art.

[0006] The technical solution of this invention is: A partial shading maximum power point tracking method based on spiking neural networks includes: The output voltage and current of the photovoltaic array are obtained only through voltage and current sensors. The output power is calculated, and the voltage change, current change, and power change are extracted as dynamic gradient features. The dynamic gradient features are encoded into a multidimensional spatiotemporal pulse sequence using a Delta modulator. A pulse is emitted when the amplitude of the variable change exceeds a preset quantization threshold; otherwise, it remains silent. Using the multidimensional spatiotemporal pulse sequence as input, an event-driven forward inference is performed using a spiking neural network based on a leakage current integral triggering model. The synaptic weights are accumulated to the neuron membrane potential by a pure adder tree, and a pulse is output when the membrane potential exceeds the firing threshold. Reward and penalty signals are generated based on the power change. The synaptic weights are updated online based on the pulse timing-dependent plasticity rule. When the power decreases and the ratio of current change to voltage change exceeds the shading judgment threshold, a penalty signal is generated to adjust the weights in reverse. The statistical output layer pulse frequency is mapped to the increment of the DC-DC converter pulse width modulation duty cycle, driving power regulation.

[0007] Furthermore, the Delta modulator employs a positive and negative dual-channel decoupling coding mechanism: let the input variable in the k-th sampling period be... Internally maintains the reference tracking level for discrete states. Define tracking error for and The difference, let the quantization threshold be... The following coding logic is executed in each clock cycle: If E(k) is not less than The ascending channel encoder outputs a discrete pulse at the current moment, and simultaneously updates the reference level to the current value plus... ; If E(k) is not greater than The descent channel encoder outputs a discrete pulse at the current moment, while the reference level is updated to the current value minus... ; If the absolute value of E(k) is less than Both channels remain silent and the reference level remains unchanged; The quantization threshold It is significantly greater than the peak-to-peak value of the high-frequency switching ripple of the inductor current at the current duty cycle of the DC-DC converter.

[0008] Furthermore, the membrane potential update equation for any neuron i in the spiking neural network at discrete time step t is: , In the formula, Let be the membrane potential of neuron i at time t; For a fixed leakage current constant; The synaptic weights from the previous neuron j to the current neuron i; The input pulse of the preceding neuron j at time t is 0 or 1; when At that time, neuron i fires an output pulse and its membrane potential is forcibly cleared to zero in the next clock cycle.

[0009] Furthermore, the pulse frequency mapping to duty cycle increments employs push-pull dual-channel decoding, and the output layer includes neurons that increase the duty cycle. and neurons with reduced duty cycle Its discrete mapping equation is: , In the formula, This represents the duty cycle increment for the k-th time window; The preset hardware mapping gain coefficient; and They are respectively and The total number of pulses issued within this time window.

[0010] Furthermore, the generation condition for the exploratory penalty signal is that it simultaneously satisfies: Power change over two consecutive sampling periods Furthermore, the power decrease exceeds the preset power decrease threshold, and the ratio of current to voltage change rate... Exceeding the preset shading threshold; During the effective period of the penalty signal, strong suppression updates are performed on synaptic connections with a pulse firing time difference Δt < 0 between preceding and following neurons to eliminate the weight bias of local extrema, while synaptic connections with Δt > 0 are temporarily not updated to preserve the degree of freedom of exploration; When power recovers and increases, the penalty signal is removed and switched to a positive reward signal.

[0011] Furthermore, the shading determination threshold is set to: the absolute value of the ratio of current change to voltage change under uniform illumination is on the order of 0.01 to 0.1 A / V, and the shading determination threshold is set to the order of 1 to 10 A / V.

[0012] According to another aspect of the present invention, a partial shading maximum power point tracking system based on a spiking neural network is provided, comprising: The physical sensing and feature extraction module is used to obtain the output voltage and output current of the photovoltaic array and calculate the output power using only voltage and current sensors, and extract the voltage change, current change and power change as dynamic gradient features; The spatiotemporal pulse coding module is used to encode the dynamic gradient features into a multidimensional spatiotemporal pulse sequence through a Delta modulator. When the amplitude of the variable change exceeds a preset quantization threshold, a pulse is emitted; otherwise, it remains silent. The forward inference module of the spiking network uses a digital neuron array based on the leakage current integral triggering model and operates in an event-driven manner: when an input pulse is received, the corresponding synaptic weight is accumulated to the neuron membrane potential through a pure adder tree. The membrane potential decays with a fixed leakage current constant and outputs a pulse and resets when it exceeds the firing threshold. The online reward modulation learning module is used to generate reward and penalty signals based on power changes, and to update synaptic weights online based on pulse timing-dependent plasticity rules without the need for backpropagation computation. The pulse decoding and modulation execution module is used to count the pulse frequency of the output layer neurons within a preset time window and directly map the pulse frequency to the increment of the pulse width modulation duty cycle of the DC-DC converter.

[0013] Furthermore, in the spiking network forward inference module, synaptic weights are stored in on-chip block random access memory, and the digital neuron array is composed of a pure adder tree and a digital comparator, without relying on hardware multipliers and floating-point units; the physical perception and feature extraction module, the spatiotemporal pulse coding module, the spiking network forward inference module, the online reward modulation learning module, and the pulse decoding and modulation execution module are sequentially connected.

[0014] Furthermore, the pulse decoding and modulation execution module adopts a push-pull dual-channel decoding structure. The output layer includes two mutually exclusive output neurons: a duty cycle increasing neuron and a duty cycle decreasing neuron, which control the direction of duty cycle increase and decrease, respectively.

[0015] The beneficial effects of this invention are: 1. This invention acquires the electrical signals of the photovoltaic array using only voltage and current sensors, extracts voltage, current, and power changes as dynamic gradient features, and implicitly maps the multi-peak distortion caused by local shading to the step rate of current change relative to voltage change. It eliminates the need for additional environmental sensors such as solar irradiance and module temperature, significantly reducing the adaptive robustness of the hardware lifecycle of edge nodes.

[0016] 2. This invention updates synaptic weights online based on the pulse timing-dependent plasticity rule of reward modulation, abandoning the static modeling method that relies on offline data pre-training. It can dynamically adjust synaptic weights according to real-time power changes without global backpropagation computation. When the output characteristic curve of photovoltaic modules shifts due to factors such as aging, dust accumulation, or temperature drift, the system can continuously reshape the network weights online to adapt to the shift, giving the maximum power point tracking capability adaptive robustness throughout its entire life cycle. This overcomes the problem that offline pre-trained models are difficult to adapt to the time-varying characteristics of photovoltaic arrays.

[0017] 3. This invention employs a deterministic Delta modulator to encode continuous physical variables into multidimensional spatiotemporal pulse sequences. Compared to Poisson coding, which relies on pseudo-random number generators and consumes a large amount of logic resources, encoding can be completed using only basic comparator logic, approaching the minimum computing power of edge devices. By setting the quantization threshold to be strictly greater than the peak-to-peak value of the high-frequency switching ripple of the inductor current of the DC-DC converter at the current duty cycle, high-frequency electrical noise and minor useless disturbances are shielded during the encoding stage, ensuring that the pulse neural network is activated only by real environmental mutations.

[0018] 4. This invention utilizes a spiking neural network based on a leakage current integral triggering model for event-driven forward inference. Neurons are in an asynchronous waiting state and are activated to enter the computation cycle only when an input pulse is received. This transforms complex network topology operations into addition and accumulation operations based on pure digital logic. The hardware logic directly addresses and reads synaptic weights from the on-chip block random access memory according to the pulse index and completes the weight accumulation through a pure adder tree. This does not rely on hardware multipliers and floating-point units, making it possible to achieve microsecond-level ultra-low latency optimization control on resource-constrained edge devices such as field-programmable gate arrays or microcontrollers, effectively reducing transient energy loss when external signals change abruptly. Attached Figure Description

[0019] Figure 1 This is a flowchart of the partial shading maximum power point tracking method based on a spiking neural network according to the present invention; Figure 2 This is a schematic diagram comparing the ratio of current-voltage change rate under uniform illumination and partial shading conditions. Figure 3 This is a schematic diagram of the partial shading maximum power point tracking system based on a spiking neural network according to the present invention. Detailed Implementation

[0020] The following will be based on embodiments of the present invention. Figures 1-3 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 This embodiment provides a partial shading maximum power point tracking method based on a spiking neural network, comprising: acquiring the output voltage and output current of a photovoltaic array using only voltage and current sensors and calculating the output power; extracting voltage changes, current changes, and power changes as dynamic gradient features; encoding the dynamic gradient features into a multidimensional spatiotemporal pulse sequence using a Delta modulator; using the multidimensional spatiotemporal pulse sequence as input and performing event-driven forward inference using a spiking neural network based on a leakage current integral triggering model; updating synaptic weights online based on a pulse timing-dependent plasticity rule of reward modulation; and statistically analyzing the pulse frequency of output layer neurons within a preset time window and mapping it to duty cycle increments to drive power adjustment.

[0022] To better understand this invention, the details are as follows: First, electrical signals are acquired by voltage and current sensors deployed at the output of the photovoltaic array, power is calculated, and dynamic gradient features are extracted.

[0023] Specifically, the system acquires the current output voltage of the photovoltaic array at a preset high-frequency continuous sampling rate by using low-cost current and voltage sensors deployed at the output end of the photovoltaic array. and output current The analog signal is converted into a digital value. The controller internally multiplies the current voltage and current in real time to calculate the current output power of the photovoltaic array.

[0024] Subsequently, based on the difference between current data and historical data, the system calculates three key dynamic gradient features in real time: voltage change. Current change and power change This implicit sensing mapping process does not directly measure temperature and irradiance, but instead maps temperature drift to changes in the absolute voltage baseline and multi-peak distortion caused by local shading to the step rate of current change relative to voltage change.

[0025] It should be noted that the above implicit perceptual mapping has a clear physical basis.

[0026] The electrical characteristics of partial shading are unique and observable, and their gradient characteristics uniquely characterize the partial shading state. PV arrays typically employ a series topology of photovoltaic cells to meet the system's bus voltage requirements. Under partial shading conditions, the photocurrent generated by the shaded photovoltaic cells significantly decreases. Each cell string is connected in anti-parallel to a bypass diode. When the cell string is shaded, the diode conducts in the forward direction, providing a current path for other normal cell strings and preventing hot spot damage. The activation process of this bypass diode constitutes a nonlinear circuit topology switching behavior: when the uniform illumination changes, the photocurrent generated by all cells changes synchronously and linearly, with no bypass diode conducting. At this time, the output voltage V and current I change continuously and smoothly. The output current I of the photovoltaic array at this time can be expressed as:

[0027] in, Indicates photocurrent, Indicates reverse saturation current. Indicates series resistance. This represents the parallel resistance. When the irradiance changes uniformly... linear change Other parameters remain approximately unchanged. Taking the derivative with respect to V, we get:

[0028] This is a continuous and smooth function, whose absolute value is... The change with V is slow, with no abrupt change points.

[0029] When partial shading occurs, the aforementioned abrupt state changes at the device level cause a step-like (order-of-magnitude) drop in the array's output current. Due to inherent electrical inertia of the system (such as the charging and discharging of filter capacitors), the dynamic response of the output voltage exhibits significant hysteresis. Suppose the equivalent topology of the array undergoes a sudden change, and the output current becomes...

[0030] in, It is the photocurrent generated by a normal battery string. This is the current flowing through the bypass diode (approximately constant). The on-resistance of an ideal diode is 0. In this case:

[0031] Under uniform illumination, the photocurrent generated by all photovoltaic cells changes synchronously and linearly, and the output current I changes continuously and smoothly, with |dI / dV| on the order of approximately 0.01~0.1A / V. However, when partial shading occurs, the bypass diodes of the shaded photovoltaic cell strings conduct in the forward direction, resulting in nonlinear circuit topology switching behavior: the array output current experiences a step-like sudden drop, and due to the inherent electrical inertia of the system, the dynamic response of the output voltage exhibits significant hysteresis. Under this condition, because the on-resistance of the bypass diodes is very small, The value rises sharply to the 1-10 A / V range, an increase of 1 to 2 orders of magnitude. This characteristic cannot be replicated by uniform illumination changes, temperature drift, or load fluctuations, and thus cannot constitute a reliable physical criterion for identifying the occurrence of shading and the inflection point of the PV curve.

[0032] It should be further explained that it is not enough to simply determine whether partial shading has occurred. It is also necessary to prove that the complete structure of the multi-peak PV curve can be directly analyzed using only the voltage, current, and power timing data at the output terminal, to distinguish between local maxima and global maxima and to achieve accurate global optimization.

[0033] Specifically, we first define the structural rules of a multi-peak PV curve. Assuming a photovoltaic module consists of N independent cell strings plus bypass diode groups (typically N is three to four), under any partial shading, its PV curve can be strictly divided into N consecutive single-peak segments, each segment corresponding to a bypass diode conduction combination state. In a multi-peak PV curve, the number of peaks is less than or equal to the number of cell string groups N, and the peaks are separated by clear inflection points (corresponding to the switching points between bypass diode conduction and cutoff).

[0034] During dynamic perturbation scanning, the continuously acquired gradient sequences such as ΔP and S(k) = ΔP(k) / ΔV(k) exhibit characteristics that strictly correspond to the multi-peak structure. Specifically, within any single-peak segment, the system can determine whether the current operating point is at the rising edge, peak, or falling edge of the single-peak segment based on the sign and amplitude of ΔP and S(k)—this judgment logic constitutes the "three-segment" gradient feature analysis within the single-peak segment.

[0035] When the scan passes the inflection point, two characteristics appear simultaneously: First, a current step characteristic—due to the switching of the bypass diode state, a small change in the output voltage can cause a sharp drop or rise in current, manifested as a sudden change in ΔI / ΔV of more than an order of magnitude compared to the uniform illumination condition. This sudden change can be reliably identified by detecting whether ΔI / ΔV exceeds a preset step threshold. Second, a power-voltage slope abrupt change characteristic—before and after the inflection point, the slope dP / dV of the PV curve undergoes a sign reversal or a sudden change of an order of magnitude. The combination of these two characteristics provides a clear indication of the switching between single-peak segments.

[0036] Therefore, the system can accurately locate all inflection points and divide the complete PV curve into N consecutive single-peak segments. Although locating inflection points and delineating segments based solely on the above features cannot directly determine which single peak is the global maximum power point (because the peak power of each segment is affected by the nonlinearity of the micro-shading distribution and cannot be directly sorted by static gradient features), this gradient information provides the subsequent spiking neural network intelligent optimization module with the internal position label of the single-peak segment where the current working point is located and the trigger signal for inflection point switching.

[0037] Next, the extracted dynamic gradient features are encoded into a multidimensional spatiotemporal pulse sequence using a Delta modulator.

[0038] Specifically, the extracted continuous state variables (including V) pv I pv , and The inputs are fed in parallel into independent Delta modulator channels. Unlike Poisson coding, which relies on pseudo-random number generators and consumes a lot of logic resources, this scheme adopts a deterministic Delta modulation architecture, which can complete the encoding using only basic comparator logic, further approaching the minimum computing power of edge devices.

[0039] For the two characteristic quantities with positive and negative polarities, the output voltage gradient ΔV and current gradient ΔI of the photovoltaic array, the Delta modulator employs a dual-channel decoupled encoding mechanism. Taking the current gradient ΔI as an example, its encoding logic is as follows: Define the tracking error. Let the quantization threshold be... In each clock cycle, the hardware comparator performs differential logic: if If a significant current rise step occurs, the encoder in the rising channel will issue a discrete pulse at the current moment. Meanwhile, the internal reference level is updated to ;like If this occurs, it is determined that a sudden current drop has occurred, and the encoder of the descending channel issues a pulse. Meanwhile, the reference level is updated to ;like If the signal is not clear, it is determined to be a steady-state micro-motion or normal noise in the system. Both channels remain silent, i.e., output 0, and the reference level remains unchanged.

[0040] It should be noted that the quantization threshold The value of is strictly greater than the peak-to-peak value of the high-frequency switching ripple of the inductor current at the current duty cycle of the DC-DC converter, so as to ensure that high-frequency electrical noise and small useless disturbances are shielded during the encoding stage, so that the pulse neural network is only activated by real environmental mutations.

[0041] It should be further explained that the quantization threshold The specific value can be determined based on the circuit parameters of the DC-DC converter as follows. This embodiment uses a Boost DC-DC converter as an example, and in continuous conduction mode, the peak-to-peak value of the high-frequency switching ripple of the inductor current is... It can be estimated by the following formula:

[0042] In the formula, is the input voltage of the converter, i.e., the output voltage of the photovoltaic array; D is the current operating duty cycle; L is the inductance value of the boost inductor; The switching frequency is used. The peak-to-peak value of the ripple varies non-linearly with the duty cycle D, reaching a maximum value near D=0.5. The controller can calculate the corresponding peak-to-peak value of the ripple online based on the current real-time duty cycle, or pre-calculate the maximum peak-to-peak value of the ripple within the duty cycle scanning range. As a basis for conservative value selection.

[0043] Based on this, the quantization threshold V th_I Set as follows:

[0044] In the formula, This is a preset margin coefficient used to compensate for current sensor measurement noise, analog-to-digital conversion quantization errors, and engineering tolerances of inductance parameters. The value range is set to 1.2 to 1.5: If If the value is too small, close to 1, the random jitter of the ripple itself can easily be misjudged as an effective current step, causing the pulse neural network to generate a large number of noise pulses under steady-state conditions, resulting in unnecessary weight perturbations and duty cycle oscillations; if If the value is too large, exceeding 1.5, the quantization threshold is too high, which may mask the real but small-amplitude step characteristics of the local shading current, leading to missed detection.

[0045] For voltage gradient The quantization threshold V used th_V The margin factor can be determined based on the charging and discharging time constant of the filter capacitor and the measurement noise level of the voltage sensor, using the same margin factor method. That is, first estimate the peak-to-peak value of the intrinsic noise of the voltage measurement circuit under steady-state conditions, and then multiply it by the same margin factor. Get V th_V .

[0046] Using the above determination method, the quantization threshold V is determined. th_I and V th_V The ripple amplitude can be calculated and fixed offline based on the actual hardware parameters of the converter during the controller initialization phase, or it can be dynamically updated based on the real-time duty cycle during system operation to adapt to the fluctuation of ripple amplitude caused by changes in load or operating conditions.

[0047] Through the aforementioned positive and negative dual-channel decoupling encoding mechanism, the system will include macroscopic coordinates. With transient gradient The four-dimensional continuous state variables are transformed into eight parallel binary pulse streams, implicitly converting the originally complex physical state into a multi-dimensional spatiotemporal pulse sequence S(t) that spiking neural networks can naturally process. This multi-dimensional input architecture not only preserves the global spatial coordinates of the operating point, but also keenly captures the electrical topological changes induced by local shading, implicitly converting the originally complex physical state into a discrete pulse coding form that spiking neural networks can naturally process.

[0048] Then, using a multidimensional spatiotemporal pulse sequence as input, an event-driven forward inference is performed using a spiking neural network based on a leakage current integral triggering model.

[0049] To meet the microsecond-level low latency requirements of adaptive maximum power point tracking in photovoltaic arrays, the spiking neural network of this invention adopts a lightweight feedforward fully connected topology, specifically comprising an input layer, hidden layers, and an output layer. Regarding node size and mapping mechanism: Input layer: Contains 8 digital neurons, which are hardwired one-to-one with 8 parallel binary pulse streams generated by the previous encoding module. Specifically, input neurons 1 and 2 receive rising and falling pulse streams of the macroscopic voltage coordinate; input neurons 3 and 4 receive rising and falling pulse streams of the macroscopic current coordinate; input neurons 5 and 6 receive positive and negative abrupt pulse streams of the voltage gradient feature ΔV; and input neurons 7 and 8 receive positive and negative step pulse streams of the current gradient feature ΔI, thereby sensitively capturing the topological changes of the bypass diode caused by local shading.

[0050] Hidden layer: The number of neurons is set to 32 to balance the expressive power of multi-peak nonlinear mapping with the consumption of underlying storage resources, ensuring that synaptic weights can be fully integrated into the on-chip BRAM of the underlying chip.

[0051] Output layer: configured with two mutually exclusive neurons, corresponding to duty cycle increase and decrease instructions respectively, which directly drive the back-end PWM regulation through differential pulse firing rate.

[0052] For synaptic weight initialization, the system abandons offline pre-training and adopts a hardware-friendly zero-mean uniform distribution strategy. The network uses a low-level linear feedback shift register to generate pseudo-random numbers, which are then used to initialize all synaptic weights in the hidden and output layers. Reset for Between Fixed-point values, where the largest absolute value is... Set as the membrane potential trigger threshold V thThis strategy assigns the network a low-confidence, random, weak discharge state in the initial stage, providing initial dynamic perturbations for subsequent online adaptive optimization without causing severe system oscillations.

[0053] Specifically, each digital neuron in the network is implemented as a leakage current integral triggering model. For any neuron i in the network, at each discrete time step t, its membrane potential update equation is:

[0054] In the formula, Let be the membrane potential of neuron i at time t; To fix the leakage constant, the membrane potential is made to exhibit a natural exponential decay trend when there is no effective pulse input; The synaptic weights from the previous neuron j to the current neuron i are stored in on-chip block random access memory; is the input pulse of the preceding neuron j at time t, and its value is either 0 or 1.

[0055] It should be noted that, since the input has been discretized into a sequence of 0s and 1s, the network's forward inference is entirely transformed into an event-driven asynchronous computation mode. The forward inference occurs if and only if an input pulse is received. Only when the pulse index is reached is the corresponding neuron activated and enters the computation cycle. The hardware logic does not need to utilize the resource-intensive digital signal processor multiplier; instead, it directly addresses the corresponding synaptic weight from the on-chip block random access memory based on the pulse index. The weight value is directly added to the membrane potential of the current neuron through a pure adder tree, and a fixed leakage constant is subtracted during the operation. This extremely lightweight, purely digital logic design makes microsecond-level, ultra-low latency optimization control possible.

[0056] After the membrane potential is updated, it is immediately input in parallel to a digital comparator and compared with the system's preset dynamic firing threshold V. th Perform a comparison. If... If the neuron is fully activated, its output port immediately fires a discrete action potential on the time axis. Meanwhile, the membrane potential is forcibly reset to zero in the next clock cycle and enters the refractory period; if it does not exceed the limit, the current potential state is maintained until the next cycle.

[0057] It should be further explained that, along with the firing of the output pulse, the neuron must enter a refractory period. In the underlying hardware implementation, a multiplexer truncates the accumulation circuit, forcibly clearing the neuron's membrane potential to zero in the next clock cycle, and then restarting a new round of leakage and integration cycles. This mechanism ensures that the neuron will not be falsely triggered by the residual membrane potential accumulated in the previous cycle after firing, guaranteeing the accuracy of the pulse firing timing and the determinism of the network state evolution.

[0058] Next, the synaptic weights are updated online based on the pulse timing-dependent plasticity rule of reward modulation.

[0059] Specifically, the local reward evaluation module within the system generates a multi-dimensional reward signal based on the power change ΔP. If ΔP > 0 is detected, it indicates that the system power is increasing, and a global positive reward signal is generated. If ΔP < 0 is detected for two consecutive sampling cycles and the power decrease exceeds a preset threshold, and a current surge characteristic pulse, i.e., |ΔI / ΔV|, exceeds a preset shading judgment threshold, the system determines that the photovoltaic array is experiencing local shading and has fallen into a local power extreme value, and then triggers a global exploratory penalty signal.

[0060] It should be noted that the R-STDP in this scheme deeply couples the local temporal rules of traditional pulse timing-dependent plasticity with the global working condition reward mechanism, forming a lightweight reinforcement learning paradigm specifically adapted to local shading optimization scenarios. It can achieve microsecond-level online weight updates without global backpropagation computation. The pulse firing time difference between presynaptic and postsynaptic neurons is defined as... (in The timestamp for firing pulses to subsequent neurons. (Timestamps for firing pulses to preceding neurons). Based on the principle of causality, the system's fundamental temporal-dependent kernel function... Defined as:

[0061] In the formula, and These are the maximum adjustment constants for long-term enhancement and long-term suppression, respectively; and This is the corresponding time constant. If That is, the preceding neuron fires a pulse before the following neuron, indicating that the activity of the preceding neuron is the causal trigger for the firing of the following neuron, and this synaptic connection has a potential tendency to be strengthened; if This means that the subsequent neuron fires a pulse before the preceding neuron, indicating that there is no causal relationship between the activities of the preceding and following neurons, and that this synaptic connection has a potential tendency to be suppressed.

[0062] The system determines the final update amount of the weights using the global modulation signal R(t) output by the reward evaluation module. The synaptic weight update equation for the current discrete time step is:

[0063] When the global exploratory penalty signal is triggered, the basic timing rules of R-STDP are reversed and activated, and the weight changes. Configured as follows:

[0064] in, This is a penalty coefficient with a high multiplier.

[0065] The system is for all The synaptic connections undergo strong suppression updates, significantly reducing their weight values ​​and completely weakening erroneous neural pathways that cause the system to fall into the current local extremum, thus eliminating the weight bias formed in the early stages based on suboptimal operating conditions; for all The synaptic connections are not updated at this time, preserving the network's exploration freedom and ensuring that the system can traverse the entire voltage range of the PV curve. At the same time, the penalty signal is input to the pulse decoding and modulation execution module, which instantaneously amplifies the step size coefficient of the duty cycle increment by 5 to 10 times, enabling the system to switch from the conventional small-step fine adjustment to a large-step coarse-grained global scanning mode, quickly traversing the voltage range where local extrema are located.

[0066] When the system discovers an effective direction for power enhancement (ΔP > 0), it immediately cancels the penalty signal and switches to a global positive reward signal. At this time, the basic timing rules are positively activated, and the weight changes... Configured as follows:

[0067] In the formula, It is the positive enhancement coefficient. The weak suppression coefficient (and designed to be) ). For all The synaptic connections undergo enhancement updates, increasing their weights by a preset ratio to strengthen the correct neural pathways that guide the system towards power enhancement; for all The synaptic connections are updated with weak suppression, which slightly reduces their weight values, eliminates irrelevant interference pathways, and improves the accuracy and convergence speed of network decisions.

[0068] When the system converges to near the global maximum power point, ΔP approaches 0, and the system no longer generates any global modulation signal. The R-STDP weight update enters a dormant state. The corresponding modulation signal and weight update will only be triggered again when the PV characteristic curve shifts due to factors such as aging, dust accumulation, and temperature drift of the photovoltaic module, and a non-zero ΔP is detected.

[0069] Finally, the pulse frequencies of the output layer neurons are statistically analyzed and mapped to pulse width modulation duty cycle increments to drive power adjustment.

[0070] Specifically, the output layer of the spiking neural network is designed to contain two mutually exclusive output neurons: a duty cycle-increasing neuron and a... Compared with neurons with reduced duty cycle These represent the direction of the working point movement during optimization. Within any preset time statistics window Tw, the hardware sliding window counter independently counts... and The total number of pulses emitted is denoted as and The statistical value of the pulse frequency is directly converted into the digital increment of the pulse width modulation duty cycle, and its discrete mapping equation is:

[0071] In the formula, This represents the duty cycle increment for the k-th time window. This is the preset hardware mapping gain coefficient.

[0072] It should be noted that this direct mapping from pulse frequency to physical stride length endows the system with a natural intelligent dynamic response capability. When the system triggers an exploratory penalty signal and enters a large-span jump mode, The difference widens sharply, naturally mapping to a coarse-grained, large duty cycle increment, forcibly pulling the operating point to make large jumps to escape; as the system approaches the global maximum power point, the multidimensional gradient slows down, leading to sparse or even stopped network discharge. and The discharge frequencies all tend to zero or are in dynamic equilibrium at extremely low frequencies. It automatically converges to zero, thus achieving oscillation-free steady-state tracking. This mechanism perfectly resolves the inherent contradiction between dynamic tracking speed and steady-state oscillation in traditional optimization algorithms.

[0073] The generated pulse width modulation signal drives the DC-DC converter to complete the closed-loop control process of adjusting the output power of the photovoltaic array.

[0074] Example 2

[0075] Based on the same inventive concept as the partial shading maximum power point tracking method based on a spiking neural network in Embodiment 1 above, this embodiment provides a partial shading maximum power point tracking system based on a spiking neural network, comprising: The physical sensing and feature extraction module is used to acquire the output voltage and output current of the photovoltaic array through voltage and current sensors and calculate the output power, and extract the voltage change, current change and power change as dynamic gradient features; The physical sensing and feature extraction module is the system's data source. Its core lies in eliminating the need for macroscopic environmental sensors, relying solely on pure electrical signals to extract dynamic microscopic features reflecting localized shading and sudden environmental changes. The system uses low-cost current and voltage sensors deployed at the photovoltaic array's output end to acquire the current output voltage of the photovoltaic array at a preset high-frequency continuous sampling rate. and output current The analog signal is converted into a digital value. The controller internally multiplies the current voltage and current in real time to calculate the current output power of the photovoltaic array. Subsequently, based on the difference between current data and historical data, the system calculates three key dynamic gradient features in real time: voltage change. Current change and power change This implicit sensing mapping process does not directly measure temperature and irradiance, but instead maps temperature drift to changes in the absolute voltage baseline and multi-peak distortion caused by local shading to the step rate of current change relative to voltage change, thereby significantly reducing hardware costs.

[0076] The spatiotemporal pulse coding module is used to encode the dynamic gradient features into a multidimensional spatiotemporal pulse sequence through a Delta modulator. It compares the error between the current value of each input variable and the internal reference tracking level with a preset quantization threshold. When the change amplitude of the variable exceeds the threshold, a pulse is emitted; otherwise, it remains silent. The spatiotemporal pulse coding module is responsible for reducing the dimensionality of continuous physical variables into discrete event-driven pulse sequences, thus achieving modulus-to-dimensionality reduction. The extracted continuous state variables (including V...) pv I pv , and The input variables are fed in parallel into independent Delta modulator (or Poisson probabilistic encoder array) channels. Within each independent encoding channel, the system compares the real-time changes of the input variables with a preset quantization threshold. When the change in the variable exceeds the set threshold, it is determined that a valid state change has occurred, and the encoder outputs a discrete pulse (with a value of logic "1") on the time axis; if the change does not exceed the threshold, it remains silent (with a value of logic "0"). Through this encoding process, the continuous variable is transformed into a pulse sequence S(t) mapped by spatiotemporal features.

[0077] The forward inference module of the spiking network uses a digital neuron array based on the leakage current integral triggering model and operates in an event-driven manner: when an input pulse is received, the corresponding synaptic weight is accumulated to the neuron membrane potential through a pure adder tree. The membrane potential decays with a fixed leakage current constant and outputs a pulse and resets when it exceeds the firing threshold. The pulse network forward inference module is the core of the entire controller, employing a digital neuron array based on a Leaky Integrate-and-Fire (LIF) model for data inference. The neurons within the system are in an event-driven asynchronous waiting state, responding only when an input pulse is received. Only when the pulse index is reached is the corresponding neuron activated and enters the computation cycle. After activation, the hardware logic directly reads the corresponding synaptic weight from the internal synaptic weight memory (such as on-chip block random access memory, BRAM) based on the pulse index. Subsequently, the system does not rely on hardware multiplier DSP resources at all, but instead uses a pure adder tree to directly accumulate the weight value to the membrane potential of the current neuron. Above, and subtract a fixed leakage constant during the calculation. After the membrane potential is updated, the digital comparator compares it with the set dynamic threshold V. th The membrane potential is compared. If the membrane potential exceeds the threshold, the output layer neuron fires an action potential (output pulse) and instantly resets its membrane potential; if it does not exceed the threshold, the current potential state is maintained until the next cycle.

[0078] The online reward modulation learning module is used to generate reward or penalty signals based on power changes and to update synaptic weights online based on pulse timing-dependent plasticity rules, without the need for backpropagation computation. The online reward modulation learning module endows edge devices with full lifecycle self-evolution capabilities, including resistance to aging and mismatch prevention, through online fine-tuning. The system's internal local reward evaluation module uses power changes... Generate a multi-dimensional reward signal for the core benchmark. If detected... This indicates that the system power is increasing, and a global positive reward signal is subsequently generated; if Furthermore, accompanied by specific current surge pulses, the system determines that the photovoltaic panel is experiencing localized shading, and then triggers an exploratory penalty signal. Based on the pulse timing-dependent plasticity (R-STDP) weight update logic, the system abandons the complex global backpropagation (BP) gradient calculation. The hardware only needs to calculate the update amount directly online using the time difference between the pulses fired by the preceding and following neurons, combined with the generated reward signal, and utilizing the underlying shift register and addition / subtraction logic, and then fine-tune the synaptic weights written to the BRAM. This mechanism allows the system to reshape network weights every second based on the latest power changes.

[0079] The pulse decoding and modulation execution module is used to count the pulse frequency of the output layer neurons within a preset time window and directly map the pulse frequency to the increment of the pulse width modulation duty cycle of the DC-DC converter. It adopts a push-pull dual-channel decoding structure, and the output layer contains two mutually exclusive output neurons: a duty cycle increasing neuron and a duty cycle decreasing neuron.

[0080] The pulse decoding and modulation execution module is the final execution end of the system, responsible for smoothly converting discrete high-frequency pulses into continuous physical control quantities to drive the back-end circuits. The system uses a low-pass filter or a sliding window counter for frequency decoding, counting in real time the total number of effective pulses fired by the SNN output layer neurons within a specific time window. Subsequently, the control logic directly maps the statistically obtained pulse frequency to the pulse width modulation (PWM) duty cycle increment of the DC-DC converter (such as a Boost converter). If the received pulses are highly concentrated, it indicates that the system urgently needs to escape the current local extremum trap, and the duty cycle increment step size will automatically increase; if the pulses become sparse, it indicates that the current operating point is close to GMPP, and the duty cycle increment step size will then become extremely small, thereby achieving oscillation-free steady-state tracking. The final generated PWM signal will drive the external circuit to complete the closed-loop control process of adjusting the photovoltaic array output.

[0081] It should be noted that the physical perception and feature extraction module, the spatiotemporal pulse coding module, the pulse network forward inference module, the online reward modulation learning module, and the pulse decoding and modulation execution module are all data-connected to the processor, and each module is connected sequentially. The processor can be a central processing unit, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices. The synaptic weights in the pulse network forward inference module are stored in on-chip block random access memory (BRAM). The digital neuron array consists of a pure adder tree and digital comparators, and does not rely on hardware multipliers or floating-point units.

[0082] The specific example of the partial shading maximum power point tracking method based on spiking neural network in the aforementioned embodiment 1 is also applicable to the system of this embodiment. Through the foregoing detailed description of the method, those skilled in the art can clearly understand the specific implementation of the system in this embodiment, and will not be described in detail here.

[0083] The foregoing has shown and described the basic principles and main features of the present invention, as well as its advantages. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or basic characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description.

Claims

1. A partial shading maximum power point tracking method based on a spiking neural network, characterized in that, include: The output voltage and current of the photovoltaic array are obtained only through voltage and current sensors. The output power is calculated, and the voltage change, current change, and power change are extracted as dynamic gradient features. The dynamic gradient features are encoded into a multidimensional spatiotemporal pulse sequence using a Delta modulator. A pulse is emitted when the amplitude of the variable change exceeds a preset quantization threshold; otherwise, it remains silent. Using the multidimensional spatiotemporal pulse sequence as input, an event-driven forward inference is performed using a spiking neural network based on a leakage current integral triggering model. The synaptic weights are accumulated to the neuron membrane potential by a pure adder tree, and a pulse is output when the membrane potential exceeds the firing threshold. Reward and penalty signals are generated based on the power change. The synaptic weights are updated online based on the pulse timing-dependent plasticity rule. When the power decreases and the ratio of current change to voltage change exceeds the shading judgment threshold, a penalty signal is generated to adjust the weights in reverse. The statistical output layer pulse frequency is mapped to the increment of the pulse width modulation duty cycle of the DC-DC converter, driving power regulation.

2. The method according to claim 1, characterized in that, The Delta modulator employs a positive and negative dual-channel decoupling coding mechanism: Let the input variable in the k-th sampling period be... Internally maintains the reference tracking level for discrete states. Define tracking error for and The difference, let the quantization threshold be... The following coding logic is executed in each clock cycle: If E(k) is not less than The ascending channel encoder outputs a discrete pulse at the current moment, and simultaneously updates the reference level to the current value plus... ; If E(k) is not greater than The descent channel encoder outputs a discrete pulse at the current moment, while the reference level is updated to the current value minus... ; If the absolute value of E(k) is less than Both channels remain silent and the reference level remains unchanged; The quantization threshold It is significantly greater than the peak-to-peak value of the high-frequency switching ripple of the inductor current at the current duty cycle of the DC-DC converter.

3. The method according to claim 1, characterized in that, The membrane potential update equation for any neuron i in the spiking neural network at discrete time step t is: ， In the formula, Let be the membrane potential of neuron i at time t; For a fixed leakage current constant; The synaptic weights from the previous neuron j to the current neuron i; The input pulse of the preceding neuron j at time t is 0 or 1; when At that time, neuron i fires an output pulse and the membrane potential is forcibly cleared to zero in the next clock cycle.

4. The method according to claim 1, characterized in that, The pulse frequency mapping to duty cycle increments employs push-pull dual-channel decoding, and the output layer includes neurons with increased duty cycle. and neurons with reduced duty cycle Its discrete mapping equation is: ， In the formula, This represents the duty cycle increment for the k-th time window; The preset hardware mapping gain coefficient; and They are respectively and The total number of pulses issued within this time window.

5. The method according to claim 1, characterized in that, The exploratory penalty signal is generated under the following conditions: Power change over two consecutive sampling periods Furthermore, the power decrease exceeds the preset power decrease threshold, and the ratio of current to voltage change rate... Exceeding the preset shading threshold; During the effective period of the penalty signal, strong suppression updates are performed on synaptic connections with a pulse firing time difference Δt < 0 between preceding and following neurons to eliminate the weight bias of local extrema, while synaptic connections with Δt > 0 are temporarily not updated to preserve the degree of freedom of exploration; When power recovers and increases, the penalty signal is removed and switched to a positive reward signal.

6. The method according to claim 1, characterized in that, The shading determination threshold is set as follows: the absolute value of the ratio of current change to voltage change under uniform illumination is on the order of 0.01 to 0.1 A / V, and the shading determination threshold is set as on the order of 1 to 10 A / V.

7. A partial shading maximum power point tracking system based on a spiking neural network, characterized in that, include: The physical sensing and feature extraction module is used to acquire the output voltage and output current of the photovoltaic array through voltage and current sensors and calculate the output power, and extract the voltage change, current change and power change as dynamic gradient features; The spatiotemporal pulse coding module is used to encode the dynamic gradient features into a multidimensional spatiotemporal pulse sequence through a Delta modulator. When the amplitude of the variable change exceeds a preset quantization threshold, a pulse is emitted; otherwise, it remains silent. The forward inference module of the spiking network uses a digital neuron array based on the leakage current integral triggering model and operates in an event-driven manner: when an input pulse is received, the corresponding synaptic weight is accumulated to the neuron membrane potential through a pure adder tree. The membrane potential decays with a fixed leakage current constant and outputs a pulse and resets when it exceeds the firing threshold. The online reward modulation learning module is used to generate reward and penalty signals based on power changes, and to update synaptic weights online based on pulse timing-dependent plasticity rules without the need for backpropagation computation. The pulse decoding and modulation execution module is used to count the pulse frequency of the output layer neurons within a preset time window and directly map the pulse frequency to the increment of the pulse width modulation duty cycle of the DC-DC converter.

8. The system according to claim 7, characterized in that, In the pulse network forward inference module, synaptic weights are stored in on-chip block random access memory, and the digital neuron array is composed of a pure adder tree and a digital comparator, without relying on hardware multipliers and floating-point units; the physical perception and feature extraction module, the spatiotemporal pulse coding module, the pulse network forward inference module, the online reward modulation learning module, and the pulse decoding and modulation execution module are sequentially connected.

9. The system according to claim 7, characterized in that, The pulse decoding and modulation execution module adopts a push-pull dual-channel decoding structure. The output layer contains two mutually exclusive output neurons: a duty cycle increasing neuron and a duty cycle decreasing neuron, which control the direction of duty cycle increase and decrease, respectively.

Images